A Dynamic User Value Evaluation Model and Intelligent Recommendation Mechanism for Precision Marketing

Abstract

Live-stream commerce, characterized by real-time user-host interactions and rapidly shifting purchase intent, highlights the critical limitations of traditional static customer valuation models such as Recency, Frequency, Monetary (RFM) analysis. These conventional models, alongside standard recommendation engines optimized solely for short-term engagement metrics, fundamentally fail to capture dynamic user value trajectories and often misclassify behaviorally active but historically low-spend users. To comprehensively address this significant research gap, this study proposes a novel closed-loop, co-adaptive framework integrating a Dynamic User Value Evaluator (DUVE) and an Intelligent Recommendation Agent (IRA). Specifically, the DUVE module updates user embeddings every 15 minutes using an advanced temporal graph neural network equipped with attention-based decay mechanisms, ensuring highly accurate temporal representations. Concurrently, the IRA employs a sophisticated dual-objective deep Q-network designed to optimally balance immediate conversion rates with long-term customer lifetime value. Evaluated via a rigorous 28-day large-scale A/B test deployed on a major live-commerce platform, the proposed full system significantly outperformed the static baseline. It achieved an impressive 11.9% increase in Gross Merchandise Volume (GMV) per impression and an 18.9% higher capture rate of high-value users. Furthermore, the framework successfully raised 30-day user retention from 12.1% to 14.4% and increased the average order value by 6.2%. The empirical findings conclusively demonstrate that synchronizing real-time value assessment with adaptive recommendation strategies enables substantially more effective, profitable, and sustainable precision marketing within highly dynamic digital retail environments.