ANALYZING GAME PLAY DATA FOR ACADEMIC PERFORMANCE IMPROVEMENT: A DECISION FOREST APPROACH

Abstract

This study emphasizes the importance of comprehending the correlation between level_groups and questions prior to beginning the data training process. After completing stages 0 to 4, the game introduces a preliminary assessment checkpoint consisting of questions 1 to 3. The selection of training data entails matching data from specific level_groups to pertinent questions, thereby ensuring a tailored approach to model construction. Individual models were painstakingly developed for each query, and their accuracy was assessed; the average accuracy was 0.75585266561448. Due to the irregular distribution of 'correct' column values, the default 0.5 threshold for predicting classes may not optimize performance. To mitigate this, the F1 score, which is superior in evaluating imbalanced scenarios, was computed across multiple thresholds to identify the optimal threshold yielding the highest F1 score, about 0.63. This optimal threshold establishing a balance between precision and recall is crucial for accurate mapping of predicted probabilities to class labels 0 or 1. To effectively advance the discipline, future research should prioritize a clear research roadmap, the incorporation of emerging technologies, collaboration, and data quality.

Keywords: Game-based Learning, Model Accuracy, F1 Score.