TriSeqNet: Tri-dimensional sequential fusion network for triple relationships in CTR prediction

Abstract

Click-through rate (CTR) prediction is a core task in recommender systems. Although deep learning-based methods have achieved notable success, they primarily focus on user-item interactions while overlooking the multidimensional relationships formed during these interactions. In reality, user-item interactions form two behavioral sequences encompassing three key relationships: user-item, user-user, and item-item. To address this, we propose a Tri-dimensional Sequential Fusion Network (TriSeqNet), which comprehensively models user-item, user-user, and item-item relationships to enhance CTR prediction. TriSeqNet consists of three components: the User-to-Item Network, which captures interest evolution in user behavior sequences while computing the relevance between users and candidate items to explore user-item relationships; the User-to-User Network, which models item behavior sequences to extract global item representations and measure user correlations, capturing associations between the target user and those who have interacted with the candidate item; and the Item-to-Item Network, which learns interaction feature embeddings to explore relationships between user-interacted items and candidate items. Experimental results on three public CTR datasets show that TriSeqNet outperforms 13 baseline models, achieving AUC improvements of 2.02%, 1.52%, and 0.15%, Logloss reductions of 4.10%, 10.03%, and 1.79%, and F1-score gains of 2.87%, 2.27%, and 1.33%, compared to the best baseline.