A Recommendation System with the Use of Comprehensive Trend Indication Based on Weighted Complete Graph

Recently, Internet shopping has become widespread, websites of which are equipped with a recommendation system to help users easily find their target items from among vast product information. As a typical method to create recommendation information, collaborative filtering is used but it has a problem that recommendation results tend to be biased toward the same category. Since this study intends recommendation with a high discoverability from a large point of view of category, we define dissimilarity between products based on information on Browse Node ID held by some products in Amazon and use k-medoids to newly categorize the products. Moreover, we create a weighted complete graph with those categories as nodes and indicate the trend across different categories. The proposed system estimates and recommends a category strongly related to a category that is thought to be unknown to the user but the user will like based on information of the weighted complete graph. We evaluate the effectiveness of the proposed system through experiments with 9 undergraduate students, 12 graduate students, and 2 office workers as subjects and show that the proposed system is better in recommending unknown products to the user than existing recommendation systems.