BEAST-GB model combines machine learning and behavioral science to predict people’s decisions

A key objective of behavioral science research is to better understand how people make decisions in situations where outcomes are unknown or uncertain, which entail a certain degree of risk.

The ability to predict people’s choices in these situations could be highly advantageous, as it could help to draft effective initiatives aimed at prompting people to make better decisions for themselves and others in their community.

Researchers at Technion (Israel Institute of Technology) and various institutes in the United States recently developed a new computational model called BEAST-GB, which was found to predict people’s decisions in situations that entail risk and uncertainty.

Their proposed model, outlined in a paper published in Nature Human Behavior, combines advanced machine learning algorithms with behavioral science theory.

“Human-decision research is rich in competing theories, yet none reliably and accurately predicts human choices across contexts,” Ori Plonsky, first author of the paper, told Tech Xplore.

“To see which ideas really work, we organized CPC18, a ‘choice prediction competition’ in which anyone could submit a computational model to predict people’s decisions under risk and uncertainty. We were especially interested in knowing if data-driven machine learning, theory-driven behavioral models, or, as was our guess, a hybrid that embeds behavioral theory inside ML, would excel.”

The new machine learning model developed by Plonsky and his colleagues draws from a behavioral science framework known as BEAST (Best Estimate and Sampling Tools). This is a model based on psychological theories that were previously found to predict people’s decisions with good accuracy.

“BEAST assumes that, in choice under risk and uncertainty, people mix several strategies, such as minimizing the chances of immediate regret or hedging against worst outcomes,” explained Plonsky.

“We translated each strategy into a ‘behavioral feature,’ a concise formula that captures how sensitive a decision-maker should be to that consideration in any given choice task. We then fed these theory-based features, plus purely objective task descriptors, into Extreme Gradient Boosting (a machine learning algorithm known to be highly useful in prediction tournaments)—hence the name BEAST-GB.”

With the enhancements implemented by the researchers, the BEAST-GB model could analyze behavioral data and derive the motives driving decisions, as well as the impact of these motives in different decision-making scenarios.

Notably, BEAST-GB won the CPC18 Choice Prediction Competition in 2018, capturing 93% of predictable variation in the data it was fed, and 96% in follow-up tests utilizing a dataset that was 40 times larger.

“BEAST-GB outperformed dozens of mainstream behavioral models and purely data-driven machine learning,” said Plonsky.

“With just 2% of the training data, it has already beat a deep neural network trained on all the training data. The model even accurately predicts choices people make in new experiments it has never seen, implying it captures general human choice patterns. Finally, we used it to improve and enhance the underlying interpretable behavioral theory, so it enhances our ability to explain, not only predict, human decision making.”

This recent work highlights the promise of machine learning models that also draw from behavioral science for predicting people’s decisions and responses in real-world scenarios. In the future, BEAST-GB and other similar models could guide the design of new large-scale interventions aimed at improving people’s decisions via nudges, incentives or other behavioral science-based strategies.

Plonsky and his colleagues eventually plan to collaborate with policymakers and other parties involved in the design or implementation of behavioral science initiatives. This would allow them to test their model “in the wild,” validating its potential in real-world settings, while also yielding insight that could inform its further advancement.

“Other recent publications have suggested that human decision-making and other behaviors can be very effectively predicted using advanced data-driven machine learning methods like large language models tuned on large behavioral data,” added Plonsky.

“We now plan to continue investigating when and how BEAST-like theory can enhance such data-driven methods in predicting behavior. Specifically, we plan to extend our domain of research by including natural-language decision problems, more aligned with the real world.”

© 2025 Science X Network