A Comparative Analysis of Explainable AI Techniques for Sentiment Classification of TikTok & Tokopedia Reviews
Abstrak
Explainable Artificial Intelligence (XAI) using the LIME (Local Interpretable Model-Agnostic Explanations) method to enhance the interpretability of sentiment classification for user reviews on TikTok and Tokopedia. Using TF-IDF for feature extraction, three machine learning classifiers —Random Forest, Decision Tree, and K-Nearest Neighbours—were evaluated through K-Fold Cross-Validation. Random Forest achieved the highest classification accuracy at 89.9%, followed by Decision Tree at 88.25%, and KNN at 81.51%. The most prominent terms in positive sentiment reviews included “mantap” (16,057.23) and “bagus” (15,310.02), while negative sentiment was associated with “biaya,” “sistem,” and “ganti.” LIME provided localized, interpretable insights by highlighting the important terms that influence each prediction. In terms of positive sentiment, words such as “Tokopedia,” “update,” and “go” had strong weights, whereas negative classifications were triggered by terms like “offline” and “error.” ROC Curve analysis further confirmed Random Forest’s strong performance, showing AUC scores of 0.88 for the negative class, 0.87 for the neutral class, and 0.88 for the positive class, outperforming the other models. The network graph also identified “Tokopedia” as a central node, with frequent co-occurrence of terms like “diskon” and “pengiriman,” reflecting key user expectations. These findings demonstrate that combining interpretable AI with high-performing classifiers offers a powerful approach for sentiment analysis in digital platforms. It enables stakeholders to understand user feedback better and make data-driven decisions to improve customer satisfaction and trust.
Identifier
- DOI: https://doi.org/10.25139/inform.v10i2.10748
- OpenAlex: https://openalex.org/W7128375262