Prediction of college student performance has long been a topic of research for College Board administrators and educators concerned with university admittance standards, student development, and retention. Predictive indicators and composite rules made from high school GPA (HSGPA), college or university first year GPA (CFYGPA), standardized tests scores (SAT/ACT), in conjunction with individual subject grades have all been studied in an attempt to forecast student academic performance. Each indicator has its shortcomings leading to less informative predictive studies.;To complicate things, diversity among incoming freshmen has become an increasing concern as college officials attempt to make the scholastic experience for students more culturally and ethnically diverse. This creates a situation where by design, students are selected from high schools with different scholastic metrics and college preparatory practices thus eliminating even the possibility of an 'apples to apples' selection criterion for college performance prediction indicators for each individual student.;The task in this case is the selection of an aggregate of performance indicators (prediction rules) that will lead to a valid predictive study for the student while simultaneously grouping and evaluating students based only on other students with the most similar background. To this end, the college student prediction problem is formulated in terms of a modified Fuzzy Logic Neural Network inference system in order to dynamically extract prediction rules generated using the K-Nearest Neighbor training samples. Thus, the prediction rules selection criterion is intuitively based only on students with the most similar backgrounds. Multi-Valued Sequential Interactive Synthesis (MVSIS) is used to keep the number of rules manageable while dealing with any non-determinism that may result from conflicting rules.;Using this hybrid clustering and fuzzy neural network inference approach, meaningful prediction rules were extracted from college student grading indicators and applied effectively to predict college student performance. The minimized rule set using the K-Nearest Neighbor samples are then compared to rules generated when the entire student data sample set is used. It is shown that similar accuracy of prediction is attained using relatively small number rules generated from nearest neighbor student data as when the entire student data sample set is utilized to make predictions.
展开▼
