Predicting Human Exchange Decision-Making with Theoretically Informed Data and Machine Learning

Abstract

Abstract Artificial agents that can predict human decisions in social exchange contexts can potentially help to facilitate cooperation and promote prosocial behaviours. Modelling human decision-making is difficult in social exchange contexts where multiple contending motives inform decisions in rapidly evolving situations. We propose a mixed Theory and Data-Driven (TD2) model that is comprised of three modules: (1) a clustering algorithm that identifies strategies in interactive social exchange contexts (2) an artificial neural network that classifies an exchange decision into one of the identified strategies based on empirically defined motives and the observable differences during social exchanges, and (3) a hidden Markov model that predicts situated human decisions based on the strategies applied by humans over time. The TD2 decision-making model was trained and tested using 7,840 exchange data from "minimal group" experimental exchange games in which decisions were motivated by group ties, wealth aspiration, and interpersonal ties. The model was able to classify behaviours with 95% accuracy. Reciprocity, fairness and in-group favouritism were predicted, as separate decisions, with accuracies of 81%, 57% and 71% respectively. The performance of the model improved over time. Future work will evaluate the model in a live experiment involving Human-Agent Cooperation (HAC).