Cross-domain recommendation based on meta-networks and attention transfer

• Published in: 2024 4th International Conference on Neural Networks, Information and Communication (NNICE) pp. 5 - 9
• Main Authors: Xia, Xue, Liu, Qicheng
• Format: Conference Proceeding
• Language: English
• Published: IEEE 19.01.2024
• Subjects: Artificial neural networks; Benchmark testing; Cross-domain recommendation; Data mining; Data models; Measurement; Meta-network; Multi-headed Attention mechanisms; Prediction algorithms; Sparse matrices
• DOI: 10.1109/NNICE61279.2024.10498172

Although cross-domain recommendation systems play a crucial role in solving the data sparseness and cold start challenges in recommendation systems, current algorithms primarily rely on the user-item rating matrix for user feature modeling, leading to sparse data. Furthermore, existing cross-domain recommendation algorithms employ mapping transfer methods that result in negative migration issues. This negative migration significantly impacts the capability of cross-domain recommendation systems. Therefore, this paper presents a novel cross-domain recommendation algorithm based on meta-network and attention transfer. In the algorithm, we first utilize pre-trained user embedding information to generate user features within their respective domains. We introduce a reconstructed self-encoder to generate global user embeddings, enhancing the user representation for users in the source and target domain. Next, we leverage the user/item interaction information of the user in the source field to generate migratable user feature information through a meta-network. Finally, through the attention network, we use the user feature embeddings of the target field and the user's migratable information in the source field as the input of the attention network to generate feature embeddings specific to the target field for rating prediction. We conducted experiments on two domain pairs consisting of three Amazon datasets. The results indicate that the algorithm presented in this paper outperforms other benchmark models under MSE and MAE evaluation metrics.