Heterogenous Information Network Embedding Based Cross-Domain Recommendation System

作者：Jiangyi Yin, Yuchun Guo, Yishuai Chen · 年份：2019 · DOI：10.1109/icdmw.2019.00060 · 被引用次数：7 · 研究领域：Recommender Systems and Techniques、Advanced Graph Neural Networks、Caching and Content Delivery

The data sparsity or the cold start is a general bottleneck of recommendation service systems. Two types of state-of-art schemes address this issue via enriching knowledge. Cross-domain recommendation schemes transfer knowledge from another dense domain, but they only explore a single kind of shared knowledge. Recommendation schemes built on heterogeneous information networks (HIN) utilize knowledge implied in meta-paths connecting various objects, these methods, however, investigate meta-paths in a single domain. To improve recom-mendation performance in a highly sparse scenario, we propose a HIN Embedding based Cross-domain Recommendation (HecRec) framework, which exploits cross-domain information by establishing meta-path based HIN embeddings in both the source and the target domain and conducts personalized recommendation by integrating the obtained HIN embeddings with a rating predictor. To make the best use of cross-domain information and avoid the knowledge confliction between knowledge from different metapaths observed in real-system datasets, we adopt a concept of "overpass bridge" to integrate the HIN embeddings drawn via different meta-paths. Experiments on two public datasets, i.e., MovieLens and LibraryThing, demonstrate the capacity of HecRec on addressing data sparseness. The recommendation performance of HecRec is improved up to 6.9% in MAE and 5.5% in RMSE compared with the state-of-art schemes in the coldest cases.