Temporal Graph Learning for Dynamic Recommendation
- Publication Type:
- Thesis
- Issue Date:
- 2025
Open Access
Copyright Clearance Process
- Recently Added
- In Progress
- Open Access
This item is open access.
Dynamic recommendation, where users continuously interact with items over time, has been widely deployed in real-world applications. The burst of user-item interactions causes rapid evolution of both user representations and item representations. Compared with the conventional static recommendation that follows the one-time recommendation-evaluation paradigm, dynamic recommendation involves dynamically updating temporal representations to refresh recommendation results in a real-time multi-round manner. Temporal Graph Learning, which constructs temporal interaction graphs by their sequential information, has demonstrated its superiority in modeling temporal representations in the dynamic environment. Leveraging temporal graph learning makes it much easier to explore the evolving preferences of users and generate continuous recommendations for the dynamic recommendation scenario. Thus, this research focuses on investigating the power of temporal graph learning for dynamic recommendation.
However, designing comprehensive temporal update for user-item interaction graph remains a challenge, since existing studies on temporal graph learning are mainly centered on some single insufficient perspectives for update. Moreover, they directly conduct representation update without scrutinizing whether the interactions would truly benefit temporal graph learning, which potentially harms the performance of recommendation. Thus, uncovering the effect of both newly-arrived interactions and historical interactions poses another challenge to temporal update. In addition to the inherent challenges above, the unfairness issue among item exposures and user experiences tends to be amplified due to the evolving context in the dynamic environment. Unfortunately, most fairness methods are designed for the conventional static recommendation.
To address the mentioned challenges, this research first proposes a novel framework that comprehensively and efficiently captures user and item dynamics over time from the perspectives of inherent interaction potential, time-decay augmentation, and symbiotic local structure learning, thus fully modeling the dynamic graph evolution for recommendation. Second, this research presents a novel temporal collaboration-aware graph network by investigating the collaborative effectiveness of the newly-arrived interactions, to guide the graph evolution learning process. Third, this research constructs a novel self-correcting interaction network by dynamically filtering out the inappropriate historical interactions, to ensure the accuracy of temporal graph aggregation. In addition, this research further studies the problem of dual-side unfairness issue for the dynamic recommendation and presents an adaptive model-agnostic post-ranking method to mitigate the dynamic unfairness in temporal graph models. Extensive experiments over real-world datasets demonstrate both the effectiveness and superiority of this research, shedding light on the field of recommendation systems. This research successfully explores different aspects of the temporal graph learning for the dynamic recommendation.
Please use this identifier to cite or link to this item: