"Who is the right homeless client?": Values in Algorithmic Homelessness Service Provision and Machine Learning Research

Homelessness presents a long-standing problem worldwide. Like other welfare services, homeless services have gained increased traction in Machine Learning (ML) research. Unhoused persons are vulnerable and using their data in the ML pipeline raises serious concerns about the unintended harms and consequences of prioritizing different ML values. To address this, we conducted a critical analysis of 40 research papers identified through a systematic literature review in ML homelessness service provision research. We found that the values of novelty, performance, and identifying limitations were uplifted in these papers, whereas (in)efficiency, (low/high) cost, fast, (violated) privacy, and (homeless condition) reproducibility valuescollapse. Consequently, unhoused persons were lost (i.e., humans were deprioritized) at multi-level ML abstraction of predictors, categories, and algorithms. Our findings illuminate potential pathways forward at the intersection of data science, HCI and STS by situating humans at the center to support this vulnerable community.