Novel techniques for cross domain recommendation

Abstract

ABSTRACT newlineSearching the internet for items of interest have become a nightmare these days due newlineto the sheer amount of available data. Recommender systems have become indispensable newlinefor many e-commerce applications to tackle this information overload problem. newlineRecommender systems are of various types of which collaborative-filtering based recommender newlinesystems are the most popular in which the recommendation is carried out newlineby utilizing the observed preferences of other users who have similar likings as that of newlinethe target user. Collaborative filtering techniques like matrix factorization have been newlinedemonstrated to be highly successful wherein given a partially filled User-Item rating newlinematrix, the idea is to correctly predict the missing entries. Among different matrix newlinefactorization methods, maximum margin matrix factorization is shown to be effective newlinein predicting the unobserved ratings when very small number of observed entries are newlinegiven. While collaborative filtering techniques like matrix factorization are good at prediction, newlineas sparsity increases (very less ratings), the accuracy of prediction expectedly newlinefalls. To address the data sparsity issue, transfer learning techniques have emerged in newlinewhich the information learnt in one context is used in another. Cross-domain recommender newlinesystems or cross-domain collaborative filtering is one such method in which newlinetransfer learning is used to transfer the knowledge from dense source domain to the newlinesparse target domain so as to improve the prediction accuracy of the target domain. In newlineorder to use transfer learning, the basic assumption is that the source and target domains newlineare inherently related in some sense. Codebook Transfer (CBT) is one of the newlinepopular transfer learning methods of cross-domain collaborative filtering, in which the newlinecodebook which is the cluster-level rating pattern is learnt from source domain and is newlinetransferred to the target domain. Codebook basically captures the rating patterns in a newlinev newlinecondensed form, and it is hypothesized that the condensed rating pattern is,