Improving Operational Complaint Detection in Indonesian E-Commerce Reviews Using Interpretable TF-IDF and Linear Support Vector Machines

Abstract

Customer complaints in e-commerce app reviews can support the evaluation of operational services. However, such complaints are often written in informal language, contain more than one aspect of the issue, and are not always easily separated from general negative sentiments. This study develops a multilabel classification based on an operational complaint detection model in Indonesian e-commerce application reviews. The dataset used consists of 1,799 Shopee Indonesia reviews from the Google Play Store with 1 and 2 star ratings. Each review is annotated into five aspects of the complaint, namely CS attitude, delivery, price, product, and payment. The proposed model incorporates TF-IDF, negation-aware n-grams, seed-based features, binary relevance, and linear support vector machines. The test results showed that the hybrid model obtained a micro-F1 of 0.860 and a Hamming loss of 0.064. This value is better than the basic TF-IDF configuration on the micro-F1 and Hamming loss metrics, although the hybrid model's macro-F1 of 0.817 is almost equivalent to the baseline of 0.818. These results suggest that the integration of seed-based features can help reduce multilabel prediction errors, especially in the delivery and product aspects. In addition, the model generates phrase-based evidence that can help users search for predictive reasons more transparently.