Intelligent Churn Prediction In E-Commerce Using Conversational Analytics

Abstract

Customer retention inspection is a critical operational task for ensuring business safety, continuity, and financial reliability in modern e-commerce platforms. Traditional manual identification and standard tabular analysis methods are time-consuming, costly, and prone to miss hidden contextual warning signs within customer interactions. This paper presents a comprehensive multi-modal deep learning-based framework for automatic e-commerce customer churn identification using a late-fusion architecture of Multi-Layer Perceptrons (MLP) and Transformer-based Natural Language Processing (NLP) networks. The proposed system utilizes structured customer account profiles extracted from database tables combined with unstructured conversation text logs captured from client support transcripts. The acquired datasets undergo targeted scaling, balancing via Borderline-SMOTE, and hierarchical feature extraction to enhance prediction accuracy. A combination of deep Tabular Neural Networks and a DistilBERT transformer model is used to identify and classify risk patterns across accounts. The model effectively distinguishes between volatile, high-risk profiles and non-defective, secure customer conditions. Experimental results demonstrate that the proposed multi-modal latefusion system achieves an overall accuracy of 95% and an F1-Score of 93%, providing a highly reliable and efficient solution for real-time customer analytics, behavioral monitoring, and automated maintenance planning for enterprise marketing teams.