Agent4POI: Agentic Context-Conditioned Affordance Reasoning for Multimodal Point-of-Interest Recommendation

This collection contains several ML/IR research papers with notable privacy, robustness, and deployment security implications. Key items: DMP-MH proposes a differentially-private sanitize-then-distill pipeline to resist link-reconstruction attacks on behavior-derived semantic graphs and addresses a ‘‘Hubness Explosion’’ sensitivity problem via degree clipping and noisy mirror descent; Agent4POI introduces inference-time, context-conditioned LLM agents for POI recommendation that increase risks of contextual fingerprinting and sensitive-location inference; the HEI-informed RAG nutrition recommm