social bridges in urban purchase behavior - mit media labxdong/talk/lincolnlab_social... · 2017....

Social bridges in urban purchase behavior

Xiaowen DongMIT Media Lab

Cambridge, MA, June 2017

with Yoshihiko Suhara, Burçin Bozkaya, Vivek K. Singh, Bruno Lepri and Alex ‘Sandy’ Pentland

/22

Introduction

2

New data sources about human behavior are emerging

/22

Introduction

2


Computational social science (CSS): A paradigm shift in social science

t

/22

Introduction

2

Current population management:

- demographics- individual records- static information

The new way:- behavioral traits- collective behavior- dynamics



t

Practical impact

/22

Introduction

2

Current population management:

- demographics- individual records- static information

The new way:- behavioral traits- collective behavior- dynamics



t

Practical impact

How communication affects human decision-making?

/22

Introduction

3

/22

Introduction

3

pi /popi

di

• Classical purchase behavior models treat individual purchases separately (Huff, 1964)

/22

Introduction

3

• Study of purchase behavior influence is largely based on socio-demographics (Zeithaml, 1985)

pi /popi

di


/22

Introduction

3

• Study of purchase behavior influence is largely based on socio-demographics (Zeithaml, 1985)

pi /popi

di

• Word-of-mouth and physical exposure are powerful sources of behavioral propagation (Arndt, 1967; Bikhchandani, 1998; Algesheimer, 2005), but their effectiveness in modern city environment remains unknown


/22

Introduction• Hypothesis

4

- Physical exposure at work environment promotes idea exchange

/22


4

- Physical exposure at work environment promotes idea exchange- Individuals living in different communities but sharing similar work locations act as

social bridges between communities

/22


4

- Physical exposure at work environment promotes idea exchange- Individuals living in different communities but sharing similar work locations act as

social bridges between communities

• Test at city scale

/22

Data set• A large-scale credit card transaction data set in two cities in an OECD

country during 3 months

5

/22

Methods• Urban communities

6

/22


6

• Number of social bridges between communities

bdg(I, J) = |{i, j}|

s.t. i 2 I, j 2 J, D(Li, Lj) d

/22


6

• Number of social bridges between communities

# bridges = 4 # bridges = 4

bdg(I, J) = |{i, j}|

s.t. i 2 I, j 2 J, D(Li, Lj) d

/22

Methods• Three behavioral indexes

- choice: number of co-visited stores- temporal: similarity between temporal distributions of purchases- spending: sum of differences in median spending amount of different categories

7

/22

Methods• Three behavioral indexes

- choice: number of co-visited stores- temporal: similarity between temporal distributions of purchases- spending: sum of differences in median spending amount of different categories

7

• Remark- exclude transactions during working hours- exclude transactions at stores in home/work neighborhoods

/22

Social bridge and behavioral indexes

8

0 1 2 3 4 5 6 7 8 9−50

0

50

100

150

200

250

log2( bdg(I,J) +1)+1

mdi

ff(I,J

)

0 1 2 3 4 5 6 7 8 90.4

0.5

0.6

0.7

0.8

0.9

1


tsim

(I,J)

0 1 2 3 4 5 6 7 8 90

100

200

300

400

500


covi

sit(I

,J)

0 1 2 3 4 5 6 7 8 9−50

0

50

100

150

200

250


mdi

ff(I,J

)

0 1 2 3 4 5 6 7 8 90.4

0.5

0.6

0.7

0.8

0.9

1


tsim

(I,J)

0 1 2 3 4 5 6 7 8 90

500

1000

1500

2000

2500


covi

sit(I

,J)

City

BCi

ty A

Choice (co-visits) Temporal Spending

/22

Social bridge and purchase similarity (co-visits)• Multiple OLS regression analysis

- dependent variable (DV): # co-visits (between community pair)- independent variables (IV): # social bridges- confounding variables: population, distance, demographics, income

9

/22

Social bridge and purchase similarity (co-visits)• Multiple OLS regression analysis

- dependent variable (DV): # co-visits (between community pair)- independent variables (IV): # social bridges- confounding variables: population, distance, demographics, income

9

• Remark- entries are not independent in DV and IV- Quadratic Assignment Procedure (QAP) to test statistical significance

‣ random shuffling of communities in DV

‣ re-application of OLS

/22

Social bridge and purchase similarity (co-visits)• Regression coefficients

10

/22

Social bridge and purchase similarity (co-visits)• Regression coefficients

10

Social bridge is a stronger indicator of similar purchase behavior

/22

Social bridge and purchase similarity (co-visits)• Histogram of distance between co-visited store and co-working location

11

0 20 40 60 80 1000

1

2

3

4

5 x 104

distance (km)

freq

uenc

y co

unt

0 20 40 60 80 1000

1

2

3

4

5

6

7

8 x 105

distance (km)

freq

uenc

y co

unt

City A (62% > 2km) City B (74% > 2km)

Co-visitation is not simply due to proximity between co-visited store and co-working location

/22

Co-visits by two types of customers• Bridge customers vs. Non-bridge customers

12

/22


12

Bridge customers

/22


12

Bridge customers

Non-bridge customers

/22

Co-visits by two types of customers• Histogram of ratio of bridge customers

13

−0.1 0 0.1 0.2 0.3 0.4 0.50

1

2

3

4

5

6

7

8 x 104

ratio of bridge customers

freq

uenc

y co

unt

−0.1 0 0.1 0.2 0.3 0.4 0.50

1

2

3

4

5

6

7

8 x 104

ratio of bridge customers

freq

uenc

y co

unt

City A City B

Ratio of bridge customers are relatively small

/22

Co-visits by two types of customers• Percentage of co-visits by bridge customers

14

City A City B

A large portion of co-visits are by non-bridge customers

/22

Co-visits by two types of customers• Regression coefficients

15

/22

Co-visits by two types of customers• Regression coefficients

15

Social bridge is a indicator of similar purchase behavior even for non-bridge customers

/22

Co-visits in three merchant categories• Regression coefficients

16

/22

Co-visits in three merchant categories• Regression coefficients

16

Effect of social bridge is stronger for restaurants but weaker for supermarkets

/22

Gender difference in social bridge• Regression coefficients

17

/22

Gender difference in social bridge• Regression coefficients

17

Female-female bridges show a stronger effect

/22

Comparison with a null model

18

pi /popi

di

• Purchase choices are influenced by merchant popularity and location (Huff, 1964)

/22


18

pis =uisPs2S uis

=A↵1

s /D↵2isP

s2S(A↵1s /D↵2

is )

pi /popi

di


probability customer i visits store s

/22


18

pis =uisPs2S uis

=A↵1

s /D↵2isP

s2S(A↵1s /D↵2

is )

pi /popi

di


probability customer i visits store s

popularity of store s

distance between customer i and store s

/22

Comparison with a null model• Simulate individual purchases and co-visitation between communities• Compare the regression coefficient with the empirical one

19

/22

Comparison with a null model• Simulate individual purchases and co-visitation between communities• Compare the regression coefficient with the empirical one

19

City A City B

Effect of social bridge is not simply due to merchant popularity and location

/22

Influence of distance threshold• Regression coefficient as a function of distance d

20

/22


20

City A City B

0 0.10 0.14 0.21 0.30 0.43 0.62 0.89 1.27 1.83

−0.2

0

0.2

0.4

0.6

0.8

1

distance threshold (km)

regr

essi

on c

oeffi

cien

t

co−visits by allco−visits by bridge cus.co−visits by nonbridge cus.co−visits by all (shuffled network)co−visits by bridge cus. (shuffled network)co−visits by nonbridge cus. (shuffled network)

0 0.10 0.14 0.21 0.30 0.43 0.62 0.89 1.27 1.83

−0.2

0

0.2

0.4

0.6

0.8

1


regr

essi

on c

oeffi

cien

t


/22


20

City A City B

0 0.10 0.14 0.21 0.30 0.43 0.62 0.89 1.27 1.83

−0.2

0

0.2

0.4

0.6

0.8

1


regr

essi

on c

oeffi

cien

t


0 0.10 0.14 0.21 0.30 0.43 0.62 0.89 1.27 1.83

−0.2

0

0.2

0.4

0.6

0.8

1


regr

essi

on c

oeffi

cien

t


Peak region of blue curve (co-visits by non-bridge customers) suggests geographical constraint for social bridge effect

/22

Application: Prediction of co-visits• Three-class classification: small, medium, large amount of co-visitation• For each IV (feature), train on 20% of communities and test on the rest

80%, using LIBSVM (Chang, 2011)

21

/22

Application: Prediction of co-visits• Three-class classification: small, medium, large amount of co-visitation• For each IV (feature), train on 20% of communities and test on the rest

80%, using LIBSVM (Chang, 2011)

21

Social bridge is more efficient in predicting co-visitation than traditional factors

/22

Discussion• Social bridge captures a form of social learning due to physical exposure:

similar to “the familiar stranger” (Milgram, 1977)

22

/22



22

• Bridge customers are conceptually similar to “structural hole spanners” (Lou, 2013)

/22



22


• Easy to compute: as long as location information (social media, etc.) is available

/22



22



• No causal relation, but tested against demographics and null model based on popularity and distance (Huff, 1964)

/22



22



• No causal relation, but tested against demographics and null model based on popularity and distance (Huff, 1964)

• Strong correlation can lead to applications such as behavior prediction and stratification, campaign targeting, and resource allocation

social bridges in urban purchase behavior - mit media labxdong/talk/lincolnlab_social... · 2017....

Documents