Performance Comparison of Clustering Algorithm Based Collaborative Filtering Recommendation System

Abstract

Importance of recommendation Systems (RS), based on collaborative filtering, is escalating with exponential growth of e-commerce application, e.g., on-line shopping, internet movie and music sites, tourism, news portals, to name a few. The target of recommendation system is to predict user preferences based on their previous activities, and associating users of similar behavior. Of the two main approaches, Content Based (CB) and Collaborative Filtering (CF), CF is increasingly popular because there is no need of domain knowledge and it scales well for large datasets. CF uses archives, like user-item purchase data or user-movie review data, which are large and sparse. Traditional approaches using matrix factorization like singular valued decomposition (SVD), is inefficient, for very a large review data. It is more efficient to cluster item-vectors, where elements of an item-vector are entries from users. As the dimension of an item-vector is large and elements are sparse, conventional clustering algorithms fails. We created item-item adjacency matrix, where the elements are similarity between item-vectors. There are various clustering algorithms that work directly on the adjacency matrix. We used spectral clustering, K-means++, Agglomerative Clustering. Considering item-vectors as nodes and adjacency matrix elements as link weights, we performed graph-clustering using Louvain Algorithm, to discover groups. We used synthetic data, added different levels of noise and run different algorithms to compare their accuracies to restore the original data from the noisy one. Louvain algorithm and Spectral clustering could achieve the highest accuracies.