Predicting a person’s online shopping behavior has given rise to an industry where every click counts.
Items in a shopping cart, for example, can form the basis of a complex statistical profile developed on past behavior. As the retail environment has become more complex, so have the computational models that suggest future options in a person’s shopping algorithm.
Gunnar Epping, a cognitive science PhD researcher at Indiana University, uses nodes on IU’s research supercomputer, Carbonate, to compare and combine competing computational models of decision-making to analyze cognitive behavioral patterns like choice and response time.
A small group of online survey participants was asked to compare and select from examples of gift cards with the following criteria: dollar value, quality of restaurant, average price per meal, and distance from the person. “These four choices were selected for their high degree of variability in a participant’s subjective state when choosing a place to get dinner,” Epping said.
Gunnar Epping, PhD researcher, Indiana University Cognitive Science program
Carbonate has further given me the opportunity to improve these models as I learn more about them over the course of my academic training.
These constraints aimed to show consistencies across seemingly subjective decisions by fitting three quantitative models to participants’ choices and response times. One model is a Markov random walk, and is representative of standard sequential sampling models developed using classical probability theory. Another model is a quantum random walk, which is analogous to the Markov one except that it is developed using quantum probability theory. Finally, the open system random walk model, the main focus of this work, is a hybrid of the other two models.
“On average, there was not enough evidence to reject the Markov model in favor of the open system model, but there was enough to reject the quantum model in favor of the open system,” said Epping, who co-authored the paper with his advisor, Jerome R. Busemeyer of IU, as well as researchers at the University of Florida and Kansas University.
To describe the likelihood of a person responding at any given time, each model predicts the probability of responding for every moment in time. The curves in the figure describe how the probability of responding changes across time as predicted by each model.
The iterative nature of choice modeling research means these calculations would be difficult to verify without the high-powered computing resources at IU.
“The more traditional Markov model has been adopted by neuroscientists to evaluate which regions of the brain contribute to decision-making and has been employed in other areas of psychology, such as developmental psychology, to analyze the effects of aging on decision-making,” Epping said. “The open system model has the potential to replace the Markov model to become the primary computational model used to help understand human decision-making in a variety of fields.”
Epping shared the group’s findings at the Cognitive Science Conference in summer 2022.
“Without Carbonate, I would not have been able to carry out my research using the state-of-the-art cognitive models due to their computational complexity,” said Epping. “Carbonate has further given me the opportunity to improve these models as I learn more about them over the course of my academic training.”
