Recommendation using a clustering algorithm based on a hybrid features selection method

Abstract

The cold start problem is a potentiel problem in Recommender Systems (RSs). It concerns the inability of the system to infer recommendaation for new users or new items about wich it has not enough iformation. Specifically, when an item is new, the system may fail to perform well due to the insufficiency of available information for this item. The most common solution addressed in the literature consists in combining the content and collaborative information under a single RS. However these hybrid solutions inherit the classical problems of natural language ambiguity and don’t exploit semantic knowledge in their items representations. In this paper, we propose a hybrid RS composed of three modules to surpass those weaknesses. The first one is rested on a powerful content clustering algorithm; which uses a Hybrid Features Selection Method (HFSM). It combines statistical and semantic relevant features to get the maximum profit from the content of items. The second module is the Collaborative Filtering (CF) one, which depends only on users’ ratings. The third one combines the previous modules to solve the problem of missing values in CF approach and to handle new-item issue. The proposed hybrid Recommender is evaluated against traditional item-based CF in different settings: no cold-start situation and a simulation of a new-item scenario (an item with few/ no ratings). The conducted experiments show the ability of the proposed hybrid recommender to deliver more accurate predictions for any item and its outperformance on the classical CF approach, which fails in cold-start situations.