Trustaware Decentralized Recommender Systems TaDRS: PhD research proposal

1
Trust-aware Decentralized Recommender Systems
(TaDRS)PhD research proposal

• Phd Student Paolo Massa
• Advisor Paolo Traverso

2
Summary

• Recommender systems (RS) weaknesses
• Trust in online communities
• Addressed Problems how to exploit Trust in RSs
• Proposed solution
• Planned evaluation (experiments)
• Workplan

3
Recommender Systems (RS)

• Solution to information overload.
• Content-based
• RSs find items similar to the ones you liked in
  past
• Collaborative Filtering
• Users give ratings to items
• RS finds users similar to you (User similarity)
• Suggest you items liked by them
• CF is Simple and effective, BUT ...

In this slide should i put a graphical
explanation of CF?
4
Collaborative Filtering weaknesses

• Cold start and sparseness (97.4 sparse)
• Doesn't help the community forming
• Attackable by malicious users
• Difficult or impossible for users to control
  recommendation process
• Centralized --gt Your Profile Not reusable,
  profiles scattered, Recommendation computation
  out of control, lack of datasets for research

5
Trust in online communities

• Trust explicit rating of user on user
• E-marketplaces Ebay.com, amazon.com,
  epinions.com (fews of the dotcom survived and
  rich!)
• News sites Slashdot.org
• P2P networks eDonkey, Gnutella
• Decentralized opensource help systems Affero.org
• Network of personal weblogs (some millions
  weblogs)
• Future worlds Whuffie (Down and out in the Magic
  Kingdom)
• ...

CLAIM Trust solves CF problems
6
Decentralized Environment

• P set of n peers I set of m items
• Peers self-publish (public) info
• trust statement how much she values another peer
  PeerA.trust(p).value0.8 (in 0,1)
• rating statement how much she values an item
  PeerA.rating(i).value0.1 (in 0,1)
• about herself such as homepage, blog, photos,
  not in machine-readable format
• Every peer fetches data from other peers and
  then self-create recommendations on behalf of its
  users over these data.

7
Proposed solution

• 4 Steps strategy
• 1)propagate Trust
• To discover trustworthy peers
• 2)compute UserSimilarity(US)
• To discover similar peers
• 3)combine Trust and US
• To decide influencial peers
• 4)combine ratings of peers
• To predict items' ratings

Item1
Item2
Item3
Item4
8
Addressed Problems

• Open issue 1 Trust propagation
• Open issue 2 Combination of Ratings and Trust

9
1 Trust Propagation

• Current systems compute a Global value (Ben is
  trustable as 8/10).
• I propose computation of Subjective (Local) Trust.

Trust propagation
Problem exploit transitivity for indirect
trust. Some proposals (Advogato), no evaluations.
10
Trust Propagation Algo
1
0.9
0.6

• Input Trust statements
• Output (Subjective) Predicted Trusts
  confidence
• Max Flow graph problem
• Nodes are peers, Edges are trust statements
• Compute flow from source (ME) to sink (Ben)
• Best algo for max flow is Ford and Fulkerson
  ('56)
• Added parameters propagation horizon and
  confidence step decay.
• More Evaluation of different algorithms (with
  leave one out) on real and simulated communities

1
11
2 Combination of Ratings and Trust

• New problem no solutions
• CF automates rec. process trying to find similar
  users, forgetting explicit trust information.
• What is relationship among (computed) User
  Similarity and (direct and indirect) Trust?
• How to combine them?

12
Combination of Ratings and Trust

• Input Trusts Users Similarities ( confidence)
• Output Users Influences (for ME) confidence
• User Influence weighted sum of Inputs
• Parameters weights for US, ExpliTrust,
  PropagTrust
• (depending on CONFIDENCE)

Item1
Item2
Item3
Item4
13
Original Contributions

• Trust propagation algorithm
• Real evaluation of algorithms (experiments).
• Proposal of a trust propagation algo suited for
  Rss.
• Trust-aware Decentralized Recommender Systems
  (TaDRS)
• Introduction of new concept
• Evidence of solved CF problems
• Proposed solution for combining Trust and User
  Similarity (based on Confidence)

14
Experiments (on real and simulated communities)

• Exp1) Relation Trust/Similarity in real world (in
  allconsuming.net)
• Exp2) Test Trust propagation algos in different
  simulated communities
• Exp3) Trust solves cold start CF problem
• Exp4) Trust contradicts similarity
• Exp5) Presence of malicious peers
• Exp6) Control by trust is easier than by ratings
• Exp7) Evaluate real Accuracy and User Acceptance

15
Evaluation

• Offline on simulated (and real) communities
• Generated with different ratings and trust
  patterns
• Online
• Add Blogs on CoCoA (classic music recommender)
• Blog easy way for decentralized publishing of
  info (XML) Google bought Pyra Stanford,
  Harvard, promote use.
• Cocoa.itc.it with blogs, 1400 users'll publish
  trust/ratings
• Recommend users and items (11.000 classical
  tracks).
• Evaluate real accuracy (by feedback) and real
  user acceptance (by surveys).
• Allconsuming.net
• Blogosphere

16
Impact of Research Proposal

• RS --gt TaDRS (Trust-aware Decentralized RS)
• TaDRS solve CF sparseness problem
  (decentralization, trust confidence and
  transitivity)
• TaDRS solve cold start problem
• TaDRS support Explanation (black box problem),
  allows community forming and is easier to
  control.
• TaDRS are resistant to malicious users.
• TaDRS computation is personalized and
  decentralized
• Public database of ratings and trust (CoCoA) and
  evaluation framework for trust propagation algos.

17
Workplan
Jun 2003
Jan 2004
Mar
Sep
Dec
May
Sep 2004
RS, Trust, P2P
Algorithms
Refining
CocoaBlog MT
Blogs
CocoaBlog collect data
Community generator
CocoaBlog
RSs eval.
AllConsuming experiments
Stanford
Dublin Trinity
ConfECAI
Journal
PhD Thesis
18

• The End