Knowledge and Systems Sciences 19th International Symposium KSS 2018 Tokyo Japan November 25 27 2018 Proceedings Jian Chen
Knowledge and Systems Sciences 19th International Symposium KSS 2018 Tokyo Japan November 25 27 2018 Proceedings Jian Chen
Knowledge and Systems Sciences 19th International Symposium KSS 2018 Tokyo Japan November 25 27 2018 Proceedings Jian Chen
https://textbookfull.com/product/knowledge-and-systems-
sciences-17th-international-symposium-kss-2016-kobe-japan-
november-4-6-2016-proceedings-1st-edition-jian-chen/
https://textbookfull.com/product/prima-2018-principles-and-
practice-of-multi-agent-systems-21st-international-conference-
tokyo-japan-october-29-november-2-2018-proceedings-tim-miller/
https://textbookfull.com/product/smart-blockchain-first-
international-conference-smartblock-2018-tokyo-japan-
december-10-12-2018-proceedings-meikang-qiu/
https://textbookfull.com/product/smart-computing-and-
communication-third-international-conference-smartcom-2018-tokyo-
japan-december-10-12-2018-proceedings-meikang-qiu/
Advanced Concepts for Intelligent Vision Systems 19th
International Conference ACIVS 2018 Poitiers France
September 24 27 2018 Proceedings Jacques Blanc-Talon
https://textbookfull.com/product/advanced-concepts-for-
intelligent-vision-systems-19th-international-conference-
acivs-2018-poitiers-france-september-24-27-2018-proceedings-
jacques-blanc-talon/
https://textbookfull.com/product/web-information-systems-
engineering-wise-2018-19th-international-conference-dubai-united-
arab-emirates-november-12-15-2018-proceedings-part-ii-hakim-
hacid/
https://textbookfull.com/product/knowledge-engineering-and-
knowledge-management-21st-international-conference-
ekaw-2018-nancy-france-november-12-16-2018-proceedings-catherine-
faron-zucker/
https://textbookfull.com/product/security-standardisation-
research-4th-international-conference-ssr-2018-darmstadt-germany-
november-26-27-2018-proceedings-cas-cremers/
https://textbookfull.com/product/semantic-technology-8th-joint-
international-conference-jist-2018-awaji-japan-
november-26-28-2018-proceedings-ryutaro-ichise/
Jian Chen · Yuji Yamada
Mina Ryoke · Xijin Tang (Eds.)
Knowledge and
Systems Sciences
19th International Symposium, KSS 2018
Tokyo, Japan, November 25–27, 2018
Proceedings
123
Communications
in Computer and Information Science 949
Commenced Publication in 2007
Founding and Former Series Editors:
Phoebe Chen, Alfredo Cuzzocrea, Xiaoyong Du, Orhun Kara, Ting Liu,
Dominik Ślęzak, and Xiaokang Yang
Editorial Board
Simone Diniz Junqueira Barbosa
Pontifical Catholic University of Rio de Janeiro (PUC-Rio),
Rio de Janeiro, Brazil
Joaquim Filipe
Polytechnic Institute of Setúbal, Setúbal, Portugal
Ashish Ghosh
Indian Statistical Institute, Kolkata, India
Igor Kotenko
St. Petersburg Institute for Informatics and Automation of the Russian
Academy of Sciences, St. Petersburg, Russia
Krishna M. Sivalingam
Indian Institute of Technology Madras, Chennai, India
Takashi Washio
Osaka University, Osaka, Japan
Junsong Yuan
University at Buffalo, The State University of New York, Buffalo, USA
Lizhu Zhou
Tsinghua University, Beijing, China
More information about this series at http://www.springer.com/series/7899
Jian Chen Yuji Yamada
•
Knowledge and
Systems Sciences
19th International Symposium, KSS 2018
Tokyo, Japan, November 25–27, 2018
Proceedings
123
Editors
Jian Chen Mina Ryoke
School of Economics and Management Faculty of Business Sciences
Tsinghua University University of Tsukuba
Beijing, China Tokyo, Japan
Yuji Yamada Xijin Tang
Graduate School of Business Sciences Institute of Systems Science
University of Tsukuba CAS Academy of Mathematics and Systems
Bunkyō, Tokyo, Tokyo, Japan Sciences
Beijing, China
This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd.
The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721,
Singapore
Preface
for each submission based on the review reports from the referees, who came from
China, Japan, New Zealand, Thailand, and the USA.
We received a lot of support and help from many people and organizations. We
would like to express our sincere thanks to the authors for their remarkable contri-
butions, all the Technical Program Committee members for their time and expertise in
reviewing the papers within a very tight schedule, and the proceedings publisher
Springer for their professional help. It is the third time that the KSS proceedings are
published as a CCIS volume after successful collaboration with Springer in 2016 and
2017. We greatly appreciate our four distinguished scholars for accepting our invitation
to deliver keynote speeches at the symposium. Last but not least, we are very indebted
to the local organizers for their hard work.
We were happy to witness the thought-provoking and lively scientific exchanges in
the fields of knowledge and systems sciences during the symposium.
Organizer
Host
General Chairs
Jian Chen Tsinghua University, China
Yuji Yamada University of Tsukuba, Japan
Chonghui Guo
Setsuya Kurahashi
Abstract. Many significant policies of our society and economy are determined
by someone day after day. However, most of the plans have been discussed and
decided based on past experiences and data. Many of them estimate policy
effects by analyzing actual phenomena and data using statistical methods. For
this method called evidence-based policymaking (EBP), this lecture proposes
model-based policymaking (MBP). The MBP is designed with an agent-based
model and data science techniques, and it also called as social simulation. The
model-based approach enables to design realistic phenomena as a model and
predict the effect on unfolding future events due to hypotheses or activities that
are difficult to experiment using computer experiments. In the field of business
and sociology, data analysis as an induction method and strategy planning as a
deductive method are connected. In the lecture, I will introduce urban dynamics
model, teaching model at school, analysis of education in a family using a
genealogy in China during 500 years.
Toward Cyber Physical Innovation:
Probabilistic Modeling
for Real Field AI Applications
Yoichi Motomura
Thanaruk Theeramunkong
Generating Risk Maps for Evolution Analysis of Societal Risk Events. . . . . . 115
Nuo Xu and Xijin Tang
1 Introduction
In traditional economics, consumers are supposed to be rational in that they
behave to maximize their utilities. However, consumers do not necessarily act
as rational entities, and the principle of utility maximization may not always
be applicable to describe all consumer behaviors. When goods are discounted,
for example, consumers often buy them even if they are not in need. A ratio-
nal economic entity would not risk wasting goods as it does not support utility
maximization. To explain such seemingly irrational behavior, consumers’ mental
conditions can be considered. Mental accounting is a concept of behavioral eco-
nomics that illustrates purchase behavior influenced by mental factors. In this
research, we develop models of inter-shopping duration that take account of this
mental accounting effect.
c Springer Nature Singapore Pte Ltd. 2018
J. Chen et al. (Eds.): KSS 2018, CCIS 949, pp. 1–16, 2018.
https://doi.org/10.1007/978-981-13-3149-7_1
2 K. Miyatsu and T. Sato
2 Models
2.1 Mental Loading Model
In this study, a new concept to capture mental pressure at the time of purchase
is introduced, called mental loading. This is defined as the cumulative spend at
a point in time during two consecutive paydays. Upon being paid every month,
mental loading is cleared and increases every time a new purchase is made. This
reflects a consumer’s mental condition at purchase; in other words, the consumer
feels more pressure spending when mental loading is high. However, information
on an individual’s payday is not available for the models and differs by individual.
Let us first define Loading_Period in Eq. (1) to specify the three most com-
mon types of paydays. First, loading_period l = 1 corresponds to a payday on the
25th of every month, which would then continue until the 24th of the following
month. In Japan, the three days in Eq. (1) are representative paydays:
⎧
⎪ th th
⎨l = 1 ⇒ 25 · Nmonth − 24 · (N + 1)month
Loading_P eriod l = 2 ⇒ 5 · Nmonth − 4th · (N + 1)month
th (1)
⎪
⎩ th th
l = 3 ⇒ 17 · Nmonth − 16 · (N + 1)month
Second, let cummi,ti ,l be denoted as the cumulative spend for household i, pur-
chase occasion at ti , and loading_period of l in Eq. (2). For the first time after
payday, cummi,ti ,l is always 0 and purchase amount Mi,j is added at every
purchase until the next payday:
transl (ti )−1
j=1 (Mi,j ) tranl (ti ) = 1
cummi,ti ,l = (2)
0 tranl (ti ) = 1
Each household has its own structural parameters, which reflect the composition
of the income earner(s) in a family. For example, in the case of α∗i = (1, 0), a
household has a single income earner paid on 25th of every month. When a
household’s parameter is α∗i = (0.5, 0.5), this family should have two income
earners with paydays on the 25th and 5th of every month. Figure 1 illustrates
these two cases.
single_income
40000
double_income
30000
yen
20000
10000
0
0 10 20 30 40 50
shopping occasions
(k)
Let location parameter μi (k = 1, 2) be assumed to have a linear combina-
tion structure with explanatory vector xi,ti for household i at time ti ; then, the
parameters become time-dependent as the explanatory vector changes over time.
(k) (k) (k)
Now, μi (k = 1, 2) is replaced with μi,ti = xti,tt βi (k = 1, 2), and the likelihood
of Eq. (4) is multiplied for all shopping occasions until Ti , where Ti is the time of
the last occasion of household i in the period. R(k) {CummMi,ti , γcum,i } is also
introduced as an operator to determine to which regime it belongs. Assuming
that all households are independent, the total likelihood is derived in Eq. (5):
(1) (2) 2(1) 2(2)
p({yi,Ti } | {βi }, {βi }, {σi }, {σi }, {γcum,i }, {CummMi,Ti }, {xi,ti }) (5)
H 2 Ti
1 −1 t (k)2
= √ (k)
exp 2(k)
{log(yi,ti )−x i,tt β i }
i=1 k=1 2πσi yi,ti 2πσi
ti ∈R(k) {CummMi,t ,γinv,i }
i
Heterogeneous Mental Accounting Impact of Inter-shopping Durations 5
ψi = zit Δψ + ψ ψ
i , i ∼ N(0, Σψ ) (6)
αi = zit Δα + α α
i , i ∼ N(0, Σα ) (7)
ψi and αi are expressed as linear functions of household attribute vector
zi = (zi,1 , zi,2 , ..., zi,q )t and coefficient matrix Δψ , Δα , where q is a dimension
of the attribute vector. Σψ and Σα are the variance-covariance matrixes of ψi
and αi , respectively.
For prior distributions, these matrixes are assumed to have a natu-
ral conjugate relationship of p(Δψ , Σψ ) = p(Δψ |Σψ )p(Σψ ), p(Δα , Σα ) =
p(Δα |Σα )p(Σα ), where p(Δψ |Σψ ), p(Δα |Σα ) are multi-variable normal dis-
tributions and p(Σψ ), p(Σα ) are inverse-Wishart distributions. As a result, the
coefficients of the linear functions of the hierarchical structure are expressed in
the following equations:
2 Ti
1 −1 (k) t (k) 2
√ (k)
exp 2(k)
{log(yi,ti ) − β i xi,t }
i
k=1 2πσi yi,ti 2πσi
ti ∈R(k) {cummi,t ,γcum,i ,αi }
i
(10)
2(1) 2(2)
The parameters included in the model are {ψi }, {αi }, {σi }, {σi },
2(1) 2(2)
Δψ , Δα , Σψ , Σα , among which {ψi }, {αi }, {σi }, {σi } are heterogeneous
to the household, while Δψ , Δα , Σψ , Σα are common in all households. The
Markov Chain Monte Carlo method is applied to estimate the parameters.
Because Δψ , Δα , Σψ , Σα are structured in natural conjugate relationships
defined in Eqs. (8) and (9), these parameters can be drawn by using Gibbs
samplers. On the contrary, {ψi }, {αi } cannot formulate the conjugate structure,
and thus it needs random walk samplers of the Metropolis-Hastings algorithm.
2(1) 2(2)
σ i , σi are sampled independently by the Gibbs samplers:
(k)
expressed in inverse-Gamma form with βi in Eq. (15):
2(k) (k) ∗
σi | βi ∼ IG(vi,n /2, s∗i,n /2), k = 1, 2 (15)
∗ (k) t (k)
where vi,n v0 + ni , s∗i,n = s0 + (log(yi ) − xi t βi ) (log(yi ) − xi t βi ), v0 = 100,
s0 = 100.
For the random walk sampling of ψi , αi , candidate samples are iteratively
(r−1) 2 (r−1) 2
generated by ψi = ψi + δRW ψ
I and αi = αi + δRW α
I, where the vari-
ance in random walk noise is set as σRWψ = 10 , σRWα = 10−2 , respectively.
2 −5 2
4 Empirical Analysis
4.1 Data
We estimate the models by using scanner panel data from a retailer’s store
in Tokyo between January 1 and December 31, 2001. Whenever a household
member visits the store and makes a purchase, the details of the transaction
are recorded for each household. Altogether, 100 sample panels were randomly
selected of those who shopped more than 50 occasions and at least once every
month in that period. Table 1 shows the mean and standard deviation (SD) of
the variables and household attributes.
The explanatory variables consist of two types: household-specific and com-
mon for all households. The discount rate and items on the flyer are identical for
all households at the time of shopping, whereas the number of items purchased
on the previous shopping occasion differs by household. Unlike the brand choice
model in marketing science, which concerns particular products or categories,
we consider all the products in the store. The discount rate is defined as the
average rate of products available at all times, which is 16% in the study period.
The products examined in this study only include non-perishable food and sun-
dries. Perishable food is merchandised on market quotations, and no discount
rate is available. The number of items on the flyer includes perishable food, and
the average amount is 37.2. As each household has its own family structure and
shopping pattern, the number of items purchased per store visit differs. The
average number of items per shopping is 14.3, of which 66% are non-perishable
items. Household attributes are calculated based on holdout samples before the
study period. Average spend per shopping trip is 1,537 JPY, and 61% of house-
holds are located within walking distance of the store. The region around the
retail store has 55% of part-time workers.
By using scanner panel data from the 100 randomly selected household panels,
we estimated the heterogeneous parameters of the proposed model (Asymmetric
Model) together with the two other models: the model with no regime (Null
8 K. Miyatsu and T. Sato
Model) and the threshold model with symmetric data without the mental load-
ing variable (Symmetric Model). The Markov Chain Monte Carlo method was
applied for 50,000 iterations; 45,000 samples were discarded as the burn-in period
and the remaining 5,000 samples were used in the analysis. Table 2 summarizes
the statistics for the model valuation. Both statistics support our proposed model
(Asymmetric Model) following the Symmetric Model and Null Model.
(1) (2)
4.3 Regression Parameter (βi , βi )
Table 3 shows the statistics for the average coefficient parameters of all the house-
holds in each regime. Under Bayesian statistics, heterogeneous coefficients are
Heterogeneous Mental Accounting Impact of Inter-shopping Durations 9
estimated for each household to analyze the impact on the inter-shopping dura-
tion by household. However, in this subsection, we discuss shopping behavior
based on the average coefficients of all household panels. The intercept, which is
(1) (2)
a baseline of location parameter μ(k) , is β1 > β1 . Eliminating all the factors,
the inter-shopping duration tends to be longer in Regime 1. However, with a
(1)
negative sign of β6 , mental loading shortens this duration when it exceeds the
threshold. As a result, the inter-shopping duration tends to be shorter in Regime
(k)
1. Regarding β2 , the more items a household purchases, the longer is the dura-
tion to the next shopping occasion. When a household possesses more goods in
stock, it is rational to refrain from shopping, as there is no immediate need to
visit the store. However, this is no longer the case or a household seems to be
indifferent to the inter-shopping duration when it feels more pressure to spend.
This phenomenon is noteworthy when more non-perishable goods are procured.
(1) (2)
For the ratio of non-perishable items, β3 < β3 means that it becomes more
(1) (2)
insensitive when the regime switches to Regime 1. On the contrary, β4 < β4
implies that the discount is more sensitive in Regime 1. Lastly, the expected sign
(k)
of β5 is negative, suggesting that the duration shortens when more items are
on a flyer; however, this is only the case in Regime 2. Flyers are not as effective
in Regime 1. In general, the inter-shopping duration becomes shorter in Regime
1, where a household feels more pressure to spend. From a purchase behavior
viewpoint, households thus turn out to be more planned in Regime 1 and they
visit a store to fulfill their immediate needs at any time.
Table 3 shows the statistics of the average threshold parameters of all the house-
holds. The average threshold is 11,649 JPY, indicating that regime switching
occurs when mental loading exceeds 27.3% of monthly expenditure on average.
(1)
Understanding the relationship between γcum,i and β6,i would help explain
how mental loading affects the duration. Figure 3 shows the scatter plots of γcum,i
(1)
and β6,i . The higher the threshold, the less effective is the response of mental
loading. Households with a lower threshold perceive pressure to spend at an
earlier stage, raising the impact of mental loading. As a result, the inter-shopping
duration shortens and thus a household purchases goods more frequently to fulfill
their immediate needs.
(2)
4.5 Mental Loading Structural Parameter (αi (1) , αi )
Mental loading is the latent variable, and the structural parameters determine
the composition of income earners based on their monthly payday dates. Figure 4
∗(1) ∗(2)
shows the scatter plots of (αi , αi ). The plots on (1, 0), (0, 1), (0, 0) corre-
spond to single-income households with a loading period l = 1, 2, 3. The plots on
∗(1) ∗(2) ∗(2) ∗(1)
the αi + αi = 1, αi , and αi axes indicate double-income households
with the loading periods l = (1, 2), (2, 3), (3, 1), respectively; the others are from
10 K. Miyatsu and T. Sato
0.0
−0.2
−0.4
gamma
triple- income households. Figure 4 also summarizes the occupancy rate of the
composition of income earners among all household panels, counted based on
estimating α∗i by disregarding values below 0.01. The shares of single-, double-,
and triple- income households are 39%, 42%, and 19%, respectively. In most com-
panies in Japan, employees are paid every month; however, the payday date dif-
fers largely depending on the size and ownership status of the company: Table 3
Heterogeneous Mental Accounting Impact of Inter-shopping Durations 11
shows the statistics of the average threshold parameters of all the households.
The average threshold is 11,649 JPY, indicating that regime switching occurs
when mental loading exceeds 27.3% of monthly expenditure on average.
(1)
Understanding the relationship between γcum,i and β6,i would help explain
how mental loading affects the duration. Figure 3 shows the scatter plots of γcum,i
(1)
and β6,i . The higher the threshold, the less effective is the response of mental
loading. Households with a lower threshold perceive pressure to spend at an ear-
lier stage, raising the impact of mental loading. As a result, the inter-shopping
duration shortens and thus a household purchases goods more frequently to fulfill
their immediate needs. l = 1 is applicable for large publicly listed corporations,
l = 2 for private and small or medium-sized companies, and l = 3 for government
and public service entities. At a glance, the proportion of triple-income house-
holds seems higher. As family size has deceased in Japan, the triple- income
earners in a family are higher than expected. However, most part-timers have
two paydays a month, so this is treated as different income earners in this esti-
mation.
regime_1 regime_2
1.0
1.0
usr1 usr1
usr2 usr2
usr3 usr3
0.8
0.8
0.6
0.6
hazard
hazard
0.4
0.4
0.2
0.2
0.0
0.0
0 5 10 15 20 25 0 5 10 15 20 25
regime_1 regime_2
1.0
1.0
usr4 usr7
usr5 usr7
usr6 usr9
0.8
0.8
0.6
0.6
hazard
hazard
0.4
0.4
0.2
0.2
0.0
0.0
0 5 10 15 20 25 0 5 10 15 20 25
parameters are calculated by using the average values of the explanatory vector
and duration is the average inter-shopping duration for each regime.
As usr1–3 have shopping occasions in each regime, they have different hazard
functions, which are not monotonously increasing or decreasing but unimodal.
The location and height of a peak are different for each household. Figure 5
illustrates the hazard functions of the users. They all have sharp-pointed peaks
in Regime 1. Their purchase behavior changes once they enter Regime 1, and
they shift the mindset of paying attention to fulfill their immediate needs; thus,
the inter-shopping duration becomes shorter. Usr4–6 and usr7–9 have shopping
occasions in a single regime only. Figure 6 show the hazard functions of usr4–6 on
the left and usr7–9 on the right. Whenever a single regime remains, their hazard
Another random document with
no related content on Scribd:
How the Lord Cromwell[2006] exalted
from meane estate, was after by the
enuie of the Bishop of Winchester,
and other his complices, brought to
vntimely end, Anno Dom. 1540.
1.
2.
3.
You, that but hearing of my hated name,
Your ancient malice instantly bewray,
And for my sake your ill deserued blame
Vpon my legend publikely shall lay:
Would you forbeare to blast me with defame,
Might I so meane a priuiledge but pray,
He that three ages hath endur’d your wrong,
Heare him a little that hath heard you long.
4.
5.
6.
7.
8.
9.
11.
12.
13.
14.
15.
16.
17.
18.
19.
20.
21.
22.
23.
24.
25.
26.
28.
29.
30.
31.
32.
33.
34.
35.
36.
37.
38.
39.
40.
41.
42.
43.
45.
46.
47.
48.
49.
50.
51.
52.
53.
54.
But their great force against me wholly bent
Preuail’d vpon my purposes so farre,
That I my ruine scarsely could preuent,
So momentarie worldly fauours are,
That till the vtmost of their spight was spent,
Had not my spirit maintain’d a manly warre,
Risen they had when laid I had been low,[2051]
Vpon whose ruine after I did grow.
55.
56.
57.
58.
59.
60.
62.
63.
64.
65.
66.
67.