Predicting post-purchase regret in green appliances using BERT-enhanced sentiment representation and a SMOTE-XGBoost hybrid model

Abstract

The "green consumption gap"—the divergence between stated environmental preferences and realized post-purchase satisfaction—is particularly significant in high-involvement durable goods markets. Existing literature focuses on positive outcomes such as willingness-to-pay and brand loyalty, leaving the micro-level triggers of escalated post-purchase regret understudied. Three barriers limit progress: survivorship bias toward positive behavioral signals, the inability of shallow lexicon-based methods to capture implicit contextual semantics across heterogeneous products, and severe minority-class sparsity that degrades conventional classifiers. This paper addresses all three within a unified deep feature fusion predictive framework. A cross-category corpus of N = 2,179 verified JD.com reviews is constructed, spanning washing machines, water purifiers, and air purifiers. The target variable Y_regret = 1 is defined as a review with a star rating ≤ 3 and a voluntary follow-up comment (positive rate: 7.6%; imbalance ratio 1:12). 768-dimensional deep semantic representations are extracted from raw reviews using a pre-trained BERT model and subsequently compressed via Principal Component Analysis (PCA) to 16 dimensions to mitigate high-dimensional sparsity. These vectors are fused with category-agnostic manual features and fixed-effect control variables (brand/category one-hot vectors) to decouple cross-domain spurious correlations. SMOTE oversampling (k = 5) is applied exclusively within the training folds of a stratified 5-fold cross-validation, guaranteeing zero data leakage. The resulting BERT-enhanced SMOTE-XGBoost pipeline achieves a mean AUC of 0.9371 ± 0.0098, outperforming the manual-feature baseline by 10.16 percentage points. The ablation results demonstrate that deep contextual embeddings reshape non-linear decision boundaries, providing manufacturers with a robust early-warning system for green-churn intervention.