Predictive Modeling of Homeless Service Assignment: A Representation Learning Approach

TL;DR

This work asserts that deriving latent representations of such features is crucial in algorithmically enhancing the existing assignment decision-making process and leveraging underlying relationships between instances is crucial in algorithmically enhancing the existing assignment decision-making process.

Abstract

In recent years, there has been growing interest in leveraging machine learning for homeless service assignment. However, the categorical nature of administrative data recorded for homeless individuals hinders the development of accurate machine learning methods for this task. This work asserts that deriving latent representations of such features, while at the same time leveraging underlying relationships between instances is crucial in algorithmically enhancing the existing assignment decision-making process. Our proposed approach learns temporal and functional relationships between services from historical data, as well as unobserved but relevant relationships between individuals to generate features that significantly improve the prediction of the next service assignment compared to the state-of-the-art.

Figures (6)

• Figure 1: Overview of REPLETE. Representation learning framework first learns representations $\mathbf{A,C,V,S,R_p,R_s}$. These are used to derive features, which are subsequently input into FFNN for service assignment prediction.
• Figure 2: Scatterplot of 2--dimensional t--SNE embedding for $(a)$ one--hot encoded features and $(b)$ derived features from learned representations.
• Figure 3: Demographic parity (solid line) and equal opportunity (dotted line) of sensitive attributes (a) gender and (b) ethnicity for each service using REPLETE (blue) and PREVISE (orange). Plots for the remaining attributes are included in the Appendix section. Numerals are used in lieu of actual service names 2020hmisstandards.
• Figure 4: Demographic parity (solid line) and equal opportunity (dotted line) of sensitive attributes (a) gender, (b) ethnicity, (c) American Indian or Alaskan Native, (d) Black African American, (e) Asian, (f) Native Hawaiian or Other Pacific Islander, and (g) White for each service using REPLETE (blue) and PREVISE (orange). Numerals are used in lieu of actual service names 2020hmisstandards. Values within the range of $0.8$ to $1.2$ are consider better, with the optimal value being $1$pessach2022review.
• Figure 5: Input for (a) TRACE and PREVISE, and (b) RF, LR, FFNN$_1$. As defined in Section \ref{['sec:problem']}, $\mathcal{U}$, $\mathcal{F}$, and $p_{t_i}$ is the set of chronically homeless individuals, set of features, and service assigned at time $t_i$, respectively. Additionally, $\mathbf{F}_i$ denotes the one--hot encoded vector of feature $\mathcal{F}_i$.

Theorems & Definitions (1)

• Lemma 1