A2TN: Aesthetic-Based Adversarial Transfer Network for Cross-Domain Recommendation

Chenghua Wang, Yu Sang

Abstract: To address the long-standing data sparsity problem in recommender systems, Cross-Domain Recommendation (CDR) has been proposed to leverage the relatively richer information from a richer domain to improve the recommendation performance in a sparser domain. Therefore, enhancing the transferability of features in different domains is crucial for improving the recommendation performance. However, existing methods are usually confronted with negative transfer due to the data sparsity problem. To this end, we propose an Aesthetic-based Adversarial Transfer Network (A2TN) for CDR, which takes advantage of the transferability of aesthetic preferences to ensure the effect of feature transfer. Specifically, we first utilize an aesthetic network to extract aesthetic features and a general feature layer to embed general features, which can collaboratively capture user’s comprehensive preferences. Then, we adopt the adversarial transfer network to generate domain-dependent features for avoiding negative transfer and domain-independent features for effective knowledge transfer. Moreover, an attention mechanism is used to fuse different preferences with different weights to reflect their importance in better portraying user’s preferences. Finally, the experimental results on real-world datasets demonstrate that our proposed model outperforms the benchmark recommendation methods.