Effects of the Bipartite Structure of a Network on Performance of Recommenders

Qing-Xian Wang,Jian Li,Xin Luo,Jian-Jun Xu,Ming-Sheng Shang

DOI: https://doi.org/10.1016/j.physa.2017.11.053

IF: 3.778

2017-01-01

Physica A Statistical Mechanics and its Applications

Abstract: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.