Neural network-based systems for the predictive classification of higher education academic performance: Applications for selection and placement

CASCALLAR, E. C.; MUSSO, M. F.; KYNDT, E.

Umea

Conferencia; 13th SweSAT Conference. Selection to Higher Education ? Fairness, Efficiency and Consequences; 2010

Umea University: Deparment of Applied Educational Science

This presentation will summarize five highly successful pioneering studies using a predictive systems approach in the field of educational assessment. Because of the complexity of the variables involved in the field of education and educational assessment in particular, it is possible to take full advantage of a predictive systems approach in order to improve the quality of educational assessments.  In particular, neural networks, with their superior potential for pattern recognition and classification are particularly well suited to use available data to give a much broader expectation of academic performance, based on a wider set of predictors which lead to increase validity, accuracy, and utility of the construct and the obtained results, while at the same time increasing the accuracy of the resulting classifications. A predictive systems approach, and the resulting operational models can lead to better assessment programs, improved diagnostic and placement evaluations, better admission systems, and an opportunity for ?continuous assessment?. Two of the studies involved the modelling of expected general academic performance in higher education (first and second year students) from general cognitive, motivational, learning strategies, and background variables. In these two studies, the accuracy of the classification for various categories of expected performance was between 94-98%, even in cummulative performance data, comparing quite successefully with traditional entrance examinations.  The other studies report applications of predictive classification systems in reading, writing, and mathematics.  In all of these applications accuracy in classification was between 90-99%, and predictors also involved a wide range of motivational, self-regulation, learning strategies, cognitive and background variables.  In summary, these models are proposed as valuable alternatives to develop more comprehensive and accurate selection systems which can increase the validity and the accuracy of selection decisions in higher education, thus achieving higher levels of fairness and efficiency, while also increasing the understanding of the processes involved.