Network analysis improves interpretation of affective physiological data

Abstract

Understanding how human physiological responses to stimuli vary across individuals is critical for the fields of Affective Psychophysiology and Affective Computing. We approach this problem via network analysis. By analysing individuals' galvanic skin responses (GSRs) to a set of emotionally charged images, we model each image as a network, in which nodes are individuals and two individuals are linked if their GSRs to the given image are statistically similar. In this context, we evaluate several network inference strategies. Then, we group (or cluster) images with similar network topologies, while evaluating a number of clustering choices. We compare the resulting network-based partitions against the known arousal/valence-based ‘ground truth’ partition of the image set (which is likely noisy). While our network-based image partitions are statistically significantly similar to the ‘ground truth’ partition (meaning that network analysis correctly captures the underlying signal in the data), the network-based partitions yield insights that go beyond the ‘ground truth’ partition with respect to an independent criterion, namely in terms of latent semantic analysis (meaning that our partitions are more semantically meaningful than the ‘ground truth’ partition). Non-network-based approaches do not yield any such insights. Thus, network analysis of affective physiological data appears to improve interpretation of the data. We conclude by analysing in-depth a representative network-based image partition and discussing practical applications of the corresponding results.