Dynamic Evidence Calibration and Interest Evolution for Multimodal Video Recommendation

Abstract

Multimodal video recommendation in feed-based systems requires capturing how user interest evolves within and across videos, yet existing fine-grained methods suffer from three key limitations: (1) they treat preference matching as a static, context-free similarity computation over user history; (2) they assume user attention decays monotonically over time, ignoring the interplay between viewing fatigue and content-driven re-engagement; and (3) they rely on uniform frame sampling, which introduces semantic redundancy and transition noise. We propose DECIE (Dynamic Evidence Calibration and Interest Evolution), a candidate-aware framework that addresses these challenges through four integrated mechanisms. First, Adaptive Semantic Chunking employs hash-based temporal clustering to compress redundant frames into coherent semantic units. Second, a First-Impression Anchor derived from cover images and titles establishes an explicit prior that guides subsequent frame-level evaluation. Third, Candidate-Aware Evidence Calibration uses a lightweight hypernetwork to dynamically retrieve and historical evidence using candidate-specific chunk conditions. Fourth, a Content-Driven Dynamic Interest Evolution module models the non-monotonic transition between user immersion and boredom as a bounded state machine driven by local content feedback. Experiments on three benchmarks (MovieLens, NineRec, and MicroLens) show that DECIE consistently outperforms state-of-the-art baselines, achieving improvements of up to +11.53% in Recall@5 and +10.56% in NDCG@5. Ablation and efficiency analyses further confirm the contribution of each component and the computational benefits of semantic chunking.