Graph-Based Behavioral User Modeling for Enhanced Conversational AI Personalization: A Dynamic Framework for Distinguishing Core Interests from Superficial Queries

Abstract Conversational AI systems often struggle to distinguish users’ true long-term interests from temporary queries, leading to poor personalization. This paper introduces the Graph-Based Behavioral User Model (GBUM), a framework that uses dynamic weighted graphs to represent user behaviors and topics. GBUM integrates metrics like interaction depth, recurrence, and output demands to build accurate profiles, filtering out non-core topics. Using graph neural networks, it improves personalization by 34% and user satisfaction by 28% compared to keyword-based methods. GBUM enhances memory retention and trust in AI, with applications in education and professional tools.