Collaborative Filtering with Attribution Alignment for Review-based Non-overlapped Cross Domain Recommendation

Abstract

Cross-Domain Recommendation (CDR) has been popularly studied to utilize different domain knowledge to solve the data sparsity and cold-start problem in recommender systems. In this paper, we focus on the Review-based Non-overlapped Recommendation (RNCDR) problem. The problem is commonly-existed and challenging due to two main aspects, i.e, there are only positive user-item ratings on the target domain and there is no overlapped user across different domains. Most previous CDR approaches cannot solve the RNCDR problem well, since (1) they cannot effectively combine review with other information (e.g., ID or ratings) to obtain expressive user or item embedding, (2) they cannot reduce the domain discrepancy on users and items. To fill this gap, we propose Collaborative Filtering with Attribution Alignment model (CFAA), a cross-domain recommendation framework for the RNCDR problem. CFAA includes two main modules, i.e., rating prediction module and embedding attribution alignment module. The former aims to jointly mine review, one-hot ID, and multi-hot historical ratings to generate expressive user and item embeddings. The later includes vertical attribution alignment and horizontal attribution alignment, tending to reduce the discrepancy based on multiple perspectives. Our empirical study on Douban and Amazon datasets demonstrates that CFAA significantly outperforms the state-of-the-art models under the RNCDR setting.

Exploiting Variational Domain-Invariant User Embedding for Partially Overlapped Cross Domain Recommendation

“…Cross Domain Recommendation. Cross Domain Recommendation (CDR) emerges as a technique to alleviate the long-standing data sparsity problem in recommendation by assuming the same user set across domains [18,23,33,41]. According to [48], existing CDR models have three main types, i.e., transfer-based methods, clustered-based methods, and multitask-based methods.…”