Generating Pseudotransactions for Improving Sparse Matrix Factorization

Recent research on Recommender Systems, specifically Collaborative Filtering, has focussed on Matrix Factorization (MF) methods, which have been shown to provide good solutions to the cold start problem. However, typically the same settings are used for Matrix factorization regardless of the density of the matrix. In our experiments, we found that for MF, Root Mean Square Error (RMSE) for recommendations increases (i.e. performance drops) for sparse matrices. We propose a Two Stage MF approach so MF is run twice over the whole matrix; the first stage uses MF to generate a small percentage of pseudotransactions that are added to the original matrix to increase its density, and the second stage re-runs MF over this denser matrix to predict the user-item transactions in the testing set. We show using data from Movielens that such methods can improve on the performance of MF for sparse martrices.