Game-based Quantification of Chess Player Types

The classification of chess players based on playing styles has traditionally relied on qualitative insights from chess experts. One way is to categorize players into four main types: activists, pragmatists, theorists, and reflectors. However, the classification relies on chess experts and is not quantitatively interrogated. In this thesis, we analyzed 100,000 chess games and 8.5 million positions from renowned players to extract features that can quantify their playing styles and trained a classifier. For the feature extraction, we reproduced the board complexity metrics from two prior studies, and proposed our own methods to identify sacrifices and board pressure. We also extracted other game-level and position-level features according to chess theory to build our dataset. These features were aggregated to the player level and used to classify players into the four types. With such a big dataset, we carried out data analysis to discover insights and patterns. We performed PCA visualizations at the game level and the player level. Although no distinct patterns were found at the game level, we discovered certain patterns at the player level. Finally, we designed a player-level machine learning pipeline to predict the playing styles, reaching an accuracy of 50%, compared to the 25% baseline of random guessing. This research contributes to the data-driven analysis of chess styles, providing a foundation for future studies on player behavior and strategic tendencies within the game.