Exploiting Transfer Learning With Attention for In-Domain Top-N Recommendation

Abstract

Cross-domain recommendation has recently been extensively studied, aiming to alleviate the data sparsity problem. However, user-item interaction data in the source domain is often not available, while user-item interaction data of various types in the same domain is relatively easy to obtain. This paper proposes a recommendation method based on in-domain transfer learning (RiDoTA), which represents multi-type interactions of user-item as a multi-behavior network in the same domain, and can recommend target behavior by transferring knowledge from source behavior data. The method consists of three main steps: First, the node embedding is performed on each specific behavior network and a base network by using a multiplex network embedding strategy; Then, the attention mechanism is used to learn the weight distribution of embeddings from the above networks when transferring; Finally, a multi-layer perceptron is used to learn the nonlinear interaction model of the target behavior. Experiments on two real-world datasets show that our model outperforms the baseline methods and three state-of-art related methods in the HR and NDCG indicators. The implementation of RiDoTA is available at https://github.com/sandman13/RiDoTA.

Keywords

• Multi-behavior recommendation
• transfer learning
• attention
• multiplex network embedding