Contextual Preference Distribution Learning

A team of researchers including Benjamin Hudson, Laurent Charlin, and Emma Frejinger has introduced Contextual Preference Distribution Learning (CPDL), a novel AI pipeline designed to tackle uncertainty in decision-making caused by heterogeneous and context-dependent human preferences. The method addresses a key limitation in existing approaches like inverse optimization and choice modeling, which typically produce simple point estimates of preferences and fail to capture how those preferences shift with context. This makes them ill-suited for risk-averse planning. CPDL's innovation is a two-stage process: first, it trains a predictive model using a bounded-variance score function gradient estimator to map contextual features (like time of day or weather) to a rich, parameterizable distribution of possible human preferences, yielding a maximum likelihood estimate.

In the subsequent optimization phase, the model generates multiple plausible preference scenarios for unseen contexts, which are then used to solve downstream problems formulated as integer or linear programs. This allows for robust, risk-averse decision-making. The researchers validated CPDL in a synthetic ridesharing environment, a complex domain with highly variable user preferences. The results were striking: their approach reduced the average "post-decision surprise"—a measure of how wrong a decision turns out to be—by up to 114 times compared to a theoretically optimal risk-neutral approach with perfect predictions. It also outperformed leading risk-averse baselines by a factor of up to 25x.

This work, accepted at the CPAIOR 2026 conference, represents a significant advance in marrying machine learning with operational research. By moving beyond single-point predictions to modeling full distributions of human behavior, CPDL provides a formal framework for AI systems to make safer, more reliable decisions in real-world applications where people's choices are unpredictable and influenced by numerous factors. The methodology is broadly applicable to any domain where decisions can be framed as an optimization problem with human input, from dynamic pricing and supply chain logistics to healthcare scheduling and energy grid management.