Uncovering customer preference heterogeneity from online reviews via multi-dimensional clustering

The proliferation of online consumer reviews offers new opportunities for data-driven market segmentation. However, traditional approaches that rely on surveys or demographic profiling often fail to capture the evolving, heterogeneous nature of customer preferences. Addressing this gap, we propose a novel hybrid framework to identify distinct customer segments and infer their preferences from unstructured text. First, we integrate BERTopic and Word2Vec for robust product feature extraction, followed by a Teacher-Student distillation strategy (distilling knowledge from Qwen-Max to Qwen 2.5-7B) to extract granular feature-level sentiment. We then identify latent customer segments by applying multi-dimensional K-means clustering to a composite of Feature Mention Vectors (representing customer attention) and product metadata (reflecting choice behaviors). Finally, we model segment-specific preferences using Ordinal Logistic Regression (OLR) to identify asymmetric satisfaction drivers and classify features according to the Three-Factor Theory. Empirical validation on a large-scale dataset of iPhone and AirPods reviews from Jingdong.com proves that feature classification is neither universal nor static, as it varies with time and across customer segments. The results validate our framework’s ability to uncover preference heterogeneity and track temporal shifts across product generations, providing firms with a scalable tool for targeted product iteration and personalized marketing.