Sensitivity of Predictors in Educational Data: A Bayesian Network Model

This research investigates the application of Bayesian Networks to predict causal relationships in a dataset that captures several demographic and academic features of a group of students from a four-year public university. This educational dataset is characterized by both quantitative and qualitative variables, some of which exhibit a strong pair-wise dependence. To identify this dependence, a factorial analysis of the mixed data is conducted that allows consideration of both variable types to result in a new coordinate space that captures the variance of the data with fewer dimensions. This exploratory stage enables visualization of groups of dependent variables that may be applied for predicting outcomes of interest. It also provides a validation of the results of the Bayesian network (BN) structure modeling. The BN is learnt using bootstrapped arc strength averaging to derive a graphical relationship between variables with arcs represented by a persistence parameter representative of their occurrence in the learning process. The resulting network is shown to be characterized by two major relatively independent structures; one formed by college academic performance metrics and the second by financial, housing and student demographic variables. The prediction accuracy of the BN is evaluated using evidence from pre-college and on-going college variables. The pre-college evidence is found not to be sensitive to college degree completion outcome with only 55% accuracy rate. The on-going college evidence however, improves the prediction accuracy by 75%.