Comparison of Deep Learning Models for Sentiment Analysis of IPOT Financial App Reviews Using Convolutional Neural Network (CNN) and IndoBERT

The rapid expansion of mobile-based financial applications has generated a large volume of user reviews that contain valuable insights into user satisfaction and system performance. IPOT (Indo Premier Online Trading) is a widely used financial application in Indonesia, making sentiment analysis of its user reviews essential for evaluating its service and improving it. This study applies an experimental methodology to compare the performance of two deep learning architectures, Convolutional Neural Network (CNN) and IndoBERT, for sentiment analysis of financial application reviews. User review data were collected from the Google Play Store. Sentiment labels were automatically assigned based on user ratings, and the dataset was balanced using stratified sampling to obtain 15,000 reviews. Text preprocessing included case folding, removal of punctuation and special characters, tokenization, stopword removal, and stemming. The dataset was then split into training, validation, and testing sets, with oversampling applied only to the training data to prevent data leakage. The comparison between Convolutional Neural Networks (CNNs) and IndoBERT for sentiment analysis of IPOT financial application reviews shows that both models perform sentiment classification effectively, with different strengths across sentiment categories. The CNN model achieved higher overall accuracy (0.8113) compared to IndoBERT (0.7880), indicating strong performance in detecting dominant sentiment patterns, particularly positive sentiment. Meanwhile, IndoBERT achieved superior performance in negative and neutral sentiment classification, as evidenced by higher recall and F1 scores. The confusion matrix and error analysis results further indicate that IndoBERT is more effective at understanding contextual and nuanced language, whereas CNN is more sensitive to explicit lexical sentiment indicators.