HOUSED: 2025 Edmonton Accelerator Award winner

Impact/Innovation

We discovered a lack of solutions for addressing homelessness and the ones that do exist are reactive measures for dealing with the growing homeless population. What makes our project HOUSED unique is that we take a proactive approach by using machine learning predictions to give policy makers insight to create impactful initiatives and guidelines to help the current homeless population. We designed our model to identify which specific socioeconomic factors have the greatest impact on homelessness, giving decision-makers clear, actionable insights for assisting the homeless population. Instead of relying on black-box models, we used a linear regression approach so policymakers can easily understand how each factor contributes to the predictions, making the results transparent, accessible, and directly useful for creating informed policies.

Challenges

Our most significant challenge was collecting reliable, accurate, and sufficient data to train the model. Many datasets were incomplete, outdated, or inconsistent with our requirements. Rather than fabricating or inflating the dataset, we prioritized quality over quantity, and proceeded with trustworthy data. Another major data-related challenge was deciding which socio-economic factors to include. To address this, we conducted a detailed analysis of each available factor, examining its correlation with homelessness to identify trends, and carefully selected the most impactful variables to create an accurate and reliable dataset.

Another key challenge was navigating the ethical considerations tied to both the data and the project’s real-world impact. We held meetings with Rose Landry, Manager of AI Governance and Legal Affair at Mila, to discuss potential risks and mitigation strategies. We also conducted thorough research into best practices for ethical AI and engaged in conversations with individuals at homeless shelters, representatives from the government sector, and NGOs working directly with homeless populations. These interviews provided us with diverse perspectives into the lived experiences of the people our project aimed to support, ensuring our approach was informed and responsible.

On the technical side, one of our major hurdles was encoding the “years” feature into our linear regression model, since a standard linear regression model does not inherently understand the progression of time. After consulting with our mentor and teaching assistant, we decided to use lag features to represent time progression effectively. Additionally, we experimented with alternative models, including LSTM and Random Forest, to see if accuracy could be improved. In the end, the linear regression model outperformed the others, providing the highest accuracy for our dataset.

What we learned and accomplished

This project pushed us to grow both technically and as a team. On the technical side, we explored different machine learning concepts, learned their strengths and limitations, and implemented linear regression with careful feature engineering, residual analysis, and lag variables to improve predictions.

Working together was one of the best parts: we had a highly collaborative team where everyone contributed their strengths, from data preparation to modeling to outreach. A big milestone was meeting with local homeless shelters, which gave us real-world insight and validated that our model’s predictions reflected actual trends.

Most importantly, we learned the ethical responsibility that comes with doing data science for social issues. We focused on being transparent so people could understand how the predictions were made, worked to minimize bias in the data, and always kept in mind that every data point represents a real person. That perspective guided every decision we made in the project.