Customer Behavior Analysis Using Cascade Adaptive Feature Reconstruction And Encoding In E-Commerce Datasets

Abstract

E-commerce is the sale and purchase of goods or services through the internet, enabled by online platforms for user interaction, product reviews and transactions. Under the digital platform, customer feedback is essential in enhancing product quality, user experience and business decision-making. This paper suggests a new two-stage approach for enhancing sentiment analysis and review quality prediction in e-commerce based on Amazon product reviews. In stage one, the Hierarchical Intra-Session Behavior Adaptive Reconstructions (HBA-REC) approach preprocesses user reviews by dividing these into micro and meta sessions to identify the short- and long-term behavior patterns. HBA-REC improves data reliability by maintaining rare events, recovering truncated reviews and incorporating contextual sentiment, time and interaction dynamics. In stage two, the Cascade Adaptive Feature Reconstruction and Encoding (CAFRC) system conducts adaptive feature selection by multi-layer analysis through L1 regularization, autoencoder reconstruction loss and integrated gradients. CAFRC masks unstable or redundant features while retaining features with high semantic and predictive importance. Experimental comparisons demonstrate that HBA-REC outperforms conventional and deep learning-based preprocessing methods significantly with lower Mean Absolute Error (MAE) (0.26), Mean Squared Error (MSE) (0.22) and improved R² score (0.88). Likewise, CAFRC performs the highest classification accuracy (92.89%) compared to feature selection methods and outperforms Chi-Square, Principal Component Analysis (PCA) and L1-based methods. The combined pipeline produces a behavior-enriched high-quality feature set that enables strong and interpretable sentiment prediction models. Results indicate improved model performance, reduced overfitting and improved generalizability on e-commerce tasks. Together, HBA-REC and CAFRC constitute an integrated preprocessing and feature optimization pipeline that greatly enhances the efficiency and reliability of e-commerce analytics.