# Using Sequences of Life-events to Predict Human Lives

**Authors:** Germans Savcisens, Tina Eliassi-Rad, Lars Kai Hansen, Laust Mortensen, Lau Lilleholt, Anna Rogers, Ingo Zettler, Sune Lehmann  
**arXiv:** [2306.03009](https://arxiv.org/abs/2306.03009)  
**Date:** 2023-06-05  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2306.03009v1/

## Key points

- Machine learning has revolutionized text analysis through flexible computational models
- Transformer-based life2vec model creates embeddings of life-events in a single vector space
- The model allows for accurate predictions ranging from early mortality to personality nuances
- Extensive significance testing is performed to validate sensitivity scores of the model
- Life2vec builds complex contextual representations of health, occupation, geography, and wealth
- The model outperforms state-of-the-art baselines in predicting outcomes such as death and personality nuances
- Different aspects of life trajectories are considered based on the task at hand for predictions
- The model effectively handles complexities like missing labels and imbalanced sample sizes
- Meaningful relationships between tokens in the vocabulary are captured in embedding spaces
- Insights drawn from summaries can generate new hypotheses and serve as a starting point for causal studies

## Summary

In the age of algorithm-driven prediction, machine learning has revolutionized the analysis of text through flexible computational models. Transformer-based architectures have shown promise in making sense of various multi-variate sequences, including human lives. By leveraging comprehensive registry data for over six million individuals across decades, researchers have developed a model called life2vec that creates embeddings of life-events in a single vector space. This embedding space is robust and highly structured, allowing for accurate predictions ranging from early mortality to personality nuances. To validate the sensitivity scores of the model, extensive significance testing is performed. The model's attention to individual sequences confirms the findings discussed above and enhances interpretability. Drawing on progress from natural language processing and utilizing a massive dataset capturing events in people's lives, life2vec builds complex contextual representations of health, occupation, geography, and wealth. The transformer-based life2vec model adapts to different settings and outperforms state-of-the-art baselines in predicting outcomes such as death and personality nuances. By analyzing how the model makes these predictions, it is evident that different aspects of life trajectories are considered based on the task at hand. The model handles complexities like missing labels and imbalanced sample sizes effectively. Studying the embedding spaces reveals meaningful relationships between tokens in the vocabulary and captures ordinal features like time and income. The person embedding space condenses signals from entire life sequences into a single vector conditioned on specific prediction tasks. Insights drawn from these summaries can generate new hypotheses and serve as a starting point for causal studies. While socio-demographic factors play a significant role in human lives, predictions at an individual level have been challenging. However, with detailed data provided by models like life2vec, more accurate predictions of individual-level outcomes become possible. This advancement opens up new possibilities for understanding human behavior and improving personalized interventions based on predictive analytics.