In the fight against chronic homelessness, London, Canada, has turned to an approach that combines artificial intelligence’s predictive power with a critical need for transparency and interpretability. The HIFIS-RNN-MLP model, developed by researchers to forecast chronic homelessness, is illustrative of making convoluted artificial neural networks trustworthy and easy to explain, precisely when dealing with sensitive social issues.
Central to rendering this model interpretable is the LIME model, or Local Interpretable Model-agnostic Explanations. It changes the almost black box neural network, which allows stakeholders to glean the process behind the decisions. Relating it to all individuals in the system, LIME works toward explaining what particular elements in an individual’s history and circumstances have the most significant effect on models toward changes in chronic homelessness.
These are not just abstract data points but are shown visually, with green and red bars indicating those factors that increase or decrease the risk of experiencing chronic homelessness. This visual approach brings the model’s insight to the front line in an accessible way across caseworkers, policymakers, and other non-technically oriented stakeholders.
For even more model interpretability, the researchers used the submodular pick algorithm for model explainability within LIME. Their global view showed that the recent stay in shelters—that is, during the last 30 days—was one of the most influencing factors in predicting that one would become chronically homeless. Very surprisingly, another strongly predictive feature was the administration of a SPDAT screening questionnaire one time step back, suggesting weekly the case workers were indeed identifying high-risk individuals.
One of the most prized values of such an interpretable approach is the ability to validate known risk factors and to make discoveries in possible new insights. The model validated the importance of housing subsidies—lack of such support steers predictions toward a chronic homelessness class. It also picked up age as a substantial parameter such that those over 52 are at significantly higher risk, and those below 26 are at a lower risk, similar to other studies.
Model interpretability in the setting of homelessness prevention served several functions. Specifically, it inspires trust by stakeholders and feel that the decisions of this model are ultimately based on some logic and relevant factors; it also hopefully assures them that the model has a conscious effort to reduce biases. HIFIS-RNN-MLP will ensure the identification of when an unreasonable and irrelevant factor that seems peculiar to varying cases plays an integral role in determining homelessness. Third, it allows for collaboration between the AI system and human experts, enabling assessment workers to embed decisions produced by the model within their professional judgment.
In addition, the interpretability of the model allows for fulfilling ethical guidelines for AI deployment in the public service. One key value of the Canadian government’s Directive on Automated Decision-Making involves the importance of explanation in automated systems, a characteristic HIFIS-RNN-MLP lacks.
Finally, and maybe most interestingly, this model makes it interpretable to go in new directions for targeted interventions. Since it identifies variables spelling out the highest risk for developing into chronic homelessness, by doing so, this model identifies all of the areas in which a prior service delivery system can be prone to bettering prevention strategies. It suggests, for instance, that the emphasis on service by this model regarding recent shelter stays seems to point toward rapid intervention when people visit shelters for the first time, which could be a significant path for keeping people from getting into chronic homelessness.
