Scholarlink

<p>The success of digital marketing relies on fine-grained customer behavior profiling, yet the required data collection creates a persistent tension with privacy protection. Although Local Differential Privacy (LDP) provides rigorous guarantees, standard LDP mechanisms often suffer substantial utility loss when customer data are incomplete, heterogeneous, and governed by heterogeneous privacy preferences. This paper proposes the Personalized and Utility-Optimized LDP Framework (PU-LDP), a client-side framework for privacy-preserving customer profiling. PU-LDP integrates (i) a BiSample-MD mechanism for unbiased estimation under partial missingness, (ii) an Attribute State Encoding (ASE) scheme that preserves key-value correlations by converting structured attributes into unified categorical states, and (iii) a Stepwise Mechanism with Data Recycling to support personalized privacy budgets while reducing the utility loss caused by data fragmentation. Extensive experiments on real-world and synthetic datasets show that PU-LDP consistently outperforms representative LDP baselines in mean and frequency estimation under varying missing rates and privacy budgets. We further discuss how the resulting aggregate profiles can serve as privacy-preserving state inputs to a downstream deep reinforcement learning (DRL) marketing optimizer. The present study validates the profiling framework; the DRL optimization component is positioned as a practical deployment direction for future end-to-end evaluation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Customer behavior profiling under differential privacy protection for marketing strategy optimization

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月20日

查看全文
  • Lu Yang

摘要

The success of digital marketing relies on fine-grained customer behavior profiling, yet the required data collection creates a persistent tension with privacy protection. Although Local Differential Privacy (LDP) provides rigorous guarantees, standard LDP mechanisms often suffer substantial utility loss when customer data are incomplete, heterogeneous, and governed by heterogeneous privacy preferences. This paper proposes the Personalized and Utility-Optimized LDP Framework (PU-LDP), a client-side framework for privacy-preserving customer profiling. PU-LDP integrates (i) a BiSample-MD mechanism for unbiased estimation under partial missingness, (ii) an Attribute State Encoding (ASE) scheme that preserves key-value correlations by converting structured attributes into unified categorical states, and (iii) a Stepwise Mechanism with Data Recycling to support personalized privacy budgets while reducing the utility loss caused by data fragmentation. Extensive experiments on real-world and synthetic datasets show that PU-LDP consistently outperforms representative LDP baselines in mean and frequency estimation under varying missing rates and privacy budgets. We further discuss how the resulting aggregate profiles can serve as privacy-preserving state inputs to a downstream deep reinforcement learning (DRL) marketing optimizer. The present study validates the profiling framework; the DRL optimization component is positioned as a practical deployment direction for future end-to-end evaluation.