Quantify fying Social In Influence in in Epin inions Akshay - - PowerPoint PPT Presentation

Quantify fying Social In Influence in in Epin inions

Akshay Patil, Golnaz Ghasemiesfeh, Roozbeh Ebrahimi & Jie Gao

In Introduction

Social Network:

• Structure made up of entities & their relationships.

i.e.: Facebook, G+, Y!, etc.

Content Generation Websites:

• i.e.: Wikipedia, Youtube, Epinions, Instagram, etc.

Explosion in “online social activity”+ “content generation”:

• Large-Scale data availability.
• Quantitative research into dynamics.

Overlay of “content generation”+ “social structure”

• Study the mutual influence of content and

social structure on each other.

: : A Consumer Review Sit

ite

Epinions.com incorporates a social structure into its rating system:

• Rating system
• Users write reviews
• Users rate other user’s reviews [1-5] (1:bad ->5:Excellent)
• Social structure
• Trust other users to form “web of trust” (public)
• Distrust other users to form “block list” (private)
• We are interested in the interplay of “rating system data”

and “social structure data”.

Epinions

Dataset

Statistic

#Users 131,828 #Reviews 1,197,816 #Trust Edges 717,667 (85%) #Distrust Edges 123,705 (15%) #Ratings 12,943,546 Time Range Jan’01 to Aug’03

0.01% 2.13% 4.63% 14.30 %

78.93 %

Ratings

1 2 3 4 5

1. . Relationship Formation

Scenario:

• User A has a couple of trustees

(his “web of trust”, or “friends”)

• A’s friends have a trust/distrust

relationship with user B. Classification:

• If the majority of A’s friends trust B, we

say they collectively trust B.

• If the majority of A’s friends distrust B, we

say the collectively distrust B.

• Other wise they are neutral or in disagreement.

1 2 … n A B

1. . Relationship Formation

Question 1: Is there a correlation between the collective opinion of A’s friends about B and his future relationship with B?

• If A’s friends collectively trust B,

is A more likely to trust B as well?

• If A’s friends collectively distrust B,

is A more likely to distrust B as well?

1 2 … n A B

Raw Observations: How meaningful (statistically significant) are these results?

1. . Relationship Formation

95.09% 85.93% 93.85%

80.00% 85.00% 90.00% 95.00% 100.00%

Collectively Trust Collectively Distrust Opinionated

1. . Relationship Formation, Random Shuffle

Measure over/under representation compared to mere chance (approach by Leskovec et al.’10):

• Randomly shuffle trust/distrust edges, while maintaining

the same percentages.

• Redo “relationship formation” analysis.
• Compute “Surprise” value: 𝒕 =

𝑹−𝑭 𝑹 𝑭 𝑹 𝟐−𝒒𝟏

• Q: actual quantity of a scenario,

E[Q]: expected quantity under shuffling, p0: priori probability of the scenario.

1. . Relationship Formation, Random Sh Shuffle

Surprise is number of standard deviations by which the actual quantity differs from the expected number under the random shuffling model.

• s > 0  overrepresentation
• s < 0  underrepresentation
• s = 6  p-value ≈ 10-8
• A value of s > 6 results in excellent statistical significance.

“Agreeing with Friends”: surprise values in excess of 70! Strong correlation between a user’s friends opinions and formation of his future relationships (distrust is hidden).

1. . Relationship Formation, Random Sh Shuffle

• 120
• 90
• 60
• 30

30 60 90 Agreeing with Friends Disagreeing with Friends Neutral Friends Surprise

1. . Relationship Formation: Li Linking Habits

The dataset exhibits users with very different linking habits

• Some are very trustworthy/ trustful compared to others.

Analysis should not overlook

• The quality of the reviews (trustworthiness) written by the user.
• The degree of trustfulness of the person creating the link.

Looking through the lenses of linking habits (Leskovec et al.’10). Receptive Baseline (trustworthiness):

• Fraction of received trust links.

Generative Baseline (trustfulness):

• Fraction of given trust links.

1. . Relationship Formation: Li Linking Habits

Receptive/Generative Surprise: Number of standard deviations the quantity is above the expected number.

• If B was trusted/distrusted based solely on his trustworthiness,

Receptive Surprise = 0.

• If A made his decision based solely on his trustfulness,

Generative Surprise = 0. Observations: Receptive Surprise Generative Surprise Collectively Trust 96.76 34.99 Collectively Distrust

• 104.15
• 56.31

1. . Relationship Formation: Li Linking Habits

Collectively Trust

• Users exceed both generative and receptive baselines in

trusting and being trusted.

• This can be explained by homophily or influence of

friends. Collectively Distrust

• Users fall behind generative/receptive baselines.
• This can be explained by heterophobia or lack of context

by friends (distrust edges are hidden).

2. . Fri riend of f Friend (FoF) Dynamics

Scenario:

• User A trusts user B and user B trusts

user C.

• A does not have a trust/distrust edge

to C.

• A rates a review by C.

Question 2: Is A more likely to give a favorable rating to C’s review? Why is this about the influence of the social structure on the rating system data?

B A C

Rate

2. . Fri riend of f Friend Dynamics

B A C

Rate

B A C

Rate

B A C

Rate

B A C

Rate

FoF FoE EoF EoE

2. . FoF Dynamics: Random Shuffling

• 1500
• 1000
• 500

500 1000 1500 2000

1 2 3 4 5

Surprise

FoF

• 400
• 200

200 400 600 800 1000

1 2 3 4 5

Surprise

EoF

2. . FoF Dynamics: Random Shuffling

• 400
• 300
• 200
• 100

100 200 300 400 500

1 2 3 4 5

Surprise

FoE

• 200
• 150
• 100
• 50

50 100 150 200

1 2 3 4 5

Surprise

EoE

2. . FoF Dynamics: Rating Habits Analysis

Rating Gen./Rec. Surprise FoF EoF FoE EoE 1

• Gen. sur.
• Rec. sur.
• 43.77
• 10.91

66.03 19.80

• 8.36

57.13

• 1.43

2.76 2

• Gen. sur.
• Rec. sur.
• 627.54
• 527.77

789.36 89.58

• 108.83

206.16 26.54 4.03 3

• Gen. sur.
• Rec. sur.
• 360.72
• 181.17

2.01 10.16

• 304.42

65.53

• 124.23

5.89 4

• Gen. sur.
• Rec. sur.
• 847.21
• 370.22
• 115.23
• 3.57
• 381.94

81.06

• 190.69
• 4.27

5

• Gen. sur.
• Rec. sur.

1065.09 519.91

• 189.93
• 61.03

531.88

• 173.88

214.75

• 1.36

2. . Fri riend of f Friend Dynamics : : Summary

Distinct Trends in 2 (out of 4) scenarios,

• FoF: Shift towards assigning higher ratings to C’s review

(Specially 5).

• Homophily/Influence
• EoF: Shift towards assigning lower ratings to C’s review

(Specially 1 &2).

• Heterophobia/Influence?

In the remaining 2 scenarios (FoE & EoE), it is hard to get a solid interpretation.

3. . Building a Predictor: Corr. Analysis

Utilize FoF dynamics as features and “actual rating” as target value.

Dynamics Correlation Coefficient FoF 0.1112 EoF

• 0.0918

FoE 0.0105 EoE

• 0.0001

3. . Building a a Predictor: Pic icking Features

Class Feature Information Gain

Trust

A’s Generative Baseline 0.1595 C’s Generative Baseline 0.2291 A’s Receptive Baseline 0.1943 C’s Receptive Baseline 0.4496

Rating

• Avg. Rating given by A

0.3316

• Avg. Rating given by C

0.3776

• Avg. Rating received by A’s Reviews

0.2453

• Avg. Rating received by C’s Reviews

0.5362

FoF Dynamics

Number of FoF Paths 0.3813 Number of EoF Paths 0.1894 Number of FoE Paths 0.0119 Number of EoE Paths 0.0198

3. . Prediction Results (B (Bootstrap Aggregating)

0.65 0.7 0.75 0.8 0.85

Low (1, 2) Medium (3) High (4, 5)

ROC Area (AUC) = 0.91 Overall Accuracy = 76%

Precision Recall F-Score

Conclusion

Relationship Formation

• Random Shuffle and Linking Habits: Strong Correlation.

Exceeding both generative and receptive baselines in trusting and being trusted.

Friend of Friend Dynamics

• FoF: Shift towards assigning higher ratings to C’s review.

EoF: Shift towards assigning lower ratings to C’s review.

Building a Predictor

• This alignment can be used to predict (recommend) content that

would be likeable by a user.

• We achieve good prediction accuracy with a simple feature set.

Thank You!