This paper proposes a novel approach for estimating the statistical multimedia user preference by providing weights to multimedia contents with respective to their consumed time. The optimal weights can be obtained by training the statistical system in the sense that the mutual information between old preference and current preference is maximized. The weighting scheme can be done by partitioning a user's consumption history data into smaller sets in a time axis. With developing a mathematical derivation of our learning method, experiments were implemented for predicting the TV genre preference using 2,000 TV viewers' watching history and showed that the performance of our method is better than that of the typical method.