Attentive-Feature Transfer based on Mapping for Cross-domain Recommendation

Abstract

Recommendation systems have been widely developed for numerous applications. Existing systems may still suffer from negative transfer or cold starts. These drawbacks are essentially due to overlooking domain-specific users' personal preferences or cross-domain user-item interactions. To address these problems, we propose a cross-domain recommendation algorithm built on a mapping-based attentive feature transfer (MAFT) model. Our MAFT model utilizes matrix factorization and an attention mechanism for fine-grained modeling of user preferences. Then, overlapping cross-domain user features are combined through feature fusion. Moreover, a multilayer perceptron (MLP) is built to map the obtained user features to target-domain user features. Finally, the user-item ratings can be predicted in the target domain. We carried out experiments on the large-scale MovieLens dataset as well as the real Douban Book and Douban Movie datasets. The results show that the precision of the MAFT-based method is clearly higher than those of other cross-domain recommendation methods, especially for cold-start users with few item interactions.