Cross-Domain Sequential Recommendation via Neural Process

Haipeng Li,Jiangxia Cao,Yiwen Gao,Yunhuai Liu,Shuchao Pang

2024-10-17

Abstract:Cross-Domain Sequential Recommendation (CDSR) is a hot topic in sequence-based user interest modeling, which aims at utilizing a single model to predict the next items for different domains. To tackle the CDSR, many methods are focused on domain overlapped users' behaviors fitting, which heavily relies on the same user's different-domain item sequences collaborating signals to capture the synergy of cross-domain item-item correlation. Indeed, these overlapped users occupy a small fraction of the entire user set only, which introduces a strong assumption that the small group of domain overlapped users is enough to represent all domain user behavior characteristics. However, intuitively, such a suggestion is biased, and the insufficient learning paradigm in non-overlapped users will inevitably limit model performance. Further, it is not trivial to model non-overlapped user behaviors in CDSR because there are no other domain behaviors to collaborate with, which causes the observed single-domain users' behavior sequences to be hard to contribute to cross-domain knowledge mining. Considering such a phenomenon, we raise a challenging and unexplored question: How to unleash the potential of non-overlapped users' behaviors to empower CDSR?

Information Retrieval,Social and Information Networks

What problem does this paper attempt to address?

The paper attempts to address the problem of how to fully utilize the behavior data of non-overlapping users to improve the performance of recommendation systems in Cross-Domain Sequential Recommendation (CDSR).

### Background and Problem Description

In traditional cross-domain sequential recommendation, most methods mainly rely on the multi-domain behavior data of overlapping users to capture the correlations between different domains. However, these overlapping users only account for a small portion of the entire user group, thus methods based on overlapping users have the following issues:

1. **Assumption Bias**: Assuming that the behavior patterns of overlapping users can represent the behavior patterns of all users, which is obviously unreasonable.
2. **Performance Limitation**: Since the number of non-overlapping users is much larger than that of overlapping users, relying solely on the learning paradigm of overlapping users limits the model's performance.
3. **Knowledge Mining Difficulty**: For non-overlapping users, the lack of behavior data from other domains makes it difficult for single-domain user behavior sequences to contribute to cross-domain knowledge mining.

### Solution

To overcome the above issues, the paper proposes a new cross-domain sequential recommendation framework called CDSRNP (Cross-Domain Sequential Recommendation with Neural Processes). The main innovations of this framework include:

1. **Neural Processes (NP)**: Combining the advantages of meta-learning and stochastic processes, it establishes a direct connection between non-overlapping users and items in other domains by aligning the prior and posterior distributions of the support set and query set.
2. **Fine-Grained Interest Adaptation Layer**: Designed a fine-grained interest adaptation layer to identify the interests of relevant overlapping users, thereby providing useful information for the prediction of non-overlapping users.
3. **Comprehensive User Behavior Pattern Modeling**: Not only learns the item-item collaborative signals within a single domain but also captures the user-item interaction pattern distribution across all domains.

### Experimental Results

Experimental results show that CDSRNP significantly outperforms existing state-of-the-art methods on two real-world datasets. Especially in cases with a high proportion of non-overlapping users, the performance improvement of CDSRNP is more pronounced, demonstrating that this method can effectively utilize the behavior data of non-overlapping users to improve the overall performance of the recommendation system.

### Main Contributions

1. **New Perspective**: Proposes a new perspective by considering the behavior data of non-overlapping users to enhance cross-domain sequential recommendation.
2. **Innovative Method**: Designs a novel CDSRNP framework that can capture the complete interaction pattern distribution of both overlapping and non-overlapping users.
3. **Detailed Analysis**: Validates the effectiveness and superiority of CDSRNP through extensive experiments and detailed analysis.

In summary, by introducing neural processes and a fine-grained interest adaptation layer, this paper successfully addresses the issue of insufficient utilization of non-overlapping user behavior data in cross-domain sequential recommendation, providing new ideas and methods for the research of recommendation systems.