London, ON uses AI to fight chronic homelessness





As more and more communities turn to AI to solve problems, the conversation has shifted from if local governments can benefit from AI, to how they will put it to use.

In $ London, ON$ , AI is being rolled out to do everything from write reports to forecast water usage to identifying COVID-19 cases based on a chest X-ray (this last model is now being used in at least 20 countries).

Now, a recently introduced model uses this data-driven technology to predict whether individuals are at risk of becoming chronically homeless up to 6 months in the future. This allows shelters and prevention services to more effectively support at-risk individuals. It also could save the city millions of dollars each year. Now, London is making their model available for any government to use.

We caught up with Matt Ross, Manager of Artificial Intelligence in IT at the City of London. Not only did Matt spearhead these AI projects in London, he’s also a leader in the field overall, driving the municipal use of AI forward, while working to make sure it remains ethical, transparent, and responsible.

This conversation has been edited for length and clarity. For more of London’s AI work and best practices, listen to the Govlaunch podcast at the top of this page.



Govlaunch: You recently launched an impressive AI model that is likely going to have a big impact on social services around homelessness, can you share more about this project?

Matt Ross: The backdrop is that a few years ago, our Homeless Prevention Division did the hard work of getting all our shelters and agencies serving homeless populations using the same application. Different organizations all share the same database, which gives you a very comprehensive view of someone who uses many systems and services throughout the city.

This system is called HIFIS (Homeless Individuals and Families Information System). It’s a federal application and is used by a lot of municipalities across Canada.

We prototyped the Chronic Homelessness AI (CHAI) model which consumes all of the HIFIS data to predict the probability of someone becoming chronically homeless six months in the future. (Chronic homelessness means greater than 180 days in the shelter system in the year.)

An interesting business case behind this is that a study showed that a chronically homeless individual cost the city $135,000 a year. A chronically homeless individual in our system utilizes on average 534 days of shelter stays, whereas somebody who never becomes chronically homeless has 45 on average. So that’s 12 times as many resources used by what is actually a small percentage of our clients.

The CHAI model is extremely successful at finding these individuals. Then service providers can take those predictions and prioritize resources to stop them from becoming chronically homeless.

Govlaunch: How does it work?

Matt Ross: The model is based on demographic attributes about the person and their service usage. For example, this person is this age, doesn’t have a family, and has used a shelter 17 days in the last month, and 20 days in the last two months. The model picks up on these kinds of patterns and then provides a report with explainable AI which shows their risk probability of chronic homelessness and which features the prediction is based on.

Govlaunch: What is explainable AI?

Matt Ross: I think explainable AI will be critical for government adoption of AI in general.

It not only gives you the prediction — the probability of someone becoming chronically homeless, but then gives you the reason why. What features and fields and demographic attributes the model is relying on to make its prediction. This is useful for transparency and auditability and building trust with caseworkers and shelter heads.

Govlaunch: The AI model launched a few weeks ago — how’s it going so far?

Matt Ross: The model is now in production. They’re implementing it with case workers, and starting to train people on how to use it. There are a few early adopters at certain shelters who are really excited about it.

We won’t have enough good data for probably a year or so, once we’ve seen how it really works in practice and how successful the early interventions are. But the performance metrics are very good. It has a 92.1% recall (the accuracy at which it can identify an individual who is at risk).

Govlaunch: We’ve $ heard from your colleague Mat Daley$  about how London aims to share its innovative work with other communities. How can other local governments leverage the chronic homelessness model you’ve built?

Matt Ross: At the beginning of this project we thought, HIFIS is used by dozens of municipalities — why don’t we build this model for open source?

The model is available on Github and we’ll post it on $ our Govlaunch page$ .

It’s worth mentioning why I value open source. Not only will it eliminate some of the duplicated work that’s happening across cities, but because of it we built better code, and we built a better model. When you have a target for open source you need to think about someone who’s never seen the code before. You document it better, you build it in a more modular and extensible fashion.

Ironically, it actually ended up saving us money in our maintenance and support contract because it was so well documented that the scope and support model was very clear for the vendor.

Govlaunch: That’s great that you’ve made it so widely available.

Matt Ross: The idea is to not just build applications and deliver them, but also document any research associated with them, and make them open source for other municipalities to use. We all have similar problems. We sometimes use the exact same applications. It doesn’t make sense for us all to create these things from scratch because they’re time intensive to build and deploy well. The idea is for our lab to be a resource to Canadian, and even international, municipalities.

Govlaunch: You were able to launch this project with a team of 4 people. What roles would you say are key to a successful AI project?

Matt Ross: We had a small team of four people, plus some additional stakeholders. Small teams can be nimble, make decisions quickly, and iterate very quickly. I think with projects utilizing AI where there’s high uncertainty that’s unavoidable, you need to get to prototyping quickly and get that in front of stakeholders to make iterations based on that.

The critical people are a data scientist who builds the model; a database administrator and business systems analyst, who gets the data and data pipeline set up, and also productionizes the model; and then the business stakeholder. In our case it was the Manager of Homeless Prevention, who helped us understand data and ultimately owned the solution and implementation of the predictions. Then we had a manager (me) who has a data science background and can manage our other data scientists, database administrators and the business to translate between the three.

We then partnered with a vendor to do the ongoing maintenance and support — annual retraining, monitoring performance metrics and remediating any break/fix situations. That’s been really valuable for us.

Govlaunch: We recently shared some $ tips from Ben Gready in Edmonton$  for local governments looking to get into AI. What advice would you give to other municipalities?

Matt Ross: A few things come to mind.

One is the classic “garbage in, garbage out.” Machine learning models are only as good as the data they rely on. Focusing on data governance and data hygiene will set you up for success for future AI projects.

That being said, data doesn’t need to be perfect. You can prototype a model and see what the performance metrics are to set a baseline and then improve quality in the future if there is business value.

Another big one is that early AI projects are going to look like IT project influencing, rather than IT project management. You’re probably going to be looking at a database, seeing what has sufficient data quality for an AI implementation. Then you’re selling that to the internal division. So at the early stages, it’s IT influencing out, rather than the business coming to IT, and this is because of the gap in knowledge. IT influencing will often require a different set of skills and different team composition to be successful.

Another big one is to begin with the highest-value AI projects, rather than just the low-hanging fruit projects. This is to unlock the goodwill to do more projects. If you only go for the low hanging fruit, it’s easy to develop but there might not be that much value, meaning you’re not going to want to put in all the effort to productionize, maintain, and support it. So you might as well focus on the high value projects, since they’ll unlock the energy to do the next high value projects.

Govlaunch: Why should local governments consider using AI?

Matt Ross: We have huge datasets, and a large number of operational processes with a large volume of tasks flowing through them. They’re great candidates for optimization using AI models to really tap some of the value there.

It’s interesting because cities are just at the early stages of AI deployment. There’s a huge opportunity for sharing technology and building things collaboratively.

--

To hear more about London’s AI solutions and how other local governments can implement similar technology, listen to the Govlaunch podcast with Matt Ross at the top of this article.

For more on London’s overall culture of innovation, $ check out this story$ .