Sentiment Analysis and Topic Extraction of Customer Complaints on Tokopedia Using the SVM and Latent Dirichlet Allocation (LDA) Methods

Abstract

The growth of e-commerce transactions on Tokopedia generates massive volumes of unstructured customer review data. Most prior studies focus solely on binary sentiment polarity without uncovering the underlying root causes of customer complaints. This study aims to implement an integrated framework combining Support Vector Machine (SVM) and Latent Dirichlet Allocation (LDA) to classify sentiment polarity and extract latent complaint topics automatically. The dataset comprises 40,607 Tokopedia product reviews processed through text preprocessing without stemming and TF-IDF feature extraction limited to the top 5,000 features. To handle severe class imbalance (6.78% negative vs. 93.22% positive), the linear SVM was optimized with cost-sensitive balanced class weighting. The evaluation results demonstrate an overall accuracy of 86.43% and a negative class Recall of 71.45%, successfully isolating 1,338 complaint reviews in the test set. Subsequently, LDA topic modeling ( ) achieved a topic coherence score ( ) of 0.2523, mapping complaints into four primary operational dimensions: (1) Product Quality, Durability, and Damage; (2) Size and Physical Dimension Fit; (3) Seller Responsiveness and Service; and (4) Color Variant and Visual Mismatch. This integrated approach effectively provides granular, actionable diagnostic insights to support decision-making systems for operational and customer satisfaction enhancement in e-commerce platforms.

GS Cited Analysis