Multi-behavior aware recommendation with joint contrastive learning and reinforced negative sampling
Abstract Traditional recommendation systems usually rely on single user-item interaction information to capture the characteristics of users and items. However, in real-world applications, the interactions between users and items are far more diverse. Therefore, efficiently integrating multidimensio...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Springer
2025-06-01
|
| Series: | Complex & Intelligent Systems |
| Subjects: | |
| Online Access: | https://doi.org/10.1007/s40747-025-01970-1 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Abstract Traditional recommendation systems usually rely on single user-item interaction information to capture the characteristics of users and items. However, in real-world applications, the interactions between users and items are far more diverse. Therefore, efficiently integrating multidimensional interaction data is crucial to improve the performance of recommendation systems. The existing methods still face challenges in terms of modeling behavioral dependencies, handling skewed data distributions, and optimizing the quality of negative samples. To address these obstacles, a joint contrastive learning and reinforced negative sampling model is proposed to multi-behavior aware recommendation. First, user interaction data concerning multiple behaviors are incorporated to automatically capture and encode behavioral context information and high-order dependencies, enabling a more comprehensive understanding of user behavior patterns. Second, contrastive learning is applied to the user data acquired under different behaviors to construct enhanced personalized user behavior representations. Additionally, a similarity-based strategy using the dot product operation is introduced to identify hard negative samples with high similarity to users, thereby strengthening the item-side representations. The model is optimized through multi-task learning. Extensive experiments conducted on three real datasets, i.e., IJCAI, Tmall, and Retail, demonstrate that the proposed method improves upon the performance of the existing multi-behavior baseline models by 14.5%, 6.0%, and 1.1% in terms of hit ratio and by 19.7%, 9.4%, and 2.3% in terms of the normalized discounted cumulative gain, respectively, verifying the effectiveness of the new approach. |
|---|---|
| ISSN: | 2199-4536 2198-6053 |