Scalability Improvement of Recommender Systems using influential Users Extracted from Social Network’s Structure

Collaborative filtering is one the well-known methods in recommender systems that is widely used in e-commerce sites. This method benefits from users’ ratings records that had similar views with the target user, to recommend a product to the target user. Heavy computations of collaborative filtering, for predicting users’ ratings, has made scalability one of the main issues of collaborative filtering. Utilizing influential users is one of the solutions to overcome this issue. In this solution, instead of using all the users, ratings prediction process is performed by influential users’ preferences. Influential user is a user that omission of his records from the recommender system reduces prediction accuracy more than the others. Although this solution reduces ratings prediction’s time complexity, however, computing user’s influence itself has a noticeable time complexity. This paper presents a new method for computing users’ influence from users’ social network. The indicated social network is the implicit network between users, which its links show the degree of similarity between ratings’ records of the users. Users’ influence in this network is estimated by centrality measures and just by utilizing social network’s graph structure. Thus, the time complexity of computing influence reduces to a linear function. Yet, our evaluations show that despite reducing time complexity, decrease in accuracy is not much and the proposed method has a similar accuracy to previous methods. Hence, the proposed method is a suitable solution for large scale recommender systems.

Keywords:

recommender system , collaborative filtering , social network , social network analysis , users’ graph , influential users , scalability