Transcript KDD12-eTrust - Arizona State University

eTrust: Understanding Trust
Evolution in an Online World
Jiliang Tang, Huiji Gao and Huan Liu
Computer Science and Engineering
Arizona State University
August 12-16, 2012 KDD2012
Data Mining and Machine Learning Lab
Atish Das Sarma
eBay Research Lab
eBay Inc.
Trust and Its Evolution
• Trust plays an important role in helping online
users collect reliable information
– Abundant research on static trust for making good
decisions and finding high quality content
• However, trust evolves as people interact and time
passes by
– It is necessary to study its evolution
– Its study can advance online trust research for trust
related applications
Our Contributions
1. We identify the differences of trust study in
physical and online worlds
2. We investigate how to study online trust
evolution
3. We show if this study can help improve the
performance of trust related applications
Research in Physical and Online Worlds
• Trust evolution in a physical world
- Step 1: inviting a group of participants ( a small group)
- Step 2: recording their sociometric information
- Step 3: recording conditions or situations for the change
• Differences encountered in an online world
- Users are world-widely distributed
- Sociometric information on trust is unavailable
- Passive observation is the modus operandi to gather data
Studying Online Trust Evolution
• Overcoming the challenge of passive observation
– Where can we find relevant data for trust study (an issue
about environment)
– How can we infer about the information about trust (an
issue about methodology)
• Modeling online trust evolution
– How to incorporate social theories mathematically
• Evaluating the gain of trust evolution study
– Rating prediction and trust prediction
Online Rating System
time t
Online Rating System
time t
time t+1
Online Rating System
time t
Temporal Information
time t+1
Social Science theories
• Correlations between rating and user
preference
- Dynamics of rating
• Correlations between user preference and
trust
- Drifting user preferences
Methodology for Trust Evolution
Social theories
Temporal
information,
rating etc
Social theories
Dynamics of
user preference
Rating Prediction
Online Rating
System
Trust
Evolution
Our Framework: eTrust
Components of eTrust
Part 1
Part 4
Part 2
Part 3
Part 1: Modeling Rating via User Preference
• Rating is related to user preference and item
characteristic
-
-
t
i
is the preference of i-th user in time t, q j is the
characteristic of j-th item and K is the number of latent
facets of items
p
Part 2: Modeling Rating via Trust Network
• People is likely to be influenced by their trust networks
Trust strength
between i-th and
v-th users in the
k-th facet
Decaying the earlier rating
Part 3: Modeling Trust and User preference
• Modeling the correlation between trust and
user preference
t
is preference similarity vector in the k-th facet and
ivk
s
is a user specific bias
bi
Part 4: Modeling Change of User Preference
• Modeling the change of user preference
c is a function to control how user preference change, λ
controls the speed of change
Experiments
• Datasets
• Findings from the study of trust evolution
• Can eTrust help improve trust related
applications?
- Rating Prediction
- Trust Prediction
Experiments
• Datasets
• Findings from the study of trust evolution
• Can eTrust help improve trust related
applications?
- Rating Prediction
- Trust Prediction
Datasets
• Epinions
- Product review sites
- Statistics
http://www.public.asu.edu/~jtang20/datasetcode
/truststudy.htm
Splitting the Dataset
• Epinions is separated into 11 timestamps
11thJan, 11thJan,
2001,
2002,
T1
T2
11thJan,
2009,
…….
11thJan,
2010,
T10
T11
Experiments
• Datasets
• Findings from the study of trust evolution
• Can eTrust help improve trust related
applications?
- Rating Prediction
- Trust Prediction
Speed of Change of Trust
• The evolution speed of an open triad is 6.12
times of that of a closed triad
User preferences drift over time
The speed of change varies with
people and facets
Experiments
• Datasets
• Findings from the study of trust evolution
• Can eTrust help improve trust related
applications?
- Rating Prediction
- Trust Prediction
Experiments
• Datasets
• Findings from the study of trust evolution
• Can eTrust help improve trust related
applications?
- Rating Prediction
- Trust Prediction
Applications of eTrust:
Rating Prediction
• Given ratings before T, we predict ratings in
T+1 as,
Testing Datasets
• We further divide data in T11 into two testing
datasets
- N: the ratings involved in new items or new users(10.06%)
- K: the remaining ratings
Comparison of Rating Prediction
Experiments
• Datasets
• Findings from the study of trust evolution
• Can eTrust help improve trust related
applications?
- Rating Prediction
- Trust Prediction
Applications of eTrust:
Trust Prediction
• The likelihood of trust establishing is
estimated as,
Testing Datasets
• We also divide data in T11 into two testing
datasets
- E: trust relations established among existing users
- N: trust relations involved in new users (23.51%)
Comparison of Trust Prediction
Future Work
• Seek more applications for eTrust
- Ranking evolution
- Recommendation systems
- Helpfulness prediction
• Generalize eTrust to other online worlds
- e-commerce
Questions
Acknowledgments: This work is, in part, sponsored by ARO via a grant (#025071).
Comments and suggestions from DMML members and reviewers are greatly
appreciated.