Hybrid deep learning for e-commerce customer satisfaction classification
Array, cilt.31, 2026 (ESCI, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 31
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.array.2026.101153
- Dergi Adı: Array
- Derginin Tarandığı İndeksler: Emerging Sources Citation Index (ESCI), Scopus, Compendex, Directory of Open Access Journals
- Anahtar Kelimeler: Customer satisfaction classification, E-commerce analytics, GRU-LSTM, Hybrid deep learning, Imbalanced data, SMOTE
- Ankara Yıldırım Beyazıt Üniversitesi Adresli: Evet
Özet
Customer satisfaction is a determinant of retention, platform reputation, and competitiveness in e-commerce. This study develops a leakage-aware and imbalance-aware framework for binary customer satisfaction classification using records from the Olist e-commerce dataset. Review scores of 1-2 were classified as dissatisfied, whereas scores of 3-5 were classified as satisfied. Following preprocessing, the analytical dataset comprised 108,449 observations. Review-text variables were excluded to reduce the risk of target-related information leakage. The feature set incorporated product, payment, seller, customer-location, fulfillment, and order-status attributes, together with total order value, payment-based transaction residual, profit-margin ratio, and product volume. Eight conventional machine learning algorithms (Decision Tree, Logistic Regression, Naïve Bayes, Random Forest, Classification and Regression Trees, CatBoost, XGBoost, and LightGBM) were compared with CNN, RNN, MLP, GRU-LSTM, and LSTM-GRU models. The data were divided using a stratified 80:20 train-test split. SMOTE was applied exclusively to the processed training partition, while the independent test set retained its original class distribution. Recurrent architectures processed reshaped tabular inputs and were interpreted as nonlinear representation-learning models rather than temporal forecasting systems. GRU-LSTM achieved the highest satisfied-class performance, with 86.12% accuracy, 86.42% precision, 99.18% recall, and 92.36% F1-score. CNN, LSTM-GRU, and Naïve Bayes produced comparable F1 scores. However, dissatisfied-customer recognition remained limited: GRU-LSTM achieved 76.03% precision, 14.35% recall, and 24.14% F1-score. Macro F1-score, weighted F1-score, and balanced accuracy were 58.25%, 81.86%, and 56.76%, respectively. The findings demonstrate that high aggregate performance can conceal weak minority-class recognition. The study emphasizes class-sensitive evaluation for retrospective transaction-stage customer satisfaction classification and motivates stage-specific, category-aware, and cost-sensitive approaches.