Mapping Life Trajectories: A Machine Learning Approach to Clustering Multidimensional Life Courses

• In Event: 20469 - Early Life Course Dynamics

Abstract

Taking a life-course perspective, this paper uses novel machine-learning methods to identify typologies of life trajectories by classifying multi-dimensional life-trajectory data with transformer embeddings. This paper also demonstrates how demographic characteristics predict trajectory typology memberships and how typology memberships are associated with later-life outcomes. Using data from 1970 British Cohort Study, I first establish a two-dimensional life trajectory data combining individuals’ partnership histories and work/education histories. I apply transformer-based embeddings to capture complex temporal dependencies and dynamics from multi-dimensional trajectory data. I then apply K-Means and Gaussian Mixture Models (GMM) clustering methods to classify individuals based on learned representations. The choice of embedding models, clustering methods, and hyperparameters is optimized via key performance metrics. Preliminary results identify four female clusters and five male clusters with distinct work-life trajectories. Qualitatively, the four female clusters can be described as “late bloomers”, “independent & steady”, “marriage & caregiving first”, and “career-family jugglers”. The five male clusters can be described as “early cohabitation with continuous struggles”, “cohabiting careerists”, “late committers, long learners”, “traditional breadwinners”, and “young settlers”. Among them, the “independent & steady” and “marriage & caregiving first” women are associated with worse economic security and physical and mental health outcomes at the age of 46. “Early cohabitation with continuous struggles” men are associated with worse economic security and physical and mental health outcomes at the age of 46. Another key finding is that GMM performs better for women while K-Means performs better for men in clustering life trajectories. It might suggest fundamental gendered differences in life-course patterns—with women having more fluid, overlapping, and nonlinear trajectories than men. This research makes both a methodological contribution as one of the first studies to apply transformer-based embeddings to life trajectory data, and a substantive contribution by advancing the understanding of the gendered typologies of behavioral patterns and life trajectories.

Author

• Zerui Tian, University of Oxford