Social Trust-Aware Recommendation System: A T-index Approach Alireza Zarghami Soude Fazeli Nima Dokoohaki Mihhail Matskin Presented at Workshop on Web Personalization, Reputation and Recommender Systems ( WPRRS’09 ) In conjunction with IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology ( WI’09 and IAT’09). September 2009

Motivation Contribution T-Index + TopTrustee Approach Framework Ontologies Trust Calculation Metric Choice Trust Transposure Trust Propagation Recommendation Prediction Experiment Coverage, MAE, Indegree Conclusion Agenda

Motivation

Contribution

T-Index + TopTrustee

Approach

Framework

Ontologies

Trust Calculation

Metric Choice

Trust Transposure

Trust Propagation

Recommendation Prediction

Experiment

Coverage, MAE, Indegree

Conclusion

Memory-based Utilize the entire user-item data to predict likeness; NNR, Pearson, statistical approach Model-Based Clustering, Bayesian, Rule based, Probabilistic Approach Trust-Based √ Correlation between trust and similarity (proved by Golbeck, Massa/Avesani) Collaborative Filtering heuristics

Memory-based

Utilize the entire user-item data to predict likeness;

NNR, Pearson, statistical approach

Model-Based

Clustering, Bayesian, Rule based, Probabilistic Approach

Trust-Based √

Correlation between trust and similarity (proved by Golbeck, Massa/Avesani)

Model a recommendation system Utilizes a distributed trust-based CF Utilizes Semantic Web Ontology to deal with heterogeneous networks of users and items Ability to traverse the trust networks to collect Recommendations To have better coverage and prediction accuracy in short traversal by optimizing the trust network maintenance mechanism Contribution

Model a recommendation system

Utilizes a distributed trust-based CF

Utilizes Semantic Web Ontology to deal with heterogeneous networks of users and items

Ability to traverse the trust networks to collect Recommendations

To have better coverage and prediction accuracy in short traversal by optimizing the trust network maintenance mechanism

Our work is based upon two main ideas: T-index A measure inspired by Hindex to discover the agents within our trust network who provide trust values higher or equal to T. TopTrustee A list, which provides information about users who might not be accessible within a predefined maximum path length. TopTrustee List= (m) raters who provide Highest T-index values. TopTrustee/ T-index

Our work is based upon two main ideas:

T-index

A measure inspired by Hindex to discover the agents within our trust network who provide trust values higher or equal to T.

TopTrustee

A list, which provides information about users who might not be accessible within a predefined maximum path length.

TopTrustee List=

(m) raters who provide Highest T-index values.

TopTrustee Idea Depicted An example of TopTrustee Finding trustworthy users across the trust network even outside the traversal path length limit

An example of TopTrustee

Finding trustworthy users across the trust network even outside the traversal path length limit

T-index Idea Depicted An example of T-index Indegree (Ua) = 7 Indegree (Ub) = 5 T-index (Ua) = 2 T-index (Ub) = 4

An example of T-index

Indegree (Ua) = 7

Indegree (Ub) = 5

T-index (Ua) = 2

T-index (Ub) = 4

Ontological Model Framework can deal with heterogeneous networks of user and item in a distributed manner Users from different groups can be hosted by different servers possibly located in different organization for sake of privacy, accessible by their URI

Framework can deal with heterogeneous networks of user and item in a distributed manner

Users from different groups can be hosted by different servers possibly located in different organization for sake of privacy, accessible by their URI

User Ontology Relationship: Top-n Trustees Rank Relation: History of rating T-index: User's T-index

Relationship:

Top-n Trustees

Rank Relation:

History of rating

T-index:

User's T-index

Item Ontology Ontological Item Profile

Ontological Item Profile

Choice of trust metric (common case) Trust value defined as a decimal value [0,1] For users who find each other through TopTrustee list, calculated directly based on their common item in two steps: Transpose their values to have same rating scale Calculate their mutual Trust Trust Metric and Calculation

Choice of trust metric

(common case) Trust value defined as a decimal value [0,1]

For users who find each other through TopTrustee list, calculated directly based on their common item in two steps:

Transpose their values to have same rating scale

Calculate their mutual Trust

Transposure of Trustee Rating Different users have different scales for rating, Row : Truster Column : Trustee tr(5)=4.43 Trustee' rating of 5 is considered as 4.43 for Truster to calculate trust

Different users have different scales for rating,

Row : Truster

Column : Trustee

tr(5)=4.43 Trustee' rating of 5 is considered as 4.43 for Truster to calculate trust

After transposing rating values of trustee to the same scale rating of truster, we compute their mutual trust value based on this formula* Formula calculates the sum of their differences in rating values for common items divided by the number of truster's item multiplied by the maximum rating value. * N. Lathia, S. Hailes and L. Capra. “Trust-based collaborative filtering”, in IFIPTM 2008: Joint iTrust and PST Conferences on Privacy, Trust Management and Security, P.14, London, 2008. Trust Computation

After transposing rating values of trustee to the same scale rating of truster, we compute their mutual trust value based on this formula*

Formula calculates the sum of their differences in rating values for common items divided by the number of truster's item multiplied by the maximum rating value.

* N. Lathia, S. Hailes and L. Capra. “Trust-based collaborative

filtering”, in IFIPTM 2008: Joint iTrust and PST Conferences on

Privacy, Trust Management and Security, P.14, London, 2008.

Basic approach For users who has no direct trust relationship, we propagate trust by multiplying trust values of the nodes are located in the path between them. Trust Propagation

Basic approach

For users who has no direct trust relationship, we propagate trust by multiplying trust values of the nodes are located in the path between them.

Collecting Recommendation for users Recommendations for a particular user are collected by asking from its direct or indirect neighbors through traverals. Limiting traversal length Trust threshold √ Path length For instance: Um is more trustworthy than Ug for Ua

Recommendations for a particular user are collected by asking from its direct or indirect neighbors through traverals.

Limiting traversal length

Trust threshold √

Path length

For instance:

Um is more trustworthy than Ug for Ua

Recommendation – traversal path length we just collect recommendations from short traversal length , so all traversals are limited to a predefined maximum traversal path length. If the maximum defined as 3, traversal can not go further than Um regardless of trust value

we just collect recommendations from short traversal length , so all traversals are limited to a predefined maximum traversal path length.

If the maximum defined as 3, traversal can not go further than Um regardless of trust value

Recommendation Prediction Prediction of the Recommendations collected from direct or indirect neighbors are done by the weighted average of their rating based on their trust values calculated either through computation or propagation

Prediction of the Recommendations collected from direct or indirect neighbors are done by the weighted average of their rating based on their trust values calculated either through computation or propagation

MovieLens http://www.movielens.org/ 100,000 rating of 5-point scale 943 users and 1682 movies Rating are sorted according to their timestamps 80% of rating used to build the network 20% of rating used to test the recommendations Experiment - Dataset

MovieLens

http://www.movielens.org/

100,000 rating of 5-point scale

943 users and 1682 movies

Rating are sorted according to their timestamps

80% of rating used to build the network

20% of rating used to test the recommendations

Parameters N number of neighbors per user (2,3,5,10,20,50) M number of TopTrustee per item (2,3,5,7) With or without T-index (0,100) Trust threshold is defined as 0.1 Maximum path length of traversal is defined as 3 Experiments MAE Coverage Indegree distribution of most trustworthy users Experiment Types / Parameters

Parameters

N number of neighbors per user (2,3,5,10,20,50)

M number of TopTrustee per item (2,3,5,7)

With or without T-index (0,100)

Trust threshold is defined as 0.1

Maximum path length of traversal is defined as 3

Experiments

MAE

Coverage

Indegree distribution of most trustworthy users

Coverage

MAE

Indegree distribution most trustworthy users

Trust network visualization Configuration: n=3 m=3 T-index=0

Trust network visualization Configuration: n=3 m=3 T-index=100

Designed an ontological model to model heterogeneous networks of users and items Introduced TopTrustee list to enhance the process of discovering neighbors Introduced T-index as a measure of trustworthiness which can improve the Coverage and MAE in short traversal path length, especially for small size of neighbors T-index can improve trust network structure by increasing the number of well connected clusters Conclusion

Designed an ontological model to model heterogeneous networks of users and items

Introduced TopTrustee list to enhance the process of discovering neighbors

Introduced T-index as a measure of trustworthiness

which can improve the Coverage and MAE in short traversal path

length, especially for small size of neighbors

T-index can improve trust network structure by increasing the number of well connected clusters

Thanks  Contacts: ----------- Alireza Zarghami http://www.isk.kth.se/~zarghami/ Soude Fazeli http://www.isk.kth.se/~soude/ Nima Dokoohaki http://web.it.kth.se/~nimad/ Misha Matskin http://www.idi.ntnu.no/~misha/ Questions

Thanks 

Contacts:

-----------

Alireza Zarghami

http://www.isk.kth.se/~zarghami/

Soude Fazeli

http://www.isk.kth.se/~soude/

Nima Dokoohaki

http://web.it.kth.se/~nimad/

Misha Matskin

http://www.idi.ntnu.no/~misha/

Nima Dokoohaki Software and Computer Systems ( SCS ), Department of Electronics ,Computer and Software Systems ( ECS ), School of Information and Communications Technology ( ICT ), Royal Institute of Technology ( KTH ), Stockholm, Sweden Office: +46 (0) 8 790 4149 Cell : +46 (0) 76 269 76 30 Fax: +46 (0) 8 751 1793 Email : nimad@kth.se http://web.it.kth.se/~nimad/ Presenter

Nima Dokoohaki

Software and Computer Systems ( SCS ), Department of Electronics ,Computer and

Software Systems ( ECS ), School of Information and Communications Technology ( ICT ), Royal Institute of Technology ( KTH ), Stockholm, Sweden Office: +46 (0) 8 790 4149 Cell : +46 (0) 76 269 76 30 Fax: +46 (0) 8 751 1793 Email : nimad@kth.se

http://web.it.kth.se/~nimad/