A Refined SVD Algorithm for Collaborative Filtering

TL;DR

This paper introduces a refined SVD algorithm for collaborative filtering that leverages K-means clustering to improve the initialization of missing ratings, leading to better prediction accuracy.

Contribution

It proposes a novel combination of K-means clustering with SVD to enhance the initialization process in collaborative filtering.

Findings

Outperforms existing initialization techniques in accuracy.

Improves the convergence speed of SVD-based predictions.

Demonstrates effectiveness on sparse rating matrices.

Abstract

Collaborative filtering tries to predict the ratings of a user over some items based on opinions of other users with similar taste. The ratings are usually given in the form of a sparse matrix, the goal being to find the missing entries (i.e. ratings). Various approaches to collaborative filtering exist, some of the most popular ones being the Singular Value Decomposition (SVD) and K-means clustering. One of the challenges in the SVD approach is finding a good initialization of the unknown ratings. A possible initialization is suggested by [1]. In this paper we explain how K-means approach can be used to achieve the further refinement of this initialization for SVD. We show that our technique outperforms both initialization techniques used separately.