EVENT DOSSIER
A paper published on September 7, 2026, on arXiv proposed the “Continual Graph Memory” approach. This research aims to address the performance degradation in adaptive recommendation systems caused by changes in user intentions. By introducing the Continual Graph Memory mechanism, the system can dynamically update and retain key information in the user behavior graph, thereby more accurately capturing trends in user intentions and improving the accuracy and adaptability of recommendations in dynamic environments.
本文提出 CGM-Rec，一种用于应对意图漂移的持续图记忆自适应推荐框架。该系统将图状态视为可写内存，包含保守更新的语义图记忆和快速反应的 episodic lesson 记忆两个互补组件，测试期间模型参数冻结，仅通过记忆写入实现适应。实验采用冻结参数、单路重排序协议，在多个推荐设置下验证了 CGM-Rec 的性能。结果表明，CGM-Rec 在多数指标上优于神经及 LLM 基线；在 Bundle 的采样候选重排序任务中，其 HR@1 较最强 LLM 基线提升 29.58%，并在 ML-100K 数据集上以 0.5941 的 HR@5 超越 K-RagRec。