Recommendation engine based on derived wisdom for more similar item neighbors

Collaborative filtering (CF) is a popular and widely accepted recommendation technique. CF is an automated form of word-of-mouth communication between like-minded or similar users. The search for these similar users as neighbors from a large user population challenges the scalability of the user based CF approach. As a remedy, an item based CF, pre-computes pairwise item similarities to identify item neighbors. However, data sparsity remains here a major concern, as most of the neighbors of the given item might not be rated by the active user. Consequently, in the traditional item based CF approach, the neighborhood comprises of distant item neighbors having relatively low similarities which in turn affects the overall recommendation quality. The current work addresses this shortcoming in the existing item based CF approach. As a solution, we propose a hybrid user–item based CF where the item neighbors having highest similarity with the given item are selected, irrespective of whether they are rated by the active user. Subsequently, to handle sparsity, missing ratings for some of these selected item neighbors are imputed by multiple linear or ordinal logistic regression. In this approach, ratings of the active user are regressed with ratings of their most similar user(s). The motivation behind this work is to rely on closer rather than distant neighbors, which despite their presence were not used for generating recommendations in the past. The efficacy of the proposed hybrid approach utilizing both user and item similarities is established by its superior predictive performance over three different datasets.