Multi-Level Graph Attention Network Contrastive Learning for Knowledge-Aware Recommendation

This batch contains arXiv papers on recommender systems, user-persona modelling, datasets, IR simulation, multimodal LLM integration, and a compression framework. Key items with security/privacy relevance: (1) UserGPT and its User Behavior Simulation Engine and HPR-Bench — synthetic but realistic user traces and curriculum SFT for persona models, which can be used to build high-fidelity profiling models and risk re-identification or amplification of sensitive traits; (2) OpenIIR — an LLM-driven, multi-agent IR simulation platform that can generate large-scale, reproducible studies of persona/社