Predicting Where: AI meets Location Intelligence

source: pixabay.com

In their recent book, “Prediction Machines”, the authors Agrawal, Gans and Goldfarb lay out a compelling and insightful proposition. They claim that the explosion in artificial intelligence in business is due to simple economics. How so? Artificial intelligence lowers the cost of prediction.

Prediction is about making bets on the future. It’s about bringing certainty to the uncertain. It’s also one of the primary strengths of artificial intelligence. By using artificial intelligence to address problems of prediction, businesses can make “better, faster and cheaper decisions”.

This is an astute and perhaps obvious conclusion when you think about it. The laws of economics rule any business model. A technology that can improve the predictive powers of a business is going to create a stir. One that is also affordable and accessible is going to be truly transformative. This is the unique perspective of the book. The true value proposition of AI is this: better prediction at a better price.

Prediction and Location Intelligence

I thought about how cheaper and better prediction applies to the world of location intelligence. Location intelligence at its core is about deriving insights from spatial information. Businesses with well-developed location intelligence capabilities use spatial information to support critical decision-making. Much of this decision-making involves making predictions.

Historically, predictive analysis in the location intelligence world has been based on traditional statistical techniques. This primarily includes regression analysis and its variants. In this type of analysis, a model is created based on a fixed set of variables and coefficients. Basically, you want to predict y based on a set of known variables a, b, c etc. The effectiveness of such a model is based on how well it predicts the future based on the average of what has occurred in the past.

The predictive power of a regression model has a lot (re: everything) to do with the variables you select to be a part of your model. For example, to predict where and when it might rain you might create a regression model based on historical rain observations in certain locations as well as other influencing variables in the same location such as elevation, time of year, temperature etc. The regression then attempts to predict the probability of rain based on current conditions by extrapolating from historical data.

This approach works fine for certain use cases. But for something as complex as predicting the rain it can quickly show its limitations. That’s because regression seeks to provide a prediction that is correct on average but is never actually correct. And that’s what we’re looking for: accurate predictions.

Another challenge with regression analysis is selecting the variables to include in the model. When you are dealing with tens of thousands of potential variables (such as what might influence rain) this can be an intractable task for the model creator. The benefit of AI-based machine learning is that the selection of variables and combinations of variables is determined by the AI itself. This can result in many unexpected combinations of variables that result in superior predictive power compared to the regression approach.

Machine learning excels at finding relationships between seemingly unrelated variables. The improvements in computing power and the advancement of AI techniques have made this possible.

Prediction in Practice

The marriage of AI and location intelligence is not a recent phenomenon. There are many documented applications of AI to location-based problems (read here for one) and volumes of scientific literature on the topic. In fact, a whole domain of “Geo.AI” has emerged in recent years.

What’s intriguing, and what the “Prediction Machines” book focuses on, is how AI-based prediction, specifically, can help businesses with decision-making. And what I’m interested in is how this applies to location-specific decisions. What decisions do businesses make that are primarily “where” questions? And which of these are actually problems of prediction?

Fortunately, Agrawal et al. provide a tool to help with the “what” and the “how” of using AI to make business predictions — the AI Canvas. And it works for location-specific problems too.

I won’t dwell on the specifics of the AI Canvas. You can do that here. My interest is using this tool to see what it would take to use AI-based prediction to address location-specific business problems.

Let’s look at an example everyone cares about: crime prevention. Most of us live in cities. And we know that cities are usually trying to prevent and reduce crime. We also know that crime usually happens somewhere. It’s inherently a location-specific problem. And it’s a problem that has far-reaching social consequences for the residents of the community. Anticipating where crime will happen, before it happens, would be one step closer to preventing it. In essence, crime prevention is a problem of prediction. Specifically, it’s about predicting location.

We know what we want to predict, but what will AI contribute? How will it affect the actions required to prevent crime? What data do we need to make the prediction? By using the AI Canvas we can identify the key requirements for using predictive AI to support crime prevention. Here’s what it looks like.

There’s much to consider when using AI to make predictions. Especially about such a sensitive issue. And especially when location is added to the mix. This is one of the benefits of the AI Canvas. It helps to think through the problem. For instance, failing to prevent criminal activity is bad but wasting resources and potentially harassing innocent people is even worse in a lot of cases. A false positive is likely worse than a false negative. How sure do we need to be in our model before we take it seriously and act?

This demonstrates why the “judgment” aspect of the decision is crucial. And, for now, this is in many cases still the domain of human decision makers. I say for now because, in the future, human judgment about the trade-off of social consequences could give way to AI. That’s a discussion for another day.

Another issue is the input and training datasets. When we talk about the declining cost of prediction, it’s due in large part to improvements in computing power and AI algorithms. However, the datasets needed to train and use the model are not necessarily cheap to obtain. In the example above, I identified demographic data and real-time sensor-based data. Acquiring this data, in some cases, is still quite costly. In a lot of AI use cases, data will actually be the limiting factor. The cost of prediction, particularly for location-specific problems, will continue to drop with more and better solutions for data capture and acquisition.

Finding Opportunities

Using the AI Canvas you can really hone in on opportunities to leverage AI for prediction. An enlightening exercise would be to run through the critical decisions your business makes and see which can be cast as prediction problems. For those in the location intelligence trenches, a way to know if a prediction problem is location-specific is if the prediction statement takes a spatial form. Some examples:

The opportunities abound. But if there is a business challenge that is inherently about location and can be recast as a problem of prediction there’s likely an AI opportunity in play.

This is an exciting time for businesses embracing AI and location intelligence. The declining cost of prediction means it’s easier than ever to look at the decisions that impact a business’s customers, community or operations and respond in a way and at a speed that makes a positive difference. And you don’t need AI to predict that!

This article was originally posted on LinkedIn