A Machine Learning Model To Predict Human Life Outcomes

Abstract

Can we predict human life? This study presents a machine learning model for predicting six outcomes of children belonging to disadvantaged, or fragile families. It is motivated by recent work developing models with enhanced predictive power as well as previous attempts from researchers to model fragile families. The model is based on data from the Fragile Families & Child and Wellbeing Study, a data set of over 4,000 families from cities across the U.S. that aims to determine how over 12,000 variables about the children and their parents, schools, and overall environments collected over 15 years can impact life outcomes. The outcomes are child grit, GPA, family eviction, job layoff, material hardship and job training. My approach relies on existing data science techniques -- imputation of missing data, elimination of low variance features, prediction via ridge and logistic regression models -- and proposes and evaluates a new model -- a neural network pre-trained on synthetic data created from Gaussian noise. This approach creates a flexible model with reasonable bias for human behavior and enhanced predictive power. Results revealed that the proposed model outperformed benchmark models for four of the six outcomes by up to 28% and outperformed existing researchers’ models for grit and GPA by up to 40%.