Optimasi Klasifikasi Keluhan Pengguna E-Wallet Berbasis Deep Learning IndoBERT untuk Mendukung Manajemen Layanan TI

Abstract

Abstract.

Purpose: The increasing use of e-wallets in Indonesia has led to an increasingly diverse number of user complaints related to digital services, necessitating a classification method capable of quickly and accurately grouping complaints to support IT Service Management. This study aims to optimize the classification of e-wallet user complaints using IndoBERT and examine its use in supporting service desk and incident management processes.

Method/Design/Approach: The study used a quantitative approach and research data were obtained from user reviews of e-wallet applications on the Google Play Store platform, consisting of seven categories: Transactions, Top Up, Balance, Login, Verification, Applications, and Customer Service, with 100 data points each. The dataset was stratified into 560 training data points (80%), 70 validation data points (10%), and 70 testing data points (10%). The research stages included preprocessing, labeling, IndoBERT tokenization, fine-tuning, hyperparameter optimization, and evaluation using accuracy, precision, recall, F1-score, and confusion matrix.

Results/Findings: Experimental results show that IndoBERT Baseline achieved 98.57% accuracy, while IndoBERT Optimization achieved 100% accuracy, precision, recall, and F1-score. The Confusion Matrix shows that all 70 test data sets were correctly classified.

Novelty/Originality/Value: The contribution of this research lies in optimizing the IndoBERT model to classify e-wallet complaints based on IT service problem categories and utilizing the classification results to support IT Service Management processes, specifically complaint categorization, incident grouping, and service desk.