House Price Prediction

House Price Prediction Using Machine Learning Techniques
Ammar Alyousfi
2018
December
1
Contents
1 Introduction 4
1.1 Goals of the Study . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.2 Paper Organization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2 Literature Review 4
2.1 Stock Market Prediction Using Bayesian-Regularized Neural Networks . . . . . . . . 4
2.2 Stock Market Prediction Using A Machine Learning Model . . . . . . . . . . . . . . . 5
2.3 House Price Prediction Using Multilevel Model and Neural Networks . . . . . . . . 6
2.4 Composition of Models and Feature Engineering to Win Algorithmic Trading Chal-
lenge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.5 Using K-Nearest Neighbours for Stock Price Prediction . . . . . . . . . . . . . . . . . 8
3 Data Preparation 10
3.1 Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
3.2 Reading the Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.3 Getting A Feel of the Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12
3.4 Data Cleaning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.4.1 Dealing with Missing Values . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.5 Outlier Removal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.6 Deleting Some Unimportant Columns . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4 Exploratory Data Analysis 27

4.1 Target Variable Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
4.2 Correlation Between Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
4.2.1 Relatioships Between the Target Variable and Other Varibles . . . . . . . . . . 30
4.2.2 Relatioships Between Predictor Variables . . . . . . . . . . . . . . . . . . . . . 37
4.3 Feature Engineering . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.3.1 Creating New Derived Features . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.3.2 Dealing with Ordinal Variables . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.3.3 One-Hot Encoding For Categorical Features . . . . . . . . . . . . . . . . . . . 42
5 Prediction Type and Modeling Techniques 43

5.0.1 1. Linear Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43
5.0.2 2. Nearest Neighbors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.0.3 3. Support Vector Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.0.4 4. Decision Trees . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.0.5 5. Neural Networks . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
5.0.6 6. Random Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
5.0.7 7. Gradient Boosting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6 Model Building and Evaluation 45

6.1 Feature Scaling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
6.2 Splitting the Dataset . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
6.3 Modeling Approach . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
6.3.1 Searching for Effective Parameters . . . . . . . . . . . . . . . . . . . . . . . . . 47
6.4 Performance Metric . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
6.5 Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
2
6.5.1 Linear Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
6.5.2 Nearest Neighbors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
6.5.3 Support Vector Regression . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
6.5.4 Decision Tree . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 52
6.5.5 Neural Network . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
6.5.6 Random Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
6.5.7 Gradient Boosting . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
7 Analysis and Comparison 57

7.1 Performance Interpretation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 58
7.2 Feature Importances . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
7.2.1 XGBoost . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
7.2.2 Random Forest . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
7.2.3 Common Important Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . 63
8 Conclusion 65
9 References 65
3
1 Introduction
Thousands of houses are sold everyday. There are some questions every buyer asks himself like:
What is the actual price that this house deserves? Am I paying a fair price? In this paper, a machine
learning model is proposed to predict a house price based on data related to the house (its size,
the year it was built in, etc.). During the development and evaluation of our model, we will show
the code used for each step followed by its output. This will facilitate the reproducibility of our
work. In this study, Python programming language with a number of Python packages will be
used.
1.1 Goals of the Study

The main objectives of this study are as follows:
• To apply data preprocessing and preparation techniques in order to obtain clean data
• To build machine learning models able to predict house price based on house features
• To analyze and compare models performance in order to choose the best model
1.2 Paper Organization

This paper is organized as follows: in the next section, section 2, we examine studies related to
our work from scientific journals. In section 3, we go through data preparation including data
cleaning, outlier removal, and feature engineering. Next in section 4, we discuss the type of our
problem and the type of machine-learning prediction that should be applied; we also list the pre-
diction techniques that will be used. In section 5, we choose algorithms to implement the tech-
niques in section 4; we build models based on these algorithms; we also train and test each model.
In section 6, we analyze and compare the results we got from section 5 and conclude the paper.
2 Literature Review
In this section, we look at five recent studies that are related to our topic and see how models were
built and what results were achieved in these studies.
2.1 Stock Market Prediction Using Bayesian-Regularized Neural Networks

In a study done by Ticknor (2013), he used Bayesian regularized articial neural network to predict
the future operation of financial market. Specifically, he built a model to predict future stock
prices. The input of the model is previous stock statistics in addition to some financial technical
data. The output of the model is the next-day closing price of the corresponding stocks.
The model proposed in the study is built using Bayesian regularized neural network. The
weights of this type of networks are given a probabilistic nature. This allows the network to
penalize very complex models (with many hidden layers) in an automatic manner. This in turn
will reduce the overfitting of the model.
The model consists of a feedforward neural network which has three layers: an input layer, one
hidden layer, and an output layer. The author chose the number of neurons in the hidden layer
based on experimental methods.The input data of the model is normalized to be between -1 and
1, and this opertion is reversed for the output so the predicted price appears in the appropriate
scale.
4
Figure 1: Predicted vs. actual price
The data that was used in this study was obtained from Goldman Sachs Group (GS), Inc. and
Microsoft Corp. (MSFT) . The data covers 734 trading days (4 January 2010 to 31 December 2012).
Each instance of the data consisted of daily statistics: low price, high price, opening price, close
price, and trading volume. To facilitate the training and testing of the model, this data was split
into training data and test data with 80% and 20% of the original data, respectively. In addition
to the daily-statistics variables in the data, six more variables were created to reflect financial
indicators.
The performance of the model were evaluated using mean absolute percentage error (MAPE)
performance metric. MAPE was calculated using this formula:
∑ri=1 (abs(yi − pi )/yi )

MAPE = × 100 (1)
r
where pi is the predicted stock price on day i, yi is the actual stock price on day i, and r is the
number of trading days.
When applied on the test data, The model achieved a MAPE score of 1.0561 for MSFT part,
and 1.3291 for GS part. Figure 1 shows the actual values and predicted values for both GS and
MSFT data.
2.2 Stock Market Prediction Using A Machine Learning Model

In another study done by Hegazy, Soliman, and Salam (2014), a system was proposed to predict
daily stock market prices. The system combines particle swarm optimization (PSO) and least
square support vector machine (LS-SVM), where PSO was used to optimize LV-SVM.
The authors claim that in most cases, artificial neural networks (ANNs) are subject to the over-
fitting problem. They state that support vector machines algorithm (SVM) was developed as an
alternative that doesn’t suffer from overfitting. They attribute this advantage to SVMs being based
on the solid foundations of VC-theory. They further elaborate that LS-SVM method was refor-
mulation of traditional SVM method that uses a regularized least squares function with equality
5
Figure 2: The structure of the model used
constraints to obtain a linear system that satisfies Karush-Kuhn-Tucker conditions for getting an
optimal solution.
The authors describe PSO as a popular evolutionary optimization method that was inspired
by organism social behavior like bird flocking. They used it to find the optimal parameters for LS-
SVM. These parameters are the cost penalty C, kernel parameter γ, and insensitive loss function
ϵ.
The model proposed in the study was based on the analysis of historical data and technical
financial indicators and using LS-SVM optimized by PSO to predict future daily stock prices. The
model input was six vectors representing the historical data and the technical financial indicators.
The model output was the future price. The model used is represented in Figure 2.
Regarding the technical financial indicators, five were derived from the raw data: relative
strength index (RSI), money flow index (MFI), exponential moving average (EMA), stochastic os-
cillator (SO), and moving average convergence/divergence (MACD). These indicators are known
in the domain of stock market.
The model was trained and tested using datasets taken from https://finance.yahoo.com/. The
datasets were from Jan 2009 to Jan 2012 and include stock data for many companies like Adobe
and HP. All datasets were partitioned into a training set with 70% of the data and a test set with
30% of the data. Three models were trained and tested: LS-SVM-PSO model, LS-SVM model, and
ANN model. The results obtained in the study showed that LS-SVM-PSO model had the best
performance. Figure 3 shows a comparison between the mean square error (MSE) of the three
models for the stocks of many companies.
2.3 House Price Prediction Using Multilevel Model and Neural Networks
A different study was done by Feng and Jones (2015) to preduct house prices. Two models were
built: a multilevel model (MLM) and an artificial neural network model (ANN). These two models
were compared to each other and to a hedonic price model (HPM).
The multilevel model integrates the micro-level that specifies the relationships between houses
within a given neighbourhood, and the macro-level equation which specifies the relationships
between neighbouhoods. The hedonic price model is a model that estimates house prices using
6
Figure 3: MSE comparison
some attributes such as the number of bedrooms in the house, the size of the house, etc.
The data used in the study contains house prices in Greater Bristol area between 2001 and 2013.
Secondary data was obtained from the Land Registry, the Population Census and Neighbourhood
Statistics to be used in order to make the models suitable for national usage. The authors listed
many reasons on why they chose the Greater Bristol area such as its diverse urban and rural blend
and its different property types. Each record in the dataset contains data about a house in the area:
it contains the address, the unit postcode, property type, the duration (freehold or leasehold), the
sale price, the date of the sale, and whether the house was newly-built when it was sold. In total,
the dataset contains around 65,000 entries. To enable model training and testing, the dataset was
divided into a training set that contains data about house sales from 2001 to 2012, and a test set
that contains data about house sales in 2013.
The three models (MLM, ANN, and HPM) were tested using three senarios. In the first senario,
locational and measured neighbourhood attributes were not included in the data. In the second
senario, grid references of house location were included in the data. In the third senario, measured
neighbourhood attributes were included in the data. The models were compared in goodness of
fit where R2 was the metric, predictive accuracy where mean absolute error (MAE) and mean
absolute percentage error (MAPE) were the metrics, and explanatory power. HPM and MLM
models were fitted using MLwiN software, and ANN were fitted using IBM SPSS software. Figure
4 shows the performance of each model regarding fit goodness and predictive accuracy. It shows
that MLM model has better performance in general than other models.
7
Figure 4: Model performance comparison
2.4 Composition of Models and Feature Engineering to Win Algorithmic Trading

Challenge
A study done by de Abril and Sugiyama (2013) introduced the techniques and ideas used to win
Algorithmic Trading Challenge, a competition held on Kaggle. The goal of the competition was
to develop a model that can predict the short-term response of order-driven markets after a big
liquidity shock. A liquidity shock happens when a trade or a sequence of trades causes an acute
shortage of liquidity (cash for example).
The challenge data contains a training dataset and a test dataset. The training dataset has
around 754,000 records of trade and quote observations for many securities of London Stock
Exchange before and after a liquidity shock. A trade event happens when shares are sold or
bought, whereas a quote event happens when the ask price or the best bid changes.
A separate model was built for bid and another for ask. Each one of these models consists of K
random-forest sub-models. The models predict the price at a particular future time.
The authors spent much effort on feature engineering. They created more than 150 features.
These features belong to four categories: price features, liquidity-book features, spread features
(bid/ask spread), and rate features (arrival rate of orders/quotes). They applied a feature selection
algorithm to obtain the optimal feature set (Fb ) for bid sub-models and the optimal feature set (Fa )
of all ask sub-models. The algorithm applied eliminates features in a backward manner in order
to get a feature set with reasonable computing time and resources.
Three instances of the final model proposed in the study were trained on three datasets; each
one of them consists of 50,000 samples sampled randomly from the training dataset. Then, the
three models were applied to the test dataset. The predictions of the three models were then
averaged to obtain the final prediction. The proposed method achieved a RMSE score of 0.77
approximately.
2.5 Using K-Nearest Neighbours for Stock Price Prediction

Alkhatib, Najadat, Hmeidi, and Shatnawi (2013) have done a study where they used the k-nearest
neighbours (KNN) algorithm to predict stock prices. In this study, they expressed the stock pre-
diction problem as a similarity-based classification, and they represented the historical stock data
as well as test data by vectors.
The authors listed the steps of predicting the closing price of stock market using KNN as
follows:
• The number of neaerest neighbours is chosen
8
Figure 5: Prediction performance evaluation
Figure 6: AIEI lift graph
• The distance between the new record and the training data is computed
• Training data is sorted according to the calculated distance
• Majority voting is applied to the classes of the k nearest neighbours to determine the pre-
dicted value of the new record
The data used in the study is stock data of five companies listed on the Jordanian stock ex-
change. The data range is from 4 June 2009 to 24 December 2009. Each of the five companies has
around 200 records in the data. Each record has three variables: closing price, low price, and high
price. The author stated that the closing price is the most important feature in determining the
prediction value of a stock using KNN.
After applying KNN algorithm, the authors summarized the prediction performance evalua-
tion using different metrics in a the table shown in Figure 5.
The authors used lift charts also to evaluate the performance of their model. Lift chart shows
the improvement obtained by using the model compared to random estimation. As an example,
the lift graph for AIEI company is shown in Figure 6. The area between the two lines in the graph
is an indicator of the goodness of the model.
Figure 7 shows the relationship between the actual price and predicted price for one year for
the same company.
9
Figure 7: Relationship between actual and predicted price for AIEI
3 Data Preparation
In this study, we will use a housing dataset presented by De Cock (2011). This dataset describes
the sales of residential units in Ames, Iowa starting from 2006 until 2010. The dataset contains a
large number of variables that are involved in determining a house price. We obtained a csv copy
of the data from https://www.kaggle.com/prevek18/ames-housing-dataset.
3.1 Data Description

The dataset contains 2930 records (rows) and 82 features (columns).
Here, we will provide a brief description of dataset features. Since the number of features is
large (82), we will attach the original data description file to this paper for more information about
the dataset (It can be downloaded also from https://www.kaggle.com/c/house-prices-advanced-
regression-techniques/data). Now, we will mention the feature name with a short description of
its meaning.
Feature Description
MSSubClass The type of the house involved in the sale
MSZoning The general zoning classification of the sale
LotFrontage Linear feet of street connected to the house
LotArea Lot size in square feet
Street Type of road access to the house
Alley Type of alley access to the house
LotShape General shape of the house
LandContour House flatness
Utilities Type of utilities available
LotConfig Lot configuration
LandSlope House Slope
Neighborhood Locations within Ames city limits
10
Feature Description
Condition1 Proximity to various conditions
Condition2 Proximity to various conditions (if more than one is
present)
BldgType House type
HouseStyle House style
OverallQual Overall quality of material and finish of the house
OverallCond Overall condition of the house
YearBuilt Construction year
YearRemodAdd Remodel year (if no remodeling nor addition, same as
YearBuilt)
RoofStyle Roof type
RoofMatl Roof material
Exterior1st Exterior covering on house
Exterior2nd Exterior covering on house (if more than one material)
MasVnrType Type of masonry veneer
MasVnrArea Masonry veneer area in square feet
ExterQual Quality of the material on the exterior
ExterCond Condition of the material on the exterior
Foundation Foundation type
BsmtQual Basement height
BsmtCond Basement Condition
BsmtExposure Refers to walkout or garden level walls
BsmtFinType1 Rating of basement finished area
BsmtFinSF1 Type 1 finished square feet
BsmtFinType2 Rating of basement finished area (if multiple types)
BsmtFinSF2 Type 2 finished square feet
BsmtUnfSF Unfinished basement area in square feet
TotalBsmtSF Total basement area in square feet
Heating Heating type
HeatingQC Heating quality and condition
CentralAir Central air conditioning
Electrical Electrical system type
1stFlrSF First floor area in square feet
2ndFlrSF Second floor area in square feet
LowQualFinSF Low quality finished square feet in all floors
GrLivArea Above-ground living area in square feet
BsmtFullBath Basement full bathrooms
BsmtHalfBath Basement half bathrooms
FullBath Full bathrooms above ground
HalfBath Half bathrooms above ground
Bedroom Bedrooms above ground
Kitchen Kitchens above ground
KitchenQual Kitchen quality
TotRmsAbvGrd Total rooms above ground (excluding bathrooms)
Functional Home functionality
Fireplaces Number of fireplaces
11
Feature Description
FireplaceQu Fireplace quality
GarageType Garage location
GarageYrBlt Year garage was built in
GarageFinish Interior finish of the garage
GarageCars Size of garage (in car capacity)
GarageArea Garage size in square feet
GarageQual Garage quality
GarageCond Garage condition
PavedDrive How driveway is paved
WoodDeckSF Wood deck area in square feet
OpenPorchSF Open porch area in square feet
EnclosedPorch Enclosed porch area in square feet
3SsnPorch Three season porch area in square feet
ScreenPorch Screen porch area in square feet
PoolArea Pool area in square feet
PoolQC Pool quality
Fence Fence quality
MiscFeature Miscellaneous feature
MiscVal Value of miscellaneous feature
MoSold Sale month
YrSold Sale year
SaleType Sale type
SaleCondition Sale condition
3.2 Reading the Dataset

The first step is reading the dataset from the csv file we downloaded. We will use the read_csv()
function from Pandas Python package:
import pandas as pd
import numpy as np
dataset = pd.read_csv("AmesHousing.csv")
3.3 Getting A Feel of the Dataset

Let’s display the first few rows of the dataset to get a feel of it:
# Configuring float numbers format

pd.options.display.float_format = '{:20.2f}'.format
dataset.head(n=5)
12
Order PID MS SubClass MS Zoning Lot Frontage Lot Area
1 526301100 20 RL 141.00 31770
2 526350040 20 RH 80.00 11622
3 526351010 20 RL 81.00 14267
4 526353030 20 RL 93.00 11160
5 527105010 60 RL 74.00 13830
Street Alley Lot Shape Land Contour Utilities Lot Config

Pave NaN IR1 Lvl AllPub Corner
Pave NaN Reg Lvl AllPub Inside
Pave NaN IR1 Lvl AllPub Corner
Pave NaN Reg Lvl AllPub Corner
Pave NaN IR1 Lvl AllPub Inside
Land Slope Neighborhood Condition 1 Condition 2 Bldg Type House Style

Gtl NAmes Norm Norm 1Fam 1Story
Gtl NAmes Feedr Norm 1Fam 1Story
Gtl Gilbert Norm Norm 1Fam 2Story
Overall Qual Overall Cond Year Built Year Remod/Add Roof Style Roof Matl
6 5 1960 1960 Hip CompShg
5 6 1961 1961 Gable CompShg
5 5 1997 1998 Gable CompShg
Exterior 1st Exterior 2nd Mas Vnr Type Mas Vnr Area Exter Qual Exter Cond
BrkFace Plywood Stone 112.00 TA TA
VinylSd VinylSd None 0.00 TA TA
Wd Sdng Wd Sdng BrkFace 108.00 TA TA
BrkFace BrkFace None 0.00 Gd TA
VinylSd VinylSd None 0.00 TA TA
13
Foundation Bsmt Qual Bsmt Cond Bsmt Exposure BsmtFin Type 1 BsmtFin SF 1
CBlock TA Gd Gd BLQ 639.00
CBlock TA TA No Rec 468.00
CBlock TA TA No ALQ 923.00
CBlock TA TA No ALQ 1065.00
PConc Gd TA No GLQ 791.00
BsmtFin Type 2 BsmtFin SF 2 Bsmt Unf SF Total Bsmt SF Heating Heating QC

Unf 0.00 441.00 1080.00 GasA Fa
LwQ 144.00 270.00 882.00 GasA TA
Unf 0.00 406.00 1329.00 GasA TA
Unf 0.00 1045.00 2110.00 GasA Ex
Unf 0.00 137.00 928.00 GasA Gd
Central Air Electrical 1st Flr SF 2nd Flr SF Low Qual Fin SF Gr Liv Area
Y SBrkr 1656 0 0 1656
Y SBrkr 896 0 0 896
Y SBrkr 1329 0 0 1329
Y SBrkr 2110 0 0 2110
Y SBrkr 928 701 0 1629
Bsmt Full Bath Bsmt Half Bath Full Bath Half Bath Bedroom AbvGr Kitchen AbvGr
1.00 0.00 1 0 3 1
0.00 0.00 1 0 2 1
0.00 0.00 1 1 3 1
1.00 0.00 2 1 3 1
0.00 0.00 2 1 3 1
Kitchen Qual TotRms AbvGrd Functional Fireplaces Fireplace Qu Garage Type

TA 7 Typ 2 Gd Attchd
TA 5 Typ 0 NaN Attchd
Gd 6 Typ 0 NaN Attchd
Ex 8 Typ 2 TA Attchd
TA 6 Typ 1 TA Attchd
14
Garage Yr Blt Garage Finish Garage Cars Garage Area Garage Qual Garage Cond
1960.00 Fin 2.00 528.00 TA TA
1961.00 Unf 1.00 730.00 TA TA
1958.00 Unf 1.00 312.00 TA TA
1968.00 Fin 2.00 522.00 TA TA
1997.00 Fin 2.00 482.00 TA TA
Paved Drive Wood Deck SF Open Porch SF Enclosed Porch 3Ssn Porch Screen Porch
P 210 62 0 0 0
Y 140 0 0 0 120
Y 393 36 0 0 0
Y 0 0 0 0 0
Y 212 34 0 0 0
Pool Area Pool QC Fence Misc Feature Misc Val Mo Sold

0 NaN NaN NaN 0 5
0 NaN MnPrv NaN 0 6
0 NaN NaN Gar2 12500 6
0 NaN NaN NaN 0 4
0 NaN MnPrv NaN 0 3
Yr Sold Sale Type Sale Condition SalePrice

2010 WD Normal 215000
2010 WD Normal 105000
2010 WD Normal 172000
2010 WD Normal 244000
2010 WD Normal 189900
Now, let’s get statistical information about the numeric columns in our dataset. We want to
know the mean, the standard deviation, the minimum, the maximum, and the 50th percentile (the
median) for each numeric column in the dataset:
dataset.describe(include=[np.number], percentiles=[.5]) \
.transpose().drop("count", axis=1)
15
mean std min 50% max
Order 1465.50 845.96 1.00 1465.50 2930.00
PID 714464496.99 188730844.65 526301100.00 535453620.00 1007100110.00
MS SubClass 57.39 42.64 20.00 50.00 190.00
Lot Frontage 69.22 23.37 21.00 68.00 313.00
Lot Area 10147.92 7880.02 1300.00 9436.50 215245.00
Overall Qual 6.09 1.41 1.00 6.00 10.00
Overall Cond 5.56 1.11 1.00 5.00 9.00
Year Built 1971.36 30.25 1872.00 1973.00 2010.00
Year Remod/Add 1984.27 20.86 1950.00 1993.00 2010.00
Mas Vnr Area 101.90 179.11 0.00 0.00 1600.00
BsmtFin SF 1 442.63 455.59 0.00 370.00 5644.00
BsmtFin SF 2 49.72 169.17 0.00 0.00 1526.00
Bsmt Unf SF 559.26 439.49 0.00 466.00 2336.00
Total Bsmt SF 1051.61 440.62 0.00 990.00 6110.00
1st Flr SF 1159.56 391.89 334.00 1084.00 5095.00
2nd Flr SF 335.46 428.40 0.00 0.00 2065.00
Low Qual Fin SF 4.68 46.31 0.00 0.00 1064.00
Gr Liv Area 1499.69 505.51 334.00 1442.00 5642.00
Bsmt Full Bath 0.43 0.52 0.00 0.00 3.00
Bsmt Half Bath 0.06 0.25 0.00 0.00 2.00
Full Bath 1.57 0.55 0.00 2.00 4.00
Half Bath 0.38 0.50 0.00 0.00 2.00
Bedroom AbvGr 2.85 0.83 0.00 3.00 8.00
Kitchen AbvGr 1.04 0.21 0.00 1.00 3.00
TotRms AbvGrd 6.44 1.57 2.00 6.00 15.00
Fireplaces 0.60 0.65 0.00 1.00 4.00
Garage Yr Blt 1978.13 25.53 1895.00 1979.00 2207.00
Garage Cars 1.77 0.76 0.00 2.00 5.00
Garage Area 472.82 215.05 0.00 480.00 1488.00
Wood Deck SF 93.75 126.36 0.00 0.00 1424.00
Open Porch SF 47.53 67.48 0.00 27.00 742.00
Enclosed Porch 23.01 64.14 0.00 0.00 1012.00
3Ssn Porch 2.59 25.14 0.00 0.00 508.00
Screen Porch 16.00 56.09 0.00 0.00 576.00
Pool Area 2.24 35.60 0.00 0.00 800.00
Misc Val 50.64 566.34 0.00 0.00 17000.00
Mo Sold 6.22 2.71 1.00 6.00 12.00
Yr Sold 2007.79 1.32 2006.00 2008.00 2010.00
SalePrice 180796.06 79886.69 12789.00 160000.00 755000.00
From the table above, we can see, for example, that the average lot area of the houses in our
dataset is 10,147.92 ft2 with a standard deviation of 7,880.02 ft2. We can see also that the minimum
lot area is 1,300 ft2 and the maximum lot area is 215,245 ft2 with a median of 9,436.5 ft2. Similarly,
we can get a lot of information about our dataset variables from the table.
16
Then, we move to see statistical information about the non-numerical columns in our dataset:
dataset.describe(include=[np.object]).transpose() \
.drop("count", axis=1)
unique top freq

MS Zoning 7 RL 2273
Street 2 Pave 2918
Alley 2 Grvl 120
Lot Shape 4 Reg 1859
Land Contour 4 Lvl 2633
Utilities 3 AllPub 2927
Lot Config 5 Inside 2140
Land Slope 3 Gtl 2789
Neighborhood 28 NAmes 443
Condition 1 9 Norm 2522
Condition 2 8 Norm 2900
Bldg Type 5 1Fam 2425
House Style 8 1Story 1481
Roof Style 6 Gable 2321
Roof Matl 8 CompShg 2887
Exterior 1st 16 VinylSd 1026
Exterior 2nd 17 VinylSd 1015
Mas Vnr Type 5 None 1752
Exter Qual 4 TA 1799
Exter Cond 5 TA 2549
Foundation 6 PConc 1310
Bsmt Qual 5 TA 1283
Bsmt Cond 5 TA 2616
Bsmt Exposure 4 No 1906
BsmtFin Type 1 6 GLQ 859
BsmtFin Type 2 6 Unf 2499
Heating 6 GasA 2885
Heating QC 5 Ex 1495
Central Air 2 Y 2734
Electrical 5 SBrkr 2682
Kitchen Qual 5 TA 1494
Functional 8 Typ 2728
Fireplace Qu 5 Gd 744
Garage Type 6 Attchd 1731
Garage Finish 3 Unf 1231
Garage Qual 5 TA 2615
Garage Cond 5 TA 2665
Paved Drive 3 Y 2652
Pool QC 4 Gd 4
Fence 4 MnPrv 330
17
unique top freq
Misc Feature 5 Shed 95
Sale Type 10 WD 2536
Sale Condition 6 Normal 2413
In the table we got, count represents the number of non-null values in each column, unique
represents the number of unique values, top represents the most frequent element, and freq rep-
resents the frequency of the most frequent element.
3.4 Data Cleaning

3.4.1 Dealing with Missing Values
We should deal with the problem of missing values because some machine learning models don’t
accept data with missing values. Firstly, let’s see the number of missing values in our dataset.
We want to see the number and the percentage of missing values for each column that actually
contains missing values.
# Getting the number of missing values in each column

num_missing = dataset.isna().sum()
# Excluding columns that contains 0 missing values
num_missing = num_missing[num_missing > 0]
# Getting the percentages of missing values
percent_missing = num_missing * 100 / dataset.shape[0]
# Concatenating the number and perecentage of missing values
# into one dataframe and sorting it
pd.concat([num_missing, percent_missing], axis=1,
keys=['Missing Values', 'Percentage']).\
sort_values(by="Missing Values", ascending=False)
Missing Values Percentage

Pool QC 2917 99.56
Misc Feature 2824 96.38
Alley 2732 93.24
Fence 2358 80.48
Fireplace Qu 1422 48.53
Lot Frontage 490 16.72
Garage Cond 159 5.43
Garage Qual 159 5.43
Garage Finish 159 5.43
Garage Yr Blt 159 5.43
Garage Type 157 5.36
Bsmt Exposure 83 2.83
BsmtFin Type 2 81 2.76
18
Missing Values Percentage
BsmtFin Type 1 80 2.73
Bsmt Qual 80 2.73
Bsmt Cond 80 2.73
Mas Vnr Area 23 0.78
Mas Vnr Type 23 0.78
Bsmt Half Bath 2 0.07
Bsmt Full Bath 2 0.07
Total Bsmt SF 1 0.03
Bsmt Unf SF 1 0.03
Garage Cars 1 0.03
Garage Area 1 0.03
BsmtFin SF 2 1 0.03
BsmtFin SF 1 1 0.03
Electrical 1 0.03
Now we start dealing with these missing values.
Pool QC The percentage of missing values in Pool QC column is 99.56% which is very high.
We think that a missing value in this column denotes that the corresponding house doesn’t have
a pool. To verify this, let’s take a look at the values of Pool Area column:
dataset["Pool Area"].value_counts()
Pool Area
0 2917
561 1
555 1
519 1
800 1
738 1
648 1
576 1
512 1
480 1
444 1
368 1
228 1
144 1
We can see that there are 2917 entries in Pool Area column that have a value of 0. This verfies
our hypothesis that each house without a pool has a missing value in Pool QC column and a value
of 0 in Pool Area column. So let’s fill the missing values in Pool QC column with "No Pool":
dataset["Pool QC"].fillna("No Pool", inplace=True)
19
Misc Feature The percentage of missing values in Pool QC column is 96.38% which is very
high also. Let’s take a look at the values of Misc Val column:
dataset["Misc Val"].value_counts()
Misc Val
0 2827
400 18
500 13
450 9
600 8
700 7
2000 7
650 3
1200 3
1500 3
4500 2
2500 2
480 2
3000 2
12500 1
300 1
350 1
8300 1
420 1
80 1
54 1
460 1
490 1
3500 1
560 1
17000 1
15500 1
750 1
800 1
900 1
1000 1
1150 1
1300 1
1400 1
1512 1
6500 1
455 1
620 1
20
We can see that Misc Val column has 2827 entries with a value of 0. Misc Feature has 2824
missing values. Then, as with Pool QC, we can say that each house without a “miscellaneous
feature” has a missing value in Misc Feature column and a value of 0 in Misc Val column. So
let’s fill the missing values in Misc Feature column with "No Feature":
dataset['Misc Feature'].fillna('No feature', inplace=True)
Alley, Fence, and Fireplace Qu According to the dataset documentation, NA in Alley, Fence,
and Fireplace Qu columns denotes that the house doesn’t have an alley, fence, or fireplace. So
we fill in the missing values in these columns with "No Alley", "No Fence", and "No Fireplace"
accordingly:
dataset['Alley'].fillna('No Alley', inplace=True)

dataset['Fence'].fillna('No Fence', inplace=True)
dataset['Fireplace Qu'].fillna('No Fireplace', inplace=True)
Lot Frontage As we saw previously, Lot Frontage represents the linear feet of street con-
nected to the house. So we assume that the missing values in this column indicates that the house
is not connected to any street, and we fill in the missing values with 0:
dataset['Lot Frontage'].fillna(0, inplace=True)
Garage Cond, Garage Qual, Garage Finish, Garage Yr Blt, Garage Type, Garage Cars, and
Garage Area According to the dataset documentation, NA in Garage Cond, Garage Qual, Garage
Finish, and Garage Type indicates that there is no garage in the house. So we fill in the missing
values in these columns with "No Garage". We notice that Garage Cond, Garage Qual, Garage
Finish, Garage Yr Blt columns have 159 missing values, but Garage Type has 157 and both
Garage Cars and Garage Area have one missing value. Let’s take a look at the row that contains
the missing value in Garage Cars:
garage_columns = [col for col in dataset.columns if col.startswith("Garage")]

dataset[dataset['Garage Cars'].isna()][garage_columns]
Garage Type Garage Yr Blt Garage Finish Garage Cars

2236 Detchd nan NaN nan
Garage Area Garage Qual Garage Cond

2236 nan NaN NaN
We can see that this is the same row that contains the missing value in Garage Area, and that
all garage columns except Garage Type are null in this row, so we will fill the missing values in
Garage Cars and Garage Area with 0.
We saw that there are 2 rows where Garage Type is not null while Garage Cond, Garage Qual,
Garage Finish, and Garage Yr Blt columns are null. Let’s take a look at these two rows:
21
dataset[~pd.isna(dataset['Garage Type']) &
pd.isna(dataset['Garage Qual'])][garage_columns]
Garage Type Garage Yr Blt Garage Finish Garage Cars

1356 Detchd nan NaN 1.00
2236 Detchd nan NaN nan
Garage Area Garage Qual Garage Cond

1356 360.00 NaN NaN
2236 nan NaN NaN
We will replace the values of Garage Type with "No Garage" in these two rows also.
For Garage Yr Blt, we will fill in missing values with 0 since this is a numerical column:
dataset['Garage Cars'].fillna(0, inplace=True)

dataset['Garage Area'].fillna(0, inplace=True)
dataset.loc[~pd.isna(dataset['Garage Type']) &

pd.isna(dataset['Garage Qual']), "Garage Type"] = "No Garage"
for col in ['Garage Type', 'Garage Finish', 'Garage Qual', 'Garage Cond']:
dataset[col].fillna('No Garage', inplace=True)
dataset['Garage Yr Blt'].fillna(0, inplace=True)
Bsmt Exposure, BsmtFin Type 2, BsmtFin Type 1, Bsmt Qual, Bsmt Cond, Bsmt Half Bath,
Bsmt Full Bath, Total Bsmt SF, Bsmt Unf SF, BsmtFin SF 2, and BsmtFin SF 1 According to
the dataset documentation, NA in any of the first five of these columns indicates that there is no
basement in the house. So we fill in the missing values in these columns with "No Basement".
We notice that the first five of these columns have 80 missing values, but BsmtFin Type 2 has 81,
Bsmt Exposure has 83, Bsmt Half Bath and Bsmt Full Bath each has 2, and each of the others
has 1. Let’s take a look at the rows where Bsmt Half Bath is null:
bsmt_columns = [col for col in dataset.columns if "Bsmt" in col]

dataset[dataset['Bsmt Half Bath'].isna()][bsmt_columns]
Bsmt Qual Bsmt Cond Bsmt Exposure BsmtFin Type 1 BsmtFin SF 1

1341 NaN NaN NaN NaN nan
1497 NaN NaN NaN NaN 0.00
22
BsmtFin Type 2 BsmtFin SF 2 Bsmt Unf SF Total Bsmt SF
1341 NaN nan nan nan
1497 NaN 0.00 0.00 0.00
Bsmt Full Bath Bsmt Half Bath

1341 nan nan
1497 nan nan
We can see that these are the same rows that contain the missing values in Bsmt Full Bath,
and that one of these two rows is contains the missing value in each of Total Bsmt SF, Bsmt Unf
SF, BsmtFin SF 2, and BsmtFin SF 1 columns. We notice also that Bsmt Exposure, BsmtFin Type
2, BsmtFin Type 1, Bsmt Qual, and Bsmt Cond are null in these rows, so we will fill the missing
values in Bsmt Half Bath, Bsmt Full Bath, Total Bsmt SF, Bsmt Unf SF, BsmtFin SF 2, and
BsmtFin SF 1 columns with 0.
We saw that there are 3 rows where Bsmt Exposure is null while BsmtFin Type 1, Bsmt Qual,
and Bsmt Cond are not null. Let’s take a look at these three rows:
dataset[~pd.isna(dataset['Bsmt Cond']) &

pd.isna(dataset['Bsmt Exposure'])][bsmt_columns]

66 Gd TA NaN Unf 0.00

66 Unf 0.00 1595.00 1595.00
1796 Unf 0.00 725.00 725.00
2779 Unf 0.00 936.00 936.00

66 0.00 0.00
1796 0.00 0.00
2779 0.00 0.00
We will fill in the missing values in Bsmt Exposure for these three rows with "No". According
to the dataset documentation, "No" for Bsmt Exposure means “No Exposure”:
23
Let’s now take a look at the row where BsmtFin Type 2 is null while BsmtFin Type 1, Bsmt
Qual, and Bsmt Cond are not null:
dataset[~pd.isna(dataset['Bsmt Cond']) &

pd.isna(dataset['BsmtFin Type 2'])][bsmt_columns]

444 Gd TA No GLQ 1124.00

444 NaN 479.00 1603.00 3206.00

444 1.00 0.00
We will fill in the missing value in BsmtFin Type 2 for this row with "Unf". According to the
dataset documentation, "Unf" for BsmtFin Type 2 means “Unfinished”:
for col in ["Bsmt Half Bath", "Bsmt Full Bath", "Total Bsmt SF",
"Bsmt Unf SF", "BsmtFin SF 2", "BsmtFin SF 1"]:
dataset[col].fillna(0, inplace=True)
dataset.loc[~pd.isna(dataset['Bsmt Cond']) &

pd.isna(dataset['Bsmt Exposure']), "Bsmt Exposure"] = "No"
dataset.loc[~pd.isna(dataset['Bsmt Cond']) &
pd.isna(dataset['BsmtFin Type 2']), "BsmtFin Type 2"] = "Unf"
for col in ["Bsmt Exposure", "BsmtFin Type 2",

"BsmtFin Type 1", "Bsmt Qual", "Bsmt Cond"]:
dataset[col].fillna("No Basement", inplace=True)
Mas Vnr Area and Mas Vnr Type Each of these two columns have 23 missing values. We
will fill in these missing values with "None" for Mas Vnr Type and with 0 for Mas Vnr Area. We
use "None" for Mas Vnr Type because in the dataset documentation, "None" for Mas Vnr Type
means “None” (i.e. no masonry veneer):
dataset['Mas Vnr Area'].fillna(0, inplace=True)

dataset['Mas Vnr Type'].fillna("None", inplace=True)
24
Electrical This column has one missing value. We will fill in this value with the mode of this
column:
dataset['Electrical'].fillna(dataset['Electrical'].mode()[0], inplace=True)
Now let’s check if there is any remaining missing value in our dataset:
dataset.isna().values.sum()
This means that our dataset is now complete; it doesn’t contain any missing value anymore.
3.5 Outlier Removal

In the paper in which our dataset was introduced by De Cock (2011), the author states that there
are five unusual values and outliers in the dataset, and encourages the removal of these outliars.
He suggested plotting SalePrice against Gr Liv Area to spot the outliers. We will do that now:
from matplotlib import pyplot as plt

import seaborn as sns
plt.scatter(x=dataset['Gr Liv Area'], y=dataset['SalePrice'],

color="orange", edgecolors="#000000", linewidths=0.5);
plt.xlabel("Gr Liv Area"); plt.ylabel("SalePrice");
25
We can clearly see the five values meant by the authour in the plot above. Now, we will
remove them from our dataset. We can do so by keeping data points that have Gr Liv Area less
than 4,000. But first we take a look at the dataset rows that correspond to these unusual values:
outlirt_columns = ["Gr Liv Area"] + \
[col for col in dataset.columns if "Sale" in col]
dataset[dataset["Gr Liv Area"] > 4000][outlirt_columns]
Gr Liv Area Sale Type Sale Condition SalePrice

1498 5642 New Partial 160000
1760 4476 WD Abnorml 745000
1767 4316 WD Normal 755000
2180 5095 New Partial 183850
2181 4676 New Partial 184750
Now we remove them:

dataset = dataset[dataset["Gr Liv Area"] < 4000]
To avoid problems in modeling later, we will reset our dataset index after removing the outlier
rows, so no gaps remain in our dataset index:
26
dataset.reset_index(drop=True, inplace=True)
3.6 Deleting Some Unimportant Columns

We will delete columns that are not useful in our analysis. The columns to be deleted are Order
and PID:
dataset.drop(['Order', 'PID'], axis=1, inplace=True)
4 Exploratory Data Analysis

In this section, we will explore the data using visualizations. This will allow us to understand the
data and the relationships between variables better, which will help us build a better model.
4.1 Target Variable Distribution

Our dataset contains a lot of variables, but the most important one for us to explore is the target
variable. We need to understand its distribution. First, we start by plotting the violin plot for the
target variable. The width of the violin represents the frequency. This means that if a violin is the
widest between 300 and 400, then the area between 300 and 400 contains more data points than
other areas:
sns.violinplot(x=dataset['SalePrice'], inner="quartile", color="#36B37E");
27
We can see from the plot that most house prices fall between 100,000 and 250,000. The dashed
lines represent the locations of the three quartiles Q1, Q2 (the median), and Q3. Now let’s see the
box plot of SalePrice:
sns.boxplot(dataset['SalePrice'], whis=10, color="#00B8D9");
This shows us the minimum and maximum values of SalePrice. It shows us also the three
quartiles represented by the box and the vertical line inside of it. Lastly, we plot the histogram of
the variable to see a more detailed view of the distribution:
sns.distplot(dataset['SalePrice'], kde=False,
color="#172B4D", hist_kws={"alpha": 0.8});
plt.ylabel("Count");
28
4.2 Correlation Between Variables
We want to see how the dataset variables are correlated with each other and how predictor vari-
ables are correlated with the target variable. For example, we would like to see how Lot Area and
SalePrice are correlated: Do they increase and decrease together (positive correlation)? Does one
of them increase when the other decrease or vice versa (negative correlation)? Or are they not
correlated?
Correlation is represented as a value between -1 and +1 where +1 denotes the highest positive
correlation, -1 denotes the highest negative correlation, and 0 denotes that there is no correlation.
We will show correlation between our dataset variables (numerical and boolean variables only)
using a heatmap graph:
fig, ax = plt.subplots(figsize=(12,9))
sns.heatmap(dataset.corr(), ax=ax);
29
We can see that there are many correlated variables in our dataset. Wwe notice that Garage
Cars and Garage Area have high positive correlation which is reasonable because when the
garage area increases, its car capacity increases too. We see also that Gr Liv Area and TotRms
AbvGrd are highly positively correlated which also makes sense because when living area above
ground increases, it is expected for the rooms above ground to increase too.
Regarding negative correlation, we can see that Bsmt Unf SF is negatively correlated with
BsmtFin SF 1, and that makes sense because when we have more unfinished area, this means
that we have less finished area. We note also that Bsmt Unf SF is negatively correlated with Bsmt
Full Bath which is reasonable too.
Most importantly, we want to look at the predictor variables that are correlated with the target
variable (SalePrice). By looking at the last row of the heatmap, we see that the target variable
is highly positively correlated with Overall Qual and Gr Liv Area. We see also that the target
variable is positively correlated with Year Built, Year Remod/Add, Mas Vnr Area, Total Bsmt
SF, 1st Flr SF, Full Bath, Garage Cars, and Garage Area.
4.2.1 Relatioships Between the Target Variable and Other Varibles

High Positive Correlation Firstly, we want to visualize the relationships between the target vari-
able and the variables that are highly and positively correlated with it, according to what we saw
30
in the heatmap. Namely, these variables are Overall Qual and Gr Liv Area. We start with the re-
latioship between the target variable and Overall Qual, but before that, let’s see the distribution
of each of them. Let’s start with the target variable SalePrice:
We can see that most house prices fall between 100,000 and 200,000. We see also that there is
a number of expensive houses to the right of the plot. Now, we move to see the distribution of
Overall Qual variable:
sns.distplot(dataset['Overall Qual'], kde=False,

color="#172B4D", hist_kws={"alpha": 1});
31
We see that Overall Qual takes an integer value between 1 and 10, and that most houses have
an overall quality between 5 and 7. Now we plot the scatter plot of SalePrice and Overall Qual
to see the relationship between them:
plt.scatter(x=dataset['Overall Qual'], y=dataset['SalePrice'],

plt.xlabel("Overall Qual"); plt.ylabel("SalePrice");
32
We can see that they are truly positively correlated; generally, as the overall quality increases,
the sale price increases too. This verfies what we got from the heatmap above.
Now, we want to see the relationship between the target variable and Gr Liv Area variable
which represents the living area above ground. Let us first see the distribution of Gr Liv Area:
sns.distplot(dataset['Gr Liv Area'], kde=False,

33
We can see that the above-ground living area falls approximately between 800 and 1800 ft2.
Now, let us see the relationship between Gr Liv Area and the target variable:

34
The scatter plot above shows clearly the strong positive correlation between Gr Liv Area and
SalePrice verifying what we found with the heatmap.
Moderate Positive Correlation Next, we want to visualize the relationship between the target
variable and the variables that are positively correlated with it, but the correlation is not very
strong. Namely, these variables are Year Built, Year Remod/Add, Mas Vnr Area, Total Bsmt
SF, 1st Flr SF, Full Bath, Garage Cars, and Garage Area. We start with the first four. Let us
see the distribution of each of them:
fig, axes = plt.subplots(1, 4, figsize=(18,5))

fig.subplots_adjust(hspace=0.5, wspace=0.6)
for ax, v in zip(axes.flat, ["Year Built", "Year Remod/Add",
"Mas Vnr Area", "Total Bsmt SF"]):
sns.distplot(dataset[v], kde=False, color="#172B4D",
hist_kws={"alpha": 0.8}, ax=ax)
ax.set(ylabel="Count");
35
Now let us see their relationships with the target variable using scatter plots:
x_vars = ["Year Built", "Year Remod/Add", "Mas Vnr Area", "Total Bsmt SF"]
g = sns.PairGrid(dataset, y_vars=["SalePrice"], x_vars=x_vars);
g.map(plt.scatter, color="orange", edgecolors="#000000", linewidths=0.5);
Next, we move to the last four. Let us see the distribution of each of them:

for ax, v in zip(axes.flat, ["1st Flr SF", "Full Bath",
"Garage Cars", "Garage Area"]):
hist_kws={"alpha": 0.8}, ax=ax);
ax.set(ylabel="Count");
36
And now let us see their relationships with the target variable:
x_vars = ["1st Flr SF", "Full Bath", "Garage Cars", "Garage Area"]
g = sns.PairGrid(dataset, y_vars=["SalePrice"], x_vars=x_vars);
g.map(plt.scatter, color="orange", edgecolors="#000000", linewidths=0.5);
From the plots above, we can see that these eight variables are truly positively correlated with
the target variable. However, it’s apparent that they are not as highly correlated as Overall Qual
and Gr Liv Area.
4.2.2 Relatioships Between Predictor Variables

Positive Correlation Apart from the target variable, when we plotted the heatmap, we discov-
ered a high positive correlation between Garage Cars and Garage Area and between Gr Liv
Area and TotRms AbvGrd. We want to visualize these correlations also. We’ve already seen the
distribution of each of them except for TotRms AbvGrd. Let us see the distribution of TotRms
AbvGrd first:
sns.distplot(dataset['TotRms AbvGrd'], kde=False,

37
Now, we visualize the relationship between Garage Cars and Garage Area and between Gr
Liv Area and TotRms AbvGrd:
plt.rc("grid", linewidth=0.05)
h1 = axes[0].hist2d(dataset["Garage Cars"],
dataset["Garage Area"],
cmap="viridis");
axes[0].set(xlabel="Garage Cars", ylabel="Garage Area")
plt.colorbar(h1[3], ax=axes[0]);
h2 = axes[1].hist2d(dataset["Gr Liv Area"],
dataset["TotRms AbvGrd"],
cmap="viridis");
axes[1].set(xlabel="Gr Liv Area", ylabel="TotRms AbvGrd")
plt.colorbar(h1[3], ax=axes[1]);
plt.rc("grid", linewidth=0.25)
38
We can see the strong correlation between each pair. For Garage Cars and Garage Area, we
see that the highest concentration of data is when Garage Cars is 2 and Garage Area is approxi-
mately between 450 and 600 ft2. For Gr Liv Area and TotRms AbvGrd, we notice that the highest
concentration is when Garage Liv Area is roughly between 800 and 2000 ft2 and TotRms AbvGrd
is 6.
Negative Correlation When we plotted the heatmap, we also discovered a significant negative
correlation between Bsmt Unf SF and BsmtFin SF 1, and between Bsmt Unf SF and Bsmt Full
Bath. We also want to visualize these correlations. Let us see the distribution of these variables
first:

for ax, v in zip(axes.flat, ["Bsmt Unf SF", "BsmtFin SF 1", "Bsmt Full Bath"]):
hist_kws={"alpha": 0.8}, ax=ax);
ax.set(ylabel="Count")
Now, we visualize the relationship between each pair using scatter plots:
39
axes[0].scatter(dataset["Bsmt Unf SF"], dataset["BsmtFin SF 1"],
axes[0].set(xlabel="Bsmt Unf SF", ylabel="BsmtFin SF 1");
axes[1].scatter(dataset["Bsmt Unf SF"], dataset["Bsmt Full Bath"],
axes[1].set(xlabel="Bsmt Unf SF", ylabel="Bsmt Full Bath");
From the plots, we can see the negative correlation between each pair of these variables.
We will use the information we got from exploratory data analysis in this section, we will use
it in feature engineering in the next section.
4.3 Feature Engineering

In this section, we will use the insights from Exploratory Data Analysis section to engineer the
features of our dataset.
4.3.1 Creating New Derived Features

Firstly, we noticed a high positive correlation between the target variable SalePrice and each of
Overall Qual and Gr Liv Area. This gives an indication that the latter two features are very
important in predicting the sale price. So, we will create polynomial features out of these features:
For each one of these features, we will derive a feature whose values are the squares of original
values, and another feature whose values are the cubes of original values. Moreover, we will
create a feature whose values are the product of our two features values:
for f in ["Overall Qual", "Gr Liv Area"]:

dataset[f + "_p2"] = dataset[f] ** 2
dataset[f + "_p3"] = dataset[f] ** 3
dataset["OverallQual_GrLivArea"] = \
dataset["Overall Qual"] * dataset["Gr Liv Area"]
40
Also, we noticed that there are some predictor features that are highly correlated with each
other. To avoid the Multicollinearity problem, we will delete one feature from each pair of highly
correlated predictors. We have two pairs: the first consists of Garage Cars and Garage Area, and
the other consists of Gr Liv Area and TotRms AbvGrd. For the first pair, we will remove Garage
Cars feature; from the second pair, we will remove TotRms AbvGrd feature:
dataset.drop(["Garage Cars", "TotRms AbvGrd"], axis=1, inplace=True)
4.3.2 Dealing with Ordinal Variables

There are some ordinal features in our dataset. For example, the Bsmt Cond feature has the fol-
lowing possible values:
print("Unique values in 'Bsmt Cond' column:")

print(dataset['Bsmt Cond'].unique().tolist())
Unique values in 'Bsmt Cond' column:

['Gd', 'TA', 'No Basement', 'Po', 'Fa', 'Ex']
Where “Gd” means “Good”, “TA” means “Typical”, “Po” means “Poor”, “Fa” means “Fair”,
and “Ex” means “Excellent” according to the dataset documentation. But the problem is that
machine learning models will not know that this feature represents a ranking; it will be treated as
other categorical features. So to solve this issue, we will map each one of the possible values of
this feature to a number. We will map "No Basement" to 0, "Po" to 1, "Fa" to 2, "TA" to 3, "Gd" to
4, and "Ex" to 5.
The ordinal features in the dataset are: Exter Qual, Exter Cond, Bsmt Qual, Bsmt Cond,
Bsmt Exposure, BsmtFin Type 1, BsmtFin Type 2, Heating QC, Central Air, Kitchen Qual,
Functional, Fireplace Qu, GarageFinish, Garage Qual, Garage Cond, Pool QC, Land Slope
and Fence. We will map the values of each of them to corresponding numbers as described for
Bsmt Cond above and in accordance with the dataset documentation:
mp = {'Ex':4,'Gd':3,'TA':2,'Fa':1,'Po':0}
dataset['Exter Qual'] = dataset['Exter Qual'].map(mp)
dataset['Exter Cond'] = dataset['Exter Cond'].map(mp)
dataset['Heating QC'] = dataset['Heating QC'].map(mp)
dataset['Kitchen Qual'] = dataset['Kitchen Qual'].map(mp)
mp = {'Ex':5,'Gd':4,'TA':3,'Fa':2,'Po':1,'No Basement':0}
dataset['Bsmt Qual'] = dataset['Bsmt Qual'].map(mp)
dataset['Bsmt Cond'] = dataset['Bsmt Cond'].map(mp)
dataset['Bsmt Exposure'] = dataset['Bsmt Exposure'].map(
{'Gd':4,'Av':3,'Mn':2,'No':1,'No Basement':0})
mp = {'GLQ':6,'ALQ':5,'BLQ':4,'Rec':3,'LwQ':2,'Unf':1,'No Basement':0}
dataset['BsmtFin Type 1'] = dataset['BsmtFin Type 1'].map(mp)
dataset['BsmtFin Type 2'] = dataset['BsmtFin Type 2'].map(mp)
dataset['Central Air'] = dataset['Central Air'].map({'Y':1,'N':0})
41
dataset['Functional'] = dataset['Functional'].map(
{'Typ':7,'Min1':6,'Min2':5,'Mod':4,'Maj1':3,
'Maj2':2,'Sev':1,'Sal':0})
dataset['Fireplace Qu'] = dataset['Fireplace Qu'].map(
{'Ex':5,'Gd':4,'TA':3,'Fa':2,'Po':1,'No Fireplace':0})
dataset['Garage Finish'] = dataset['Garage Finish'].map(
{'Fin':3,'RFn':2,'Unf':1,'No Garage':0})
dataset['Garage Qual'] = dataset['Garage Qual'].map(
{'Ex':5,'Gd':4,'TA':3,'Fa':2,'Po':1,'No Garage':0})
dataset['Garage Cond'] = dataset['Garage Cond'].map(
{'Ex':5,'Gd':4,'TA':3,'Fa':2,'Po':1,'No Garage':0})
dataset['Pool QC'] = dataset['Pool QC'].map(
{'Ex':4,'Gd':3,'TA':2,'Fa':1,'No Pool':0})
dataset['Land Slope'] = dataset['Land Slope'].map(
{'Sev': 2, 'Mod': 1, 'Gtl': 0})
dataset['Fence'] = dataset['Fence'].map(
{'GdPrv':4,'MnPrv':3,'GdWo':2,'MnWw':1,'No Fence':0})
4.3.3 One-Hot Encoding For Categorical Features

Machine learning models accept only numbers as input, and since our dataset contains categori-
cal features, we need to encode them in order for our dataset to be suitable for modeling. We will
encode our categorical features using one-hot encoding technique which transforms the categor-
ical variable into a number of binary variables based on the number of unique categories in the
categorical variable; each of the resulting binary variables has only 0 and 1 as its possible values.
Pandas package provides a convenient function get_dummies() that can be used for performing
one-hot encoding on our dataset.
To see what will happen to our dataset, let us take for example the variable Paved Drive which
indicates how the driveway is paved. It has three possible values: Y which means for “Paved”, P
which means “Partial Pavement”, and N which means “Dirt/Gravel”. Let us take a look at Paved
Drive value for the first few rows in our dataset:
dataset[['Paved Drive']].head()
Paved Drive
0 P
1 Y
2 Y
3 Y
4 Y
Now, we perform one-hot encoding:
dataset = pd.get_dummies(dataset)
Let us see what has happened to the Paved Drive variable by looking at the same rows above:
42
pavedDrive_oneHot = [c for c in dataset.columns if c.startswith("Paved")]
dataset[pavedDrive_oneHot].head()
Paved Drive_N Paved Drive_P Paved Drive_Y

0 0 1 0
1 0 0 1
2 0 0 1
3 0 0 1
4 0 0 1
We can see for example that a value of P in the original Paved Drive column is converted to 1
in Paved Drive_P and zeros in Paved Drive_N and Paved Drive_Y after one-hot encoding.
All categorical column are converted in the same way.
Now, after we have cleaned and prepared our dataset, it is ready for modeling.
5 Prediction Type and Modeling Techniques

In this section, we choose the type of machine learning prediction that is suitable to our problem.
We want to determine if this is a ragression problem or a classification problem. In this project, we
want to predict the price of a house given information about it. The price we want to predict is a
continuous value; it can be any real number. This can be seen by looking at the target vatiable in
our dataset SalePrice:
dataset[['SalePrice']].head()
SalePrice
0 215000
1 105000
2 172000
3 244000
4 189900
That means that the prediction type that is appropriate to our problem is regression.
Now we move to choose the modeling techniques we want to use. There are a lot of techniques
available for regression problems like Linear Regression, Ridge Regression, Artificial Neural Net-
works, Decision Trees, Random Forest, etc. In this project, we will test many modeling techniques,
and then choose the technique(s) that yield the best results. The techniques that we will try are:
5.0.1 1. Linear Regression

This technique models the relationship between the target variable and the independent variables
(predictors). It fits a linear model with coefficients to the data in order to minimize the residual
sum of squares between the target variable in the dataset, and the predicted values by the linear
approximation.
43
Figure 8: Predicting who survived when the Titanic sank
5.0.2 2. Nearest Neighbors

Nearest Neighbors is a type of instance-based learning. For this technique, the model tries to find
a number (k) of training examples closest in distance to a new point, and predict the output for
this new point from these closest neighbors. k can be a user-defined number (k-nearest neighbors),
or vary based on the local density of points (radius-based neighbors). The distance metric used to
measure the closeness is mostly the Euclidean distance.
5.0.3 3. Support Vector Regression

Support vector machines (SVM) are a set of methods that can be used for classification and regres-
sion problems. When they are used for regression, we call the technique Support Vector Regres-
sion.
5.0.4 4. Decision Trees

For this technique, the goal is to create a model that predicts the value of a target variable by
learning simple decision rules inferred from the data features. An example of a simple decision
tree for predicting who survived when the Titanic sank is shown in Figure 8:
5.0.5 5. Neural Networks

Neural network is a machine learning model that tries to mimic the way of working of the biolog-
ical brain. A neural network consists of multiple layers. Each layer consists of a number of nodes.
The nodes of each layer are connected to the nodes of adjacent layers. Each node can be activated
44
Figure 9: A neural network
or not based on its inputs and its activation function. An example of a neural network is shown in
Figure 9:
5.0.6 6. Random Forest

Bagging is an ensemble method where many base models are used with a randomized subset of
data to reduce the variance of a the base model.
5.0.7 7. Gradient Boosting

Boosting is also an ensemble method where weak base models are used to create a strong model
that reduces bias and variance of the base model.
Each one of these techniques has many algorithmic implementation. We will choose algorithm(s)
for each of these techniques in the next section.
6 Model Building and Evaluation

In this part, we will build our prediction model: we will choose algorithms for each of the tech-
niques we mentioned in the previous section. After we build the model, we will evaluate its
performance and results.
6.1 Feature Scaling

In order to make all algorithms work properly with our data, we need to scale the features in
our dataset. For that, we will use a helpful function named StandardScaler() from the popular
Scikit-Learn Python package. This function standardizes features by subtracting the mean and
scaling to unit variance. It works on each feature independently. For a value x of some feature F,
the StandardScaler() function performs the following operation:
45
Figure 10: train_test_split() operation
x−µ
z=
s
where z is the result of scaling x, µ is the mean of feature F, and s is the standard deviation of
F.
from sklearn.preprocessing import StandardScaler
scaler = StandardScaler()
# We need to fit the scaler to our data before transformation
dataset.loc[:, dataset.columns != 'SalePrice'] = scaler.fit_transform(
dataset.loc[:, dataset.columns != 'SalePrice'])
6.2 Splitting the Dataset

As usual for supervised machine learning problems, we need a training dataset to train our model
and a test dataset to evaluate the model. So we will split our dataset randomly into two parts,
one for training and the other for testing. For that, we will use another function from Scikit-Learn
called train_test_split():
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(

dataset.drop('SalePrice', axis=1), dataset[['SalePrice']],
test_size=0.25, random_state=3)
We specified the size of the test set to be 25% of the whole dataset. This leaves 75% for the
training dataset. Now we have four subsets: X_train, X_test, y_train, and y_test. Later we
will use X_train and y_train to train our model, and X_test and y_test to test and evaluate
the model. X_train and X_test represent features (predictors); y_train and y_test represent the
target. From now on, we will refer to X_train and y_train as the training dataset, and to X_test
and y_test as the test dataset. Figure 10 shows an example of what train_test_split() does.
6.3 Modeling Approach

For each one of the techniques mentioned in the previous section (Linear Regression, Nearest
Neighbor, Support Vector Machines, etc.), we will follow these steps to build a model:
46
• Choose an algorithm that implements the corresponding technique
• Search for an effective parameter combination for the chosen algorithm
• Create a model using the found parameters
• Train (fit) the model on the training dataset
• Test the model on the test dataset and get the results
6.3.1 Searching for Effective Parameters

Using Scikit-Learn, we can build a decision-tree model for example as follows:
model = DecisionTreeRegressor(max_depth=14, min_samples_split=5, max_features=20)
We can do this but to probably achieve a better performance if we choose better values for
the parameters max_depth, min_samples_split, and max_features. To do so, we will examine
many parameter combinations and choose the combination that gives the best score. Scikit-Learn
provides a useful function for that purpose: GridSearchCV(). So for the example above, we will
do the following:
parameter_space = {
"max_depth": [7, 15],
"min_samples_split": [5, 10],
"max_features": [30, 45]
}
clf = GridSearchCV(DecisionTreeRegressor(), parameter_space, cv=4,

scoring="neg_mean_absolute_error")
clf.fit(X_train, y_train)
The code above will test the decision-tree model using all the parameter combinations. It will
use cross validation with 4 folds and it will use the mean absolute error for scoring and comparing
different parameter combinations. At the end, it will provide us with the best parameter combi-
nation that achieved the best score so we can use it to build our model.
Sometimes, when the number of parameter combinations is large, GridSearchCV() can
take very long time to run. So in addition to GridSearchCV(), we will sometimes use
RandomizedSearchCV() which is similar to GridSearchCV() but instead of using all parameter
combinations, it picks a number of random combinations specified by n_iter. For the example
above, we can use RandomizedSearchCV() as follows:
clf = RandomizedSearchCV(DecisionTreeRegressor(), parameter_space, cv=4,
scoring="neg_mean_absolute_error", n_iter=100)
This will make RandomizedSearchCV() pick 100 parameter combinations randomly.
6.4 Performance Metric

For evaluating the performance of our models, we will use mean absolute error (MAE). If ŷi is the
predicted value of the i-th element, and y is the corresponding true value, then for all n elements,
RMSE is calculated as:
1 n
n i∑
MAE(y, ŷ) = |yi − ŷi | .
=1
47
6.5 Modeling
6.5.1 Linear Regression
For Linear Regression, we will choose three algorithmic implementations: Ridge Regression and
Elastic Net. We will use the implementations provided in the Scikit-Learn package of these algo-
rithms.
1. Ridge Regression This model has the following syntax:
Ridge(alpha=1.0, fit_intercept=True, normalize=False, copy_X=True,

max_iter=None, tol=0.001, solver=auto, random_state=None)
Firstly, we will use GridSearchCV() to search for the best model parameters in a parameter
space provided by us. The parameter alpha represents the regularization strength, fit_intercept
determines whether to calculate the intercept for this model, and solver controls which solver to
use in the computational routines.
from sklearn.model_selection import GridSearchCV

from sklearn.linear_model import Ridge
parameter_space = {
"alpha": [1, 10, 100, 290, 500],
"fit_intercept": [True, False],
"solver": ['svd', 'cholesky', 'lsqr', 'sparse_cg', 'sag', 'saga'],
}
clf = GridSearchCV(Ridge(random_state=3), parameter_space, n_jobs=4,

cv=3, scoring="neg_mean_absolute_error")
print("Best parameters:")
print(clf.best_params_)
Best parameters:
{'alpha': 290, 'fit_intercept': True, 'solver': 'cholesky'}
We defined the parameter space above using reasonable values for chosen parameters. Then
we used GridSearchCV() with 3 folds (cv=3). Now we build our Ridge model with the best
parameters found:
ridge_model = Ridge(random_state=3, **clf.best_params_)
Then we train our model using our training set (X_train and y_train):
ridge_model.fit(X_train, y_train);
Finally, we test our model on X_test. Then we evaluate the model performance by comparing
its predictions with the actual true values in y_test using the MAE metric as we described above:
48
from sklearn.metrics import mean_absolute_error
y_pred = ridge_model.predict(X_test)
ridge_mae = mean_absolute_error(y_test, y_pred)
print("Ridge MAE =", ridge_mae)
Ridge MAE = 15270.463549642733
2. Elastic Net This model has the following syntax:
ElasticNet(alpha=1.0, l1_ratio=0.5, fit_intercept=True, normalize=False,

precompute=False, max_iter=1000, copy_X=True, tol=0.0001,
warm_start=False, positive=False, random_state=None, selection=cyclic)
Firstly, we will use GridSearchCV() to search for the best model parameters in a parame-
ter space provided by us. The parameter alpha is a constant that multiplies the penalty terms,
l1_ratio determines the amount of L1 and L2 regularizations, fit_intercept is the same as
Ridge’s.
from sklearn.linear_model import ElasticNet
parameter_space = {
"alpha": [1, 10, 100, 280, 500],
"l1_ratio": [0.5, 1],
"fit_intercept": [True, False],
}
clf = GridSearchCV(ElasticNet(random_state=3), parameter_space,

n_jobs=4, cv=3, scoring="neg_mean_absolute_error")
Best parameters:
{'alpha': 280, 'fit_intercept': True, 'l1_ratio': 1}
parameters found:
elasticNet_model = ElasticNet(random_state=3, **clf.best_params_)
elasticNet_model.fit(X_train, y_train);
49
y_pred = elasticNet_model.predict(X_test)
elasticNet_mae = mean_absolute_error(y_test, y_pred)
print("Elastic Net MAE =", elasticNet_mae)
Elastic Net MAE = 14767.90981933659
6.5.2 Nearest Neighbors

For Nearest Neighbors, we will use an implementation of the k-nearest neighbors (KNN) algo-
rithm provided by Scikit-Learn package.
The KNN model has the following syntax:
KNeighborsRegressor(n_neighbors=5, weights=uniform, algorithm=auto,

leaf_size=30, p=2, metric=minkowski, metric_params=None,
n_jobs=None, **kwargs)
space provided by us. The parameter n_neighbors represents k which is the number of neigh-
bors to use, weights determines the weight function used in prediction: uniform or distance,
algorithm specifies the algorithm used to compute the nearest neighbors, leaf_size is passed to
BallTree or KDTree algorithm. It can affect the speed of the construction and query, as well as the
memory required to store the tree.
from sklearn.neighbors import KNeighborsRegressor
parameter_space = {
"n_neighbors": [9, 10, 11,50],
"weights": ["uniform", "distance"],
"algorithm": ["ball_tree", "kd_tree", "brute"],
"leaf_size": [1,2,20,50,200]
}
clf = GridSearchCV(KNeighborsRegressor(), parameter_space, cv=3,

scoring="neg_mean_absolute_error", n_jobs=4)
Best parameters:
{'algorithm': 'ball_tree', 'leaf_size': 1, 'n_neighbors': 10, 'weights': 'distance'}
parameters found:
50
knn_model = KNeighborsRegressor(**clf.best_params_)
knn_model.fit(X_train, y_train);
y_pred = knn_model.predict(X_test)
knn_mae = mean_absolute_error(y_test, y_pred)
print("K-Nearest Neighbors MAE =", knn_mae)
K-Nearest Neighbors MAE = 22780.14347886256
6.5.3 Support Vector Regression

For Support Vector Regression (SVR), we will use one of three implementations provided by the
Scikit-Learn package.
The SVR model has the following syntax:
SVR(kernel=rbf, degree=3, gamma=auto_deprecated, coef0=0.0, tol=0.001,

C=1.0, epsilon=0.1, shrinking=True, cache_size=200, verbose=False, max_iter=-1)
space provided by us. The parameter kernel specifies the kernel type to be used in the algorithm,
degree represents the degree of the polynomial kernel poly, gamma is the kernel coefficient for
rbf, poly and sigmoid kernels, coef0 is independent term in kernel function, and C is the penalty
parameter of the error term.
from sklearn.model_selection import RandomizedSearchCV

from sklearn.svm import SVR
parameter_space = \
{
"kernel": ["poly", "linear", "rbf", "sigmoid"],
"degree": [3, 5],
"coef0": [0, 3, 7],
"gamma":[1e-3, 1e-1, 1/X_train.shape[1]],
"C": [1, 10, 100],
}
clf = GridSearchCV(SVR(), parameter_space, cv=3, n_jobs=4,

scoring="neg_mean_absolute_error")
51
Best parameters:
{'C': 100, 'coef0': 3, 'degree': 5, 'gamma': 0.004132231404958678, 'kernel': 'poly'}
we used GridSearchCV() with 3 folds (cv=3). Now we build our Support Vector Regression model
with the best parameters found:
svr_model = SVR(**clf.best_params_)
svr_model.fit(X_train, y_train);
y_pred = svr_model.predict(X_test)
svr_mae = mean_absolute_error(y_test, y_pred)
print("Support Vector Regression MAE =", svr_mae)
Support Vector Regression MAE = 12874.92786950232
6.5.4 Decision Tree

For Decision Tree (DT), we will use an implementations provided by the Scikit-Learn package.
The Decision Tree model has the following syntax:
DecisionTreeRegressor(criterion=mse, splitter=best, max_depth=None,

min_samples_split=2, min_samples_leaf=1,
min_weight_fraction_leaf=0.0, max_features=None,
random_state=None, max_leaf_nodes=None, min_impurity_decrease=0.0,
min_impurity_split=None, presort=False)
space provided by us. The parameter criterion specifies the function used to measure the quality
of a split, min_samples_split determines the minimum number of samples required to split an
internal node, min_samples_leaf determines the minimum number of samples required to be at
a leaf node, and max_features controls the number of features to consider when looking for the
best split.
from sklearn.tree import DecisionTreeRegressor
parameter_space = \
{
"criterion": ["mse", "friedman_mse", "mae"],
"min_samples_split": [5, 18, 29, 50],
"min_samples_leaf": [3, 7, 15, 25],
"max_features": [20, 50, 150, 200, X_train.shape[1]],
52
}
clf = GridSearchCV(DecisionTreeRegressor(random_state=3), parameter_space,

cv=3, scoring="neg_mean_absolute_error", n_jobs=4)
Best parameters:
{'criterion': 'mse', 'max_features': 242, 'min_samples_leaf': 7, 'min_samples_split': 18}
we used GridSearchCV() with 3 folds (cv=3). Now we build our Decision Tree model with the
best parameters found:
dt_model = DecisionTreeRegressor(**clf.best_params_)
dt_model.fit(X_train, y_train);
y_pred = dt_model.predict(X_test)
dt_mae = mean_absolute_error(y_test, y_pred)
print("Decision Tree MAE =", dt_mae)
Decision Tree MAE = 20873.949425979506
6.5.5 Neural Network

For Neural Network (NN), we will use an implementations provided by the Scikit-Learn package.
The Neural Network model has the following syntax:
MLPRegressor(hidden_layer_sizes=(100, ), activation=relu, solver=adam,

alpha=0.0001, batch_size=auto, learning_rate=constant,
learning_rate_init=0.001, power_t=0.5, max_iter=200, shuffle=True,
random_state=None, tol=0.0001, verbose=False, warm_start=False,
momentum=0.9, nesterovs_momentum=True, early_stopping=False,
validation_fraction=0.1, beta_1=0.9, beta_2=0.999, epsilon=1e-08,
n_iter_no_change=10)
space provided by us. The parameter hidden_layer_sizes is a list where its ith element repre-
sents the number of neurons in the ith hidden layer, activation specifies the activation function
for the hidden layer, solver determines the solver for weight optimization, and alpha represents
L2 regularization penalty.
53
from sklearn.neural_network import MLPRegressor
parameter_space = \
{
"hidden_layer_sizes": [(7,)*3, (19,), (100,), (154,)],
"activation": ["identity", "logistic", "tanh", "relu"],
"solver": ["lbfgs"],
"alpha": [1, 10, 100],
}
clf = GridSearchCV(MLPRegressor(random_state=3), parameter_space,

cv=3, scoring="neg_mean_absolute_error", n_jobs=4)
Best parameters:
{'activation': 'identity', 'alpha': 1, 'hidden_layer_sizes': (154,), 'solver': 'lbfgs'}
we used GridSearchCV() with 3 folds (cv=3). Now we build our Neural Network model with the
nn_model = MLPRegressor(**clf.best_params_)
nn_model.fit(X_train, y_train);
y_pred = nn_model.predict(X_test)
nn_mae = mean_absolute_error(y_test, y_pred)
print("Neural Network MAE =", nn_mae)
Neural Network MAE = 15656.581467633143
6.5.6 Random Forest

For Random Forest (RF), we will use an implementations provided by the Scikit-Learn package.
The Random Forest model has the following syntax:
RandomForestRegressor(n_estimators=warn, criterion=mse, max_depth=None,
min_samples_split=2, min_samples_leaf=1,
min_weight_fraction_leaf=0.0, max_features=auto,
max_leaf_nodes=None, min_impurity_decrease=0.0,
min_impurity_split=None, bootstrap=True, oob_score=False,
n_jobs=None, random_state=None, verbose=0, warm_start=False)
54
Firstly, we will use GridSearchCV() to search for the best model parameters in a parame-
ter space provided by us. The parameter n_estimators specifies the number of trees in the for-
est, bootstrap determines whether bootstrap samples are used when building trees. criterion,
max_depth, min_samples_split, min_samples_leaf, max_features are the same as those of the
decision tree model.
from sklearn.ensemble import RandomForestRegressor
parameter_space = \
{
"n_estimators": [10, 100, 300, 600],
"criterion": ["mse", "mae"],
"max_depth": [7, 50, 254],
"min_samples_split": [2, 5],
"min_samples_leaf": [1, 5],
"max_features": [19, 100, X_train.shape[1]],
"bootstrap": [True, False],
}
clf = RandomizedSearchCV(RandomForestRegressor(random_state=3),
parameter_space, cv=3, n_jobs=4,
scoring="neg_mean_absolute_error",
n_iter=10, random_state=3)
Best parameters:
{'n_estimators': 600, 'min_samples_split': 2, 'min_samples_leaf': 1, 'max_features': 19, 'max_d
we used GridSearchCV() with 3 folds (cv=3). Now we build our Random Forest model with the
rf_model = RandomForestRegressor(**clf.best_params_)
rf_model.fit(X_train, y_train);
y_pred = rf_model.predict(X_test)
rf_mae = mean_absolute_error(y_test, y_pred)
print("Random Forest MAE =", rf_mae)
Random Forest MAE = 14506.456657559198
55
6.5.7 Gradient Boosting
For Gradient Boosting (GB), we will use the renowned XGBoost implementations.
XGBoost model has the following syntax:
XGBRegressor(max_depth=3, learning_rate=0.1, n_estimators=100, silent=True,

objective='reg:linear', booster='gbtree', n_jobs=1, nthread=None,
gamma=0, min_child_weight=1, max_delta_step=0, subsample=1,
colsample_bytree=1, colsample_bylevel=1, reg_alpha=0, reg_lambda=1,
scale_pos_weight=1, base_score=0.5, random_state=0, seed=None,
missing=None, importance_type='gain', **kwargs)
Firstly, we will use GridSearchCV() to search for the best model parameters in a param-
eter space provided by us. The parameter max_depth sets the maximum depth of a tree,
learning_rate represents the step size shrinkage used in updating weights, n_estimators spec-
ifies the number of boosted trees to fit, booster determines which booster to use, gamma speci-
fies the minimum loss reduction required to make a further partition on a leaf node of the tree,
subsample is subsample ratio of the training instances; this subsampling will occur once in every
boosting iteration, colsample_bytree specifies the subsample ratio of columns when construct-
ing each tree, colsample_bylevel specifies the subsample ratio of columns for each split, in each
level, reg_alpha is L1 regularization term, and reg_lambda is L2 regularization term.
from xgboost import XGBRegressor
parameter_space = \
{
"max_depth": [4, 5, 6],
"learning_rate": [0.005, 0.009, 0.01],
"n_estimators": [700, 1000, 2500],
"booster": ["gbtree",],
"gamma": [7, 25, 100],
"subsample": [0.3, 0.6],
"colsample_bytree": [0.5, 0.7],
"colsample_bylevel": [0.5, 0.7,],
"reg_alpha": [1, 10, 33],
"reg_lambda": [1, 3, 10],
}
clf = RandomizedSearchCV(XGBRegressor(random_state=3),
parameter_space, cv=3, n_jobs=4,
scoring="neg_mean_absolute_error",
random_state=3, n_iter=10)
Best parameters:
{'subsample': 0.3, 'reg_lambda': 3, 'reg_alpha': 33, 'n_estimators': 2500, 'max_depth': 6, 'lea
56
we used GridSearchCV() with 3 folds (cv=3). Now we build our XGBoost model with the best
parameters found:
xgb_model = XGBRegressor(**clf.best_params_)
xgb_model.fit(X_train, y_train);
y_pred = xgb_model.predict(X_test)
xgb_mae = mean_absolute_error(y_test, y_pred)
print("XGBoost MAE =", xgb_mae)
XGBoost MAE = 12556.67524760929
7 Analysis and Comparison

In the previous section, we created many models: for each model, we searched for good parame-
ters then we constructed the model using those parameters, then trained (fitted) the model to our
training data (X_train and y_train), then tested the model on our test data (X_test) and finally,
we evaluated the model performance by comparing the model predictions with the true values in
y_test. We used the mean absolute error (MAE) to evaluate model performance.
Using the results we got in the previous section, we present a table that shows the mean abso-
lute error (MAE) for each model when applied to the test set X_test. The table is sorted ascend-
ingly according to MAE score.
Model MAE
XGBoost 12556.68
Support Vector Regression (SVR) 12874.93
Random Forest 14506.46
Elastic Net 14767.91
Ridge 15270.46
Neural Network 15656.38
Decision Tree 20873.95
K-Nearest Neighbors (KNN) 22780.14
We also present a graph that visualizes the table contents:
x = ['KNN', 'Decision Tree', 'Neural Network', 'Ridge',

'Elastic Net', 'Random Forest', 'SVR', 'XGBoost']
y = [22780.14, 20873.95, 15656.38, 15270.46, 14767.91,
14506.46, 12874.93, 12556.68]
57
colors = ["#392834", "#5a3244", "#7e3c4d", "#a1484f",
"#c05949", "#d86f3d", "#e88b2b", "#edab06"]
fig, ax = plt.subplots()
plt.barh(y=range(len(x)), tick_label=x, width=y, height=0.4, color=colors);
ax.set(xlabel="MAE (smaller is better)", ylabel="Model");
By looking at the table and the graph, we can see that XGBoost model has the smallest MAE,
12556.68 followed by Support Vector Regression model with a little larger error of 12974.93. Af-
ter that, Random Forest and Elastic Net models come with similar errors: 14506.46 and 14767.91
respectively. Then come Ridge and Neural Network models with close errors: 15270.46 and
15656.38 respectively. Then comes Decision Tree model with MAE of 20873.95, and at last, the
K-Nearest Neighbors model with an error of 22780.14.
So, in our experiment, the best model is XGBoost and the worst model is K-Nearest Neighbors.
We can see that the difference in MAE between the best model and the worst model is significant;
the best model has almost half of the error of the worst model.
7.1 Performance Interpretation

We chose the mean absolute error (MAE) as our performance metric to evaluate and compare
models. MAE presents a value that is easy to understand; it shows the average value of model
error. For example, for our XGBoost model, its MAE is 12556.68 which means that on average,
XGBoost will predict a value that is bigger or smaller than the true value by 12556.68. Now to
understand how good this MAE is, we need to know the range and distribution of the data. In our
case, we need to see the values of the target variable SalePrice which contains the actual house
prices. Let’s see the violin plot, box plot, and histogram of SalePrice in our dataset:
58
sns.violinplot(x=dataset['SalePrice'], inner="quartile", color="#36B37E");
sns.boxplot(dataset['SalePrice'], whis=10, color="#00B8D9");
59
60
From the three plots above, we can understand the distribution of SalePrice. Now let’s get
some numerical statistical information about it:
y_train.describe(include=[np.number])
SalePrice
count 2193.00
mean 179846.69
std 79729.38
min 12789.00
25% 128500.00
50% 159895.00
75% 214000.00
max 625000.00
We can see that the mean is 179,846.69 and the median is 159,895. We can see also that the
first quartile is 128,500; this means that 75% of the data is larger than this number. Now looking
at XGBoost error of 12,556.68, we can say that an error of about 12,000 is good for data whose
mean is 159,895 and whose 75% of it is larger than 128,500.
61
7.2 Feature Importances
Some of the models we used provide the ability to see the importance of each feature in the dataset
after fitting the model. We will look at the feature importances provided by both XGBoost and
Random Forest models. We have 242 features in our data which is a big number, so we will take a
look at the 15 most important features.
7.2.1 XGBoost
Let’s discover the most important features as determined by XGBoost model:
xgb_feature_importances = xgb_model.feature_importances_
xgb_feature_importances = pd.Series(
xgb_feature_importances, index=X_train.columns.values
).sort_values(ascending=False).head(15)
fig, ax = plt.subplots(figsize=(7, 5))

sns.barplot(x=xgb_feature_importances,
y=xgb_feature_importances.index,
color="#003f5c");
plt.xlabel('Feature Importance');
plt.ylabel('Feature');
7.2.2 Random Forest

Now, let’s see the most important features as for Random Forest model:
62
rf_feature_importances = rf_model.feature_importances_
rf_feature_importances = pd.Series(
rf_feature_importances, index=X_train.columns.values
).sort_values(ascending=False).head(15)
fig, ax = plt.subplots(figsize=(7,5))
sns.barplot(x=rf_feature_importances,
y=rf_feature_importances.index,
color="#ffa600");
plt.xlabel('Feature Importance');
plt.ylabel('Feature');
7.2.3 Common Important Features

Now, let us see which features are among the most important features for both XGBoost and
Random Forest models, and let’s find out the difference in their importance regarding the two
models:
common_imp_feat = [x for x in xgb_feature_importances.index

if x in rf_feature_importances.index]
commImpFeat_xgb_scores = [xgb_feature_importances[x]
for x in common_imp_feat]
commImpFeat_rf_scores = [rf_feature_importances[x]
for x in common_imp_feat]
63
ind = np.arange(len(commImpFeat_xgb_scores))
width = 0.35
fig, ax = plt.subplots()
ax.bar(ind - width/2, commImpFeat_xgb_scores, width,
color='#003f5c', label='XGBoost');
ax.bar(ind + width/2, commImpFeat_rf_scores, width,
color='#ffa600', label='Random Forest')
ax.set_xticks(ind);
ax.set_xticklabels(common_imp_feat);
ax.legend();
plt.xticks(rotation=90);
64
8 Conclusion
In this paper, we built serveral regression models to predict the price of some house given some
of the house features. We eveluated and compared each model to determine the one with highest
performance. We also looked at how some models rank the features according to their importance.
In this paper, we followed the data science process starting with getting the data, then cleaning
and preprocessing the data, followed by exploring the data and building models, then evaluating
the results and communicating them with visualizations.
As a recommendation, we advise to use this model (or a version of it trained with more recent
data) by people who want to buy a house in the area covered by the dataset to have an idea about
the actual price. The model can be used also with datasets that cover different cities and areas
provided that they contain the same features. We also suggest that people take into consideration
the features that were deemed as most important as seen in the previous section; this might help
them estimate the house price better.
9 References
Alkhatib, K., Najadat, H., Hmeidi, I., & Shatnawi, M. K. A. (2013). Stock price prediction using k-
nearest neighbor (kNN) algorithm. International Journal of Business, Humanities and Technology,
3(3), 32-44.
de Abril, I. M., & Sugiyama, M. (2013). Winning the kaggle algorithmic trading challenge with
the composition of many models and feature engineering. IEICE transactions on information and
systems, 96(3), 742-745.
Feng, Y., & Jones, K. (2015, July). Comparing multilevel modelling and artificial neural net-
works in house price prediction. In Spatial Data Mining and Geographical Knowledge Services
(ICSDM), 2015 2nd IEEE International Conference on (pp. 108-114). IEEE.
Hegazy, O., Soliman, O. S., & Salam, M. A. (2014). A machine learning model for stock market
prediction. arXiv preprint arXiv:1402.7351.
Ticknor, J. L. (2013). A Bayesian regularized artificial neural network for stock market forecast-
ing. Expert Systems with Applications, 40(14), 5501-5506.
De Cock, D. (2011). Ames, Iowa: Alternative to the Boston housing data as an end of semester
regression project. Journal of Statistics Education, 19(3).
65

House Price Prediction

Uploaded by

Copyright:

Available Formats

House Price Prediction

Uploaded by

Document Information

Original Description:

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

House Price Prediction

Uploaded by

Copyright:

Available Formats

House Price Prediction Using Machine Learning Techniques

4 Exploratory Data Analysis 27

5 Prediction Type and Modeling Techniques 43

6 Model Building and Evaluation 45

7 Analysis and Comparison 57

1.1 Goals of the Study

1.2 Paper Organization

2.1 Stock Market Prediction Using Bayesian-Regularized Neural Networks

∑ri=1 (abs(yi − pi )/yi )

2.2 Stock Market Prediction Using A Machine Learning Model

2.4 Composition of Models and Feature Engineering to Win Algorithmic Trading

2.5 Using K-Nearest Neighbours for Stock Price Prediction

• The number of neaerest neighbours is chosen

Figure 6: AIEI lift graph

3.1 Data Description

3.2 Reading the Dataset

3.3 Getting A Feel of the Dataset

# Configuring float numbers format

Street Alley Lot Shape Land Contour Utilities Lot Config

Land Slope Neighborhood Condition 1 Condition 2 Bldg Type House Style

BsmtFin Type 2 BsmtFin SF 2 Bsmt Unf SF Total Bsmt SF Heating Heating QC

Kitchen Qual TotRms AbvGrd Functional Fireplaces Fireplace Qu Garage Type

Pool Area Pool QC Fence Misc Feature Misc Val Mo Sold

Yr Sold Sale Type Sale Condition SalePrice

unique top freq

3.4 Data Cleaning

# Getting the number of missing values in each column

Missing Values Percentage

Now we start dealing with these missing values.

dataset['Misc Feature'].fillna('No feature', inplace=True)

dataset['Alley'].fillna('No Alley', inplace=True)

dataset['Lot Frontage'].fillna(0, inplace=True)

garage_columns = [col for col in dataset.columns if col.startswith("Garage")]

Garage Type Garage Yr Blt Garage Finish Garage Cars

Garage Area Garage Qual Garage Cond

Garage Type Garage Yr Blt Garage Finish Garage Cars

Garage Area Garage Qual Garage Cond

dataset['Garage Cars'].fillna(0, inplace=True)

dataset.loc[~pd.isna(dataset['Garage Type']) &

dataset['Garage Yr Blt'].fillna(0, inplace=True)

bsmt_columns = [col for col in dataset.columns if "Bsmt" in col]

Bsmt Qual Bsmt Cond Bsmt Exposure BsmtFin Type 1 BsmtFin SF 1

Bsmt Full Bath Bsmt Half Bath

dataset[~pd.isna(dataset['Bsmt Cond']) &

Bsmt Qual Bsmt Cond Bsmt Exposure BsmtFin Type 1 BsmtFin SF 1

BsmtFin Type 2 BsmtFin SF 2 Bsmt Unf SF Total Bsmt SF

Bsmt Full Bath Bsmt Half Bath

dataset[~pd.isna(dataset['Bsmt Cond']) &

Bsmt Qual Bsmt Cond Bsmt Exposure BsmtFin Type 1 BsmtFin SF 1

BsmtFin Type 2 BsmtFin SF 2 Bsmt Unf SF Total Bsmt SF

Bsmt Full Bath Bsmt Half Bath

dataset.loc[~pd.isna(dataset['Bsmt Cond']) &

for col in ["Bsmt Exposure", "BsmtFin Type 2",

dataset['Mas Vnr Area'].fillna(0, inplace=True)

3.5 Outlier Removal

from matplotlib import pyplot as plt

plt.scatter(x=dataset['Gr Liv Area'], y=dataset['SalePrice'],

Gr Liv Area Sale Type Sale Condition SalePrice