KMS Chongqing Institute of Green and Intelligent Technology, CAS
Effects of the bipartite structure of a network on performance of recommenders | |
Wang, Qing-Xian1; Li, Jian2; Luo, Xin2; Xu, Jian-Jun2; Shang, Ming-Sheng2 | |
2018-02-15 | |
摘要 | Recommender systems aim to predict people's preferences for online items by analyzing their historical behaviors. A recommender can be modeled as a high-dimensional and sparse bipartite network, where the key issue is to understand the relation between the network structure and a recommender's performance. To address this issue, we choose three network characteristics, clustering coefficient, network density and user-item ratio, as the analyzing targets. For the cluster coefficient, we adopt the Degree-preserving rewiring algorithm to obtain a series of bipartite network with varying cluster coefficient, while the degree of user and item keep unchanged. Furthermore, five state-of-the-art recommenders are applied on two real datasets. The performances of recommenders are measured by both numerical and physical metrics. These results show that a recommender's performance is positively related to the clustering coefficient of a bipartite network. Meanwhile, higher density of a bipartite network can provide more accurate but less diverse or novel recommendations. Furthermore, the user-item ratio is positively correlated with the accuracy metrics but negatively correlated with the diverse and novel metrics. (C) 2017 Elsevier B.V. All rights reserved. |
关键词 | Bipartite network Clustering coefficient Network density User-item ratio Recommender system |
DOI | 10.1016/j.physa.2017.11.053 |
发表期刊 | PHYSICA A-STATISTICAL MECHANICS AND ITS APPLICATIONS |
ISSN | 0378-4371 |
卷号 | 492页码:1257-1266 |
收录类别 | SCI |
WOS记录号 | WOS:000423495100104 |
语种 | 英语 |