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

A Human-Centered Review of Algorithms in Homelessness Research

“…In the study, frontline caseworkers and unhoused individuals expressed concern about the utility of using administrative data to train and build the model and questioned the construct validity of proxy outcomes used by the algorithm. In a related vein, Showkat et al [123] examined values embedded in homelessness algorithms and found contextual circumstances of unhoused persons were lost through the machine learning model development pipeline.…”

“…Our study builds on human-centered SIGCHI work by Kuo et al [69] and Showkat et al [123], which calls to the forefront, the need to empower human perspectives in homelessness algorithm design. One way to address this need is to critically examine the design of these models through a review of literature on decision-making algorithms in homelessness research using a human-centered lens.…”

“…Even the two models in our corpora that used unsupervised ML methods (K-means clustering) focused on identifying service provider clusters [133] and different types of at-risk families entering the shelter system [48]. Notwithstanding the ethical issues that arise from typecasting individuals and potentially reinforcing historical injustices [32,69,123], research on homelessness has long found that homelessness is caused, sustained and resolved by a confluence of individual, community, and societal-level factors such as availability of public benefits and family conflict/abuse [86,100]. Thus, algorithms that seek to categorize individuals may merely provide a narrow snapshot of an individual's circumstances without fully considering the community and welfare systems they live in.…”

“…Within homelessness, government agencies have increasingly adopted algorithms to predict one's risk of experiencing homelessness or related harms [63,70,131,137]. HCI scholars have been at the forefront of examining algorithmic systems in high-stakes decision-making environments such as homelessness, education, child welfare, and online behavior with an interest in understanding the social impact of these algorithm [23,57,62,69,76,105,111,123]. Works have examined potential harms and limitations of these systems and investigated issues around bias and fairness by incorporating stakeholder perceptions [22,73,112].…”

“…Karusala et al [57] and Kuo et al [69] studied stakeholder perceptions of algorithms used in homelessness and found it failed to account for contextual case details and interactive dynamics between caseworkers and clients. Showkat et al [123] studied values encoded in homelessness algorithms, and found human context became lost through computational abstractions (i.e., data preprocessing and model parameter tuning).…”