Research map / Reinforcement learning / Visual and representation RL
Active Geospatial Search for Efficient Tenant Eviction Outreach
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.
Proposes active geospatial search (AGS) and a hierarchical RL variant (HAGS) that sequentially picks parcels to canvass for tenant eviction risk under a travel-cost budget, outperforming greedy, random, and conventional active search baselines on St. Louis eviction data.
Problem
Identifying at-risk tenant properties for eviction-prevention outreach under a limited canvassing budget with travel costs, where eviction risk is unknown a priori and must be learned while searching over thousands of parcels.
Approach
Combine a prediction module (updated by BCE on observed query labels) with a REINFORCE-trained search policy, and scale to large areas via a two-level hierarchy that first selects a region and then a parcel within it, with shared parameters across regions.
Key results
- In large-area uniform-cost search with 5% positive rate, HAGS finds 25.75 targets at budget 50 versus 22.25 for the best baseline (Greedy Adaptive), with improvements of about 3-17% over the most competitive baseline.
- With distance-based costs, HAGS improves over the most competitive baseline by roughly 42% to 70%.
- In small-area ablation, AGS with both modalities (2.24 ANT at 5% rate, budget 15) slightly outperforms tabular-only (1.97) and visual-only (1.84).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use hierarchical HAGS rather than flat AGS | AGS · Uniform-cost search, 5% and 10% positive rates | ANT: 10.25 vs 8.75; 13.25 vs 10.5 | Helped |
| use hierarchical HAGS rather than flat AGS | AGS · Distance-based search, 5% and 10% positive rates | ANT: 2.00 vs 1.25; 2.25 vs 1.50 | Helped |
| add visual features to the tabular-only model (ablation) | AGS-TAB · Small-area search, 2.5%, 5%, and 10% positive rates | ANT: 1.008 -> 1.052; 1.972 -> 2.240; 4.216 -> 4.372 | Helped |
| replace tabular-only features with visual-only features (ablation) | AGS-TAB · Small-area search, 2.5%, 5%, and 10% positive rates | ANT: 1.008 vs 0.932; 1.972 vs 1.840; 4.216 vs 3.496 | Hurt |
Surprising findings
- Flat AGS designed for small areas performs worse than the simple Greedy Adaptive heuristic at large scale, indicating the hierarchy is needed.
- Greedy (exploitation-only) baselines often find almost no targets under distance-based costs at low positive rates.
Limitations
- Evaluation uses a single mid-sized region (St. Louis) and simulated bootstrapped positive rates rather than live outreach.
- Improvements over baselines are sometimes small (3-17%) in uniform-cost settings and the ablation gains from visual data are modest.
- Only the authors' own eviction data is used, so generalization to other cities is untested.
Experimental setup
Datasets: St. Louis City and County eviction court filings and municipal property records ACS neighborhood data NAIP satellite imagery (June 2022, 60cm)
Benchmarks: St. Louis eviction filings (Oct-Dec 2022 test period)
Metrics: average number of targets (ANT) found within budget
Baselines: Random Greedy Greedy by unit count Greedy Adaptive Conventional active search (Jiang et al. 2017) AGS
Builds on: REINFORCE active search (Jiang et al. 2017) visual active search sarkar2023partially (GAS)
Closest papers by meaning
- I Open at the Close: A Deep Reinforcement Learning Evaluation of Open Streets Initiatives (AAAI 2024 · similarity 0.80)
- Occupancy-based Policy Gradient: Estimation, Convergence, and Optimality (NeurIPS 2024 · similarity 0.80)
- Preventing Eviction-Caused Homelessness through ML-Informed Distribution of Rental Assistance (AAAI 2024 · similarity 0.80)
- Dynamic Neighborhood Construction for Structured Large Discrete Action Spaces (ICLR 2024 · similarity 0.80)
- Diffusion Transformers as Open-World Spatiotemporal Foundation Models (NeurIPS 2025 · similarity 0.79)
- Hierarchical Optimization via LLM-Guided Objective Evolution for Mobility-on-Demand Systems (NeurIPS 2025 · similarity 0.79)
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.