The Overfitting in Probabilistic Latent Semantic Models

The symmetric EM algorithm is proposed for probabilistic latent semantic analysis in collaborative filtering. The algorithm allows to reveal the latent interest profiles of both users and items, then to easily construct high-quality similarity measures of all required types: user–user, item–item, and item–user. The advantage of the proposed approach is that different profiles are consistent to each other due to symmetry of the algorithm. To estimate the quality of profiles and similarity measures empirically we use a sample of labeled items and the kNN classifier. Experiment show that the excessive optimization is redundant and can lead to overfitting.