Resource-Efficient Sentiment Classification of App Reviews Using a CNN-BiLSTM Hybrid Model

Abstract

This study evaluates the performance of a hybrid convolutional neural network and bidirectional long short-term memory (CNN + BiLSTM) model for sentiment classification on user reviews from the Spotify mobile application. The primary aim is to explore whether competitive results can be achieved without relying on transformer-based architectures, which often require substantial computational resources. The proposed CNN + BiLSTM model combines local feature extraction with sequential context modeling and is benchmarked against traditional machine learning and simpler deep learning models, including a Random Forest classifier enhanced with polarity features, a standalone CNN, and a fully connected DNN. Sentiment labels were binary (positive or negative) and directly provided in the dataset without being inferred from star ratings. The dataset was balanced to avoid class skew. Experimental results indicate that the CNN + BiLSTM model achieves moderate improvements over the baseline models, with an accuracy of 0.8861 and an F1-score of 0.8691. While it does not surpass the highest-performing transformer-based methods reported in the literature, it performs comparably to several of them, despite having a lower computational footprint. Analyses of ROC curves, confusion matrices, and training dynamics further contextualize the model’s performance, showing strengths in classifying negative sentiments and convergence efficiency. To address overfitting, early stopping and dropout layers were employed as regularization techniques. The study contributes to the ongoing discourse on resource-efficient sentiment analysis by showing that hybrid architectures may offer a practical balance between model complexity and performance in specific application domains.