Magic barrier estimation models for recommended systems under normal distribution

摘要

Real data are usually imperfect. The inherent nature of data determines the magical barrier of a machine learning task. In this paper, we propose three normal distribution models to estimate the magic barrier of recommender systems in terms of mean absolute error (MAE). The first model assumes that the users’ ratings are all subject to the same normal distribution. The second assumes that there are different appropriate standard deviation settings for different rating levels. The third divides users into different groups and assumes that the settings for the standard deviations are related to the rating levels and the groups of users. In this way, the latter models are more realistic than the former ones. Experimental results on three well-known datasets show that three models are consistent since the estimated values are close to each other. Popular recommendation algorithms also approach the magic barriers closely.