Predictive Modeling of Homeless Service Assignment: A Representation Learning Approach

Khandker Sadia Rahman, Charalampos ChelmisAAAI 2025 · DOI

TL;DR

This paper introduces REPLETE, a novel representation learning framework for predicting homeless service assignments. By capturing temporal, functional, and individual relationships within historical data, REPLETE significantly outperforms existing methods, offering a more accurate and potentially fairer approach to resource allocation in homelessness services.

What Problem Does This Paper Address?

Homelessness is a growing societal challenge, and effectively assigning individuals to the right services is crucial for successful reintegration. Existing machine learning approaches often struggle with the categorical nature of administrative data used in this domain. This categorical data, when one-hot encoded, leads to high dimensionality and fails to capture the rich, underlying relationships between different features and services. This gap hinders the development of accurate predictive models that can truly understand an individual's needs and trajectory within the homeless system. The paper specifically targets individuals experiencing chronic homelessness, aiming to improve the prediction of their next service assignment.

Methods

The core of REPLETE is its representation learning framework. Instead of relying on traditional one-hot encoding, REPLETE learns latent representations for services, features, and individuals. This is achieved through a novel optimization function that incorporates three key contexts:

  • Temporal Context: Captures how the timing of service assignments influences their meaning. It uses Non-negative Matrix Factorization (NMF) to learn representations for services and time units based on their co-occurrence.
  • Functional Context: Models the idea that subsequent services are assigned to address unmet needs from preceding ones. It learns relationships between service pairs using symmetric NMF based on transitional probabilities.
  • Individual Context: Accounts for the fact that individuals with similar characteristics and histories are likely to benefit from similar services. It learns representations for individuals and clusters them based on their features and past service assignments.

These learned representations are then used to derive new features, which are fed into a feed-forward neural network (FFNN) for the final service assignment prediction. The optimization process uses the Alternating Direction Method of Multipliers (ADMM) to handle the non-convexity of the objective function.

Datasets

The study utilizes an anonymized dataset from CARES of NY, containing 18,817 records from 6,011 chronically homeless individuals in the Capital Region of New York State, spanning the period from 2012 to 2018. The dataset includes information on 9 different homeless services. The data was split into a 70-30 train-test set.

Key Insights and Ideas

  • Beyond One-Hot Encoding: The paper strongly advocates for moving beyond one-hot encoding for categorical features in this domain. Learned representations are shown to capture richer relationships, leading to better performance.
  • Holistic Context Matters: The success of REPLETE highlights the importance of considering multiple facets of an individual's journey: when they received services (temporal), why they received them in sequence (functional), and who they are similar to (individual).
  • Derived Features Enhance Separability: Visualizations (t-SNE) demonstrate that the features derived from learned representations offer significantly better separability for distinguishing between service labels compared to one-hot encoded features.
  • Robustness to History Length: The model's performance remains consistent across different history lengths, suggesting it can effectively learn from varying amounts of past data.

Key Results

REPLETE significantly outperforms several state-of-the-art baselines, including TRACE and PREVISE, as well as standard machine learning models like RF, LR, Transformer, FFNN, RNN, and LSTM.

  • Accuracy: REPLETE achieved an accuracy of 0.832, outperforming the best baseline (LSTM at 0.802).
  • F₁ Score: REPLETE achieved an F₁ score of 0.567, also outperforming the best baseline (LSTM at 0.513).
  • Statistical Significance: Statistical tests confirm that REPLETE's performance is significantly better than PREVISE and LSTM.
  • Bias Evaluation: The approach demonstrates a lower level of bias compared to PREVISE, particularly for high-frequency services, indicating a step towards more equitable service assignment.

Limitations and Future Work

  • Geographical Limitation: The dataset is limited to the Capital Region of New York State, which might affect generalizability.
  • Data Limitations: The dataset does not record service availability or capacity, which are crucial for real-world deployment.
  • Future Work: The authors plan to incorporate fairness constraints directly into the objective function to develop truly "unbiased" and "fair" predictive models. They also suggest exploring additional dimensions for service assignment and investigating the impact of service availability.

Summary updated

BibTeX
@inproceedings{rahman2025predictive,
  author = {Khandker Sadia Rahman and Charalampos Chelmis},
  title = {{Predictive Modeling of Homeless Service Assignment: A Representation Learning Approach}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i27.35052},
}
Figure