심층 융합 기반 다중 대상 교차 도메인 추천

초록

Cross-domain recommendation systems transfer knowledge across different domains to improve their performance in a target domain. However, they suffer from the problem of “negative transfer,” in which transferred knowledge operates as noise in a rich domain. In turn, it decreases their recommendation performance. This paper proposes a novel Deep Fusion-based Multi-target Cross-Domain Recommendation named DFMCDR. By exploiting Doc2Vec, one of the famous word embedding techniques, we fuse temporal-sequentially and transfer knowledge across domains user-wise to model users and items. In addition, a deep neural network structure is introduced to effectively learn the linearity and non-linearity of user-item interactions and integrate them to predict users’ preference possibility. Extensive experiments with three domain (i.e., restaurant, hotel, and attraction) datasets from TripAdvisor, one real-world online review service, demonstrate that DFMCDR outperforms the state-of-the-art algorithms for single and cross-domain recommendations. Furthermore, an additional experiment shows that DFMCDR can be effectively and efficiently adapted to multi-target cross-domain recommendation fusing for more domains.

키워드