Incorporating Community Detection and Clustering Techniques into Collaborative Filtering Model

Abstract

Recommender systems seek to predict the ‘rating’ or ‘preference’ that user would give to an item based upon historical grading datasets.At the era of IT, the Internet has the ability to reflect the social network, which makes possible to incorporate community detection and clustering algorithms to improve their performance. That is, we can discover the hidden information present in the social network by using community detection algorithms, and use clustering technique to reveal users’ preference on the basis of their behavioral history. Such information is used as feedback to our recommender system to improve the prediction accuracy. Based on this consideration, in this research, we revise the well-known SVD (Singular Value Decomposition) algorithm andhave made two contributions. First, we derive groups by overlapping the communities with the clusters, and feed them as implicit feedback to SVD++. In addition, we introduce a matrix, referred to as difference matrix, and use it as an input to our algorithm. The resulting model is referred to as difference-SVD. We test all the models on Baidu's dataset, which is a large data set on movie recommendation. The results show that difference-SVD is quite promising: While both SVD++ and difference-SVD show higher accuracy than SVD, difference-SVD outperforms SVD++ in that it runs fasterthan SVD++.