Machine Learning for Discovery: Uncovering Novel Preference Configurations in Online Dating

• In Event: Middle-Range Theorizing and Its Empirical Tests

Abstract

Scholars of assortative mating have long treated romantic pairings as the joint product of preferences and constraints. Online dating data have provided new empirical leverage by recording unreciprocated overtures, making it possible to observe partner selection closer to the point of choice. Despite this advance, three implicit assumptions embedded in the standard modeling approach may be limiting what we can learn from the data that platforms make available.
First, existing studies focus on a small set of theoretically motivated profile variables—typically race, gender, age, education, and income—leaving unexamined the rich attitudinal and lifestyle data that platforms collect. Second, main-effects models assume that each attribute contributes independently to a target's attractiveness, ruling out the possibility that daters evaluate trait bundles conjunctively. Third, the functional relationship between attributes and messaging decisions may be more complex than linear models can capture, involving similarity-dependence, threshold effects, or asymmetric penalties.
We propose a graduated modeling framework that progressively relaxes these assumptions across tiers of increasing flexibility. Tier 1 replicates standard logistic regression on conventional profile attributes. Tier 2 employs Explainable Boosting Machines and other interpretable machine learning methods that learn nonlinear shape functions and automatically detect pairwise interactions while preserving transparency. Tier 3 fits dual-encoder neural networks that can capture arbitrary functions of chooser and target attributes, establishing a ceiling on predictive performance. Each tier is tested with and without an expanded feature set including attitudinal match questions and messaging style measures, drawn from a major U.S. online dating platform.
If more flexible models outperform the standard approach, the interpretable models will reveal what was missed: which trait combinations, attitudinal dimensions, and nonlinear patterns matter for partner choice but fall outside the purview of existing theories and methods.

Authors

• Galen Hall, University of Michigan-Ann Arbor

• Jett Rosen, University of Michigan

• Elizabeth Bruch, University of Michigan-Ann Arbor