Research map / Learning theory and algorithms / Fairness
Preventing Eviction-Caused Homelessness through ML-Informed Distribution of Rental Assistance
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds ML models on county and state administrative data plus eviction court records to proactively rank tenants facing eviction by 12-month homelessness risk, validated retrospectively, in shadow mode, and with a planned RCT.
Problem
Rental assistance programs allocate scarce funds reactively (first-come-first-served, call-based), so they miss tenants at high risk of future homelessness who never apply. The paper asks whether predictive models can prioritize tenants facing eviction by need while staying fair across race and gend
Approach
Score all tenants with an active eviction filing (not only those who call for help) with supervised classifiers trained on linked administrative and court data using temporal validation, select the top 100 per month, and evaluate efficiency (precision@100), effectiveness (recall@100), equity (TPR ratios), and the share of overlooked people found.
Key results
- Random Forest and Logistic Regression reach average precision@100 of 0.20, about 20% above the best heuristic (previous homelessness, 0.15) and about 10x the random baseline (0.02).
- RF identifies on average 28% of the missed group (people who became homeless without applying or receiving aid), versus 23% for the previous-homelessness heuristic and 4% for random selection.
- In shadow mode deployment, 22 of 100 predicted individuals used homelessness services within 12 months (precision about 0.20-0.22), and 17 were missed by the current system.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Prioritize the top 100 eviction-filing tenants using RF or LR instead of the previous-homelessness heuristic | Rental assistance prioritization in Allegheny County · Allegheny County temporal validation | precision@100: 0.15 -> 0.20 | Helped |
| Prioritize the top 100 tenants using RF rather than random selection | Rental assistance prioritization in Allegheny County · Allegheny County temporal validation | precision@100: 0.02 -> 0.20 | Helped |
| Use RF to identify tenants overlooked by the current process | Rental assistance prioritization in Allegheny County · Missed group | percentage of missed group found: 23% for B1; 28% for RF | Helped |
| Use RF to identify tenants overlooked by the current process | Rental assistance prioritization in Allegheny County · Shadow mode deployment | overlooked individuals becoming homeless identified: 17 | Helped |
| Use RF for tenants with previous homelessness rather than first-time homelessness | Random forest (RF) · Allegheny County temporal validation | recall: 55% ... but only 4% | Mixed |
| Prioritize using RF or LR rather than current process | Rental assistance prioritization in Allegheny County · Race subgroup | Black-to-white true positive rate ratio: 1.34 and 1.14 | Helped |
| Prioritize using RF or LR | Rental assistance prioritization in Allegheny County · Gender subgroup | female-to-male true positive rate ratio: 0.9 and 0.87 | Hurt |
Surprising findings
- First-time homelessness is much harder to predict: RF average recall is 55% for people with prior homelessness but only 4% for first-time cases.
- Over 70% of applicants and over 80% of people receiving current rental assistance had no prior homelessness, suggesting the existing process mostly serves people without a homelessness history.
Limitations
- Label bias: homelessness service use is a proxy, so individuals who are unsheltered or couch-surfing are missed by the outcome definition.
- Historical data is conditioned on past allocations, so no counterfactual outcomes; the approach is an augmentation of the current process and the RCT is still planned rather than completed.
- Eligibility requirements were not modeled; women are slightly under-served by RF and LR (TPR ratio about 0.9 and 0.87).
Experimental setup
Datasets: Allegheny County ACDHS administrative data (demographics, evictions, program interactions, public housing, mental and be Allegheny County court eviction records
Benchmarks: Allegheny County ACDHS historical temporal validation splits (Jan 2012 to Aug 2023) Shadow mode deployment Sept 2022 to Aug 2023
Metrics: precision@100 recall@100 fraction of missed group found true positive rate ratio (equality of opportunity) by race a
Baselines: Previous homelessness heuristic (B1) Baserate random selection (B2) Earliest OFP heuristic (B3)
Builds on: temporal validation equality of opportunity
Closest papers by meaning
- Evaluating Index-based Treatment Allocation in Underresourced Communities (AAAI 2025 · similarity 0.81)
- Predictive Modeling of Homeless Service Assignment: A Representation Learning Approach (AAAI 2025 · similarity 0.81)
- Active Geospatial Search for Efficient Tenant Eviction Outreach (AAAI 2025 · similarity 0.80)
- Allocation Requires Prediction Only if Inequality Is Low (ICML 2024 · similarity 0.79)
- Discretionary Trees: Understanding Street-Level Bureaucracy via Machine Learning (AAAI 2024 · similarity 0.78)
- The Value of Prediction in Identifying the Worst-Off (ICML 2025 · similarity 0.76)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.