Answers are grounded in this paper and cite their sources.
Dynamic interaction network (DIN), consisting of a stream of timestamped user-item interactions, is prevalent in real-world online scenarios, e.g., dynamic multi-round recommendation. Existing studies have learned the evolving representations of users and items from the perspectives of newly-arrived and historical interactions. However, most of them assume the historical interactions would reflect the user/item's actual representation evolution, without scrutinizing whether they benefit the representation learning. In addition, they neglect the negative excitement of the newly-arrived interactions on learning representation. In this paper, we propose the CREAM, a dynamic self- C orrecting inte R action n E twork with neg A tive excite M ent. Specifically, CREAM consists of four core components: (i) warm initialization, which generates flexible embeddings for the cold-start users and items; (ii) adaptive dual-side corrector, which dynamically and adaptively corrects the histories for both the user side and item side; (iii) temporal excitement learning, which learns the temporal negative impact of new interactions; and (iv) update and prediction, which utilizes both historical and new interactions for representation update and then conducts future prediction by integrating temporal and static representations. Extensive experiments on three public datasets demonstrate the proposed CREAM outperforms state-of-the-art baselines for the future interaction prediction task.
No takes yet. Share an insight, caveat, or question.
Tang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: