Demographical Analysis of US Homelessness and Predictions Based on Long Short-Term Memory Model

Abstract

Homelessness is fast becoming one of the most troublesome issues in United States (U.S.) in terms of the uprising total number of homeless population and potential social instability that it may cause. In previous studies, efforts mainly focused on the characteristics of specific subgroups out of the overall homeless. In this paper, more attention is paid to determine major factors related to homelessness in U.S. using correlation and a model is proposed to predict the tendency of change in the total number for homeless population for each state in U.S. by applying advanced machine learning method of Long Short-Term Memory (LSTM). The study suggests that each state would be primarily troubled by homeless subgroups with different characteristics. Its results also imply that crime rates may relate to with the trend of change in total homeless population and the dominant concerns regarding to the categories of crime differ on a state basis. The study further suggests that after optimization, the LSTM model with interpolation or with multi-dimensions are the best approaches to predict the overall change in the number of homeless populations for each state in the U.S.