상세정보

Cross-domain recommendation is one way to alleviate the inherent sparsity problem in recommender systems. Typically, leveraging source domain data helps improving recommendation performance in the target domain. Transfer learning or domain adaptation methods are well-suited for cross-domain recommendation tasks, by transferring the shared user features.
While the important user feature may vary from domain to domain, it is very challenging to model this concept, especially when there is no meta information available.
In this paper, we therefore propose a novel method of using multi-channel embedding with domain-aware attention (MCXA). MCXA allows different domains to focus on user features differently and enables transfer learning.
In addition to domain-aware features from multi-channel embedding, MCXA can capture domain-invariant user features as well with adversarial domain classification loss. The experimental result demonstrates that multi-channel approach helps to extract domain-aware features effectively and MCXA outperforms baseline methods by a remarkable margin. Furthermore, we provide a result of using the entire data from the target domain for its practical use. We also provide analysis on domain-aware attention weights in terms of domain relatedness.

1 Introduction
2 Related Work
2.1 Cross-domain Recommender Systems 
2.2 Domain Adaptation
2.3 Multi-channel Approaches 
3 Method
3.1 Embedding layer: Multi-channel user embedding with attention 
3.2 Score layer: Shared MLP layers 
3.3 Prediction layer: Domain-specific learning and domain classification
3.4 Joint learning
4 Experiments
4.1 Dataset 
4.2 Metrics
4.3 Baselines
4.4 Experimental Setting
5 Results
5.1 Performance of different cross-domain methods (RQ1)
5.2 Effectiveness of multi-channel (RQ2)
5.3 Analysis on attention weights (RQ3)
6 Conclusion