TRUNCATED SVD WITH CONDITIONAL COLLABORATIVE FILTERING RECOMMENDER SYSTEM

Abstract

Recommender Systems are designed to suggest personally to each of the users the items they are most likely to like. Sparsity of the User Matrix, cold-start and scalability of the algorithm to real world are the key issues that a good recommender system aims to address. The collaborative filtering based recommendation systems are based on the classical Bayesian equation of conditional probability. They can be either item based or user based, depending on the strategy by which neighbourhood is found. Singular Value decomposition is the matrix factorization technique, that aims at reducing the resource (memory) requirements with minimum possible information loss, measured by frobenius norm. Truncated SVD involves keeping only the significant eigen vectors and thus only their corresponding rows and columns. The proposed Hybrid algorithm aims at combining, experimentally, the knowledge about the preferences of user from the truncated svd approach and the collaborative filtering. This essentially translates to the problem of determining the 2 constants which act as weights for the combination. The accuracy calculation of the results can be inferred from the error rates computed from the results, thus we use several methods of computing error, so that the wholesome view. RMSE, MAE and REM are three of the error evaluation variations.