Three essays on dating and marriage using deep learning

Abstract

This dissertation consists of three empirical studies that apply deep learning and Artificial Intelligence (AI) to the field of economics in dating and marriage. Each essay addresses a distinct question at the intersection of romantic relationships and economic outcomes, unified by methodological innovation. Chapter 2 investigates whether people tend to pair with partners who are genetically similar and how such genetic assortative mating influences household outcomes. Using polygenic scores from genomic data, I find strong evidence that couples are genetically similar in traits like subjective well-being, Age at First Birth, and Number of Children Ever Born. Notably, couples that show similar genetic tendency for higher well-being and delayed childbearing accumulate greater wealth, and they are less likely to divorce, while those with similar genetic traits toward larger families have more children but lower wealth. These findings imply that genetic assortative mating may cause economic advantages or disadvantages, and even lead to wealth inequality and demographic trends. Chapter 3 examines the role of physical attractiveness in dating and marital stability by utilizing deep learning-based facial recognition on a unique dataset of celebrity couples. We construct beauty scores and facial similarity measures from facial images of celebrities to analyze beauty premium in relationship dynamics. Physically more attractive individuals, especially women, tend to have more number of partners, however, the beauty effect remains only temporary for long-term relationships. In both marriage and short-term relationships, beauty yields greater opportunities initially, but the effect of education and number of past partners remain longer in relationship stability. For instance, we find that more attractive female celebrities have more partners but often shorter relationship durations, which suggest that appearance may negatively influence the relationship duration when no other forms of compatibility are supported. We also detect patterns in partner selection, such as same-race pairings and preferences shaped by profession, while finding no support for common beliefs like couples looking alike or astrological sign compatibility influencing stability. Chapter 4 explores a novel methodology by using large language models (LLMs) to simulate human behavior in dating markets. We create an artificial sample of silicon agents, AI agents generated by LLMs, calibrated to mimic a real-world speed-dating experiment. These AI-driven agents remarkably demonstrate human-like decision-making. For example, male agents prioritize physical attractiveness in partners, whereas female agents place greater weight on intelligence, mirroring gender differences in real world. These AI agents also reproduce human patterns of same-race preference. Also, incorporating latent personality traits (extroversion and openness) generate match predictions that are in line with real world data, demonstrating strong algorithmic fidelity. This innovative approach shows that generative AI can serve as a cost-effective, controlled, and ethical tool to replace human in social science research. Overall, the three essays illustrate how deep learning tools can be integrated into economic analysis of personal relationships. It is not only about understanding the determinants of partner selection and relationship stability but also highlighting implications for economic policy. The results inform debates on inequality (by showing how “like marries like” can widen wealth gaps), on the limits of observable attributes like beauty in securing lasting partnerships, and on new research frontiers using AI. Overall, this dissertation shows that modern AI techniques can deepen empirical economic research, suggesting ways to better-informed family policies and novel methodologies for complex human decision problems.