This study uses information about in-game strategy use, identified through cluster analysis of actions in an educational video game, to make data-driven modifications to the game in order to reduce construct-irrelevant behavior. The examination of student strategies identified through cluster analysis indicated that (a) it was common for students to pass certain levels using incorrect mathematical strategies and (b) throughout the game a large number of students used order-based strategies to solve problems rather than strategies based on mathematics, making measurement of their mathematical ability difficult. To address the construct irrelevant variance produced by these issues, two minor changes were made to the game and students were randomly assigned to either the original version or the revised version. Students who played the revised version (a) solved levels using incorrect mathematical strategies significantly less often and (b) used order-based strategies significantly less often than students who played the original version. Additionally, student perception of the revised version of the game was more positive than student perception of the original version, though there were no significant differences in either in-game or paper-and-pencil posttest performance. These findings indicate that data mining results can be used to make targeted modifications to a game that increased the interpretability of the resulting data without negatively impacting student perception or performance.