Improving Accuracy Performance on CNN and BiLSTM Algorithms for Classification of Public Complaints Against Police Through WhatsAPP

The Indonesian republic police have provided services to the community by facilitating reporting via WhatsApp. However, messages sent through WhatsApp require manual identification to determine the type of offense reported by the public. Therefore, this research aims to assist in this process by applying deep learning approaches, specifically Convolutional Neural Network (CNN) and Bidirectional Long Short-Term Memory (BiLSTM). Despite the high accuracy of CNN and BiLSTM in some classifications, their performance varies depending on the dataset. This study enhances both the architecture and data preprocessing, including class balancing using SMOTE. To prevent overfitting, early stopping is applied during training. The optimization process in this study utilizes the Adam optimizer, known for its efficiency in handling large-scale data and its ability to adapt the learning rate. The experimental results indicate that while early stopping effectively mitigates overfitting, the resulting accuracy of 68% is not optimal. Without early stopping, the accuracy improves to 91 % for CNN and 98% for BiLSTM. These findings confirm that stopping models at the right moment and the choice of data augmentation and split techniques significantly impact the overall accuracy, making them crucial aspects of this research.