Portability of Predictive Academic Performance Models: An Empirical Sensitivity Analysis

Abstract

The portability of predictive models of academic performance has been widely studied in the field of learning platforms, but there are few studies in which the results of previous evaluations are used as factors. The aim of this work was to analyze portability precisely in this context, where preceding performance is used as a key predictor. Through a study designed to control the main confounding factors, the results of 170 students evaluated over two academic years were analyzed, developing various predictive models for a base group (BG) of 39 students. After the four best models were selected, they were validated using different statistical techniques. Finally, these models were applied to the remaining groups, controlling the number of different factors with respect to the BG. The results show that the models’ performance varies consistently with what was expected: as they move away from the BG (fewer common characteristics), the specificity of the four models tends to decrease.

Using Case Based Reasoning, in the Student Graduation Prediction System (CBR)

“…In the education sector, prediction models have developed to predict student performance [1], graduation success [2], student retention [3], online discussion performance [4], critical course [5], and interactive volume [6]. In the health sector, there are models predicting diabetes [7], disease diagnosis [8], risk of infectious disease [9], and hypoglycemia [10]. There are also prediction models of performance areas [11] [12][13],predicting targets [14], and predicting progress [15].…”