Scholarlink

<p>Customer journey management (CJM) has become increasingly important for firms transitioning from traditional customer relationship management (CRM), requiring analytical methods that both explain behavioral drivers and translate insights into operationally executable plans. This study proposes an explainable and operations-ready framework that integrates the pre-purchase and purchase stages, with an optional post-purchase extension via product-review analytics. The framework combines clickstream-based clustering, latent Dirichlet allocation (LDA) for mapping customer-need topics to product functions and specifications, and tree-based gradient-boosting models with Shapley additive explanations (SHAP) to identify the key determinants of click engagement, click-to-purchase behavior, and transaction value. These explainable outputs are transformed into a quality function deployment (QFD) matrix, which is subsequently incorporated into a mixed-integer linear programming (MILP) allocator constrained by budget, delivery capacity, minimum-coverage requirements, and frequency-capping rules. The empirical analysis uses enterprise-scale behavioral and transactional data from a leading Taiwanese consumer electronics firm, comprising 412,580 users and 156 source/medium pairs (March–November 2020), supplemented by 3,492,632 product-review records (January 2021–August 2022). The results reveal substantial cross-segment heterogeneity in customer value and reachable audiences, supporting the need for differentiated segment–channel strategies. In the review-analytics extension, an LDA–LightGBM configuration achieves predictive performance comparable to LSTM and CNN baselines while offering significantly faster training, enabling more frequent model updates. The MILP produces feasible and interpretable segment–channel allocations, validated through return on ad spend (ROAS), cost per acquisition (CPA), and segment-level conversion-lift analysis. Overall, the integrated SHAP–QFD–MILP pipeline demonstrates how explainable modeling outputs can be operationalized into prescriptive and auditable decisions, providing a scalable blueprint for customer-centric resource planning across products, markets, and channels.</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

An explainable AI-driven framework integrating SHAP and QFD for transitioning from CRM to CJM in customer-centric optimization

  • 期刊论文

  • 发表日期：

    2026年1月9日

查看全文
  • Tzu-Chien Wang

摘要

Customer journey management (CJM) has become increasingly important for firms transitioning from traditional customer relationship management (CRM), requiring analytical methods that both explain behavioral drivers and translate insights into operationally executable plans. This study proposes an explainable and operations-ready framework that integrates the pre-purchase and purchase stages, with an optional post-purchase extension via product-review analytics. The framework combines clickstream-based clustering, latent Dirichlet allocation (LDA) for mapping customer-need topics to product functions and specifications, and tree-based gradient-boosting models with Shapley additive explanations (SHAP) to identify the key determinants of click engagement, click-to-purchase behavior, and transaction value. These explainable outputs are transformed into a quality function deployment (QFD) matrix, which is subsequently incorporated into a mixed-integer linear programming (MILP) allocator constrained by budget, delivery capacity, minimum-coverage requirements, and frequency-capping rules. The empirical analysis uses enterprise-scale behavioral and transactional data from a leading Taiwanese consumer electronics firm, comprising 412,580 users and 156 source/medium pairs (March–November 2020), supplemented by 3,492,632 product-review records (January 2021–August 2022). The results reveal substantial cross-segment heterogeneity in customer value and reachable audiences, supporting the need for differentiated segment–channel strategies. In the review-analytics extension, an LDA–LightGBM configuration achieves predictive performance comparable to LSTM and CNN baselines while offering significantly faster training, enabling more frequent model updates. The MILP produces feasible and interpretable segment–channel allocations, validated through return on ad spend (ROAS), cost per acquisition (CPA), and segment-level conversion-lift analysis. Overall, the integrated SHAP–QFD–MILP pipeline demonstrates how explainable modeling outputs can be operationalized into prescriptive and auditable decisions, providing a scalable blueprint for customer-centric resource planning across products, markets, and channels.