Online video session progress prediction using low-rank matrix completion

Abstract

The prediction of online video session progress is useful for both optimizing and personalizing end-user experience. Our approach for online video recommendation is to use the session progress information instead of using a traditional rating system. We approach the prediction of session progress as a matrix completion problem, and complete the session progress matrix using noisy low-rank matrix completion (NLMC). Events collected from the end-user video sessions are tracked and logged in a server. We process a large number of logs, represent them as a partially observed user by video matrix, and use regularized nuclear norm minimization for matrix completion. Our initial results show improvement over baseline methods of prediction using just the means. We further investigate the reason for the difference in performance for the same prediction methods between our dataset and the dataset used in the Netflix challenge. Our experiments indicate that the number of observed entries at a given sparsity is a good indicator of the performance of the Singular Value Decomposition (SVD) based matrix completion methods. This implies that the results for our dataset would further improve by either observing more entries for the same set of users and videos or by including new users or videos at the same sparsity level. Moreover, we introduce an algorithm to generate submatrices of any required sparsity and size from a given matrix to fairly compare algorithm performances on datasets of varying characteristics.