MIT Researchers Upgrade Preference Prediction Models with 'Power of Three' Concept
Researchers at MIT have introduced a significant enhancement to random utility models (RUMs), a nearly century-old framework used to predict human preferences. This upgrade addresses a long-standing limitation in how these models account for correlations between choices, moving beyond traditional pairwise comparisons. The findings could improve predictions in various real-world scenarios, from urban planning to consumer behavior.
MIT researchers have developed an upgrade to random utility models (RUMs), a mathematical framework for describing and predicting human preferences. The original concept was introduced by American psychologist L. L. Thurstone in his 1927 paper, “A law of comparative judgment.” Thurstone's work, a cornerstone of psychometrics, proposed that individuals select options based on perceived highest value, even without assigning a precise numerical score.
RUMs are widely applied in government and industry to forecast human behavior in hypothetical situations. Examples include predicting transportation choices after a road closure or optimizing the allocation of public funds for the common good. Despite their widespread use and decades of refinement, the models have historically had limitations.
A key deficiency identified by the MIT team lies in the common practice of estimating RUMs primarily from pairwise comparisons. In this traditional approach, individuals choose between two options (e.g., A or B). While this method is cognitively straightforward for participants, it makes it challenging to identify correlations between numerous choices, as the standard RUM application often assumes the utilities derived from options are independent.
The new research, presented in April at the International Conference on Learning Representations in Rio de Janeiro, Brazil, uncovers fundamental facts that suggest more can be derived from these models. The paper, co-authored by Yeshwanth Cherapanamjeri, Gabriele Farina, Constantinos Daskalakis, and Sobhan Mohammadpour, indicates that moving beyond simple pairwise comparisons is crucial for a more accurate understanding of preferences. The researchers' work aims to account for the interconnectedness of choices, which the source suggests is related to considering "the power of three."
According to MIT News AI, this research offers a major upgrade to the established framework, potentially leading to more sophisticated and reliable predictions across various fields.
