Comparison of LSTM and TCN Models for Customer Churn Prediction Based on Sentiment and Transaction Data

Abstract

This study investigates the combined use of customer review sentiment analysis and transaction history to predict customer churn on the Balimall Market e-commerce platform. The dataset includes 41,519 reviews labeled with positive and negative sentiments and 48 transaction samples labeled as churn or non-churn based on RFM method. Two deep learning models, Long Short-Term Memory (LSTM) and Temporal Convolutional Network (TCN), are applied in parallel for each analysis path. Data pre-processing includes filtering, cleaning, tokenizing, normalization, sentiment labeling, as well as feature engineering and churn labeling. Evaluation using accuracy, precision, recall, F1-score, and confusion matrix metrics shows that TCN excels with 91.55% accuracy on sentiment analysis and 91.67% on churn prediction, while LSTM achieves 86.35% and 86.67% respectively. Segment analysis shows that 47.30 % of users express negative sentiment yet remain active, 51.69 % express positive sentiment and remain active , 0.54 % express negative sentiment and churn, and 0.48 % express positive sentiment and churn. This finding demonstrates that negative sentiment alone does not necessarily lead to churn; instead, the greatest churn risk arises in negative sentiment churners and positive sentiment churners. Expert validation confirmed the reliability of both models, with the recommendation of using a hybrid to combine the advantages of each architecture. The results of this study are expected to help Baliyoni Group design a more targeted customer retention strategy and improve customer satisfaction by examining these segment conditions.