Improving Social Recommendations by applying a Personalized Item Clustering Policy
0 likes1,193 views
This document presents a personalized item clustering approach for social recommendations. It aims to improve recommendation accuracy, novelty, and diversity. The approach clusters items based on a user's social network consumption patterns. It then constructs item networks and performs random walks to select recommendation items from each cluster. An evaluation on an Epinions dataset found the approach outperformed baseline and traditional recommender systems in novelty and diversity while achieving comparable accuracy.
Improving Social Recommendations by applying a Personalized Item Clustering Policy
- 1. Introduction Algorithmic Components Experiments Summary Improving Social Recommendations by applying a Personalized Item Clustering policy Georgios Alexandridis, Georgios Siolas Andreas Stafylopatis School of Electrical and Computer Engineering National Technical University of Athens 15780 Zografou, Athens, Greece The 5th ACM RecSys Workshop on Recommender Systems & the Social Web (RSWeb 2013) Alexandridis, Siolas, Stafylopatis Social Recommendations via Personalized Item Clustering
- 2. Introduction Algorithmic Components Experiments Summary Problem Statement Intuition Objectives & System Outline Problem Statement Human taste is influenced by many factors People tend to consume items that are not alike Pure CF or item-based approaches quite often miss those peculiarities of human taste Recommender Systems should be able to identify connections between seemingly uncorrelated items that are of interest, though, to a particular user In this way, the overall user satisfaction is expected to increase because the recommended items would be novel compared to what has been previously consumed the list of recommended items would be more diverse compared to the list of items returned by pure CF or item-based techniques Alexandridis, Siolas, Stafylopatis Social Recommendations via Personalized Item Clustering
- 3. Introduction Algorithmic Components Experiments Summary Problem Statement Intuition Objectives & System Outline Intuition Homophily: In social networks, people establish bonds predominantly with people they share common interests with In social RS, people follow/befriend people with similar taste We share evaluations with our peers on common subsets of items Those items have some characteristics in common Even if they’re considered to be uncorrelated by standard similarity techniques Intuition: locate common consumption patterns of items in the subsets and of other items in the system Alexandridis, Siolas, Stafylopatis Social Recommendations via Personalized Item Clustering
- 4. Introduction Algorithmic Components Experiments Summary Problem Statement Intuition Objectives & System Outline Objectives & System Outline Socially-aware personalized item clustering recommendation system Main Objectives: Make recommendations that are accurate, novel and diverse System Outline 1 Place the items that a specific user has evaluated into clusters according to the rating behavior of the members of his Personal Network Peers in his/her social network Similar peers 2 For each cluster Construct the Item Consumption Network Perform a Random Walk on the aforementioned network and return the most visited nodes 3 Merge the returned nodes of each cluster and return N recommendations Alexandridis, Siolas, Stafylopatis Social Recommendations via Personalized Item Clustering
- 5. Introduction Algorithmic Components Experiments Summary The Personal Network The Item-to-Item Adjacency Matrix Personalized Clustering Item Consumption Network The Random Walk on the ICN The Personal Network The Personal Network of user u Neighbors in the social network that bear a similarity to u Other similar users Other users in the social network (e.g “friends-of-friends”) that are similar to u Other users in the social network Similarity is measured by readily-applied indices in RS literature e.g. Pearson, Cosine, Manhattan Alexandridis, Siolas, Stafylopatis s7 u7 t3,7 u3 s3 , t3 u5 t2,1 t2 ut t3,5 u2 t2,4 s1 , t1 t1,4 s6 u4 u1 u6 Social Recommendations via Personalized Item Clustering
- 6. Introduction Algorithmic Components Experiments Summary The Personal Network The Item-to-Item Adjacency Matrix Personalized Clustering Item Consumption Network The Random Walk on the ICN The Item-to-Item Adjacency Matrix i1 i1 0 items having been evaluated i2 0 by u above a relevance threshold i3 0 A = i4 3 Elements: i5 0 ai,j denotes the frequency that items i and j have been i6 4 evaluated together by u’s i7 0 Rows and Columns i2 0 0 0 0 0 0 1 i3 0 0 0 0 3 2 0 i4 3 0 0 0 0 8 0 i5 0 3 0 0 0 0 4 i6 4 0 2 8 0 0 0 i7 0 1 0 0 4 0 0 peers in his/her Personal Network By definition, matrix A is square and symmetric Alexandridis, Siolas, Stafylopatis Social Recommendations via Personalized Item Clustering
- 7. Introduction Algorithmic Components Experiments Summary The Personal Network The Item-to-Item Adjacency Matrix Personalized Clustering Item Consumption Network The Random Walk on the ICN Personalized Clustering 4 Matrix A: adjacency matrix of an undirected graph i1 3 i4 i7 1 i2 2 8 Nodes: items consumed by u Edges: frequency of items having been accessed together by u’s peers Perform spectral clustering on this graph to locate clusters of items accessed together i6 4 3 i5 Alexandridis, Siolas, Stafylopatis Social Recommendations via Personalized Item Clustering i3
- 8. Introduction Algorithmic Components Experiments Summary The Personal Network The Item-to-Item Adjacency Matrix Personalized Clustering Item Consumption Network The Random Walk on the ICN Item Consumption Network Nodes: items Black: members of the cluster Gray: other items, accessed by u’s peers along with members of the cluster s1 (10) number in parenthesis are total evaluations by all users Edges: frequency of common access by u’s peers The ICN graph is connected and non-bipartite assumes the properties of a symmetric time-reversible finite Markov chain Alexandridis, Siolas, Stafylopatis i2 (30) 2 5 i4 (80) s2 (20) 4 i1 (50) 3 i3 (40) 3 s3 (15) Social Recommendations via Personalized Item Clustering
- 9. Introduction Algorithmic Components Experiments Summary The Personal Network The Item-to-Item Adjacency Matrix Personalized Clustering Item Consumption Network The Random Walk on the ICN The Random Walk on the ICN Random walks on connected, non-bipartite graphs reach their steady-state distribution regardless of the staring node Modified Random Walk on the ICN graph Return the most visited non-seed nodes as recommendations Modified: next node is not sampled uniformly at random but depends on the edge weight number of evaluations of both the current and the following node Alexandridis, Siolas, Stafylopatis Social Recommendations via Personalized Item Clustering
- 10. Introduction Algorithmic Components Experiments Summary Datasets Evaluation Methodology Reference Systems Results Datasets Performance Evaluation on the Epinions dataset Medium-sized dataset 50k users 140k items 664k ratings 487k trust statements Very sparse dataset as measured by the rating’s density and the clustering coefficient of the trust network (power law distributions) Ratings are skewed towards the upper scale (4-5) by a ratio of 1 to 3 Behavioral phenomenon of users predominantly rating items they’ve both consumed and liked Any naive RS that would blindly recommend any item with a high score would perform satisfactory! Alexandridis, Siolas, Stafylopatis Social Recommendations via Personalized Item Clustering
- 11. Introduction Algorithmic Components Experiments Summary Datasets Evaluation Methodology Reference Systems Results Evaluation Methodology Evaluation Objectives Accuracy of predictions Coverage of predictions Qualitative criteria for the list of recommended items How novel they are (in terms of what has already been consumed) How diverse they are from one another Evaluation Metrics 1 2 3 4 Root Mean Square Error (RMSE) Ratings’ Coverage Distance-based Item Novelty Intra-list Diversity Alexandridis, Siolas, Stafylopatis Social Recommendations via Personalized Item Clustering
- 12. Introduction Algorithmic Components Experiments Summary Datasets Evaluation Methodology Reference Systems Results Reference Systems Baseline Systems UserMean ItemMean Traditional Recommender Systems Collaborative Filtering Item-based Recommendation Trust Aggregation RS MoleTrust (with propagation horizon up to 3) TidalTrust Alexandridis, Siolas, Stafylopatis Social Recommendations via Personalized Item Clustering
- 13. Introduction Algorithmic Components Experiments Summary Datasets Evaluation Methodology Reference Systems Results Results Results on the Epinions Dataset (for a list of 5 recommended items) Performance Metrics RMSE Coverage Novelty A. Baseline A.1 ItemMean A.2 UserMean B. Collaborative Filtering B.1 Manhattan Similarity (All Neighbors) C. Item-Based Recommendation C.1 Manhattan Similarity (All Similar Items) D. Trust-based Approaches D.1 MoleTrust-1 D.2 MoleTrust-2 D.3 MoleTrust-3 D.4 TidalTrust E. Our Recommender E.1 Personalized Item Clustering Alexandridis, Siolas, Stafylopatis Diversity 1.09 1.20 86.43% 98.58% 11.89% 9.70% 24.23% 19.42% 1.07 79.57% 20.11% 56.23% 1.20 39.29% 16.86% 45.26% 1.23 1.16 1.12 1.08 25.58% 56.52% 70.89% 74.67% 29.16% 32.31% 42.13% 45.38% 43.62% 54.02% 56.65% 59.17% 1.05 58.17% 53.11% 63.04% Social Recommendations via Personalized Item Clustering
- 14. Introduction Algorithmic Components Experiments Summary Summary We proposed a novel social RS based on personalized item clustering Our approach yields satisfactory results on most of our evaluation objectives (accuracy, novelty, diversity) It could be further improved in the personal network formation phase and the clustering algorithm Alexandridis, Siolas, Stafylopatis Social Recommendations via Personalized Item Clustering