Dual-channel adaptive user-intent routing alignment for multimodal recommendation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Graph-based multimodal recommendation systems leverage visual and textual item features to alleviate data sparsity, yet three systematic limitations persist: textual modeling is confined to item metadata, neglecting review-derived sentiment signals; user-side representations rely on ID embeddings without personalized multimodal alignment; and the standard BPR objective applies uniform gradient pressure irrespective of sample quality. We propose DAURA ( D ual-Channel A daptive U ser-Intent R outing A lignment ), a plug-in enhancement framework comprising three modules: (1) an LLM Review Sentiment Channel that distills affective signals from 5-core user reviews into an independent textual representation, complementing metadata-derived semantic features; (2) an Adaptive Dual-track User-Centric Alignment (AD-UCA) module that constructs per-user multimodal profiles via graph aggregation and employs a lightweight intent router to assign personalized modality-preference weights at the loss level; and (3) a Dual-core Adaptive Margin BPR (DAM) module that fuses sentiment polarity and semantic density to generate item-level adaptive ranking margins. Experiments on three Amazon benchmark datasets and the Yelp Open Dataset show that DAURA consistently outperforms all baselines across product and restaurant recommendation domains, while ablation studies validate the contribution of each module.