SVD-initialised K-means clustering for collaborative filtering recommender systems

K-means is a popular partitional clustering algorithm used by collaborative filtering recommender systems. However, the clustering quality depends on the value of K and the initial centroid points and consequently research efforts have instituted many new methods and algorithms to address this problem. Singular value decomposition (SVD) is a popular matrix factorisation technique that can discover natural clusters in a data matrix. We use this potential of SVD to solve the K-means initialisation problem. After finding the clusters, they are further refined by using the rank of the matrix and the within-cluster distance. The use of SVD based initialisation for K-means helps to retain the cluster quality and the cluster initialisation process gets automated.