Scholarlink

<p>The rapid expansion of online retail underscores the critical need for precise customer segmentation to drive personalized marketing, reduce churn, and boost lifetime value. This study develops an end-to-end, highly interpretable segmentation pipeline encompassing advanced feature engineering, dimensionality reduction, exhaustive hyperparameter tuning, and robust validation to reveal stable, actionable customer groups in a large, real-world UK online-retail dataset (541,909 records). We augment the classic RFM (Recency, Frequency, Monetary) framework with: TPAC TF-IDF embeddings of item descriptions, holiday-purchase flags, and exponential recency decay; CACV net monetary value and cancellation ratios. After outlier filtering on RFM scores, we apply PCA (230 dimensions) and compare ten clustering methods (selected to represent major algorithmic paradigms: centroid-based [K-Means], probabilistic [GMM], hierarchical [BIRCH, Agglomerative], density-based [DBSCAN, OPTICS, HDBSCAN], graph-based [Spectral], message-passing [Affinity Propagation], and mode-seeking [Mean Shift]). We perform a full grid search per algorithm using a 'safe' silhouette scorer (ignoring noise) and also report Davies-Bouldin and Calinski-Harabasz indices. Temporal stability is assessed via adjusted Rand indices across time splits, and cluster interpretability is enhanced through SHAP-based feature importance analyses. By integrating textual, temporal, and cancellation behaviors into segmentation followed by systematic tuning and multi-metric validation our pipeline delivers superior cluster quality and actionable business insights compared to prior work. Segments directly enable strategic interventions: 'High-Decay Loyalists' (precision = 0.92) receive VIP retention offers yielding 22–31% ROI lift, while 'At-Risk Cancellers' (recall = 0.89) trigger targeted win-back campaigns. We also demonstrate a reproducible framework for selecting both model and feature set. DBSCAN (ε = 0.3, min_samples = 3 on 10 PCA components) achieved the best silhouette score (0.986), markedly exceeding the 0.72 benchmark in the literature. Agglomerative clustering (average linkage, 2 clusters) scored 0.776, while OPTICS and Spectral Clustering also outperformed classical Gaussian- or centroid-based models. A temporal ARI above 0.8 confirms cluster stability.</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

A shap-enhanced PCA-DBSCAN framework for interpretable retail customer segmentation and strategic insight

  • 期刊论文

  • 发表日期：

    2025年10月13日

查看全文
  • M. Martha Sucharitha,
  • J. P. Senthil Kumar,
  • G. Sateesha,
  • M. V. Ram Prasad,
  • Md. Shaik Amzad Basha

摘要

The rapid expansion of online retail underscores the critical need for precise customer segmentation to drive personalized marketing, reduce churn, and boost lifetime value. This study develops an end-to-end, highly interpretable segmentation pipeline encompassing advanced feature engineering, dimensionality reduction, exhaustive hyperparameter tuning, and robust validation to reveal stable, actionable customer groups in a large, real-world UK online-retail dataset (541,909 records). We augment the classic RFM (Recency, Frequency, Monetary) framework with: TPAC TF-IDF embeddings of item descriptions, holiday-purchase flags, and exponential recency decay; CACV net monetary value and cancellation ratios. After outlier filtering on RFM scores, we apply PCA (230 dimensions) and compare ten clustering methods (selected to represent major algorithmic paradigms: centroid-based [K-Means], probabilistic [GMM], hierarchical [BIRCH, Agglomerative], density-based [DBSCAN, OPTICS, HDBSCAN], graph-based [Spectral], message-passing [Affinity Propagation], and mode-seeking [Mean Shift]). We perform a full grid search per algorithm using a 'safe' silhouette scorer (ignoring noise) and also report Davies-Bouldin and Calinski-Harabasz indices. Temporal stability is assessed via adjusted Rand indices across time splits, and cluster interpretability is enhanced through SHAP-based feature importance analyses. By integrating textual, temporal, and cancellation behaviors into segmentation followed by systematic tuning and multi-metric validation our pipeline delivers superior cluster quality and actionable business insights compared to prior work. Segments directly enable strategic interventions: 'High-Decay Loyalists' (precision = 0.92) receive VIP retention offers yielding 22–31% ROI lift, while 'At-Risk Cancellers' (recall = 0.89) trigger targeted win-back campaigns. We also demonstrate a reproducible framework for selecting both model and feature set. DBSCAN (ε = 0.3, min_samples = 3 on 10 PCA components) achieved the best silhouette score (0.986), markedly exceeding the 0.72 benchmark in the literature. Agglomerative clustering (average linkage, 2 clusters) scored 0.776, while OPTICS and Spectral Clustering also outperformed classical Gaussian- or centroid-based models. A temporal ARI above 0.8 confirms cluster stability.