DiMA Accepted at AAAI 2026 (Oral)

We are pleased to announce that our paper DiMA (Distinguishing Resident and Tourist Preferences via Multi-Modal LLM Alignment for Out-of-Town Cross-Domain Recommendation) has been accepted at AAAI 2026 as an Oral presentation!

Authors: Fan Zhang, Jinpeng Chen, Tao Wang, Huan Li, Senzhang Wang, Feifei Kou, Ji Ye, Kaimin Wei, Zhenye Yang

Highlights (from the abstract): DiMA targets reranking in out-of-town (OOT) recommendation, addressing two key bottlenecks — cross-modal reasoning difficulty and resident/tourist preference shift — with an end-to-end multi-modal LLM alignment and teacher-student distillation framework:

• Multi-modal alignment: unify POI images and text into structured semantic labels via MLLM/LLM for cross-modal comparison and efficient downstream inference.
• Preference disentanglement: a Chain-of-Thought (CoT) reasoning flow where the “teacher” LLM distinguishes and synthesizes “tourist-mode” preferences from both home-city and travel-city histories to rerank candidate POIs.
• Efficient deployment: two-stage SFT→DPO training of a compact “student” model that learns and can surpass the teacher’s reranking logic, enabling low-latency inference and significantly improved OOT reranking.
• Experiments: DiMA significantly improves NDCG@K/MAP@K over strong baselines on real multi-city data.

Stay tuned for the paper and code release!