earticle

영어

A Collaborative Filtering-based recommender system collects all of customers’ transaction data and determines relevant customer set, called neighborhood, which are determined by representative measures such as Cosine or Pearson-correlation and then generates product recommendation lists from the transactions of neighborhood. Though such a procedure has been known to a very effective method, its computational overhead can be prohibitive when the customer base is large. At the same time, when the sparse level of the transaction data is high, it brings deterioration in the quality of recommendations. The paper proposes the use of a customer network, for recommendations that accommodate the large-scale and sparsity nature of the transaction dataset. What is proposed in this study is a more active form of social network application, recommender network, utilizing the fast diffusion and information sharing capability of social network. From the literature of bipartite graph, we formulate CF-based recommendation task as a network problem, and then we propose a microscopic process governing the link strength of dynamic networks as a recommendation process. In order to validate the effectiveness and the efficiency of the proposed method, we build a recommendation network for the product recommendation using real product transaction data and compare it against the traditional system based on collaborative filtering. Experiment results show that the microscopic process of the recommendation network is computationally more efficient than, but as accurate as, the global optimization process of the traditional recommender system.