PRODUCT SALES PREDICTION USING XGBOOST WITH FEATURE IMPORTANCE ANALYSIS FOR ADVERTISING MEDIA EVALUATION

Abstract

Product sales prediction plays a crucial role in supporting data-driven marketing strategies and optimizing advertising expenditures. Although previous studies have demonstrated the effectiveness of machine learning techniques for sales forecasting, most of them primarily focus on prediction accuracy and provide limited insights into the contribution of individual advertising channels to sales performance. This limitation reduces the interpretability and practical value of predictive models for business decision-making. Therefore, this study proposes a product sales prediction framework using Linear Regression as a baseline model and XGBoost Regression combined with Feature Importance Analysis for advertising media evaluation. The novelty of this study lies in integrating predictive modeling and interpretable analysis within a single framework, enabling both accurate sales prediction and the identification of influential advertising factors. Hyperparameter optimization and five-fold cross validation were employed to improve model reliability and robustness. Experimental results show that Linear Regression outperformed XGBoost, achieving an R² score close to 1.0, while XGBoost achieved an R² score of 0.953 with a mean cross-validation R² score of 0.950, indicating stable predictive performance. Feature Importance Analysis revealed that Affiliate Marketing was the most influential factor, followed by Billboards and Social Media. These findings contribute to marketing analytics by providing interpretable insights that support advertising budget optimization and more effective data-driven business decision-making.