preferencebased

Preferencebased, often written as preference-based, refers to approaches that use preferences rather than explicit numerical utilities to guide decisions, learn models, or optimize outcomes. It is used across decision theory, machine learning, economics, and human–computer interaction.

In decision making and optimization, systems collect ordinal judgments—rankings or pairwise comparisons—about options. From these, they construct an order or a latent utility with ordinal information and select the top alternatives, or optimize with respect to a latent preference model. Methods include pairwise comparison, rank aggregation, and Bayesian preference elicitation, as well as preference-based reinforcement learning.

In machine learning, preference-based learning learns from user preferences to rank items or predict choices. Tasks include learning to rank, ranking-based recommender systems, and Bayesian optimization with preferences. Active learning can query users for the most informative comparisons to reduce sample complexity.

Advantages of preferencebased methods include natural handling of subjective criteria and insensitivity to scale, with the ability to integrate multiple criteria via induced preferences. Limitations include data intensity, susceptibility to noisy or inconsistent preferences, issues of collinearity, and the need for careful elicitation design. Evaluation often relies on rank-based metrics rather than absolute error, which can complicate interpretation.

Applications span personalized recommendations, search ranking, product configuration, and policy selection in multi-criteria settings, as well as interactive systems that adapt to user choices.