A Probabilistic Perspective Model for Recommendation Considering Long Tail Effect

Behavior-based recommendation algorithm is one class of the most important methods in recommendation systems. A variety of models are researched for a long time, such as collaborative filtering, graph-based models, matrix factorization and so on. Different characteristics in many aspects of these methods are also be analyzed. In this study, we design a new similarity measure from the perspective of probabilistic, considering the so-called long tail effect hiding in human behaviors. We also compare this model with cosine similarity and hybrid spreading algorithm in graph models from both theoretical and experimental aspects. It proves that our approach performs better than the other two models in several guidelines.