Effectivecollaborative movie recommender system using asymmetric user similarity and matrix factorization

Recommender systems are becoming ubiquitous these days to advise important products to users. Conventional collaborative filtering methods suffer from sparsity, scalability, and cold start problem. In this work, we have implemented a novel and improved method of recommending movies by combining the asymmetric method of calculating similarity with matrix factorization and Tyco (typicality-based collaborative filtering). The asymmetric method describes that similarity of user A with B is not the similar as the similarity of B with A. Matrix factorization shows items (movies) as well as users by vectors of factors derived from rating pattern of items (movies). In Tyco clusters of movies of the same genre are created, and typicality degree (a measure of how much a movie belongs to that genre) of each movie in that cluster was considered and subsequently of each user in a genre was calculated. The similarity between users was calculated by using their typicality in genres rather than co-rated items. We had combined these methods and employed Pearson correlation coefficient method to calculate similarity to optimize results when compared to cosine similarity, Linear Regression to make predictions that gave better results. In this research work stochastic gradient descent is also used for optimization and regularization to avoid the problem of over fitting. All these approaches together provide better prediction and handle problems of sparsity, cold start, and scalability well as compared to conventional methods. Experimental results confirm that our HYBRTyco gives improved results than Tyco regarding mean absolute error (MAE)and mean absolute percentage error (MAPE), especially on the sparse dataset.