Item-Based Collaborative Filtering Recommendation

Abstract

Recommender systems apply kno wledge disco v ery tec hniques to the p roblem of making p ersonalized recommendations for information, pro d ucts or services d uring a liv e i n teraction. These s ystems, e sp ecially the k-nearest neigh bo r c o l l a bo r a - tiv e ltering based ones, a re ac hieving widespread success o n the W eb. The t remendous gro wth i n the amoun t o f a v ail- able information and t he n um b e r o f v isitors to W eb sites i n recen ty ears p oses s ome k ey c hallenges for recommender sys- tems. These a re: pro d ucing high qualit y recommendations, p e rforming man y recommendations p e r second f or millions of users a nd items a nd ac hieving high co v erage in the face o f data sparsit y . In traditional c ollab o rativ e ltering systems the amoun t o f w ork increases with the n um be r o f partici- pan ts in the s ystem. New r ecommender system tec hnologies are needed that can quic kly pro duce high qualit y recom- mendations, e v en for v ery l arge-scale problems. T o address these issues w e ha v e explored item-based collab orativ e l- tering tec hniques. Item-based tec hniques rst a nalyze the user-item matrix t o iden tify relationships b et w een di eren t items, and t hen u se these relationships to indirectly compute recommendations for users. In this pap er w e a nalyze di eren t i tem-based recommen- dation generation algorithms. W e l o o k i n to di eren t t e c h- niques for computing item-item similarities (e.g., item-item correlation vs. cosine similarities b e t w een i tem v ectors) and di eren tt e c hniques for obtaining recommendations f rom t hem (e.g., w eigh ted s um vs. regression m o d el). Finally , w e e x - p e rimen tally e v aluate our results and compare t hem t o the basic k-nearest neigh bo r approac h. Our exp erimen ts sug- gest that item-based algorithms pro vide dramatically b e tter p e rformance t han user-based algorithms, while a t the same time pro viding b etter qualit y t han the b e st a v ailable user- based a lgorithms.