<![CDATA[The rapid growth of online chess platforms has generated large-scale structured game data that can be utilized for data-driven analysis. In blitz mode games, match outcomes are categorized into win, lose, and draw; however, the distribution of these outcomes is inherently imbalanced, with draw representing a small minority of the dataset. This study aims to evaluate the effectiveness of cost-sensitive learning through balanced class weighting in improving classification performance on an imbalanced large-scale blitz chess dataset. A total of 100,000 rated blitz games were extracted from the Lichess open database and processed through preprocessing, feature extraction, and stratified data splitting. Three supervised learning algorithms - Support Vector Machine (SVM), Decision Tree, and Random Forest - were implemented. Model performance was evaluated using Macro F1-score as the primary metric, along with accuracy and 5-fold stratified cross-validation. The results indicate that without cost-sensitive learning, the recall for the minority class (draw) approaches zero despite achieving higher overall accuracy (0.54). In contrast, applying balanced class weighting significantly improves minority class detection, increasing recall for draw up to 0.73 with a Macro F1-score of approximately 0.40, although overall accuracy decreases to 0.45. This demonstrates the trade-off between global performance and minority class sensitivity. Feature importance analysis further reveals that move count is the most influential predictor of match outcomes. These findings confirm that imbalance-aware learning plays a critical role in large-scale chess outcome classification and highlight the importance of appropriate evaluation metrics in imbalanced datasets]]>
Copyrights © 2026
