1. Trust-Based Rating Prediction for Recommendation in Web 2.0 Collaborative Learning Social Software ITHET 29th April – 1st May 2010, Cappadocia, Turkey Na Li, Sandy El Helou, Denis Gillet Real-Time Coordination and Distributed Interaction Systems (ReAct) Automatic Control Lab, Swiss Federal Institute of Technology in Lausanne
2. Outline Introduction Collaborative Learning Domain 3A Interaction Model Trust-Based Rating Prediction Approach Evaluation and Results Conclusion and Future Work
3. Introduction Web 2.0 social software A large amount of user generated content New challenge: selection of useful resources RSS Feeds Pictures Pictures Wiki Pages Documents Videos
4. Introduction Rating systems Evaluate quality of content in open environment Provide recommendation for different users
5. Introduction Rating systems – application level Rating systems – academic research level TidalTrust (J. Golbeck), MoleTrust(P. Massa) User explicitly specifies a trust value towards another user Build trust network, Random walk in trust network Personalized rating prediction
10. Trust-Based Rating Prediction Approach Direct trust (DT): derived from a particular type of relationship W (Membership): weight of “membership” relationship N (Alice, Membership): number of group activities Alice joins Is Member of Advanced Algorithms Group Activity Alice
11. Trust-Based Rating Prediction Approach Trust propagation Propagation distance (PD) Bob Commented by Article Rated by Sara Create Is Member Has Member French Learning Activity Luis Alice Rate Video Rated by Ben Created by Jack Propagate Propagate Propagate PD
13. Trust-Based Rating Prediction Approach Rating prediction from a user to an item Using user’s “Web of Trust” People in “Web of Trust” are seen as trustable Average of all the rating scores given by trustable people, weighted by their trust value
14. Evaluation and Results Using Remashed data set 50 users, 6000 items, 3000 tags and 450 ratings “Leave-one-out” method Compare “predicted score – actual score” deviation of trust-based prediction and simple average
15. Evaluation and Results Change parameters Weights for relationships doesn’t make a significant difference in rating prediction Increasing size of trust network might add noise, lead to bigger prediction error
16. Conclusion and Future Work Propose a trust-based rating prediction approach, inferring trust in an implicit way Provide personalized rating prediction so as to evaluate user-generated content in collaborative learning environment Future deploy and evaluation will be conducted in a collaborative learning platform, namely Graaasp(graaasp.epfl.ch)
Trust propagates layer by layer, until reaching the propagate distance we predefine.A “Web of Trust” is constructed in this way.
Different weights and propagate distances are tried.On this test set, the change of trust weights for relationships doesn’t make a significant difference in the results of rating prediction.We get an optimal propagate distance value, which indicates that, instead of improving the prediction results, increasing the size of trust network might add noise, leading to bigger prediction error.
Different weights and propagate distances are tried.On this test set, the change of trust weights for relationships doesn’t make a significant difference in the results of rating prediction.We get an optimal propagate distance value, which indicates that, instead of improving the prediction results, increasing the size of trust network might add noise, leading to bigger prediction error.