CDR
The traditional CDR method utilizes a scoring matrix from two perspectives—knowledge aggregation and knowledge transfer—to complete information transmission for cold start users. (Wu et al., 2020; Li et al., 2022). Existing work has integrated users’ Ds and Dt product rating matrices into a unified joint matrix to share user factors, thus achieving cross-domain integration of knowledge. When different fields are considered as a whole, however, these collaborative filtering-based works can be plagued by serious data sparsity issues (Wei, 2021).
In order to solve the problem of difficult session modeling during the CDR process due to complex situations, Zhang, Hua, et al. (2023) proposed the multi-channel interaction model (DCMI), an interactive model for cross-domain personalized recommendation using dual cross-domain session information. This model has high recommendation performance and considerable accuracy. However, due to data sparsity issues, the model lacks sufficient user preference data for model training and prediction. Based on the parameter sharing method, some parameters of the Ds are shared in the Dt and jointly learned to update these parameters. Li and Tuzhilin (2020) suggested a CDR model: deep dual transfer cross-domain recommendation (DDTCDR). In this model, the core of CoNet is the cross connected unit of the hidden layer, which learns the weights of different layers in two domains to achieve parameter sharing. However, this method has difficulty to capturing dynamic changes in user preferences and cannot adapt to real-time changes in user preferences. By directly sharing the Ds data with the Dt, Hong et al. (2020) proposed a corresponding recommendation model called cross-domain deep neural network (CD-DNN) based on data sharing. This method can enrich the Dt data and thus achieve good recommendation performance. Due to the lack of sufficient interactive information in user preferences in new fields, though, this method has a cold start problem. Implementing CDR from a graph perspective, Yang et al. (2021) developed a deep multi-graph embedding (DMGE) recommendation model by constructing interactive heterogeneous graphs between two domains. This model can combine graph convolutional neural networks to transform recommendation tasks into graph link prediction tasks. Yet, the interpretability of this model is poor, making it difficult for users and domain experts to understand the model’s recommendation results, which may lead to users' distrust of and dissatisfaction with the recommendation results. By utilizing auxiliary information to enhance user representation, Zhao et al. (2020) created a CDR framework via aspect transfer network for cold-start users (CATN), which performs cross-domain aspect-level feature matching during rating prediction, resulting in good recommendation results. However, this method requires a large amount of computational resources and time to train and predict; as a result, the model might be unable to respond to user requests in real-time, which undermines the user experience.