Evidence Conflict Analysis Approach to Obtain an Optimal Feature Set for Bayesian Tutoring Systems

Abstract

Identifying the appropriate features for constructing a Bayesian student model is crucial to ensure that the model is always optimal. Feature sets can be identified via two types of feature selection algorithms: (i) algorithms that return a discrete set of features, and (ii) algorithms that rank features from the highest to the lowest importance with respect to a class label. To determine the optimal feature set from the second type of feature selection algorithm has always been a challenge, mainly because indifference in overall predictive accuracies between feature sets often occurs. In this light, this paper proposes evidence conflict analysis approach to tackle the challenges. This approach analyzes the conflicts in evidence when a Bayesian Network is employed as a student model. To demonstrate the proposed method, the experiments in this study had utilized two datasets that were transformed from 244 students’ log data. The empirical findings suggested that evidence conflict analysis can differentiate the performance of feature sets having the same overall predictive accuracy.