On the paradigms of learning analytics: Machine learning meets epistemology
Shayan Doroudi
Abstract
Baker et al. (2021) recently proposed using a philosophical framework to classify learning analytics research in terms of four paradigms. Here I build on their theme of reflecting on philosophical differences in different approaches to learning analytics. I first present two limitations of their classification, which raise questions for how to best classify different approaches in learning analytics. In an attempt to resolve these questions, I draw upon the bias-variance tradeoff from machine learning and show how different learning analytics approaches can be viewed in terms of their positions on the tradeoff. However, I claim that this is not enough, as we must also be attuned to the underlying epistemologies behind different approaches. I claim a constructivist epistemology for learning analytics has been missing, which could, in part, explain Baker et al.'s (2021) observation that constructivist work has been relatively absent in established learning analytics research communities. Drawing on prior work from different fields, I present a sketch of what a constructivist data science philosophy might look like and how it could help advance learning analytics. Sitting at the nexus of the learning sciences and machine learning, the field of learning analytics is in a unique position to theorize about philosophy and epistemology; this paper encourages us to pursue more work in such a direction.