Analyzing consumer-product graphs: Empirical findings and applications in recommender systems

• 成果类型: Article
• 署名作者: Huang, Zan; Zeng, Daniel D.; Chen, Hsinchun
• 署名单位: Pennsylvania Commonwealth System of Higher Education (PCSHE); Pennsylvania State University; Pennsylvania State University - University Park; University of Arizona
• 刊物名称: MANAGEMENT SCIENCE
• ISSN/ISSBN: 0025-1909
• DOI: 10.1287/mnsc.1060.0619
• 发表日期: 2007
• 页码: 1146-1164
• 关键词: random graph theory; consumer-purchase behavior; topological features; recommender systems; collaborative filtering

摘要

We apply random graph modeling methodology to analyze bipartite consumer-product graphs that represent sales transactions to better understand consumer purchase behavior in e-commerce settings. Based on two real-world e-commerce data sets, we found that such graphs demonstrate topological features that deviate significantly from theoretical predictions based on standard random graph models. In particular, we observed consistently larger-than-expected average path lengths and a greater-than-expected tendency to cluster. Such deviations suggest that the consumers' product choices are not random even with the consumer and product attributes hidden. Our findings provide justification for a large family of collaborative filtering-based recommendation algorithms that make product recommendations based only on previous sales transactions. By analyzing the simulated consumer-product graphs generated by models that embed two representative recommendation algorithms, we found that these recommendation algorithm-induced graphs generally provided a better match with the real-world consumer-product graphs than purely random graphs. However, consistent deviations in topological features remained. These findings motivated the development of a new recommendation algorithm based on graph partitioning, which aims to achieve high clustering coefficients similar to those observed in the real-world e-commerce data sets. We show empirically that this algorithm significantly outperforms representative collaborative filtering algorithms in situations where the observed clustering coefficients of the consumer-product graphs are sufficiently larger than can be accounted for by these standard algorithms.