Population validity for educational data mining models: A case study in affect detection

Abstract

Information and communication technology (ICT)‐enhanced research methods such as educational data mining (EDM) have allowed researchers to effectively model a broad range of constructs pertaining to the student, moving from traditional assessments of knowledge to assessment of engagement, meta‐cognition, strategy and affect. The automated detection of these constructs allows EDM researchers to develop intervention strategies that can be implemented either by the software or the teacher. It also allows for secondary analyses of the construct, where the detectors are applied to a data set that is much larger than one that could be analyzed by more traditional methods. However, in many cases, the data used to develop EDM models are collected from students who may not be representative of the broader populations who are likely to use ICT. In order to use EDM models (automated detectors) with new populations, their generalizability must be verified. In this study, we examine whether detectors of affect remain valid when applied to new populations. Models of four educationally relevant affective states were constructed based on data from urban, suburban and rural students using ASSISTments software for middle school mathematics in the Northeastern United States. We found that affect detectors trained on a population drawn primarily from one demographic grouping do not generalize to populations drawn primarily from the other demographic groupings, even though those populations might be considered part of the same national or regional culture. Models constructed using data from all three subpopulations are more applicable to students in those populations than those trained on a single group, but still do not achieve ideal population validity—the ability to generalize across all subgroups. In particular, models generalize better across urban and suburban students than rural students. These findings have important implications for data collection efforts, validation techniques, and the design of interventions that are intended to be applied at scale.

Advanced, Analytic, Automated (AAA) Measurement of Engagement During Learning

“…We have selected 15 studies to emphasize key dimensions of the measurement approach, including sensor-free vs. sensor-based measurement, annotations by the self vs. external observers, unimodal vs. multimodal sensing, lab vs. classroom research, learning activities with varying levels of interactivity, and different validation methods. We prioritized studies that can be considered as pioneering in the field, such as the first study showcasing multimodal engagement measurement in real-world classrooms (Arroyo et al, 2009), the first study emphasizing generalizability beyond the individual (Ocumpaugh, Baker, Gowda, Heffernan, & Heffernan, 2014), or the first person-independent automated measure of mind wandering (Author, year). We acknowledge that our choice of case studies is both subjective and incomplete, but our goal is to provide an overview of a promising new approach rather than review a well-established paradigm.…”

A Review and Meta-Analysis of Multimodal Affect Detection Systems

“…Furthermore, no notable efforts were made to assess generalizability across tasks, situational contexts, datasets, and cultures. This is particularly important since emerging data suggests that models trained on individuals from one demographic do not necessarily generalize to another [Ocumpaugh et al 2014].…”

“…Fifth, the standard procedure of collecting labeled data to train supervised classifiers is inherently limited due to the manual affect annotation process, thereby resulting in small datasets (in terms of number of unique individuals). It is unlikely that this approach will lead to models that generalize at large [Ocumpaugh et al 2014]; hence, it might be useful to consider semisupervised learning approaches that only require a small subset of the training data to be annotated. Furthermore, crowd-sourcing techniques might be useful alternatives to current cumbersome annotation methods that simply do not scale to larger datasets [McDuff et al 2012].…”

Item Ordering Effects with Qualitative Explanations using Online Adaptive Tutoring Data

“…The data set analyzed in this paper comes from the publicly available ASSISTments 2012-2013 data set which has been used for numerous other research projects [5,6]. In this paper, we focus on the Skill Builder sequences (SBs) used in ASSISTments, where a sequence contains a set of problems assumed to address one concept which we refer to as a knowledge component (KC).…”