Examining the relationship between student performance and video interactions

TL;DR

This study investigated whether video interaction data could predict student performance on lab assessments, finding that simple logistic regression models with clickstream features were ineffective, highlighting the complexity of predicting academic success.

Contribution

The paper provides an empirical evaluation of using clickstream features from instructional videos to predict student performance, revealing limitations of basic models in this context.

Findings

Logistic regression models could not predict student performance

Adding contextual features did not improve prediction accuracy

Highlights the complexity of linking video interactions to assessment outcomes

Abstract

In this work, we attempted to predict student performance on a suite of laboratory assessments using students' interactions with associated instructional videos. The students' performance is measured by a graded presentation for each of four laboratory presentations in an introductory mechanics course. Each lab assessment was associated with between one and three videos of instructional content. Using video clickstream data, we define summary features (number of pauses, seeks) and contextual information (fraction of time played, in-semester order). These features serve as inputs to a logistic regression (LR) model that aims to predict student performance on the laboratory assessments. Our findings show that LR models are unable to predict student performance. Adding contextual information did not change the model performance. We compare our findings to findings from other studies and explore caveats to the null-result such as representation of the features, the possibility of underfitting, and the complexity of the assessment.