Trust Model for High Quality of Recommendations

1
Trust Model for High Quality of Recommendations

• G. Lenzini, N. Sahli, and H. Eertink
• (Telematica Instituut, NL)

SECRYPT, special session, Porto, July 2008
2
Opening
3
Ratings and Recommender/Review Systems

• Recommender systems aim to predict the rating
  that a user would give to an unknown item (as if
  he had indeed tasted, used, tried it)

4
Recommender Systems

• Recommender systems three main categories
• Content based the prediction estimated from the
  ratings that the user has given to similar
  items
• items are similar on content-based factors (tags,
  keywords, ontologies)
• Collaborative (filtering) based the prediction
  estimated from the ratings that similar users
  have given to the item
• users are similar on taste likelihood
  calculated upon common rated items
• Hybrid

5
Trust and Collaborative Filtering

• To overcome the limitation of current recommender
  systems (i.e., sparsity and accuracy) very recent
  proposals suggest to substitute the user
  similarity with trust.
• P. Massa, P. Aversani, Trust-aware Recommender
  SystemsRECSYS 2007
• N. Lathia, S. Hailes, L. Capra, Trust-based
  Collaborative FilteringIFIPTM 2008
• DellAmico, L. Capra, SOFIA Social Filtering for
  Robust Recommendations, IFIPTM 2008
• D. Quercia, today
• The experimental results are positive.
  Rummble.com uses trust-based recommendation with
  commercial scope.

6
Epinions.com
7
Epinions.com
8
Our motivation
9
Virtual Communities

• We were working on virtual communities in
  e-commerce applications (i.e., recommender and
  reviews systems).
• Virtual communities size may increases quite
  fast. Trust becomes fuzzy quite fast too.

10
Flixter.com
11
Virtual Communities Networks of Trust questions

• How to provide specific solutions to maintain
  trust relationships in those community? (e.g.,
  autonomous)
• How to increase the trustworthiness of members
  towards the community and the information they
  find there? (e.g., increase personalization)
• What features can be advantageous in the design
  of a trustworthy virtual community (e.g.,
  agent-based, mobility)?
• How to improve current recommender system that
  are based on virtual communities (e.g., by
  improving the quality of recommendation)?

12
Quality vs Usefulness

• How to distinguish between a not useful
  recommendation (but coming from a trusted
  recommender) from a recommendation of doubt
  honesty?
• Recommenders experiences might have maturated in
  different contexts. Recommenders may have tastes
  that are completely different from ours.
• That is sufficient/correct to label them as
  untrustworthy?

13
In practice Peer Review of Justification
?
14
Our Proposal
15
Solution for High Quality of Recommendation

• We designed a framework for an hybrid
  recommender/reviews where trust and other
  mechanisms are used to achieved high quality of
  recommendations
• Key concepts
• Trust Model
• Architecture (skipped in the talk, look into the
  paper)

16
Key Concepts
17
Virtual Agora, TRat, TRec
Embedded
Delegate
Items
Recommenders
TRec
TRat
network of (un)trusted recommender
registrer of (un)trusted items
Virtual Agora
18
Trust Model
19
Trust Model

• Aim build/use/update TRat(A) and TRec(A)
• Notation
• In TRat(A), agents-items
• In TRec(A), agents-agents (recommenders)
• temporary and eventual, e.g.,

20
Virtual Agora, TRat, TRec
Embedded
Delegate
Items
Recommenders
TRec
TRat
network of (un)trusted recommender
register of (un)trusted items
Virtual Agora
21
Detail of TRat(A), items

• A rating that a user gives to an item is
  calculated, at a certain time, in a certain
  context, by using a combination of the following
  strategies
• content-based (past experience on the similar
  items, in the same or similar context)
• collaborative filtering approaches (ratings from
  similar users, same or similar items, same or
  similar context)
• trust-based approaches (ratings from trusted
  users, same of similar items, same or similar
  context)
• Recommended ratings are selected/weighted upon
  their quality
• Outputs are merged and recommenders and their
  recommendations are stored (from temporary to
  eventual)

22
On High quality of recommendation

• quality trust in the source ? analysis of
  justification

23
TRat(A), items Recommendation

• A accepts Ds recommendation only if Ds
  trustworthiness combined with an evaluation of
  the justification that D has given for his
  recommendation is above a certain threshold.
• Ds justification is a set of arguments
  supporting the rating gave for each aspects
  (e.g., food, ambience, service)
• Ds arguments are evaluated against As way of
  reasoning by running an argumentation protocol

24
Argumentation Protocol
An argumentation protocol is a composition of
dialogue games (primitives assert, attack,
defend, challenge, justify, accept, refuse, or
declare undefined) Logic-based, efficient,
implementation of argumentation protocols are
available in the literature (J. Bentahar and J.J.
Meyer, 2007)
25
Example (informal version)

• Olga
• why? (asking for ground)
• I may not like the place (stating a
  counter-argument)
• since traditional cooking may be not clean
  (ground for the counter-argument)
• is not for that that I am willing to pay a price
  (alternative counter-arguments)
• (refuse the argument)

• Paul
• I love that place (claim)
• They serve traditional food, cooked in the
  traditional way.(grounds for a claim)
• why? (asking for ground)
• yes, sometimes, it is the price you pay for
  discovering new tastes (undercutting
  counter-argument)
• Ok, I agree

26
Running an Argumentation Protocol
A and D run a protocol to argue on the arguments
that D has given for each aspect of its
recommendation. Output of the protocol a value of
As argumentation trust in Ds arguments
27
Argumentation Trust
Nau argument accepted or undefined Nr
argument refused N Nr Nau
28
Consequences

• Ds arguments can be so strong to have Ds
  recommendation accepted (by As) despite Ds
  trust as a recommender is not so strong
• (after a real experience) if Ds recommendation
  was indeed a good one, As trust in D increases.
• Ds arguments are so weak to have Ds
  recommendation refused (by A) despite Ds trust
  as recommender is high.
• (after a real experience) if Ds recommendation
  was not a good one, Ds trust is not affected
  because that recommendation was not accepted
  anyhow.
• Trust is dynamic

29
Virtual Agora, TRat, TRec
Embedded
Delegate
Items
Recommenders
TRec
TRat
network of (un)trusted recommender
register of (un)trusted items
Virtual Agora
30
TRec(A), recommenders

• As builds/maintains its trust in D by using a
  combination of the following strategies
• evaluation of Ds reputation (as a recommender)
  according to As past experience
• direct evaluation of D by content-based
  strategies (referral trust bootstrap)
• check between Ds given recommendations and As
  direct experience w.r.t. items recommended by D

31
Conclusion andFuture Directions
32
Features of our solution

• Context-awareness
• Unobtrusiveness
• Usefulness
• Quality
• Privacy and Subjectiveness
• Mobility
• Low Traffic

33
On going work Duine Toolkit

• We have already implemented a prototype JADEX
  (Jadex 2008) as a development environment, which
  handles BDI concept.
• In order to commercialise our solution and make
  it useful for the market, we are currently
  integrating our approach to a set of well-known
  techniques.
• Duine Toolkit (M. Van Setten et al, 2004),
  developed in our Institute, is a framework for
  hybrid recommender which makes available a
  number of prediction techniques and allows them
  to be combined dynamically

34
On going, future work

• Have the solution implemented in a review site
• Evaluation by return of business-based metrics
• Mobility and automatic context capture with
  IYOUIT

35
Not(Questions) ? Thanks(gabriele.lenzini_at_telin.
nl)