A Human-Centered Review of Algorithms in Homelessness Research

Title of the Paper

A Human-Centered Review of Algorithms in Homelessness Research

Paper Information

• Subject Area: Humanitarian crises, homelessness, and algorithm applications
• Keywords: Algorithmic decision-making, fairness, resource allocation, homelessness risk assessment, algorithm design, human-centered computing, social impact

Research Background and Issues

• Key Issues or Challenges Identified by the Authors: Homelessness affects approximately 1.6 billion people globally and continues to worsen with urban population growth and increasing pressure on social services. Government agencies have begun adopting data-driven models to predict individuals' risk of homelessness and allocate limited resources. However, many decision-making algorithms lack fairness, are prone to bias against homeless individuals, and overlook the complex ecological factors in real-world environments.
• Importance of the Research: Homelessness profoundly impacts the quality of life for impoverished populations and intersects with areas such as health, child welfare, and criminal justice. Properly designed decision-making algorithms can more efficiently allocate resources and improve service quality, but poorly designed algorithms may lead to unfair outcomes due to bias or design flaws.
• Motivation and Related Work:
  • HCI researchers have conducted in-depth studies on algorithmic decision-making in high-risk environments, such as education, child welfare, and online behavior governance.
  • Existing algorithmic studies often fail to capture behavioral interactions and detailed case-specific information, resulting in unfair resource allocation and biased generalized data.
  • This study adopts a "human-centered design" framework to critically examine the design structures and social impacts of existing algorithms.

Solution

• Proposed Methods or Solutions:
  • Systematic literature review: The authors reviewed 57 papers published between 1998 and 2023, analyzing decision-making algorithms in the homelessness domain.
  • Using the human-centered algorithm design framework (HCAD) to evaluate the computational methods, predictive variables, and target outcomes of algorithms, as well as their ecological validity in real-world contexts.
• Innovations:
  • The human-centered evaluation perspective emphasizes that algorithm design is not solely about technical performance but also involves social fairness, judicial significance, and human impact.
  • Uniquely reveals the tension between fairness and interpretability in algorithms, while highlighting that resource allocation models often rely on simulated data, which cannot be effectively applied under real-world conditions.
• Implementation Steps and Key Techniques:
  • Categorizing computational techniques, including generalized linear models (GLM), machine learning (ML), deep learning (DL), and optimization techniques.
  • Structuring predictive variable analysis using coding methods (e.g., demographic data, service needs, health risks).
  • Cross-analyzing the relationships between target variables and computational methods to uncover design biases and limitations in the models.

Research Findings

• Specific Findings:
  • Most algorithms focus only on "who" the homeless individuals are (e.g., demographic data and health conditions) while neglecting systemic issues such as "what services are needed" and "whether services are available."
  • Identified problems in homelessness algorithms, including unrealistic simulated data, neglect of structural social factors, and oversimplified intensity metrics.
  • Only a small portion of studies addressed fairness and bias issues (9 papers, accounting for 15.7% of the total).
• Advantages:
  • Provides more systematic insights into algorithm design, promoting research on algorithm fairness and real-world adaptability.
  • Proposes a human-centered reflective framework to help academia focus on the needs of low-income populations.
• Experimental or Evaluation Results:
  • Predictive models predominantly focus on risk assessment (53.4%), offering strong explanatory power but low ecological validity.
  • Resource allocation models (25.9%) employ complex optimization techniques but face challenges in reliable real-world deployment.
• Limitations and Future Directions:
  • Limitations include: research concentrated in North America, excluding global variations in homelessness policies and environments; data primarily sourced from administrative records rather than grassroots cases.
  • Suggestions for future research:
    • Encourage cross-disciplinary collaboration and human-centered design practices.
    • Develop resource allocation algorithms using real-world data instead of simulated data.
    • Establish guidelines for selecting predictive variables to ensure algorithms are fair and client-needs-oriented.

Additional Points

• Summary Design Guidelines:
  • Emphasize the involvement of affected individuals to define algorithm feasibility.
  • Evaluate data sources and predictive criteria from a theoretical perspective.
  • Adopt human-centered narrative strategies to redefine resource allocation systems and foster interdisciplinary dialogue and collaborative design.
  • Advocate for a shift toward AI design paradigms that prioritize data quality.