Predicting Chronic Homelessness at an Emergency Homeless Shelter in Calgary using Neural Network models and Time-Stamped data records

Abstract

Chronic homeless shelter users are some of the most vulnerable individuals of the homeless population. Identifying chronic homeless shelter users is a challenge as most of the individuals accessing emergency shelters are non-chronic and existing definitions used for identifying shelter users take too long to triage chronic clients. A machine learning solution is proposed in this thesis to accurately predict chronic homeless clients of an emergency shelter using time-stamped client access data records in shorter time frames. A data-driven pre-processing approach is used to develop machine learning models with high precision and high sensitivity for the variable size multivariate time-stamped dataset. Several machine learning models and feature sets are explored for different shelter interaction timelines and compared using statistical classification metrics. The models are also evaluated using shelter access statistics to provide insights from a social impact perspective and select models that reduce the risk of incorrect predictions.