Panel Paper:
Using Machine Learning to Target Assistance: Identifying Tenants at Risk of Landlord Harassment
*Names in bold indicate Presenter
How should governments help these tenants? One model would be reactive rights enforcement: waiting for tenants to recognize that an action like cutting of their hot water is a rights violation and taking the steps to report them. Yet tenants' rights are hollow if the low-income tenants who need assistance the most are unaware of their rights. So rather than waiting for tenants to "come to the city" - that is, waiting for tenants to proactively recognize a rights violation by their landlord, decide to report it, and navigate the correct bureaucracies in search of assistance{the Administration created the Tenant Support Unit (TSU) in July of 2015. De Blasio described the goals of TSU, housed within the Mayor's Public Engagement Unit (PEU), as one of bringing the city to the doors of tenants: "When it comes to protecting tenants and aordable housing, we don't wait for a 311 call to come in. We have teams knocking on doors in fast-changing neighborhoods to solve problems then and there. This is a new strategy that's helping us keep New Yorkers in their homes and ght displacement before it happens." (Mayor Bill de Blasio, 2016)
The present paper asks how we can use large-scale data and machine learning to help the Tenant Support Unit do proactive rights enforcement. TSU sends outreach specialists to visit buildings where tenants face high risks of landlord harassment. The agency faces a daunting task, with over 6000 buildings and over 140,000 tenants in their outreach areas. How should they decide which buildings to visit first? Currently, there is wide variation in the likelihood that a particular visit to a building helps the agency discover landlord harassment. We use large-scale data and over 800 machine learning models to explore whether machine learning-guided prioritization outperforms their expert judgment at finding cases of landlord harassment.
The initial results show promising improvements - with the same number of visits, the agency can find over 1000 more cases of landlord harassment. Focusing on gradient boosting, the best-performing model, we show how the results not only lead to improved efficiency but seem also to lead to equitable prioritization. The model flags two types of neighborhoods as having high-risk buildings - 1) neighborhoods with rapid demographic change, where landlord harassment might be driven by strong financial incentives to convert units to market rate ones and 2) neighborhoods with persistent poverty, where landlord harassment may result less from market-rate conversion incentives and more from continued tenant-landlord power asymmetries.
