1. Liang Xiang, Quan Yuan, Shiwan Zhao, Li Chen, Xiatian Zhang, Qing Yang and Jimeng Sun Institute of Automation Chinese Academy of Sciences, IBM Research Department of Computer Science, Hong Kong Baptist11University IBM T.J. Watson Research Center Temporal Recommendation on Graphs via Long- and Short-term Preference Fusion KDD 2010 Intelligent Database Systems Lab. School of Computer Science & Engineering Seoul National University Center for E-Business Technology Seoul National University Seoul, Korea Presented by Sung Eun, Park 11/04/2010

2. Copyright  2010 by CEBT Contents  Introduction  STG Construction  Recommendation on STG Injected Preference Fusion Temporal Personalized Random Walk Complexity Analysis  Experiment  Conclusion 2

3. Copyright  2010 by CEBT Introduction  Motivation Overall behavior of a user may be determined by her long-term interest But at any given time, a user is also affected by her short-term interest due to transient events, such as new product releases and special personal occasions such as birthdays  Goal Explicitly model users’ long-term and short-term preferences for recommendation on implicit datasets 1. How to represent the two types of user preferences? 2. How to model the interaction between the two types of preferences? 3

4. Copyright  2010 by CEBT STG Construction  Transform “ triples” into “ and ” by dividing the time into bins and binding the bins with corresponding users  Session-based Temporal Graph (STG) a directed bipartite graph G(U, S, I,E,w) where U denotes the set of user nodes, S is the set of session nodes, and I is the set of item nodes, E is the set of edges and weight – Session node Dividing the time slices into bins and binding the bins with corresponding users 4

5. Copyright  2010 by CEBT STG Construction item1item2item3item4 user11110 user20101 5 u1u1 u2u2 i1i1 i2i2 i3i3 i4i4 Session2 (u 1,T 2 ) Session1 (u 1,T 1 ) Session3 (u 2,T 0 )  u’s long-term preferences user node v u connects to all items viewed by user u, denoted as N(u)  u’s short-term preferences v ut only connects to items user u viewed at time t, denoted as N (u, t)  if we start walking from the user node v u, we will pass through N(u) and then reach unknown items which are similar to items in N(u)  if we walk from the session node v ut, we will reach unknown items similar to items in N(u, t)

6. Copyright  2010 by CEBT Injected Preference Fusion(IPF) 6  An algorithm based on STG, which balancing the impact of long- term and short-term preferences when making recommendation  Basic Idea Injected preferences into both user node(β) and session node(1-β) Then in propagation process, the preferences were propagated to an unknown item node Finally the nodes which get most preferences will be recommended to current user

7. Copyright  2010 by CEBT Injected Preference Fusion(IPF) 7 u1u1 u2u2 i1i1 i2i2 i3i3 i4i4 Session2 (u 1,T 2 ) Session1 (u 1,T 1 ) Session3 (u 2,T 0 ) Preference β Preference 1-β useritemsession ηuηu ηsηs Making recommendation for U1 at time T1 : Session Temporal Graph(STG) STG edge weight definition 1 1

8. Copyright  2010 by CEBT  Datasets(Bookmarking)  Evaluation Metric Hit Ratio : put the latest item of each user into test set, then generate a list of N(N=10)items for everyone at time t. If the test item appears in the recommendation list, we call it hit  Compared Algorithms Temporal User KNN Temporal Item KNN Temporal Personalized Random walk Injected Preferences Fusion ( IPF ) 8 CiteULIkeDelicious User46078,861 Item16,0543,257 User-item pair109,34659,694 Sparsity99.85%99.97%

9. Copyright  2010 by CEBT Experiments  β‘s impact-balance the injected preferences on user and session node Optimal results were get when β belongs to [0.1,0.6] Proves the impacts of both long-term and short-term preferences in making good recommendation β=0, only short-term β =1, only long-term β =0, only short-term β =1, only long-term CiteULike Delicious

10. Copyright  2010 by CEBT Experiments  η‘s impact-balance the injected preferences on user and session node Proves that effectiveness of balancing long-term and short-term preferences in propagation process Since X-axis is the logarithm value, it means the optimal hit ratio can be get for a wide range of η η=0, item-item connected only by session CiteULike Delicious η=1, item-item connected only by user η=0, item-item connected only by session η=1, item-item connected only by user

11. Copyright  2010 by CEBT  Session size‘s impact on Hit Ratio of IPF The result is not very sensitive to the size of time window Users’ interest on research topics(CiteULike)drift more slowly than interests on browsing web pages(Delicious) Experiments

12. Copyright  2010 by CEBT  Overall Accuracy Comparison – TItemKNN : Time weighted item-based CF – TUserKNN : Time weighted user-based DF – TPRW : Temporal Personalized Random Walk – MS-IPF : Multi Source IPF  User Temporal Item KNN as baseline On CiteULike, MS-IPF improves TItemKNN up to 15.02% On Delicious, MS-IPF improves TItemKNN up to 34.45% Experiments CiteULikeDelicious

13. Copyright  2010 by CEBT Conclusion  propose a Session-based Temporal Graph (STG) which incorporates temporal information to model long-term and short- term preferences simultaneously  Based on STG framework, we propose a new algorithm, Injected Preference Fusion (IPF), to balance the impacts of users’ long- term and short-term preferences for accurate recommendation.  To further demonstrate the model generality, we extend the personalized Random Walk for temporal recommendation based on STG.

14. Q&A Thank you 14