Talk Title: Offline risk score and policy learning for responsible allocation of scarce housing to people experiencing homelessness

We study two problems in homeless services provision: the problem of learning vulnerability scoring rules that accurately predict the risk of an adverse outcome for people experiencing homelessness; and the problem of learning policies for matching people experiencing homeless to very scarce housing resources. We consider several challenges related to the self-reported and observational nature of the data available to help learn these models, the presence of unobserved confounders, and the occurrence of distribution shifts caused by improvements to the administration and wording of the survey used to collect the data. We propose solutions that meld optimization, machine learning, and causal inference to address these challenges. We demonstrate their effectiveness on data from the homeless management information system. This work is the result of long term partnerships with homeless service providers and policy-makers in Los Angeles and Missouri.