Cascading Relevance-driven Recommendation Network for CTR Prediction in Trigger-Introduced Recommendation

E-commerce platforms increasingly feature Trigger-Introduced Recommendation (TIR) scenarios. In this setup, a user clicks an item on the homepage—referred to as the trigger item—which conveys a strong, albeit vague and implicit, instant interest. The user is then directed to an undertaking page displaying relevant target items. While traditional recommendation models capture historical user preferences well, they often fail to explicitly leverage the unique relevance between the trigger item and candidate target items. This limitation degrades the immersive shopping experience. To bridge this gap, the authors introduce the Cascading Relevance-driven Recommendation Network (CRRN), designed to jointly process personalized user behavior histories alongside explicit trigger-target interactions.

The CRRN framework integrates three primary components to enhance relevance learning and click-through rate (CTR) prediction. First, the Trigger-Target Interaction (TTI) layer extracts explicit and implicit cross-features between the trigger and target items using a personalized double-layer gating network. Second, the Cascading Interest Fusion (CIF) module predicts the user's trigger intention probability and uses cascading attention blocks alongside cosine similarity to dynamically fuse instant and personalized interests. Third, a Category-assisted Pairwise Loss (CPL) explicitly refines trigger relevance at a granular level by establishing partial order relationships guided by item category associations.

Extensive evaluations on both large-scale industrial datasets from Tmall and public benchmark datasets demonstrate that CRRN outperforms recent state-of-the-art recommendation models, achieving significant gains in AUC and RelaImpr metrics. Ablation studies confirm that each component—the interaction layer, the cascading interest fusion, and the category-guided loss—contributes meaningfully to overall accuracy. Furthermore, online A/B testing validates the practical effectiveness and robustness of the model in real-world industrial deployment.