Predicting general academic performance and identifying the differential contribution of participating variables using artificial neural networks

Abstract

Many studies have explored the contribution of different factors from diverse theoretical perspectives to the explanation of academic performance. These factors have been identified as having important implications not (Everson, 1995;Garson, 1998

Approaches to learning as predictors of academic achievement: Results from a large scale, multi-level analysis

“…Approaches to learning may mediate some relationships, such as those between self-efficacy and performance (Musso et al, 2013). Intelligence, working memory capacity, attention, anxiety management and cultural capital are also likely to form part of the picture of what influences academic achievement (Musso et al, 2013). Given this complexity -and the ways in which grades conflate various aspects of achievement -the modest relationships found between approaches and academic outcomes in the present study are to be expected.…”

“…The work on creating more constructively aligned curricula has focused mainly on assessments which would reward a deep approach, whereas organised effort is more likely to be of value in any learning situation. Approaches to learning may mediate some relationships, such as those between self-efficacy and performance (Musso et al, 2013). Intelligence, working memory capacity, attention, anxiety management and cultural capital are also likely to form part of the picture of what influences academic achievement (Musso et al, 2013).…”

Using New Models to Analyze Complex Regularities of the World

“…Since we share the enthusiasm to promote new methods, we also feel that is of utmost importance to perform the analyses with these new methods using extremely high methodological standards. In this respect, we find some of the procedures followed in the recent article by Musso, Kyndt, Cascallar and Dochy (2013) problematic. Before discussing about these issues in detail, we wish to indicate that we agree with Edelsbrunner and Schneider's (2013) previous commentary on this article where they state that there are other data analysis techniques with similar properties than ANNs, but without the drawbacks.…”

“…(Reading the paper sometimes makes you feel that the authors claim otherwise.) While the linear discriminant model DA1 used in the Musso et al (2013) paper has about 2*18 + 18*18 = 360 parameters, the neural network model has 18*15*2 = 540 parameters. The difference in number of parameters is not huge, but all the parameters of the MLP are used for modeling the conditional distribution, while the parameters in the linear discriminant model also take care of modeling the relationships between variables.…”

“…It would have been much more meaningful to compare the MLP to some discriminative classifier such as a logistic regression, or better yet, some sparse version of it such as the group lasso with interaction terms (Meier, van de Geer, & Bühlmann, 2008) that would make interpreting the model and the variable (group) importances much easier. Furthermore, the Musso et al (2013) paper is very unclear about how the variable importances have been calculated. The attempt to follow the references only lead to the statements like "this has been implemented in software X" or to an unpublished technical report by one of the authors.…”