2015_06_EDM - Sensor-Free or Sensor-Full: A Comparison of Data Modalities in Multi-Channel Affect Detection
Computational models that automatically detect learners’ affective states are powerful tools for investigating the interplay of affect and learning. Over the past decade, affect detectors—which recognize learners’ affective states at run-time using behavior logs and sensor data—have advanced substantially across a range of K-12 and postsecondary education settings. Machine learning-based affect detectors can be developed to utilize several types of data, including software logs, video/audio recordings, tutorial dialogues, and physical sensors. However, there has been limited research on how different data modalities combine and complement one another, particularly across different contexts, domains, and populations. In this paper, we describe work using the Generalized Intelligent Framework for Tutoring (GIFT) to build multi-channel affect detection models for a serious game on tactical combat casualty care. We compare the creation and predictive performance of models developed for two different data modalities: 1) software logs of learner interactions with the serious game, and 2) posture data from a Microsoft Kinect sensor. We find that interaction-based detectors outperform posture-based detectors for our population, but show high variability in predictive performance across different affect. Notably, our posture-based detectors largely utilize predictor features drawn from the research literature, but do not replicate prior findings that these features lead to accurate detectors of learner affect.