Improved Collaborative Filtering Recommendation Algorithm By Leveraging Product Relationships

With the increasing amount of information on Web 2.0, recommender systems have become a prevalent tool to help satisfy users’ need of personalization. Collaborative filtering(CF) is widely applied in recommender systems. To further improve the accuracy of recommendation, many algorithms have been proposed by incorporating user relationships, such as social network based recommendation algorithms. However, the effectiveness of product relationships to improve the accuracy of recommendation has been ignored by many researchers.This study aims to incorporate product relationships into traditional CF based recommendation algorithms to further improve the accuracy of recommendation.Generally, there are two types of product relationships: implicit and explicit product relationships. First, traditional approaches which exploit the similarity methods to mine implicit product relationships have several limitations. For instance, the similarity of two products computed by the similarity methods is symmetric. Besides, they cannot consider the relationships among more than two products. To address these limitations, we adopt the adaptive association rule technique to mine implicit product relationships, including one-to-one and many-to-one. Then, we leverage such kind of implicit product relationships as the regularizer, and incorporate it into the matrix factorization model. In the meanwhile, to further investigate the distinct influence of different types of implicit product relationships on improving the accuracy of recommendation, we devise four different strategies to select the implicit product relationships.Furthermore, in real e-commerce applications, there are explicit relationships among products. Typically, products having similar features will be classified into a same category, whereas products having dissimilar features will be divided into different categories. Moreover, considering the two different cases where one product may belong to only one category, namely one-to-one product category relationships, or multiple categories, namely one-to-many product categories relationships, we propose a novel matrix factorization based product recommendation model. Different from the existingwork, we design the category-specific user and product latent factor vectors to more accurately describe the user and product latent factors under different categories.Furthermore, the explicit product relationships have been utilized as the regularizer to constrain the learning of product latent feature vector.Lastly, we conduct extensive and comprehensive experiments on four real-world datasets to evaluate the effectiveness of the proposed model.