An Evaluation of Tabular Neural Network Approaches for Human Affective State Classification from Physiological Signals
Abstract
Despite advances in machine learning approaches in application domains such as computer vision, natural language processing, and speech recognition, to name a few, there is some uncertainty regarding the viability of neural network approaches applied to tabular data the type of data stored in a (row, column) table format. The standard approach for building machine learning classification methods on tabular data is in the form of decision trees (DTs). However, a recent study comparing different neural network architectures to DT-based methods found that neural network approaches, in many cases, outperformed DT-based methods when evaluated on 40 different tabular data sets, suggesting there are now viable neural network approaches for tabular data. In this report, we described our initial results evaluating tabular neural network approaches for human affective state classification. We used AutoGluon-Tabular, an open-source machine learning framework built for tabular data to develop models to predict high/low arousal or positive/negative valence using features extracted from electrocardiograms and galvanic skin responses obtained from three publicly available data sets. Our classification results were only marginally above random chance, suggesting that subject-independent cross-data set human affective state classification with peripheral physiological signals remains a significant challenge.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 01, 2022
- Accession Number
- AD1182171
Entities
People
- David Chhan
- Vernon J. Lawhern
Organizations
- United States Army Research Laboratory