Beyond Biosensors: A Supervised Machine Learning Framework For Classification of Synthetically Defined Emotional States Inference Using Behavioural And Interaction-Based Proxies In Digital

Abstract

Affective computing is becoming a field that needs scalable solutions to the inference of classifying synthetically defined emotional states without the necessities of using physiological measurements. The work suggests a classifier-free, hardware-agnostic supervised machine learning system, based on behavioral and interaction based proxies, such as typing dynamics, click latency, response patterns and navigation behavior, to classify emotions in digital UIs. An artificial behavioral dataset has been created and also tested on Logistic Regression, Random Forest, Gradient Boosting, and LSTM. Comparable classification performance was shown in experimental results, whereby, Logistic Regression and LSTM achieved the highest accuracy of 72.5% with LSTM rating the highest ROC-AUC of 0.876. The results suggest that behavioral signals are meaningful alternatives to affective computing with biosensors, and can be used to scale affective-aware digital applications based on adaptable and intelligent machine learning systems.