Classification of Public Complaints Basedon Text Mining Using Modified K-Nearest Neighbor, Naïve Bayes and C4.5 Algorithm

Abstract

To improve public services, accuracy and acceleration are needed in classifying the types of complaints so that complaints can immediately get a response from the relevant regional apparatus. This public complaint data is in text form and is not balanced in each category of regional apparatus, so we contribute to research to compare the performance of different text mining-based classification algorithms. In addition, we also tested the resampling method to overcome imbalanced data. In the final stage, testing is carried out using a multiclass confusion matrix table to show accuracy, precision, recall, and f1-score. The test results show the highest value in the Naïve Bayes algorithm with the ComplementNB model without resampling data, which is 89.58% accuracy, 86.72% precision, 82.40% recall, 84.09% f1-score. However, all scores decreased when combined with SMOTE resampling of 83.66% accuracy, 67.79% precision, 80.35% recall, 71.68% f1-score. ComplementNB can be an alternative model in the classification of public complaints with imbalanced datasets