Forecasting the Number of Tenants At-Risk of Formal Eviction: A Machine Learning Approach to Inform Public Policy

Maryam Tabar, Wooyong Jung, Amulya Yadav, Owen Wilson Chavez, Ashley Flores, Dongwon LeeIJCAI 2022 · DOI

TL;DR

This paper introduces MARTIAN, a novel multi-view deep learning model that forecasts the number of tenants at risk of formal eviction at the census tract level, months in advance. By integrating diverse data sources, MARTIAN offers a more accurate and timely prediction tool, crucial for optimizing eviction prevention programs and informing public policy.

What Problem Does This Paper Address?

Eviction in the U.S. is a critical societal issue with far-reaching consequences, including material hardship, increased health risks, and exacerbated poverty and housing inequality. While policymakers have allocated resources for eviction prevention, the effectiveness of these programs is hampered by a lack of accurate, forward-looking data on which tenants are at risk. Current forecasting methods are often too coarse in temporal resolution (e.g., yearly) or fail to adequately leverage the variety of available data sources. This gap prevents efficient resource allocation, leaving vulnerable populations at risk.

Methods

The core of this work is the MARTIAN (Multi-view model forcAsting the number of Tenants at-rIsk of formal evictioN) model. MARTIAN employs a multi-view deep neural network architecture designed to integrate data with different spatial and temporal resolutions. It comprises three distinct "views":

  • View 1: Uses a Long Short-Term Memory (LSTM) network to process historical time-series data of eviction filings within a census tract.
  • View 2: Employs another LSTM network to learn patterns from monthly labor statistics data.
  • View 3: Utilizes a Multi-Layer Perceptron (MLP) to process static features from American Community Survey (ACS) data.

The outputs from these three views are concatenated and fed into an output layer to forecast the number of tenants at risk of eviction. This multi-view approach allows the model to capture complex relationships across different data modalities.

Datasets

The study utilizes three primary data sources for training and evaluation:

  • Eviction Filing Records: Detailed information on eviction cases filed in Dallas County, TX, since 2019, provided by the Child Poverty Action Lab (CPAL). This dataset is used to define the target variable (number of eviction filings) and as an input feature.
  • Labor Statistics: Monthly data from the U.S. Bureau of Labor Statistics, providing insights into the economic status of the "Dallas-Fort Worth-Arlington" metropolitan area. This data is used for its temporal resolution and relevance to employment status.
  • American Community Survey (ACS): Annual data from the U.S. Census Bureau, providing demographic and housing characteristics at the census tract level. This dataset offers a broader, albeit less frequent, view of community conditions.

The paper does not introduce new benchmark datasets but uses real-world data from Dallas County, TX.

Key Insights and Ideas

  • The Power of Multi-View Learning: MARTIAN's success highlights the benefit of integrating diverse data sources with varying resolutions. Treating these sources as separate "views" and then combining their learned representations proves more effective than a monolithic approach.
  • Temporal Resolution Matters: The need for monthly forecasts is critical for timely intervention. Traditional yearly forecasts are insufficient for proactive policy implementation.
  • Data Source Contributions: Even data with lower temporal resolution (ACS) or broader spatial coverage (labor statistics for the metro area) provides valuable signals for fine-grained, monthly predictions at the census tract level.
  • Generalizability is Achievable: The model's ability to perform well on unseen census tracts demonstrates its potential for wider application beyond the training region.

Key Results

  • Superior Predictive Performance: MARTIAN consistently outperforms a wide range of baseline models (including Ridge, SVM, XGBoost, Random Forest, LightGBM, MLP, LSTM, GRU, and TabNet) across different forecast horizons (n=1, 2, 3 months).
  • Quantitative Gains: On average, MARTIAN achieves 5.00% lower RMSE and 0.88% higher Spearman correlation compared to the best-performing baseline model. This indicates both more accurate predictions and better preservation of the ranking of census tracts by eviction risk.
  • Ablation Study Findings: The ablation study confirms the importance of each data source:
  • Removing eviction filing data (View 1) severely degrades performance, showing its indispensability.
  • Removing labor statistics (View 2) leads to a noticeable increase in RMSE and decrease in Spearman.
  • Removing ACS data (View 3) also impacts performance, indicating its utility for providing a "big picture" context.
  • Cross-Region Test Success: MARTIAN demonstrates strong generalizability, outperforming baselines by 3.32% lower RMSE and 3.77% higher Spearman on unseen census tracts, suggesting its applicability to new regions.

Limitations and Future Work

  • Data Limitations: The paper notes that ACS data is not available at a monthly resolution, and labor statistics are provided at the metropolitan level, not the census tract level. While MARTIAN effectively incorporates these, future work could explore even finer-grained or more timely data if available.
  • Court Decisions: The eviction filing records do not include final court decisions, meaning the model forecasts the risk of formal eviction filing, not necessarily the final outcome.
  • Usability Review: The usability of MARTIAN is currently under review by subject matter experts, indicating that real-world deployment and feedback are ongoing.
  • Future Directions: The authors suggest that MARTIAN can be used to generate heatmaps of forecasted eviction risk (as shown in Figure 2) and to inform more effective resource allocation plans. They are actively discussing deployment with organizations like Texas Housers and CPAL to maximize the tool's impact.

Summary updated

BibTeX
@inproceedings{tabar2022forecasting,
  author = {Maryam Tabar and Wooyong Jung and Amulya Yadav and Owen Wilson Chavez and Ashley Flores and Dongwon Lee},
  title = {{Forecasting the Number of Tenants At-Risk of Formal Eviction: A Machine Learning Approach to Inform Public Policy}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/719},
}
Figure