Recommendation with Generalized Logistic Transformation

Many recommender systems explicitly or implicitly assume that rating data are normally distributed. This assumption is handy, but often does not hold in practice, resulting in system underperformance. In this paper, we design a recommendation algorithm embedding a new distribution model. First, we introduce a generalized logistic transformation and a parameter estimator Minimum Absolute Skewness Estimator (MASE) to obtain generalized-Gaussian distributed data. Second, we propose a new model, namely generalized logit-generalized-normal (GLG-normal) distribution to fit the observed frequency distribution. Finally, we design GLG-N probabilistic matrix factorization (GPMF) recommendation algorithm. Experiments were undertaken on the 3 subsets of Jester. Results show that 1) GLG-normal captures the essence of the frequency distribution, and 2) GPMF is 5% better than PMF in terms of MAE, and significantly outperforms some other schemas.