Symmetrization and overfitting in probabilistic latent semantic analysis

An algorithm is proposed for revealing latent user’s interests from the observable protocol of users behavior, e.g., site visits. The algorithm combines the ideas of customer environment analysis and probabilistic latent semantic analysis. A quality criterion based on the classification of preliminarily labeled sites is introduced to optimize the algorithm parameters and compare algorithms. The experiments show that the quality has an optimum by the essential parameters of the algorithm, however the attempt of too precise optimization can lead to overfitting.