Evaluating Classification Models for Predicting Product Success in Indonesian E-Commerce

The intense competition within the Indonesian e-commerce landscape presents a significant challenge for sellers in forecasting product performance. This study offers a unique contribution by systematically comparing seven machine learning classification algorithms to predict product success across Indonesia's three largest platforms: Shopee, Tokopedia, and Lazada. The primary objective is to identify the most effective algorithm for predicting whether a product's sales will surpass the market median. The methodology involved aggregating and preprocessing a dataset of 3,673 product listings. Product success was defined as a binary variable based on sales volume exceeding the dataset's median. Seven models, including Logistic Regression, KNN, SVM, and tree-based ensembles like Random Forest, XGBoost, and LightGBM, were trained and optimized using a 5-fold cross-validated GridSearchCV. Evaluation was based on accuracy, ROC AUC, and F1-score. The results demonstrate a clear performance hierarchy, with tree-based ensemble models achieving superior results. Random Forest emerged as the premier model, attaining an accuracy of 83.2% and an AUC of 0.907. A subsequent feature importance analysis revealed that shop_followers and price were the most significant predictors of success. This finding has crucial practical implications, particularly for Micro, Small, and Medium Enterprises (MSMEs), by providing a data-driven framework for decision-making. The model enables them to focus resources on actionable strategies—building seller reputation and optimizing pricing—to enhance their competitiveness effectively.