Comparison of Machine Learning Models To Provide Preliminary Forecasts of Real Estate Prices

Journal of Housing and the Built Environment (2022) 37:2079–2114
https://doi.org/10.1007/s10901-022-09937-1
ARTICLE
Comparison of machine learning models to provide

preliminary forecasts of real estate prices
Jui‑Sheng Chou1 · Dillon‑Brandon Fleshman1 · Dinh‑Nhat Truong1,2
Received: 24 January 2021 / Accepted: 31 January 2022 / Published online: 15 March 2022
© The Author(s), under exclusive licence to Springer Nature B.V. 2022
Abstract
Real estate is one of the most critical investments in the household portfolio, and repre-
sents the greatest proportion of wealth of the private households in highly developed coun-
tries. This research provides a succinct review of machine learning techniques for predict-
ing house prices. Data on dwelling transaction prices in Taipei City were collected from
the real price registration system of the Ministry of the Interior, Taiwan. Four well-known
artificial intelligence techniques—Artificial Neural Networks (ANNs), Support Vector
Machine, Classification and Regression Tree, and Linear Regression- were used to develop
both baseline and ensemble models. A hybrid model was also built and its predictive per-
formance compared with those of the individual models in both baseline and ensemble
schemes. The comprehensive comparison indicated that the particle swarm optimization
(PSO)-Bagging-ANNs hybrid model outperforms the other models that are proposed
herein as well as others that can be found in the literature. The provision of multiple pre-
diction models allows users to determine the most suitable one, based on their background,
needs, and comprehension of machine learning, for predicting house prices.
Keywords House price forecasting · Data mining · Machine learning · Ensemble method ·
Particle swarm optimization · Hybrid model
* Jui‑Sheng Chou
[email protected]
Dillon‑Brandon Fleshman
[email protected]
Dinh‑Nhat Truong
[email protected]
1
Department of Civil and Construction Engineering, National Taiwan University of Science
and Technology, Taipei, Taiwan
2
Department of Civil Engineering, University of Architecture Ho Chi Minh City, Ho Chi Minh City,
Vietnam
13
Vol.:(0123456789)
2080 J.-S. Chou et al.
1 Introduction
The Taipei metropolitan area is located in Northern Taiwan and is one of the most
important economic centers in Taiwan. It covers an area of 2224.1 km2, accounting
for 6.1% of the area of Taiwan (Wang et al., 2018). The area includes a population
of approximately 6.7 million, accounting for 30% of Taiwan’s total population (Wang
et al., 2018). Taipei is one of the world’s most densely populated cities, with an aver-
age population density of 9942 people per square kilometer (Li et al., 2016). Taipei is
also one of the world’s most expensive cities. Taipei’s house price-to-income ratio has
risen sharply from just 6.4 in 2004 to about 15.8 in 2020, according to the country’s
Ministry of Interior (MOI)—higher than London (8.6), New York (5.9), Toronto (9.9),
Sydney (11.8) or Vancouver (13) (Delmendo, 2021).
Real estate is one of the most crucial investments in the household portfolios in
Taiwan. (Chiang et al., 2015). Housing not only provides a place to live but is also a
critical investment that affects quality of life. The price of a house is of considerable
interest to its stakeholders (homeowners, appraisers, developers, investors, and others).
However, due to the information asymmetry in house markets, people commonly expe-
rience financial losses as a result of their investment. The forecasting of house prices
has a significant role in the making of decision by stakeholders to actualize the poten-
tial of their investment (Kouwenberg & Zwinkels, 2014).
Traditionally, appraisers estimate the value of housing from explicit market infor-
mation, ultimately assessing prices based on their subjective evaluations. Accordingly,
the appraised price will inevitably be influenced by manual interference. To resolve
this issue, a widely used technique, machine learning (ML), has been introduced (Fan
et al., 2018; Liu & Liu, 2019). Several studies have proposed effective and efficient
machine learning techniques for predicting house prices. However, to the best of our
knowledge, a comprehensive comparison among the available machine learning mod-
els to provide preliminary forecasts of real estate prices has not been reported earlier
in the literature.
To fill the research gap, this work investigates the popular machine learning models
for real estate valuation. Four single (baseline) machine learning models, linear regres-
sion (LR), classification and regression tree (CART), support vector regression (SVR),
and artificial neural network (ANNs), as well as three ensemble models, involving vot-
ing, bagging, and stacking, are compared for house valuation. Moreover, hybrid models
are developed by introducing a metaheuristic algorithm—particle swarm optimization
(PSO) to possibly improve upon the performance of aforementioned machine learners.
Two software packages (WEKA and MATLAB) are utilized to develop above models
for the potential real estate estimators. WEKA (Waikato Environment for Knowledge
Analysis), available at https://www.cs.waikato.ac.nz/ml/weka/, is an advanced collection
of machine learning algorithms and preprocessing techniques that has been designed to
implement practical applications; it provides, among other features, various methods for
transforming and preprocessing the input data and for making an attribute selection, as
well as for classification, clustering and regression tasks and for a statistical evaluation
of the resulting learning schemes. WEKA is a very user-friendly tool and it can also be
easily interfaced with some of the most commonly used programming languages for
machine learning tasks, such as Java, Python and R (Merlini & Rossini, 2021). MAT-
LAB, the language of technical computing platform, is available in academia, industry
and research laboratories easily (Job et al., 2021).
13
Comparison of machine learning models to provide preliminary… 2081
Particularly, WEKA is aimed at practitioners who do not have a programming back-

ground. MATLAB is used by those with an adequate background in machine learn-
ing, and it offers metaheuristic coding that provides more accurate estimation of house
prices. The main objective of this research is to find the most efficient and accurate pre-
dictive models for the current dwelling dataset by carrying out multiple machine learn-
ing techniques.
2 Literature review
2.1 Traditional methods for valuing real estate
The cost approach, the income approach and the sales comparative approach are three
traditional methods of valuing real estate. The cost approach determines the value of
real estate in the following steps. Firstly, the site is appraised as if it were vacant. Sec-
ondly, the cost of the improvements on the site are assessed. Finally, the value of the site
under its optimal use is determined with depreciation subtracted. The method is suitable
for properties that do not generate incomes, such as schools, and can be compared with
only a few buildings in the surrounding area.
The sales comparative approach values properties by comparison with similar
properties that were recently involved in open market transactions. As every property
is unique, any characteristic may impact its estimated value. The superiority of this
method is that it reflects market value using a simple computation (Adetiloye & Eke,
2014; Kontrimas & Verikas, 2011).
The income approach derives market value from the estimated future income that
would be generated by the property, discounted to obtain the present value. The gross
income multiplier (GIM) and discounted cash-flow analysis (DCF) are commonly used
techniques to relate estimate income to market value. The GIM method estimates the
value of an investment by dividing the sale price by gross annual income. DCF analy-
sis estimates the value of an investment based on its future cash flow. By utilizing the
discount rate, DCF analysis seeks the present value of expected future cash flow, which
is used to assess the potential investment. This method, unlike the cost approach, is suit-
able for commercial offices or leased apartments that are used to generate income.
2.2 Hedonic price theory
Rosen (1974) (Rosen, 1974) defined hedonic prices as “the implicit prices of attributes
and are revealed to economic agents from observed prices of differentiated products and
the specific quantities of characteristics associated with them”. In other words, an item
can be valued by using its characteristics, allowing appraisers to obtain its hedonic price
by adding up the implicit values of its homogeneous characteristics. This theory implies
a regressive relationship between the item and its attributes, in which each attribute dis-
tinctly contribute to the hedonic price of the item. Compressing distinguished character-
istics into one dimension to avoid complexity and the intractability of a multi-commod-
ity is one of the advantages of hedonic price theory (Xiao, 2017).
Hedonic price theory has been widely used in real estate evaluation. It allows
appraisers to decompose house price into values of individual attributes (such as living
13
area, size of plot, number of rooms, number of bathrooms, location, parking facilities,
and location). House price normally fluctuates with the size of the plot, which is one of
the attribute that most strongly affects the property value. Despite the wide use of the
hedonic house price model and the traditional methods introduced above, some issues
such as multi-linearity and explanatory variable interactions limit its accuracy in prop-
erty valuation (Limsombunchai et al., 2004). Therefore, many researchers have intro-
duced techniques, such as those based on artificial intelligence, to solve these problems.
2.3 Use of machine learning in house price prediction
The development of machine learning accelerated with increases in computing power and
its availability (Chaphalkar & Sandbhor, 2013). Machine learning is implemented to con-
struct models that mimic human reasoning; consequently, it can deduce new facts from
historical data and respond adaptively to changes of previously obtained information (Chou
& Bui, 2014; Chou & Truong, 2021). The practicality of machine learning has been estab-
lished in various fields and machine learning has been much used in academic fields and
forecasting. The use of machine learning for forecasting residential values has been studied
in the literature since the 1990s (Chaphalkar & Sandbhor, 2013). Owing to its ability to
efficiently solve problems that humans are not able to answer right away, machine learning-
based models are feasible alternatives to stakeholders for predicting house prices given the
nonlinearity of the relevant data (Barzegar et al., 2016).
In recent years, numerous machine learning-based methods have been used to forecast
in property values, as they can find functional relationships in historical data. ANNs, SVR,
CART and LR are the most fundamental and commonly used machine learning algorithms
for prediction (Chou et al., 2018). In 2004, Limsombunchai et al. (2004) compared the
predictive accuracy of the hedonic price model with that of the ANNs model in house price
prediction. ANNs has been proven to be superior to the hedonic house price model (Lim-
sombunchai et al., 2004). Wu (2017) utilized various feature extraction algorithms and fea-
ture selection methods to build an SVR predicting model for a similar purpose (Wu, 2017).
Fan et al. (2006) applied CART to avoid such potential problems as market disequilibrium
that affect the hedonic-based regression approach (Fan et al., 2006). Varna et al. (2018)
proposed LR and ANNs models for house price prediction, enhanced using a boosting
algorithm (Varma et al., 2018). Therefore, this research exploits ANNs, SVR, CART and
LR to develop baseline and ensemble models in WEKA.
Various studies have used metaheuristic methods to find the optimal hyperparameters of
machine learning techniques and ultimately enhance the predictive accuracy (Claesen & De
Moor, 2015). PSO has been proven to yielding more satisfactory results than other optimiza-
tion algorithms, and it is based on uncomplicated concepts and is easy to implement (Alfi-
yatin et al., 2017; Wang et al., 2014). Wang et al. (2014) introduced the PSO algorithm to
optimize the hyperparameters of the support vector machine (SVM) model. Their proposed
PSO-SVM model had better forecast real estate values than did SVM or back propagation
neural network (Wang et al., 2014). Alfiyantin et al. (2017) utilized the PSO algorithm to
select influential variables and to determine their optimal coefficients in prediction. Their
research proved that PSO can be combined with regression analysis to minimize the predic-
tion error (Alfiyatin et al., 2017). This work introduces PSO, which is simple and efficient
and provides robust optimization for building hybrids of baseline and ensemble models.
13
3 Methodology
3.1 Data visualization and preprocessing
The box plot is a common technique for visualizing and interpreting data (Williamson
et al., 1989), which are transformed to five statistical values—upper extreme, upper quar-
tile, median, lower quartile, and lower extreme. Numerical data are thus plotted graphi-
cally, allowing researchers to comprehend quickly the broad pattern of the data and to see
the patterns that are concealed in the dataset, such as the outliers (Potter, 2006). Outliers
are observations that are assumed to follow a different distribution from the majority of the
dataset. They are also likely to have a profound impact on the data analysis, resulting in
erroneous prediction (Schwertman et al., 2004).
The threshold of outliers is based on four boundaries—two inner fences and two outer
fences. The data between inner fences and outer fences are designated as mild outliers,
while the data that exceed the outer fences are designated as extreme outliers (Dawson,
2011). Since a boxplot not only enables the easy identification of outliers, but also provides
decisive information to eliminate them or not so as to prevent distortion of the measures
of spread and sensitivity in prediction (Schwertman et al., 2004), this research employs a
boxplot technique to identify and exclude the extreme outliers to make more accurate and
stable predictions for approximately normal distributed house price datasets.
Feature selection is another statistical technique used to enhance predictive performance
and the time and cost-effectiveness of prediction (Guyon & Elisseeff, 2003). The three cat-
egories for feature selection are filters, wrappers and embedded techniques. Filters extract
features without involving learning techniques, leading to high computational efficiency.
Wrappers utilize learning algorithms to evaluate the influential attributes in the feature
space. Despite the fact that wrappers tend to outperform filters, as the feature space grows,
the computational burden increases, so wrappers are rarely used in practice. Finally, the
embedded techniques are the combination of filters and wrappers (Hira & Gillies, 2015;
Urbanowicz et al., 2018).
In accordance with our research objectives and considering computational efficiency,
this work uses a feature selection method in WEKA, called ReliefF. ReliefF is a weight-
based algorithm, called ReliefFAttributeEval in the WEKA platform, and is used in con-
junction with Ranker search algorithm to obtain a rank list (Fallahi & Jafari, 2011). It
updates the weight by randomly sampling instances from training data, and subsequently
evaluates the performance by distinguishing the nearest sample instances of selected attrib-
utes from the same and different classes (Huang et al., 2009). ReliefF can handle both nom-
inal and continuous features, identifying their interactions, and managing missing data; it
is therefore considered one of the most beneficial and significant filter-based algorithms
(Palma-Mendoza et al., 2018).
3.2 Baseline prediction models
3.2.1 Artificial neural networks
Artificial neural networks are inspired by human brains, and were originally designed to
duplicate human learning. A basic ANN structure has three main layers—an input data
layer, a hidden layer, and an output layer (Bahia, 2013; Peter & Raglend, 2017). The
13
input data layer is the layer to which collected data are input; the data are then transferred
through the hidden layer, and finally the training result or prediction is output to the output
layer. More specifically, a neural network is built from numerous interconnected neurons
with an activation mechanism by which information is transferred between them. Each
neuron is associated with a weight and bias adjustment from the input connection. An acti-
vation function sums the weighted inputs and biases, and the value thus obtained is trans-
ferred to the output neuron (Bahia, 2013; Dey et al., 2019; Khamis & Kamarudin, 2014).
3.2.2 Support vector machine for classification and regression
Support Vector Machine (SVM) is a supervised learning algorithm, which was orig-
inally developed by Vapnik (1995) (Vapnik, 1995). It can both be utilized to solve
classification and regression problems. When target variables include categori-
cal data, the support vector machine for classification can be used to determine the
optimal hyper-plane to distinguish among different classes (Cheng et al., 2019; Chou
et al., 2018, 2021a). When the target variables are continuous data, the support vector
machine for regression (SVR) is used. SVR is a variation of SVM that maps its inputs
onto n-dimensional feature space, and then fits them by a nonlinear kernel function
(Chou et al., 2016). The general model for feature space can be given by the following
equation.
n
∑
f (x, 𝜔) = 𝜔i gi (x) + b (1)
j=1
where wi is a weight vector; gi (x) is a set of non-linear transformations; [ b is a bias

] term.
To measure the quality of estimation, a loss function L𝜀 = y, f (x, 𝜔) is applied,
where
{
[ ] 0 if |y − f (x, 𝜔)| ≤ 𝜀
L𝜀 = y, f (x, 𝜔) = (2)
|y − f (x, 𝜔)| − 𝜀 otherwise
SVR implements an ε-insensitive loss function in the hyperplanes to concurrently mini-

mize ‖𝜔‖2 and reduce the complexity of the model. Kernel functions such as linear, radial
basis, polynomial, and sigmoid functions can be used to find support vector on the func-
tion surface during training. Owing to its superiority, the radial basis function (RBF) [i.e.,
‖x − xk ‖ )] has been proven effective for SVR (Chou et al., 2017) and
K(x, xk ) = exp(−𝛾 ‖ ‖2
was used as the kernel function in this study.
3.2.3 Classification and regression tree
Classification and regression tree (CART) is a popular tree-structured analysis tech-

nique for use with either categorical variables (classification tree) or numerical vari-
ables (regression tree). Particularly, CART reacts sensitively to slight changes in the
training dataset (Erdal & Karakurt, 2013). Each internal node in CART has a parti-
tioning rule for testing an attribute. Each branch is represented by the previous result
of the test. Eventually, the terminal nodes display a class label or prediction. CART
is pruned to optimize the total error by reducing the number of branches through a
repeating operation that is determined by the Gini impurity (Chou et al., 2016, 2021b).
13
The Gini impurity indicates the likelihood that a randomly selected sample is labeled
incorrectly, as follows.
J
∑
Gini = 1 − p2i (3)
i=1
where i = categorical variable; J = number of categories and pi = probability that i-th cat-
egory is chosen.
3.2.4 Linear regression
The purpose of linear regression is to find functional relationship between response variables
and explanatory variables (Fumo & Rafe Biswas, 2015). When only one explanatory variable
is considered in the prediction process, it is called univariate linear regression. When more
the one explanatory variables is considered, the process is called multivariate or simple linear
regression. The house price dataset involves a numerical response variable, Y, and multiple
explanatory variables, Xi, which will be the inputs of the multivariate regression model dur-
ing the prediction process. The general form of multivariate linear regression is,
Y = 𝛽0 + 𝛽1 X1 + 𝛽2 X2 + … + 𝛽i Xi + 𝜀 (4)
where Y = response variable; X1, X2,….., Xi = explanatory variables; i = number of varia-
bles; 𝛽0 = constant variable; 𝛽1 , 𝛽2 , … , 𝛽i = regression coefficient; and finally, 𝜀 = error term.
3.3 Ensemble prediction models
3.3.1 Voting
Voting is the simplest ensemble prediction model, both conceptually and with respect to
implementation (Chou & Tran, 2018). In prediction problems, voting averages the outputs of
target models for further analysis. In this study, the voting method was managed to generate
11 ensemble models using two, three or four individual techniques based on the aforemen-
tioned ANNs, SVR, CART, and LR models. The two-technique ensembles were ANNs + LR,
ANNs + SVR, ANNs + CART, SVR + CART, SVR + LR, and CART + LR. The three-tech-
nique ensembles were ANNs + SVR + CART, ANNs + SVR + LR, ANNs + CART + LR, and
SVR + CART + LR. The four-technique ensemble was ANNs + SVR + CART + LR.
3.3.2 Bagging
Bagging is an acronym for bootstrap aggregation, which is a universal ensemble method

and one of the earliest to be developed (Erdal & Karakurt, 2013). Bagging promotes model
diversity by applying bootstrap technique to generate randomly several independent sub-
data from the original training dataset, and trains each model individually using the target-
ing machine learning technique. Eventually, they are aggregated to build a homogeneous
model for prediction by either averaging the outputs of the models or outputting the results
with majority votes. This work presents four homogeneous bagging ensembles that use
machine learning techniques; they are ANNs ensemble, SVR ensemble, CART ensemble
and LR ensemble.
13
3.3.3 Stacking
Stacking is also called stacked generalization; it is a hierarchically structured ensemble

learning method that was proposed by Wolpert (1992). In the first hierarchy, the assem-
blies of base-learners can be customized to train the samples. Subsequently, the output of
all base-learners are “stacked” and input to the meta-learner. The meta-learner is utilized
to aggregate the patterns extracted from the base-learners and a prediction is ultimately
made (Qian & Rasheed, 2007). Since Wolpert had proven that SVR outperforms other
supervised learning techniques in recognizing patterns (Wolpert, 1992), SVR is used as
the meta-learner for all 11 heterogeneous combinations of base-learners as described in the
voting scheme.
3.4 Hybrid models
Hybrid models have been proved to outperform markedly single and ensemble learners
both empirically and theoretically, especially when they are used to deal with classification
problems, complicated regression, and high-dimensional data. Hybrid models are widely
applied to a diverse range of real world issues, such as financial forecasting, medical diag-
nosis, and person recognition. They can significantly improve adaptability as well as the
quality of reasoning (Kazienko et al., 2013). Various studies have found that hybrid mod-
els, which incorporate metaheuristic methods and machine learners, are likely to have a
favorable predictive accuracy. However, its effectiveness in house price prediction is still
unknown. Therefore, in this study, metaheuristic algorithms are introduced to improve the
performance of predictive models.
A metaheuristic algorithm can be regarded as a general-purpose heuristic method, and it
assists a particular heuristic to escape from local optima and guide it toward favorable regions
in the search domain, which contain robust solutions (Hammouche et al., 2010). Countless
metaheuristic algorithms for optimizing hyperparameters of machine learners have been imple-
mented. The most commonly used metaheuristic techniques, to name a few, are ant colony,
genetic algorithm and grid search. Each has its weaknesses. The disadvantage of ant colony is
it frequently traps in local optimums (Wang & Yao, 2001). The genetic algorithm requires intri-
cate implementation and a distinct crossover or mutation to be designed for each situation
(Elbeltagi et al., 2005). Grid search has a very large computation time (Do et al., 2007).
Another popular metaheuristic algorithm is particle swarm optimization, which can be
used for comprehensive global optimization; conceptually simple; is easily implemented
and converges rapidly (Wang et al., 2014). Hence, rather than using the aforementioned
metaheuristic techniques, PSO is utilized to optimize the machine learners herein to
develop hybrid models. PSO is an adaptive algorithm that was first proposed by Kennedy
and Eberhart in 1995 (Kennedy & Eberhart, 1995), and based on the social-psychological
metaphor of bird flocking or fish schooling (Ghasemi et al., 2019).
In the PSO algorithm, a population of particles are placed stochastically in the search
space, and each particle represents a feasible solution. Every particle evaluates its objective
function at its present location. Subsequently, the velocities of particles are updated in a
manner that depends on its best position and the best position of the swarm; new positions
of all particles are ultimately obtained from the average of their velocities. The aforemen-
tioned procedure iterates until the stopping criterion is met (Armaghani et al., 2017). The
velocity and location of the particles are determined and updated using Eqs. (5) and (6).
13
vi (t + 1) = 𝜑 × vi (t) + (c1 × rand() × (pbest

i − pi (t))) + (c2 × rand() × (pgbest − pi (t)))
(5)
pi (t + 1) = pi (t) + vi (t) (6)
where vi (t + 1) = updated velocity of the ith particle; φ = inertia weight; c1 and c2 = weight-
ing coefficients for individual best and global best positions respectively; pi (t) = position of
the ith particle at time t; pbest
i
= best known position of ith particle so far, and pgbest = best
position of any particle in the swarm until now. The rand () = a function that generates the
value of a uniformly distributed random variable ∈ [0, 1].
3.5 Model reliability and performance evaluation
3.5.1 Cross‑fold validation
Cross-fold validation is utilized to compare the predictive accuracy of two or more models
with reliability. Initially, k-fold cross-validation randomly splits the dataset into k individual
and equally sized subsets. One of the partitioned subsets is used for validation while the rest of
them are applied to generate the training model. The aforementioned procedure is repeated k
times to complete the cross-validation. Kohavi (1995) proposed that when the value of k is 10,
the cross—validation is analytically valid, and the variance, bias and computation efficiency
are all optimal (Kohavi, 1995).
3.5.2 Statistical indicators
3.5.2.1 Pearson correlation coefficient (R) The Pearson correlation coefficient is a measure
of the correlation between two variables; it has value from −1 to 1. A value closer to 1
indicates a more positive correlation between the variables. A value closer to −1 indicates
a higher negative correlation between the variables, representing the two variables have a
higher negative correlation. A correlation coefficient of 0 indicates no linear relationship
between the variables. Equation (7) defines Person’s R.
∑ �∑ ��∑ � �
n y.y� − y y
R= �
�∑ � �∑ �2 � �∑ � �∑ �2 (7)
n y2 − y n y�2 − y�
where y = actual value; y′ = predicted value; and n = number of samples.
3.5.2.2 Mean absolute percentage error (MAPE) MAPE is also known as the mean absolute
percentage deviation, which is a statistical measure of predictive accuracy. By calculating the
ratio between the relative differences and the baseline value, it is unaffected by sample size or
units. A lower MAPE indicates greater accuracy of the model. Equation (8) displays the math-
ematical formula of MAPE.
n
1 ∑ || y� − y ||
MAPE = (8)
n i=1 || y ||
13
3.5.2.3 Root mean squared error (RMSE) RMSE represents the degree of dispersion between
predicted values and actual values; it is the square root of the average sum of the errors between
them. The RMSE is given by Eq. (9).
√
√ n
√1 ∑
RMSE = √ (y� − y)2 (9)
n i=1
3.5.2.4 Mean absolute error (MAE) MAE is the absolute average error between actual and
predicted values; and so its unit is the same as the observed values. The MAE can be com-
puted using Eq. (10).
n
1 ∑|
MAE = y − y� || (10)
n i=1 |
3.5.2.5 Synthesis index (SI) To understand the performance of each model comprehen-
sively, the synthesis index (SI) is introduced. SI is a statistical measure that is based on
the four aforementioned statistical indicators, and is computed using 1-R, MAPE, RMSE,
and MAE. The range of SI is between 0 and 1; a value close to 0 indicates that the model
has a high accuracy in average perspective. The equation for SI is as follows.
m ( )
1 ∑ Pi − Pmin,i
SI = (11)
m i=1 Pmax,i − Pmin,i
where m = number of performance measures; Pi = ith performance measure.
4 House price dataset
The dataset was purchased with granted funds from the Ministry of Science and Technol-
ogy, Taiwan, and obtained from the real price registration system (RPRS) of the Ministry
of the Interior, Taiwan. The RPRS includes the prices and many attributes in all housing-
related transactions in Taipei City, Taiwan, during the period from 2013 to 2017, with a
total of 209,402 samples. The samples undergo a data cleaning in which a box plot is used to
eliminate extreme outliers that were not considered normal appraisal processes. The features
based on literature suggestions were selected from the RPRS for subsequent analysis. In par-
ticular, the actual distance between locations of real estate and objectives, which are Yes In
My Back Yard (YIMBY) and Not In My Back Yard (NIMBY), is calculated by transforming
the coordinate TWD97, obtained from the real price registration system, to WGS84. The
end of this section will define prestigious elementary and junior high school districts.
13
4.1 Data preprocessing
4.1.1 Data cleaning
The transaction data that were provided by the RPRS include a number of transaction
targets, such as land only and parking space transactions. Also, information is missing
in cases in which the transaction is abnormal. To ensure that each datum can be used,
information that does not meet specified conditions is deleted. Only information that is
complete and consistent with normal conditions is retained.
4.1.1.1 Screening based on transaction target The purpose of this study is to predict
house prices using both internal and external features. Therefore, only transaction sam-
ples for “room + land” and “room + land + parking space” are used. A total of 13,695
samples were removed to maintain the consistency of the dataset.
4.1.1.2 Screening based on transaction period The law on registering real prices came
into effect in August, 2012. Earlier registration data were deficient. To ensure continuity of
the dates associated with valid prices as well as completeness of the data, transaction data
for Taipei City from January 2013 to December 2017 are used. A total of 8365 samples are
excluded for this reason.
4.1.1.3 Screening based on total construction floors Most building types in Taipei City are
residential buildings with various numbers of floors—under six, between 6 and 11 or over
12 floors. Such buildings are less heterogeneous than other kinds of real estate. Residential
buildings whose numbers of floors fell within these three categories were used, while others
were expunged. A total of 44,977 samples thus eliminated.
4.1.1.4 Screening based on area Property can be registered as shared upon registration of
the price, and each co-owner is free to dispose of his or her due portion. The share may be
involved in legal auctions, mergers or other transactions that are authorized under Article 23
of the Technical Rules for the Valuation of Real Estate; such transactions cannot be identi-
fied in the remarks column of RPRS database. To guarantee that all the transaction data are
associated with normal trades, a total of 183 targets with areas of under three pings (floor
area unit particularly used in Taiwan), which may involve a partial transfer of rights of com-
monly held properties, were thus removed.
4.1.1.5 Screening based on building age The value of real estate is reduced by physical,
functional or economic depreciation as its age increases. Therefore, data that are not asso-
ciated with a specified building age are removed. Furthermore, this study focuses on pre-
owned housing, and homes with an age of less than one year may be pre-sold. Therefore,
data for dwellings that are less than a year old are deleted. A total of 39,786 samples were
thus excluded.
4.1.1.6 Screening based on transaction floor If the transaction involves property on the
first floor, it involves a commercial target, which falls outside the scope of our study. In
addition, if a transaction involves more than one floor or the basement, the total transaction
price will not reflect the market. Therefore, data concerning transactions that involve the
13
first floor, multiple floors or contains the basement are removed. A total of 15,619 samples
are eliminated.
4.1.1.7 Screening based on number of rooms Owners of real estate in Taipei City often
divide their home into several rental suites, which can be classified as income-type real
estate. The value of this category of homes is difficult to compare with that of ordinary ones.
To avoid including possible rental suites, transactions that involve more than six rooms are
excluded. A total of 312 samples are thus expunged.
4.1.1.8 Screening based on main usage To ensure the homogeneity of the massive valua-
tion models that will be developed in this study, only transactions that involve “residential
usage” and “residential and commercial usage” as the main utilization are retained, while
the rest are eliminated. A total of 58,504 are thus removed.
4.1.1.9 Screening based on parking space In the RPRS, some parking spaces of inde-
pendent property owners are mandatorily registered, while others are “openly registered”,
so their cost and area need not be specified. To maintain the consistency of the data,
transactions without the registered price and the area of the parking space are excluded.
A total of 3318 samples are thus eliminated.
4.1.1.10 Screening based on special transaction Article 23 of the Technical Regula-

tions for Real Estate Valuation requires that real estate appraisers consider underlying
trading conditions when selecting a target for comparison. If a special transaction status
is specified, an appropriate adjustment should be made. If the circumstances that affect
the transaction price cannot be effectively grasped, then the transaction should not be
considered. Consequently, to ensure that the transaction price satisfies both criteria for
so-called “normal price” in the valuation technical rules, only normal transactions speci-
fied in the remarks column are considered. A total of 11,022 samples were thus expunged.
4.1.2 Removing outliers
A box plot is used to detect and eliminate extreme outliers that fall outside the outer
fences in the original dataset. Figure 1 shows the box plot of the original data-
set after cleaning. A total of 401 extreme outliers that exceed the upper outer fence
are then excluded. Following the box plot elimination, 13,220 samples remain, with
minimum = 1.00E+06, maximum = 7.41E+07, average = 2.14E+07 and standard devia-
tion = 1.33E+07. Figure 2 further presents the density of samples in each district of Tai-
pei City. The resulting samples that fit to the study scope are subsequently utilized to
develop house price prediction models.
13
# 107
Maximum = 837,300,000
10
8
Upper Outer Fence = 74,297,500
7
5 Upper Inner Fence = 51,148,750
3
Q3 = 28,000,000
Average = 24,747,459
2
Median = 18,000,000
Q1 = 12,567,500
1
Minimum = 1,000,000
0
Fig. 1 Box plot of original dataset (Unit: NTD)
4.2 Feature description and acquisition
4.2.1 Feature identification
House price are affected by geopolitical features, regional features and individual fea-
tures of the dwelling. Geopolitical features involve economic, political, social and envi-
ronmental conditions, which can be implicitly reflected by the date of the house trans-
action. Individual features designated are obtained from the information available in
the RPRS. Regional factors include Amenities and Disamenities. The features in RPRS
were compared with the influential features addressed in the literature. The selected
factors will be described below. Table 1 defines Amenities and Disamenities. Table 2
lists the 54 ultimately determined features. Table 3 provides comprehensive information
about the collection of features from the literature (Chiang et al., 2017; Chau & Chin,
2003; Du et al., 2013; Geng et al., 2015; Huang et al., 2017a; Huang et al., 2017b; Lin,
2004; Rong & Sun, 2014; Wang, 2011; Wei et al., 2010; Wen et al., 2005; Wu et al.,
2014; Yang & Su, 2011; Zhang & Zhang, 2010). The response variable and individual
features are described as below.
13
Fig. 2 Density of samples at each district of Taipei City
4.2.1.1 Response variable The response variable is the acquisition price of the real
estate (house price), which is provided in RPRS.
4.2.1.2 Characteristic features of the housing Building type: Most buildings in Taipei City
are residential buildings with fewer than six and 11, or more than 12 floors, resulting in great
homogeneity. The construction costs of various types of building, and their functions must
be considered.
Area: A building can be divided into three segments: the main building area, the aux-
iliary building area and a public-shared facilities area. The sum of these three areas, over
which the owners have land rights, is usually proportional to the house price in a nonlinear
pattern.
13
Table 1 Description of amenities and disamenities

Category Feature Description
Amenity YIMBY MRT (Mass Rapid Transit) Exits of all MRT stations in Taipei City
exit
Neighborhood park The parks with an area of over 1 hectare in Taipei City
Large park The parks with an area of over 10 hectares in Taipei City
Prestigious school The prestigious elementary and junior high schools with
full recruitment during the period from 2013 to 2017
Department store 33 department stores in Taipei City on Google Maps
Large gymnasium 12 Sports Centers in Taipei City and Taipei Gymnasium
Disamenity Funeral parlor The First and Second Funeral Parlor in Taipei City
NIMBY Airport Taipei Songshan Airport
Substation Substations located in Taipei City
High voltage tower High voltage towers located in Taipei City
Waste incineration plant Neihu, Muzha, and Beitou Waste Incineration Plants
Sewage treatment plant Neihu and Dihua Sewage Treatment Plants
Gas station Gas stations located in Taipei City
Age: As time goes on, buildings lose value owing to the weather, negative effects of
natural disasters and internal use. Hence, the age of a house is an important factor.
Number of rooms: The number and areas of living rooms, bedrooms and bathrooms
have a significant impact on the price of real estate. Different families have different needs.
Number of total floors: More floors in a building correspond to greater cost of project
development. Under ordinary circumstances, more floors correspond to a higher house
price.
Transaction floor: The transaction floor is the floor on which the real estate is located.
A higher floor of the residence provides better views.
Parking space: The presence or absence of parking spaces affects, and the number of
parking spaces is usually nonlinearly proportional to, the value of a dwelling.
Types of parking space: Whether a parking space is on the ground or uses an elevator
affects house prices.
4.2.1.3 Disamenity NIMBY Funeral parlor: Funeral ceremonies sometimes generate

loud noises, which reduce tranquility and folk customs and religious beliefs make most
people unwilling to live in the neighborhood.
Airport: In areas around airports, the noise of aircraft during take-off and landing can
be heard, and the construction and height of surrounding buildings are limited.
Substation: Residents prefer to keep their distance from the powerful electromagnetic
waves around substations.
High voltage tower: Residents prefer to keep their distance from the powerful elec-
tromagnetic waves around high voltage towers.
Waste incineration plant: Although large-scale waste incineration plants have pollu-
tion control standards, their pollution of the air has a negative impact on willingness to
purchase houses nearby.
13
Table 2 Original features available from RPRS for house price prediction
Characteristics of house Area Ping

Age Year
Building under 6 floors Yes = 1, No = 0
Building between 6-11 floors Yes = 1, No = 0
Building over 12 floors Yes = 1, No = 0
Brick building Yes = 1, No = 0
RC building Yes = 1, No = 0
SRC building Yes = 1, No = 0
Number of living rooms Value
Number of bedrooms Value
Number of bathrooms Value
Number of total floors Value
Number of parking spaces Value
Transaction floor Value
Ground parking space Yes = 1, No = 0
Elevator parking space Yes = 1, No = 0
Other types of parking space Yes = 1, No = 0
Transaction period The first half of 2013 Yes = 1, No = 0
The second half of 2013 Yes = 1, No = 0
The first half of 2014 Yes = 1, No = 0
Administrative region Beitou District Yes = 1, No = 0
Shilin District Yes = 1, No = 0
Datong District Yes = 1, No = 0
Songshan District Yes = 1, No = 0
Zhongshan District Yes = 1, No = 0
Neihu District Yes = 1, No = 0
Wanhua District Yes = 1, No = 0
Zhongzheng District Yes = 1, No = 0
Da’an District Yes = 1, No = 0
Xinyi District Yes = 1, No = 0
Nangang District Yes = 1, No = 0
Wenshan District Yes = 1, No = 0
13
Table 2 (continued)
Amenity YIMBY MRT exit Distance in

meter
Park within 1 hectare Distance in
meter
Park over 10 hectares Distance in
meter
Gymnasium Distance in
meter
Prestigious elementary school Distance in
meter
Prestigious junior high school Distance in
meter
Department store Distance in
meter
Disamenity NIMBY Funeral parlor Distance in
meter
Airport Distance in
meter
Substation Distance in
meter
High voltage tower Distance in
meter
Waste incineration plant Distance in
meter
Sewage treatment plant Distance in
meter
Hospital Distance in
meter
Sewage treatment plant: Sewage treatment plants may produce odor affecting the
prices of nearby dwellings.
Gas station: Gas stations may have environmental impacts by causing traffic jams,
noise, air pollution and a risk of explosion in nearby areas.
Hospital: Since diseases and viruses may spread, residents are not willing to dwell
adjacent to a hospital.
4.2.1.4 Amenity YIMBY MRT: A number of studies have confirmed that the distance
between a building and an MRT station influences the value of real estate. While most
studies consider the straight line distance between the two, actual walking distances is
used herein as the input distance factor.
Park: Green resources in metropolitan areas are not easy to access. Closeness to park
corresponds to higher value of a dwelling. Actual walking distance is the input distance
factor herein.
13
Table 3 Collection of features from literature
2096
Literature Region Response variable Feature
13
Yang & Su Taipei City, Taiwan House price Distance to schools, distance to large parks, distance to department stores, distance to MRT stations,
(2011) distance to large stadiums, distance to funeral parlor, distance to sewage treatment plants, distance to
temples, distance to substations, distance to waste incineration plants
Lin (2004) Taipei City, Taiwan House price Building area, age of the house, floor, building structure, administrative division, school district to which
the building belongs
Huang et al. Taipei City, Taiwan House price Administrative division, transaction season, house type, floor, total floors, age of the house, number of
(2017a) pings, whether it is close to MRT stations, whether it is close to school, whether it is close to park, dis-
tance to city center
Chiang et al. Taipei City, Taiwan House price Building area, age of the house, trading floor, residential building, trading season, distance to Youbike sites,
(2017) distance to city center, distance to urban area, distance to hospitals, distance to parks, distance to schools,
distance to MRT stations, distance to NIMBY
Wen et al. Hangzhou City, China House price Building area, distance to West Lake, internal environment, distance to business districts, traffic conditions,
(2005) garage, loft, decoration, environment, community management, history of the house, entertainment facili-
ties, transaction period, nearby universities
Wei et al. Harbin City, China House price Building area, distance to the business districts, living function, age of the house, neighboring cultural or
(2010) sports facilities, floors, whether there is a basement or building, educational facilities, decoration
Zhang & Zhang Jilin Province, China House price Building area, number of rooms, air heating system, age of the house, distance to business districts, air pol-
(2010) lution level within the residential area, distance to polluted rivers
Wang (2011) Shanghai City, China House price Administrative division, traffic conditions, building area, age of the house, orientation of the house, decora-
tion, floor, surrounding environment, management organization, library, educational facilities, manage-
ment fee, vacant period
Rong & Sun Kunming City, China House price Distance to railways, property costs, floor area ratio, distance to educational facilities, distance to bus stops,
(2014) decoration, building area
Geng Beijing City, China House price Distance to high-speed railway stations, the school district of elementary and junior high school, distance
et al. (2015) to the nearest university, distance to hospitals, distance to supermarkets, distance to parks, distance to
administrative agencies, distance to MRT stations, distance to the nearest sightseeing spot
Wu Shenzhen City, China House price Distance to the Central Business District (CBD), distance to parks, distance to schools, distance to main
et al. (2015) roads, distance to subways
Huang et al. Shanghai City, China House price Building area, population density of administrative division, whether it is close to schools, decoration, sex
(2017b) ratio, orientation of the house, floor, GDP, average wage, population density, distance to urban area
J.-S. Chou et al.
Table 3 (continued)
Literature Region Response variable Feature
Chau & Chin Worldwide – Distance to business area, whether it has sea, river or lakeshore landscape, whether there are hills, valleys
(2003) or golf course landscape, whether the viewing of house is facing obstacles, land lease period, number of
living rooms, bedrooms and bathrooms; building area, parking space, basement, courtyard, elevator or
air-conditioning system, floor, structural materials, facilities (pool, playground, etc.), residents’ income,
whether it is close to prestigious schools, whether it is close to hospitals, whether it is close to religious
gathering places, crime rate, noise disturbance, distance to shopping centers, distance to forests, environ-
mental quality
Du et al. (2013) Taiwan – Age of the house, building area, the administrative division, type of the building, total number of floors,
whether it is the first floor, the date of transaction, the number of bathrooms, structural type
Comparison of machine learning models to provide preliminary…
2097
13
Table 4 Default parameter settings of filter-based algorithm in WEKA

Algorithm Parameter Default value
ReliefFAttributeEval doNotCheckCapabilities False

numNeighbours 10
sampleSize −1
Seed 1
Sigma 2
weightDistance False
Ranker generateRanking True
numToSelect −1
startSet N/A
threshold − 1.7976931348623157E308
Fig. 3 Feature correlation matrix for the house price dataset
13
Table 5 Ultimately filtered features for house price prediction

No. Rank Var. Feature Min Max Mean Std. dev.
1 0.0713702 X1 Area 3.09155 132.9336 34.13144 14.9292

2 0.041458 X2 Age 2 55 24.22065 13.78737
3 0.0323912 X4 Number of total floors 2 33 8.821558 4.701371
4 0.0305644 X24 Number of parking spaces 0 6 0.253555 0.534043
5 0.030357 X5 Number of rooms 0 6 2.672542 0.980382
6 0.0189563 X3 Transaction floor 2 28 5.353631 3.378328
7 0.0158378 X7 Number of bathrooms 0 6 1.611422 0.674213
8 0.0143893 X48 High voltage tower 24 3622 1312.73 837.1312
9 0.010155 X49 Sewage treatment plant 131 9634 4000.833 1,872.908
10 0.0087934 X50 Waste incineration plant 286 9640 4193.257 1983.481
11 0.0071486 X51 Airport 633 10,972 5043.304 2267.859
12 0.0061982 X53 Funeral parlor 70 10,330 4115.926 2432.05
13 0.0055773 X54 Hospital 28 4598 1235.662 769.8108
14 0.0054296 X46 Prestigious elementary school 23 9110 2187.164 1630.969
15 0.0052015 X43 Department store 0 8744 2134.945 1602.708
16 0.0044702 X45 Prestigious junior high school 76 7356 1900.686 1198.961
17 0.0040854 X6 Number of living rooms 0 6 1.69357 0.575074
18 0.0040759 X42 Park over 10 hectares 88 9480 2635.467 1597.882
19 0.0040426 X41 Park within 1 hectare 3 3999 758.643 411.7716
20 0.0035121 X52 Substation 11 3135 749.7481 421.4679
21 0.0032671 X44 Gymnasium 66 7860 2004 1209.426
22 0.003145 X47 Gas station 11 2761 726.3085 491.2171
23 0.0029717 X26 Elevator parking space 0 1 0.057716 0.233214
24 0.001303 X25 Ground parking space 0 1 0.16354 0.369872
25 0.0009398 X40 MRT exit 0 7773 831.9968 710.485
26 0.0004644 X27 Other types of parking space 0 1 0.002648 0.051388
27 0.0002376 X23 SRC building 0 1 0.023071 0.150135
– – Y House price (NTD) 1,000,000 74,100,000 21,421,931 13,319,624
School: Some parents purchase houses in prestigious school districts. Closeness to

prestigious schools increases the probability of being a member of the corresponding
prestigious school district. Distance between residence and prestigious school is there-
fore one of the factors used in modeling.
Department store: Department stores in Taipei City are often located in business dis-
tricts, and homes next to such districts tend to have a higher price due to their greater
convenience.
Public sports facility: Public sports centers or gymnasiums provide sports and leisure
services. Residents who live adjacent to large sports centers have more convenient access
to leisure. Considering the homogeneity of facilities, this study is concerned only indoor
sports centers.
13
4.2.2 Feature selection using filter‑based algorithm
In feature selection, a filter-based algorithm, ReliefFAttributeEval, is used. WEKA, which

stands for Waikato Environment for Knowledge Analysis, was developed at the University
of Waikato, New Zealand. It is a machine learning package with free accessibility and a
user-friendly graphical interface; it provides diverse algorithms for data mining, including
for data pre-processing, regression, clustering, classification and visualization.
For simplicity, ReliefFAttributeEval and Ranker are set to their default parameters,
which are exhibited in Table 4. Features with a rank of greater than zero are selected as
ultimate attributes for the following studies, in which 27 features were extracted for model
construction. Figure 3 shows the feature correlation matrix for the house price dataset.
Table 5 presents the ultimately filtered 27 features for house price prediction.
4.2.3 Distance measurement method
Several ways can be used to calculate the actual walking distance between a building and
nearby amenities or disamenities. One is to input the house address or coordinates into
Google Maps and to use the path calculation API, including the Directions API or the Dis-
tance Matrix API, to find the best route to public facilities such as MRT, parks, and others.
However, at present, Google Maps only offers this service in web-browsing mode. More
than 10,000 samples were collected from RPRS, and manually collecting distance informa-
tion from the webpage would be very time-consuming and probably result in human error.
Therefore, pre-collected coordinates of various facilities, such as the 364 exits of Taipei
MRT stations, 33 department stores, 13 large parks, 129 neighborhood parks and sports
centers and 13 Taipei Gymnasiums are used herein to calculate actual walking distances.
Table 6 Prestigious elementary and junior high schools in Taipei from 2013 to 2017
Elementary school Junior high school
Taipei Municipal DunHua Elementary School Taipei Municipal Lishan Junior High School
Taipei Municipal Jian-Kang Elementary School Taipei Municipal DunHua Junior high School
Taipei Municipal BoAi Elementary School Taipei Municipal Xingya Junior High School
National Taipei University of Education Experimental Taipei Municipal Jinhua Junior High School
Elementary School
Taipei Municipal XinSheng Elementary School Taipei Municipal Long Men Junior High School
Taipei Municipal Jinhua Elementary School Taipei Municipal Zhongzheng Junior High School
Taipei Municipal RenAi Elementary School Taipei Municipal Shipai Junior High School
Taipei Mandarin Experimental Elementary School Taipei Municipal Ming-Hu Junior High School
Affiliated Experimental Elementary School of Univer- The Junior High Division of the Affiliated Senior
sity of Taipei High School of NTNU
Taipei Municipal PengLai Elementary School Taipei Municipal Chengyuan Junior High School
Taipei Municipal NanGang Elementary School Taipei Municipal Datong Junior High School
Taipei Municipal WenHua Elementary School The Affiliated High School of National Chengchi
University
Taipei Municipal Xing-An Elementary School Taipei Municipal Yangming Junior High School
Taipei Municipal Ming Hu Elementary School Taipei Municipal NanGang Junior High School
Taipei Municipal Shipai Elementary School Taipei Municipal ZhongShan Junior High School
13
Table 7 Default parameter settings of baseline models in WEKA

Baseline model Parameter WEKA
Default value
Artificial neural network Name Multilayer perceptron

Hidden layer specification N/A
Number of hidden layers 1
Number of hidden nodes (attributes + classes)/2
Number of learning iterations 500
Momentum 0.2
Type of normalize [− 1,1]
Shuffle examples True
Learning rate 0.3
Support vector machine Name SMOreg
Kernel type RBF
Regularization parameter (C) 1
ε (epsilon) 0.001
RBF γ (gamma) 0.01
Stopping criteria Tolerance: 0.001
Classification and regression tree Name REPTree
Tree depth Infinite
Number of random splits/leaf node N/A
Min number of samples/leaf node 2
Create trainer mode Single parameter
Minimum variance for split (V) 1.0E−3
Number of folds for reduced error pruning (N) 3
Seed for random data shuffling (S) 1
Pruning (P) True
Linear regression Name Linear regression
Fitting regression line method Ordinary least squares
L2 regularization weight 1.0E−8
Features selection M5-prime
Use bias N/A
Eliminate collinear features True
Tolerance N/A
Initially, the straight-line distances between buildings and each facility were calculated,
and then the three coordinates with the shortest distance were selected. Finally, informa-
tion from an Excel file with the three coordinates was inputted into the Python comput-
ing platform and the coordinate data were loaded to the Google Maps server through the
Google Maps Distance Matrix API. The server typically returned three optimal values of
distances and the shortest was chosen, and input as the actual walking distance.
13
Table 8 Default parameter Ensemble model Parameter WEKA

settings of ensemble models in
WEKA Default value
Voting Name Voting

batchSize 100
Classifiers Customized
combinationRule Average of probabilities
numDecimalPlaces 2
Seed 1
Bagging Name Bagging
bagSizePercent 100
batchSize 100
Classifier Customized
numDecimalPlaces 2
numExcutionSlots 1
numIterations 10
Seed 1
Stacking Name Stacking
batchSize 100
Classifiers Customized
metaClassifier ZeroR
numDecimalPlaces 2
numExcutionSlots 1
numFolds 10
Seed 1
4.2.4 Location of residential school
Many studies have suggested that prestigious school districts positively influence housing
prices. The definition of “school district” for the elementary and junior high schools in Tai-
pei City is not based on the shortest distance between buildings and the schools but on the
location of the residences. The government makes minor adjustments to the school district
every year, affecting the residences that belong to it, so the school districts of elementary
and junior high schools may vary.
Residences that are closer to prestigious elementary and junior high schools are more
likely to belong to the prestigious school district. Therefore, the shortest walking distance
from the building to the prestigious school is used herein as a variable. Prestigious school
used herein include elementary and junior high schools that were full during the period
from 2013 to 2017. Table 6 presents all prestigious elementary and junior high schools that
meet this criterion.
13
Bagging-ANNs
Test Data
K-Fold
Cross
Learning Data Validation House Price Dataset
30% 70%
Validation Training PSO
Determine Paramaters Initialization

(i.e., Nn, Nb)
Train
PSO Operation
Bagging-ANNs
No
Objective Function
(i.e., MAPE)
Reach Stopping
Criteria
Train Optimal Model
Yes
Calculate Performance
Measures (i.e., MAPE, RMSE) Optimal Parameters
Fig. 4 Architecture of hybrid model
5 Model development
5.1 Baseline prediction models
For ease of use and efficiency, all of the main parameters of the baseline models that are
developed in WEKA, excluding SMOreg (SVR), are set to their default values; RBFkernel
is used instead of PolyKernel as the kernel function of SMOreg (SVR). Table 7 presents
the default parameter settings for all baseline models.
WEKA provides a user-friendly graphical interface for executing an analytic procedure,
enabling users to input data; adjust the parameters of the algorithm, and eventually con-
struct models merely by clicking buttons.
13
Begin
1. Initialization stage
Subdivide the data into k subsets as learning data (training data and validation data) and test
data
Initialize number of particle, maximum iteration, inertial weight and the boundary of optimized
parameters
2. Perform K folds such that, for each fold, the following steps are performed
while (t < Max_Iteration) do
2.1 PSO operator
(1) Find individual best and global best of each particle
(2) Calculate velocity of each particle
(3) Update new location for each particle
end while
2.2 Bagging- ANNs function for validation
(1) Set the hyperparameters: ,
(2) Train model with hyperparameters ( , ) for Bagging- ANNs
(3) Evaluate the trained (optimized) model using validation data by Eq. (8)
(4) Determine the fitness function ( , ) by Eq. (12) and go to step 2.1
2.3 Have the stopping criteria been met?
(1) If the criteria have been met, go to step 3
(2) Otherwise, go to step 2.1
3. Optimized Bagging- ANNs model
(1) Incorporate the optimized parameters ( , ) into the model using learning data
(2) Calculate the average accuracy over the k test folds from Eq. (7) to Eq. (10)
(3) Save K optimal models
4. Plot stage
(1) Evaluate the post-process results and visualize the results and confirm the best solution
End
Fig. 5 Pseudocode of PSO-Bagging-ANNs
Table 9 Parameter settings of Algorithm Parameter Value

PSO-Bagging-ANNs prediction
model in MATLAB PSO-Bagging-ANNs Max iteration 40
Number of particle 30
Inertia weight 0.9
c1 2
c2 2
Number of bags [1, 10]
Number of neurons [1, 25]
Activation function tansig
Solver trainbr
13
5.2 Ensemble prediction models
In WEKA, the implementation of ensemble models is analogous to that of single models.

Thus, only the parameters of ensemble models, which are voting, bagging, and stacking
(listed in Table 8), are considered in this section.
5.3 Hybrid prediction model
The hybrid prediction model is developed in MATLAB. MATLAB, created by Math-

Works, allows users to plot functions, manipulate matrixes, execute algorithms and create
customized user interfaces. Owing to its ability to perform advanced mathematical com-
putations, MATLAB was used herein to integrate the PSO algorithm with the best model
in baseline and ensemble schemes, which is bagging ANNs, to develop a hybrid model,
namely, PSO-Bagging-ANNs. Figures 4 and 5 show the architecture and pseudo-code of
this hybrid model, respectively. The algorithmic parameters of PSO include the number of
particles, the maximum number of iterations, the inertia weight, and c1 and c2. Based on
literature reviews, a rule of thumb and numerical experiments (Kamaruddin & Ravi, 2016;
Yu & Xiaohui, 2011; Zhan et al., 2009), the following parameters were set; number of par-
ticles = 30, maximum number of iterations = 40, inertia weight = 0.9 and c1 = c2 = 2. After
the parameters of PSO were set, the algorithm was used to optimize bagging ANNs. The
objective function is presented as
( ) Training process
f Nn , Nb = MAPEValidation data (12)
where Nn is the number of neurons and Nb is the number of bags.

Optimizing the number of bags allows the bagging algorithm to search the best
number of bags and make the best predictions. However, to prevent overlapping and
overfitting when the number of bags is very large, the upper bound on the number
of bags was set to ten. The literature suggests that for ANNs, the hyperbolic tangent
activation function, as well as Bayesian regularization backpropagation solver, which
are tansig and trainbr in MATLAB respectively, are more likely than other functions
to yield superior predictions (Isa et al., 2010; Karlik & Olgac, 2011; Kaur & Salaria,
2013; Methaprayoon et al., 2007). With respect to the number of neurons in ANNs,
Armaghani et al. (2017) recommended that the use of (Ni + No )∕2 hidden nodes, where
Ni = number of inputs and No = number of outputs (Armaghani et al., 2017). Thus, con-
sidering the aforementioned threshold as well as the affordability of hardware devices,
Table 10 Prediction performance Model Performance measure

of baseline models in WEKA
R RMSE (NTD) MAE (NTD) MAPE SIoverall
(%)
ANNs 0.948 4,927,394 3,018,614 16.20 0.000 (1)

SVR 0.929 5,124,475 3,426,277 17.25 0.530 (3)
CART 0.932 4,837,640 3,214,792 16.21 0.160 (2)
LR 0.921 5,203,429 3,747,564 21.47 1.000 (4)
Bold value denotes the best performance in all baseline models; (.)
indicates the overall ranking
13
Table 11 Prediction performance of ensemble models in WEKA

Model Performance measure
R RMSE (NTD) MAE (NTD) MAPE (%) SIoverall
Voting ANNs + CART 0.956 3,930,235 2,697,942 14.02 0.141 (4)

ANNs + SVR 0.951 4,173,897 2,887,190 14.91 0.275 (11)
ANNs + LR 0.948 4,232,145 3,015,086 16.47 0.373 (17)
CART + SVR 0.945 4,426,900 2,945,696 14.56 0.353 (15)
CART + LR 0.943 4,445,766 3,056,225 16.07 0.438 (21)
SVR + LR 0.925 5,101,488 3,552,547 19.16 0.846 (25)
ANNs + CART + SVR 0.956 3,969,050 2,704,909 13.71 0.140 (3)
ANNs + CART + LR 0.954 3,994,342 2,772,285 14.53 0.191 (7)
CART + SVR + LR 0.942 4,531,509 3,100,485 16.10 0.468 (22)
ANNs + SVR + LR 0.945 4,392,866 3,078,435 16.41 0.431 (20)
ANNs + CART + SVR + LR 0.951 4,140,885 2,859,455 14.85 0.260 (10)
Bagging ANNs 0.961 3,681,032 2,518,849 13.00 0.000 (1)
SVR 0.929 5,113,746 3,418,603 17.20 0.739 (23)
CART 0.954 3,981,850 2,600,441 13.10 0.110 (2)
LR 0.921 5,204,110 3,747,883 21.46 1.000 (26)
Stacking ANNs + CART 0.951 4,138,478 2,811,604 14.62 0.246 (9)
ANNs + SVR 0.948 4,278,884 2,978,455 15.70 0.351 (14)
ANNs + LR 0.947 4,330,566 3,055,567 16.54 0.409 (19)
CART + SVR 0.945 4,399,824 2,917,823 14.42 0.343 (13)
CART + LR 0.942 4,488,062 2,989,652 14.99 0.404 (18)
SVR + LR 0.928 5,080,933 3,443,707 17.67 0.758 (24)
ANNs + CART + SVR 0.953 4,030,459 2,728,454 13.92 0.175 (5)
ANNs + CART + LR 0.953 4,056,317 2,756,743 14.19 0.197 (8)
CART + SVR + LR 0.948 4,282,978 2,984,935 15.76 0.357 (16)
ANNs + SVR + LR 0.945 4,393,659 2,916,369 14.42 0.342 (12)
ANNs + CART + SVR + LR 0.953 4,030,031 2,728,187 13.91 0.175 (6)
Bold value denotes the best performance in all ensemble models; (.) indicates the overall ranking
Table 12 Optimal PSO-Bagging- Fold Number of bags ( Nb ) Number of

ANNs parameter values in each neurons ( Nn )
fold
1 5 15
2 7 16
3 6 19
4 5 17
5 6 19
6 5 17
7 6 22
8 5 22
9 7 19
10 6 19
13
Table 13 Prediction performance of PSO-Bagging-ANNs in MATLAB

Fold Learning data Test data
R RMSE MAE (NTD) MAPE (%) R RMSE MAE (NTD) MAPE (%)
(NTD) (NTD)
1 0.972 3,153,492 2,169,272 11.29 0.961 3,656,805 2,429,986 12.15

2 0.972 3,106,810 2,144,081 11.08 0.963 3,638,305 2,321,698 11.88
3 0.974 3,044,098 2,087,657 10.87 0.969 3,315,239 2,266,778 11.53
4 0.973 3,086,624 2,117,159 11.09 0.973 3,339,196 2,315,346 11.67
5 0.974 3,020,690 2,089,355 10.90 0.974 3,473,455 2,235,193 11.11
6 0.972 3,116,705 2,141,568 11.21 0.972 3,117,154 2,181,177 11.22
7 0.976 2,917,312 2,036,154 10.78 0.976 3,675,984 2,365,362 12.18
8 0.974 3,007,578 2,073,954 10.93 0.974 3,231,708 2,214,954 11.54
9 0.974 3,034,277 2,089,044 10.91 0.974 3,436,821 2,322,935 11.76
10 0.973 3,073,726 2,107,042 11.01 0.973 3,015,495 2,085,231 10.85
Mean 0.973 3,056,131 2,105,529 11.01 0.970 3,390,016 2,273,866 11.59
Max 0.976 3,153,492 2,169,272 11.29 0.976 3,675,984 2,429,986 12.18
Min 0.972 2,917,312 2,036,154 10.78 0.961 3,015,495 2,085,231 10.85
Std. 0.001 67,051 38,942 0.16 0.005 228,937 99,080 0.43
the upper bound on the number of neurons is set to 25. Table 9 demonstrates the param-
eter settings of PSO-Bagging-ANNs.
6 Model performance and discussion
6.1 Performance in baseline models
In the first stage of the analytical process, the implemented baseline models are four sin-
gle models in WEKA; they are Multilayer Perception (ANNs), SMOreg (SVR), REPtree
(CART) and Linear Regression (LR). Their predictive accuracies are compared in terms
of R, RMSE, MAE, MAPE and SI. Table 10 presents the performance of all four base-
line models. Table 11 shows that the ANNs model in WEKA outperformed the other
baseline models. It yielded three of the best values of the measures of interest- R value
Table 14 Performance comparison among the best models in various schemes

Scheme Model Performance measure
R RMSE (NTD) MAE (NTD) MAPE (%) SIoverall
Baseline ANNs 0.948 4,927,394 3,018,614 16.20 1.000 (3)

Ensemble Bagging ANNs 0.961 3,681,032 2,518,849 13.00 0.307 (2)
Hybrid PSO-Bagging-ANNs 0.970 3,390,016 2,273,866 11.59 0.000 (1)
Bold value denotes the best performance in all models; (.) indicates the overall ranking
13
Table 15 Comprehensive Comparison of house price prediction models in literature
2108
Year Author Data description Model Performance measure Exchange rate
13
2
R (%) RMSE (NTD) MAE (NTD) MAPE (%)
2004 Limsombunchai et al. (2004) Dataset = 200 HPM 74.99 *13,070,078 – – 1 NZD = 20.34 NTD
NN 84.08 *9,134,927 – –
2011 Kontrimas & Kamarudin (2011) Dataset = 100 OLS – – *2,213,627 15.02 1 EUR = 33.92 NTD
Max Price = 29,680,000 MLP – – *3,061,469 23.30
Min Price = 6,071,680 SVR – – *2,003,172 13.62
2013 Núñez Tabales et al. (2013) Dataset = 10,124 ANNs 86.47 *1,023,615 *747,409 – 1 EUR = 33.92 NTD
HPM 79.61 *1,072,247 *769,427 –
2014 Khamis et al. (2014) Dataset = 1,047 MLR 64.40 *1,229,082 – – 1 USD = 30.42 NTD
NN 81.70 *1,093,672 – –
2015 Sarip & Hafez (2015) Dataset = 350 ANNs – – *1,367,650 - 1 MRY = 6.45 NTD
FIS *1,174,570
FLSR *1,158,480
2018 Muralidharan et al. (2018) – NN – – – 30.65 –
DTs – – – 22.84
2022 This study Dataset = 13,220 PSO- 94.09 3,390,016 2,273,866 11.59 –
Max Price = 74,100,000 Bagging-ANNs
Min Price = 1,000,000
Average Price = 21,421,931
Std. Price = 13,319,624
The value with * indicates the original value is converted according to the exchange rate April/09/2019
The values of MAE extracted from Sarip & Hafez (2015) are approximately recorded from their graphs
The values of RMSE extracted from Khamis et al. (2014) are calculated by their original MSE values
The values of R
2, RMSE, and MAE extracted from Núñez Tabales et al. (2013) are the average performance
This table only records the best value of RMSE from the research of Limsombunchai et al. (2004)
J.-S. Chou et al.
(0.948), MAE (3,018,614 NTD), and MAPE (16.20%)—along with the second best RMSE
(4,927,394 NTD), so ANNs had the lowest SI.
6.2 Performance in ensemble models
To improve predictive performance, three ensemble models in WEKA-voting, bagging and

stacking-were used. Each method aggregates the aforementioned four baseline models dur-
ing the procedure. Table 11 displays the performance results that were obtained from the
ensemble models in WEKA. The bagging ANNs ensemble model achieved the best results
of the ensemble models in WEKA, with the best values of all evaluation measures—R
value (0.961), RMSE (3,681,032 NTD), MAE (2,518,849 NTD), and MAPE (13%).
6.3 Performance of hybrid model
In the hybrid model, PSO-Bagging-ANNs, the number of bags and the number of neurons
are optimized by PSO with 30 particles and a maximum of 40 iterations. Table 12 provides
the optimal PSO-Bagging-ANNs parameter values. Table 13 displays the prediction perfor-
mance made using PSO-Bagging-ANNs in MATLAB.
6.4 Comprehensive comparison and discussion
The best baseline, ensemble, and hybrid models were identified in this investigation. The
best machine learning baseline model is ANNs, whereas the best ensemble model is bag-
ging ANNs. The hybrid model, PSO-Bagging-ANNs, was compared with ANNs and
bagging ANNs. Table 14 shows the performance of the three models in predicting house
prices.
The baseline models have the advantage of being effortless to implement, requiring
users only to select the models and parameters in the software package WEKA herein.
However, their accuracies are lower than those in the other schemes. Consequently, the
baseline models are particularly suitable for the users with very little background knowl-
edge of machine learning. For users who understand fundamental machine learning tech-
niques, ensemble models are better because of their superior predictive performance.
The PSO-Bagging-ANNs hybrid model that was demonstrated in this research is nor-
mally more accurate than the single and ensemble models because it has the advantages
of both and the disadvantages of neither. PSO-Bagging-ANNs yielded the best results
with respect to R (0.970), RMSE (3,390,016 NTD), MAE (2,273,866 NTD) and MAPE
(11.59%). Table 15 comprehensively compares models that have been proposed for predict-
ing house prices in the literature. The PSO-Bagging-ANNs hybrid model, with R (0.970)
and MAPE (11.59%), outperforms those models in literature.
The proposed PSO-Bagging-ANNs model is built in the MATLAB environment, and
requires a decent understanding of optimization technique and machine learning concepts
for execution. Although the PSO-Bagging-ANNs model performs best in house price pre-
diction, it is more time-consuming than the other models. Therefore, users must determine
their own needs and their comprehension of machine learning to select the most appropri-
ate technique for house appraisal.
13
7 Conclusion and recommendation
Real estate is one of the most essential investments in a household portfolio. House prices
considerably vary across locations and individual features so forecasting prices of dwell-
ings is significant in decision-making concerning such a portfolio. Appraisers often esti-
mate the value of housing based on explicit market information, but ultimately assess the
price subjectively, causing the appraised price inevitably to be influenced. Consequently,
the hedonic price theory was proposed. The ultimate hedonic price is obtained by valu-
ing and summing the values of the implicit characteristics of the objective. Multi-linearity
or explanatory variable interactions may limit the accuracy of the hedonic price model in
property valuation. Machine learning techniques are developed to solve the aforementioned
problems.
Samples of dwelling transactions in Taipei City were collected from RPRS. To elimi-
nate noise in the original dataset, the data were cleaned and preprocessed using a box plot.
To select features that affect house price, filter-based algorithm, executed using ReliefFAt-
tributeEval in the WEKA platform, was carried out to improve the predictions and reduce
computational cost. The resulting samples and features were used to develop predictive
models. The machine learning schemes that are used to predict the house prices are base-
line, ensemble and hybrid techniques. Four machine learning baseline techniques (namely
ANNs, SVR, CART and LR) were executed in their orginal form. All baseline models
were then combined distinctly to generate three types of ensembles, including voting, bag-
ging and stacking. The predictive accuracy of the hybrid model—PSO-Bagging-ANNs was
compared with those of baseline/ensemble models and prior studies.
Specifically, WEKA was introduced to develop baseline and ensemble models and the
PSO-Bagging-ANNs hybrid model was constructed in the MATLAB environment. The
results demonstrated that the ANNs and bagging ANNs yielded the best predictive per-
formance of any of the baseline and ensemble models, respectively. Subsequently, the
ANNs and bagging ANNs were compared with the PSO-Bagging-ANNs hybrid model.
The comprehensive comparison of all three models indicated that the PSO-Bagging-ANNs
model outperformed the others with respect to R (0.970), RMSE (3,390,016 NTD), MAE
(2,273,866 NTD) and MAPE (11.59%). A comparison of previously proposed models dem-
onstrated that the proposed PSO-Bagging-ANNs had outstanding predictive power. This
investigation comprehensively compares various popular baseline and ensemble machine
learning techniques for predicting house prices, as well as a hybrid model developed in this
study. To the best of our knowledge, no work has included such an extensive comparison.
Users can determine their most suitable method for forecasting the values of dwellings.
This study only involves features that are available in RPRS, but several investigations
have noted that views of the neighborhood and the orientation of the buildings are critical
features in evaluating the price of housing. Further studies are recommended to establish a
3D model or conduct an on-site observation to quantify the aforementioned features using
omputer vision via deep learning techniques, ultimately to obtain more comprehensive
information for predicting house prices. This work proved the feasibility of using ANNs
with a single hidden layer for predicting house prices. It is likely to achieve greater predic-
tive accuracy with more hidden layers, which allow for what is known as deep artificial
neural networks. Using a larger house price dataset than other studies, deep learning tech-
niques are reasonably expected to reduce the likelihood of overfitting. Further works with
a larger dataset and more hidden layers along with more unstructured features (e.g., text,
image, voice, and video) as well as applying newly available methods (e.g., deep learning
13
and gradient boosting decision tree-based algorithms) are needed to possibly enhance the
predictive accuracy of house valuation.
Funding Funding was provided by Ministry of Science and Technology, Taiwan.
Data availability The fundamental codes and data that support the findings of this study are available from
the corresponding author under reasonable request.
Declarations
Conflict of interest The authors declare that they have no conflict of interest.
Ethical approval This article does not contain any studies with human participants or animals performed by
any of the authors.
References
Adetiloye, K. A., & Eke, P. O. (2014). A review of real estate valuation and optimal pricing techniques.
Asian Economic and Financial Review, 4(12), 1878–1893.
Alfiyatin, A. N., Febrita, R. E., Taufiq, H., & Mahmudy, W. F. (2017). Modeling house price prediction
using regression analysis and particle swarm optimization. International Journal of Advanced Com-
puter Science and Applications, 8(10), 323–326. https://doi.org/10.14569/IJACSA.2017.081042
Armaghani, D. J., Raja, R. S. N. S. B., Faizi, K., & Rashid, A. S. A. (2017). Developing a hybrid PSO–
ANN model for estimating the ultimate bearing capacity of rock-socketed piles. Neural Computing and
Applications, 28(2), 391–405. https://doi.org/10.1007/s00521-015-2072-z
Bahia, I. S. H. (2013). A data mining model by using ANN for predicting real estate market: Comparative
study. International Journal of Intelligence Science, 3(4), 162–169. https://doi.org/10.4236/ijis.2013.
34017
Barzegar, R., Adamowski, J., & Moghaddam, A. A. (2016). Application of wavelet-artificial intelligence
hybrid models for water quality prediction: A case study in Aji-Chay River, Iran. Stochastic Environ-
mental Research and Risk Assessment, 30(7), 1797–1819. https://doi.org/10.1007/s00477-016-1213-y
Chaphalkar, N., & Sandbhor, S. (2013). Use of artificial intelligence in real property valuation. Interna-
tional Journal of Engineering and Technology, 5(3), 2334–2337.
Chau, K. W., & Chin, T. (2003). A critical review of literature on the hedonic price model. International
Journal for Housing Science and Its Applications, 27(2), 145–165.
Cheng, M.-Y., Prayogo, D., & Wu, Y.-W. (2019). A self-tuning least squares support vector machine for esti-
mating the pavement rutting behavior of asphalt mixtures. Soft Computing, 23(17), 7755–7768. https://
doi.org/10.1007/s00500-018-3400-x
Chiang, C., Han, C.-C., Chiang, Y.-M., Tsai, T.-C., Wu, F.-S., & Seng, D. (2015). Funding liquidity in the
news and housing price. Market Liquidity, 1–25. https://doi.org/10.2139/ssrn.2565340
Chiang, Y.-H., Chuang, Y.-T., & Chang, C.-O. (2017). The impact of public bike station on residential hous-
ing price in Taipei City. Transportation Planning Journal, 46(4), 399–428. https://www.AiritiLibrary.
com/Publication/Index/10177159-201712-201802050017-201802050017-399-428
Chou, J.-S., & Bui, D.-K. (2014). Modeling heating and cooling loads by artificial intelligence for energy-
efficient building design. Energy and Buildings, 82, 437–446. https://doi.org/10.1016/j.enbuild.2014.
07.036
Chou, J.-S., Ho, C.-C., & Hoang, H.-S. (2018). Determining quality of water in reservoir using machine
learning. Ecological Informatics, 44, 57–75. https://doi.org/10.1016/j.ecoinf.2018.01.005
Chou, J.-S., Ngo, N.-T., & Chong, W. K. (2017). The use of artificial intelligence combiners for modeling
steel pitting risk and corrosion rate. Engineering Applications of Artificial Intelligence, 65, 471–483.
https://doi.org/10.1016/j.engappai.2016.09.008
Chou, J.-S., & Tran, D.-S. (2018). Forecasting energy consumption time series using machine learning tech-
niques based on usage patterns of residential householders. Energy, 165, 709–726. https://doi.org/10.
1016/j.energy.2018.09.144
13
Chou, J.-S., & Truong, D.-N. (2021). Multistep energy consumption forecasting by metaheuristic optimi-
zation of time-series analysis and machine learning. International Journal of Energy Research, 45(3),
4581–4612. https://doi.org/10.1002/er.6125
Chou, J.-S., Truong, D.-N., Le, T.-L., & Thu Ha Truong, T. (2021a). Bio-inspired optimization of weighted-
feature machine learning for strength property prediction of fiber-reinforced soil. Expert Systems with
Applications, 180, 115042. https://doi.org/10.1016/j.eswa.2021.115042
Chou, J.-S., Truong, D.-N., & Tsai, C.-F. (2021b). Solving regression problems with intelligent machine
learner for engineering informatics. Mathematics, 9(6), 686. https://doi.org/10.3390/math9060686
Chou, J.-S., Yang, K.-H., & Lin, J.-Y. (2016). Peak shear strength of discrete fiber-reinforced soils computed
by machine learning and metaensemble methods. Journal of Computing in Civil Engineering, 30(6),
04016036. https://doi.org/10.1061/(ASCE)CP.1943-5487.0000595
Claesen, M., & De Moor, B. (2015). Hyperparameter search in machine learning. In MIC 2015: The XI
Metaheuristics International Conference in Agadir, Morocco, pp. 1-5. Retrieved September 7, 2021,
from https://arxiv.org/abs/1502.02127
Dawson, R. (2011). How significant is a boxplot outlier? Journal of Statistics Education, 19(2), 1–13. https://
doi.org/10.1080/10691898.2011.11889610
Delmendo, L. C. (2021). Taiwan’s house prices surging, amidst strong economic growth. Retrieved Septem-
ber 7, 2021 from https://www.globalpropertyguide.com/Asia/Taiwan/Price-History
Dey, A., Miyani, G., & Sil, A. (2019). Application of artificial neural network (ANN) for estimating reli-
able service life of reinforced concrete (RC) structure bookkeeping factors responsible for deterioration
mechanism. Soft Computing, 24(3), 2109-2123. https://doi.org/10.1007/s00500-019-04042-y
Do, H., Silverman, H. F., & Yu, Y. (2007). A real-time SRP-PHAT source location implementation using
stochastic region contraction (SRC) on a large-aperture microphone array. In 2007 IEEE International
Conference on Acoustics, Speech and Signal Processing - ICASSP ’07, pp. I-121–I-124. https://doi.org/
10.1109/ICASSP.2007.366631
Du, Y.-S., Song, F.-C., Zeng, Y.-S., Ge, J.-N., & Chen, F.-Y. (2013). Retrospective analysis of hedonic price
model in Taiwan. Quarterly Research on Land Issues, 12(2), 44–57.
Elbeltagi, E., Hegazy, T., & Grierson, D. (2005). Comparison among five evolutionary-based optimization
algorithms. Advanced Engineering Informatics, 19(1), 43–53. https://doi.org/10.1016/j.aei.2005.01.004
Erdal, H. I., & Karakurt, O. (2013). Advancing monthly streamflow prediction accuracy of CART models
using ensemble learning paradigms. Journal of Hydrology, 477, 119–128. https://doi.org/10.1016/j.
jhydrol.2012.11.015
Fallahi, A., & Jafari, S. (2011). An expert system for detection of breast cancer using data preprocessing and
bayesian network. International Journal of Advanced Science and Technology, 34, 65–70.
Fan, C., Cui, Z., & Zhong, X. (2018). House prices prediction with machine learning algorithms. In ICMLC
2018: Proceedings of the 2018 10th International Conference on Machine Learning and Computing
(pp. 6–10). https://doi.org/10.1145/3195106.3195133
Fan, G.-Z., Ong, S. E., & Koh, H. C. (2006). Determinants of house price: A decision tree approach. Urban
Studies, 43(12), 2301–2315. https://doi.org/10.1080/00420980600990928
Fumo, N., & Rafe Biswas, M. A. (2015). Regression analysis for prediction of residential energy consump-
tion. Renewable and Sustainable Energy Reviews, 47, 332–343. https://doi.org/10.1016/j.rser.2015.03.
035
Geng, B., Bao, H., & Liang, Y. (2015). A study of the effect of a high-speed rail station on spatial variations
in housing price based on the hedonic model. Habitat International, 49, 333–339. https://doi.org/10.
1016/j.habitatint.2015.06.005
Ghasemi, M., Akbari, E., Rahimnejad, A., Razavi, S. E., Ghavidel, S., & Li, L. (2019). Phasor particle swarm
optimization: A simple and efficient variant of PSO. Soft Computing, 23(19), 9701–9718. https://doi.
org/10.1007/s00500-018-3536-8
Guyon, I., & Elisseeff, A. (2003). An introduction to variable and feature selection. Journal of Machine
Learning Research, 3, 1157–1182.
Hammouche, K., Diaf, M., & Siarry, P. (2010). A comparative study of various meta-heuristic techniques
applied to the multilevel thresholding problem. Engineering Applications of Artificial Intelligence,
23(5), 676–688. https://doi.org/10.1016/j.engappai.2009.09.011
Hira, Z. M., & Gillies, D. F. (2015). A review of feature selection and feature extraction methods applied on
microarray data. Advances in bioinformatics, 2015, 198363. https://doi.org/10.1155/2015/198363
Huang, Y., McCullagh, P. J., & Black, N. D. (2009). An optimization of ReliefF for classification in large
datasets. Data & Knowledge Engineering, 68(11), 1348–1356. https://doi.org/10.1016/j.datak.2009.07.
011
13
Huang, Y.-J., Chiang, Y.-H., & Chang, C.-O. (2017a). Impact of public housing on nearby residential prop-
erty values in Taipei city. Journal of City and Planning, 44(3), 277–302. https://doi.org/10.6128/CP.
44.3.277
Huang, Z., Chen, R., Xu, D., & Zhou, W. (2017b). Spatial and hedonic analysis of housing prices in Shang-
hai. Habitat International, 67, 69–78. https://doi.org/10.1016/j.habitatint.2017.07.002
Isa, I., Saad, Z., Omar, S., Osman, M., Ahmad, K., & Sakim, H. M. (2010). Suitable MLP network activation
functions for breast cancer and thyroid disease detection. In 2010 second international conference on
computational intelligence, modelling and simulation (pp. 39–44). IEEE. https://doi.org/10.1109/CIM-
SiM.2010.93
Job, F., Mathew, D. S., Meyer, D. T., & Narbey, S. (2021). An investigation on the experimental analysis and
MATLAB simulation for dye-sensitized solar cell. Materials Today: Proceedings, 1–7. https://doi.org/
10.1016/j.matpr.2021.07.225
Kamaruddin, S., & Ravi, V. (2016). Credit card fraud detection using big data analytics: Use of PSOANN
based one-class classification. In Proceedings of the International Conference on Informatics and Ana-
lytics (pp. 1–8). ACM. https://doi.org/10.1145/2980258.2980319
Karlik, B., & Olgac, A. V. (2011). Performance analysis of various activation functions in generalized MLP
architectures of neural networks. International Journal of Artificial Intelligence and Expert Systems,
1(4), 111–122.
Kaur, H., & Salaria, D. S. (2013). Bayesian regularization based neural network tool for software effort esti-
mation. Global Journal of Computer Science and Technology, 13(2), 45–50.
Kazienko, P., Lughofer, E., & Trawiński, B. (2013). Hybrid and ensemble methods in machine learning J.
UCS special issue. Journal of Universal Computer Science, 19(4), 457–461.
Kennedy, J., & Eberhart, R. (1995). Particle swarm optimization. In Proceedings of ICNN’95— International
Conference on Neural Networks, 4, 1942–1948. https://doi.org/10.1109/ICNN.1995.488968
Khamis, A. B., & Kamarudin, N. K. K. B. (2014). Comparative study on estimate house price using statistical
and neural network model. International Journal of Scientific & Technology Research, 3(12), 126–131.
Kohavi, R. (1995). A study of cross-validation and bootstrap for accuracy estimation and model selec-
tion. In IJCAI’95: Proceedings of the 14th international joint conference on Artificial intelligence, 2,
1137–11452.
Kontrimas, V., & Verikas, A. (2011). The mass appraisal of the real estate by computational intelligence.
Applied Soft Computing, 11(1), 443–448. https://doi.org/10.1016/j.asoc.2009.12.003
Kouwenberg, R., & Zwinkels, R. (2014). Forecasting the US housing market. International Journal of Fore-
casting, 30(3), 415–425. https://doi.org/10.1016/j.ijforecast.2013.12.010
Li, J., Liu, X., Liu, J., & Li, W. (2016). City profile: Taipei. Cities, 55, 1–8. https://doi.org/10.1016/j.cities.
2016.03.007
Limsombunchai, V., Gan, C., & Lee, M. (2004). House price prediction: hedonic price model vs. artificial
neural network. American Journal of Applied Sciences, 1(3), 193-201. https://doi.org/10.3844/ajassp.
2004.193.201
Lin, S.-J. (2004). The marginal willingness-to-pay of star public elementary and junior high school districts
in Taipei City. Journal of Housing Studies, 13(1), 15–34. https://doi.org/10.6375/JHS.200406.0015
Liu, R., & Liu, L. (2019). Predicting housing price in China based on long short-term memory incorpo-
rating modified genetic algorithm. Soft Computing, 23(22), 11829-11839. https://doi.org/10.1007/
s00500-018-03739-w
Merlini, D., & Rossini, M. (2021). Text categorization with WEKA: A survey. Machine Learning with Appli-
cations, 4, 100033. https://doi.org/10.1016/j.mlwa.2021.100033
Methaprayoon, K., Yingvivatanapong, C., Lee, W.-J., & Liao, J. R. (2007). An integration of ANN wind
power estimation into unit commitment considering the forecasting uncertainty. IEEE Transactions on
Industry Applications, 43(6), 1441–1448. https://doi.org/10.1109/TIA.2007.908203
Muralidharan, S., Phiri, K., Sinha, S. K., & Kim, B. (2018). Analysis and prediction of real estate prices: A
case of the Boston housing market. Issues in Information Systems, 19(2), 109–118. https://doi.org/10.
48009/2_iis_2018_109-118
Núñez-Tabales, J., Rey Carmona, F., & Caridad, J. (2013). Implicit prices in urban real estate valuation.
Revista de la Construcción, 12(2), 116–126. https://doi.org/10.4067/S0718-915X2013000200009
Palma-Mendoza, R.-J., Rodriguez, D., & De-Marcos, L. (2018). Distributed ReliefF-based feature selection
in Spark. Knowledge and Information Systems, 57(1), 1–20. https://doi.org/10.1007/s10115-017-1145-y
Peter, S. E., & Raglend, I. J. (2017). Sequential wavelet-ANN with embedded ANN-PSO hybrid electricity
price forecasting model for Indian energy exchange. Neural Computing and Applications, 28(8), 2277–
2292. https://doi.org/10.1007/s00521-015-2141-3
Potter, K. (2006). Methods for presenting statistical information: The box plot. Visualization of Large and
Unstructured Data Sets, GI-Edition Lecture Notes in Informatics (LNI), S-4, 97–106.
13
Qian, B., & Rasheed, K. (2007). Stock market prediction with multiple classifiers. Applied Intelligence,
26(1), 25–33. https://doi.org/10.1007/s10489-006-0001-7
Rong, L. H., & Sun, Y. M. (2014). The analysis of second-hand housing price influencing factors based on
hedonic model and WEB information. Applied Mechanics and Materials, 587–589, 2285–2289. https://
doi.org/10.4028/www.scientific.net/AMM.587-589.2285
Rosen, S. (1974). Hedonic prices and implicit markets: Product differentiation in pure competition. Journal
of Political Economy, 82(1), 34–55. https://doi.org/10.1086/260169
Sarip, A. G., & Hafez, M. B. (2015). Fuzzy logic application for house price prediction. International Jour-
nal of Property Sciences, 5(1), 24–30. https://doi.org/10.22452/ijps.vol5no1.3
Schwertman, N. C., Owens, M. A., & Adnan, R. (2004). A simple more general boxplot method for identify-
ing outliers. Computational Statistics & Data Analysis, 47(1), 165–174. https://doi.org/10.1016/j.csda.
2003.10.012
Urbanowicz, R. J., Meeker, M., La Cava, W., Olson, R. S., & Moore, J. H. (2018). Relief-based feature selec-
tion: introduction and review. Journal of biomedical informatics, 85, 189–203. https://doi.org/10.1016/j.
jbi.2018.07.014
Vapnik, V. N. (1995). The nature of statistical learning theory. Springer. https://doi.org/10.1007/
978-1-4757-2440-0
Varma, A., Sarma, A., Doshi, S., & Nair, R. (2018). House price prediction using machine learning and neu-
ral networks. In 2018 Second International Conference on Inventive Communication and Computa-
tional Technologies (ICICCT) (pp. 1936–1939). IEEE. https://doi.org/10.1109/ICICCT.2018.8473231
Wang, L. V., & Yao, G. (2001). Ultrasound-modulated laser tomography. In Saratov Fall Meeting 2000:
Optical Technologies in Biophysics and Medicine II, 4241, pp. 1–5. Retrieved September 7, 2021, from
https://doi.org/10.1117/12.431526
Wang, X. (2011). The application of SPSS in empirical research of housing hedonic price. In 2011 Inter-
national Conference on Multimedia Technology (pp. 3262–3265). IEEE. https://doi.org/10.1109/ICMT.
2011.6003072
Wang, X., Wen, J., Zhang, Y., & Wang, Y. (2014). Real estate price forecasting based on SVM optimized by
PSO. Optik, 125(3), 1439–1443. https://doi.org/10.1016/j.ijleo.2013.09.017
Wang, Y., Chen, P.-C., Ma, H.-W., Cheng, K.-L., & Chang, C.-Y. (2018). Socio-economic metabolism of
urban construction materials: A case study of the Taipei metropolitan area. Resources, Conservation
and Recycling, 128, 563–571. https://doi.org/10.1016/j.resconrec.2016.08.019
Wei, W., Guang-ji, T., & Hong-rui, Z. (2010). Empirical analysis on the housing price in Harbin City based
on hedonic model. In 2010 International Conference on Management Science & Engineering 17th
Annual Conference Proceedings (pp. 1659–1664). IEEE. https://doi.org/10.1109/ICMSE.2010.5720005
Wen, H.-Z., Sheng-hua, J., & Xiao-yu, G. (2005). Hedonic price analysis of urban housing: An empirical
research on Hangzhou, China. Journal of Zhejiang University-Science A, 6(8), 907–914. https://doi.org/
10.1631/jzus.2005.A0907
Williamson, D. F., Parker, R. A., & Kendrick, J. S. (1989). The box plot: A simple visual method to interpret
data. Annals of Internal Medicine, 110(11), 916–921. https://doi.org/10.7326/0003-4819-110-11-916
Wolpert, D. H. (1992). Stacked generalization. Neural Networks, 5(2), 241–259. https://doi.org/10.1016/
S0893-6080(05)80023-1
Wu, J., Wang, M., Li, W., Peng, J., & Huang, L. (2015). Impact of urban green space on residential hous-
ing prices: Case study in Shenzhen. Journal of Urban Planning and Development, 141(4), 05014023.
https://doi.org/10.1061/(ASCE)UP.1943-5444.0000241
Wu, J. Y. (2017). Housing price prediction using support vector regression (pp. 1–56). San José State Univer-
sity. https://doi.org/10.31979/etd.vpub-6bgs
Xiao, Y. (2017). Urban morphology and housing market. Springer. https://doi.org/10.1007/
978-981-10-2762-8
Yang, C.-H. & Su, S.-H. (2011). The impacts of housing price in YIMBY and NIMBY. Journal of Housing
Studies, 20(2), 61–80. https://doi.org/10.6375/JHS.201112.0062
Yu, H., & Xiaohui, W. (2011). PSO-based energy-balanced double cluster-heads clustering routing for wire-
less sensor networks. Procedia Engineering, 15, 3073–3077. https://doi.org/10.1016/j.proeng.2011.08.
576
Zhang, H., & Zhang, M. (2010). Environment hedonic price analysis: Evidence from Jilin city. In 2010 Sec-
ond International Conference on Communication Systems, Networks and Applications (pp. 354–356).
IEEE. https://doi.org/10.1109/ICCSNA.2010.5588741
Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.
13

Comparison of Machine Learning Models To Provide Preliminary Forecasts of Real Estate Prices

Uploaded by

Copyright:

Available Formats

Comparison of Machine Learning Models To Provide Preliminary Forecasts of Real Estate Prices

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Comparison of Machine Learning Models To Provide Preliminary Forecasts of Real Estate Prices

Uploaded by

Copyright:

Available Formats

Journal of Housing and the Built Environment (2022) 37:2079–2114

Comparison of machine learning models to provide

Jui‑Sheng Chou1 · Dillon‑Brandon Fleshman1 · Dinh‑Nhat Truong1,2

Particularly, WEKA is aimed at practitioners who do not have a programming back-

2.1 Traditional methods for valuing real estate

2.2 Hedonic price theory

2.3 Use of machine learning in house price prediction

3.1 Data visualization and preprocessing

3.2 Baseline prediction models

3.2.1 Artificial neural networks

3.2.2 Support vector machine for classification and regression

where wi is a weight vector; gi (x) is a set of non-linear transformations; [ b is a bias

SVR implements an ε-insensitive loss function in the hyperplanes to concurrently mini-

3.2.3 Classification and regression tree

Classification and regression tree (CART) is a popular tree-structured analysis tech-

3.3 Ensemble prediction models

Bagging is an acronym for bootstrap aggregation, which is a universal ensemble method

Stacking is also called stacked generalization; it is a hierarchically structured ensemble

vi (t + 1) = 𝜑 × vi (t) + (c1 × rand() × (pbest

3.5 Model reliability and performance evaluation

where y = actual value; y′ = predicted value; and n = number of samples.

where y = actual value; y′ = predicted value; and n = number of samples.

where y = actual value; y′ = predicted value; and n = number of samples.

where y = actual value; y′ = predicted value; and n = number of samples.

where m = number of performance measures; Pi = ith performance measure.

4 House price dataset

4.1.1.10 Screening based on special transaction Article 23 of the Technical Regula-

5 Upper Inner Fence = 51,148,750

Fig. 1 Box plot of original dataset (Unit: NTD)

4.2 Feature description and acquisition

Fig. 2 Density of samples at each district of Taipei City

Table 1 Description of amenities and disamenities

4.2.1.3 Disamenity NIMBY Funeral parlor: Funeral ceremonies sometimes generate

Characteristics of house Area Ping

Amenity YIMBY MRT exit Distance in

Literature Region Response variable Feature

Table 4 Default parameter settings of filter-based algorithm in WEKA

ReliefFAttributeEval doNotCheckCapabilities False

Fig. 3 Feature correlation matrix for the house price dataset

Table 5 Ultimately filtered features for house price prediction

1 0.0713702 X1 Area 3.09155 132.9336 34.13144 14.9292

School: Some parents purchase houses in prestigious school districts. Closeness to

4.2.2 Feature selection using filter‑based algorithm

In feature selection, a filter-based algorithm, ReliefFAttributeEval, is used. WEKA, which

4.2.3 Distance measurement method

Table 7 Default parameter settings of baseline models in WEKA

Artificial neural network Name Multilayer perceptron

Table 8 Default parameter Ensemble model Parameter WEKA

Voting Name Voting

4.2.4 Location of residential school

Validation Training PSO

Determine Paramaters Initialization

Fig. 4 Architecture of hybrid model

5.1 Baseline prediction models

Fig. 5 Pseudocode of PSO-Bagging-ANNs

Table 9 Parameter settings of Algorithm Parameter Value

5.2 Ensemble prediction models

In WEKA, the implementation of ensemble models is analogous to that of single models.

5.3 Hybrid prediction model

2.1 Traditional methods for valuing real estate

2.2 Hedonic price theory

2.3 Use of machine learning in house price prediction

3.1 Data visualization and preprocessing

3.2 Baseline prediction models

3.2.1 Artificial neural networks

3.2.2 Support vector machine for classification and regression

3.2.3 Classification and regression tree

3.3 Ensemble prediction models

3.5 Model reliability and performance evaluation

4 House price dataset

4.2 Feature description and acquisition

4.2.2 Feature selection using filter‑based algorithm

4.2.3 Distance measurement method

4.2.4 Location of residential school

5.1 Baseline prediction models

5.2 Ensemble prediction models

5.3 Hybrid prediction model

Voting ANNs + CART 0.956 3,930,235 2,697,942 14.02 0.141 (4)

6 Model performance and discussion

6.1 Performance in baseline models

6.2 Performance in ensemble models

6.3 Performance of hybrid model

6.4 Comprehensive comparison and discussion

7 Conclusion and recommendation