Using Social Trust for Recommender Systems

1
Using Social Trust for Recommender Systems

• Jennifer Golbeck
• Human-Computer Interaction Lab
• University of Maryland, College Park

2
How many cows in Texas?
http//www.cowabduction.com/
3
(No Transcript)
4
gt1,000,000,000
5
Introduction

• Trust is a way of doing social personalization
• Understand how to compute trust
• See applications where trust is being used for
  creating recommendations

6
Defining Trust

• Overall, trust is very complex
• Involves personal background, history of
  interaction, context, similarity, reputation,
  etc.
• Sociological definitions
• Trust requires a belief and a commitment
• E.g. Bob believes Frank will provide reliable
  information thus Bob is willing to act on that
  information
• Similar to a bet
• In the context of recommender systems, trust is
  generally used to describe similarity in opinion
• Ignores authority, correctness on facts

7
(No Transcript)
8
Trust Inference
The Goal Select two individuals - the source
(node A) and sink (node C) - and recommend to the
source how much to trust the sink.
tAC
A
B
C
tAB
tBC
9
Major Algorithms - Networks

• Advogato (Levien)
• Appleseed (Ziegler and Lausen)
• MoleTrust (Massa and Avesani)
• TidalTrust (Golbeck)

10
Advogato

• Levien 2003
• Attack resistant
• Maximum network flow based on Ford-Fulkerson
• Node capacities determined by the distance from
  the source
• This is a single source, multiple sink problem
  with capacities on the nodes
• Network flow works on a single source single sink
  problem with capacities on the edges
• The graph is transformed

11
(No Transcript)
12
Appleseed

• Ziegler and Lausen, 2004
• Based on spreading activation models
• Trust is modeled as energy
• Source node is activated through an injection of
  energy e
• e is then propagated to other nodes along edges
• All energy is fully divided among successor nodes
  wrt. their local edge weight.
• Supposing average out degrees gt 1, the closer
  the sink is to the source, and the more paths
  leading from the source to the sink, the higher
  energy flowing to sink
• Output ranking of nodes

13
TidalTrust and MoleTrust

• If the source does not know the sink, the source
  asks all of its friends how much to trust the
  sink, and computes a trust value by a weighted
  average
• Neighbors repeat the process if they do not have
  a direct rating for the sink

14
(No Transcript)
15
Exercise

• Simulator with the basics of MoleTrust and
  TidalTrust is available online
• http//zaphod.mindlab.umd.edu/SocialTrustCourse/tr
  ustDemo.html

16
What About Confidence?

• If decisions will be based on trust, we should
  know how confident we are in the trust estimate
• How do we reflect low confidence in an inferred
  value?

B
0.6
0.9
A
D
C
1
0.6
0.4
E
F
G
H
0.2
1
0.1
17
SUNNY

• Trust inference algorithm using Bayesian Networks
• Trust network is mapped into a Bayes Net
• Conditional Probability values are computed
  through profile similarity measures
• A most confident subnetwork is selected and
  trust inference is performed on that network
• Result is an inferred trust value and a
  confidence in that value

18
A Note About Distrust

• It is hard to integrate distrust into a
  propagation model
• If Alice distrusts Bob and Bob distrusts Chuck,
  should Alice trust Chuck or not?
• With no clear treatment, distrusted individuals
  could be pruned away, thus providing little
  benefit to including distrust

19
Trust from Similarity

• If there is no social network with trust values,
  how do we compute trust?
• Assuming there are underlying ratings, can we
  compute trust from similarity?
• How does this differ from collaborative filtering?

20
Exercise for the audience

• Your Favorite Movie ?????
• Your Least Favorite Movie ?????
• Some Mediocre Movie ?????
• Some Mediocre Movie ?????
• Some Mediocre Movie ?????
• Some Mediocre Movie ?????
• Some Mediocre Movie ?????
• Some Mediocre Movie ?????
• Some Mediocre Movie ?????
• Some Mediocre Movie ?????

21
Sample Profile
Knowing this information, how much do you trust
User 7 about movies?
22
Factors Impacting Trust

• Overall Similarity
• Similarity on items with extreme ratings
• Single largest difference
• Subjects propensity to trust

23
Propensity to Trust
Each layer represents a user
24
Hybrid Approaches

• If available, use trust networks and underlying
  data
• Run trust inference algorithms over both
• Integrate values to obtain better, more accurate
  results

25
Building Recommender Systems Using Trust

• Use trust as a way to give more weight to some
  users
• Trust for collaborative filtering
• Use trust in place of (or combined with)
  similarity
• Trust for sorting and filtering
• Prioritize information from trusted sources

26
Algorithms and Systems

• Advogato website using Advogato algorithm
• MoleSkiing website using MoleTrust algorithm
• FilmTrust website using TidalTrust algorithm

27
Advogato

• Peer certification of users
• Master principal author or hard-working
  co-author of an "important" free software project
• Journeyer people who make free software happen
• Apprentice someone who has contributed in some
  way to a free software project, but is still
  striving to acquire the skills and standing in
  the community to make more significant
  contributions
• Advogato trust metric applied to determine
  certification

28
Advogato Website

• http//www.advogato.org/
• Certifications are used to control permissions
• Only certified users have permssion to comment
• Combination of certifications and interest
  ratings of users blogs are used to filter posts

29
(No Transcript)
30
MoleSkiing

• http//www.moleskiing.it/ (note in Italian)
• Ski mountaineers provide information about their
  trips
• Users express their level of trust in other users
• The system shows only trusted information to
  every user
• Uses MoleTrust algortihm

31
(No Transcript)
32
(No Transcript)
33
FilmTrust

• Movie Recommender
• Website has a social network where users rate how
  much they trust their friends about movies
• Movie recommendations are made using trust
• Recommended Rating Weighted average of all
  ratings, where weight is the trust (direct or
  inferred) that the user has for the rater

34
(No Transcript)
35
Does Trust Do Well?
36
(No Transcript)
37
Benefits and Drawbacks

• Evaluation is very difficult
• Trust data is private
• Few publicly available datasets
• No good trust network simulators
• Complexity - trust inference can be time
  consuming
• Coverage - typical new user problems
• Must compute trust in others for the system to
  offer the user any benefit
• Others must be able to estimate trust in the user
  for that users voice to be heard

38
Conclusions

• Social trust can be an effective way of
  recommending user-generated content
• Requires methods for inferring trust
• Applications include aggregation, sorting, and
  filtering content
• Much work to be done in this area

39
References

• Bibliography, slides, links, examples
  athttp//www.cs.umd.edu/golbeck/recsysTrust
• Questions or commentsgolbeck_at_cs.umd.edu