Three Steps to Understand Machine Learning

[Pat_Hohl]

Machine learning (ML) is helping solve all kinds of problems, but it just sounds rather unapproachable. Most people have heard the term, and many are not sure exactly what it means. How can it help solve utility business problems? Later in this blog post, we'll look at three words that will help anyone become what a colleague of mine calls buzzword compliant in machine learning.

Given the surge in data and analytics, the term "machine learning" is appearing everywhere. It's a subset of the artificial intelligence field. That makes it sound a little like science fiction, but it is not. Modern cloud data storage and computational resources have the stunning capacity to help address real business issues, particularly with ML techniques. Machine learning uses generic programs rather than creating specific programs for each new business challenge. These programs evaluate data to identify patterns and then use those patterns to form predictions. In a sense, the machine, or computer, learns from data about past experiences and applies that learning to the problem at hand.

Machines learn the same way we humans do. Someone often tells us what to expect (algorithm), we have experiences (data), we make observations and form our own ideas about patterns (training), and then we apply those patterns to future experiences. My wife has learned to read my mood by my facial expression and location. Even though she doesn't have direct knowledge of my inner thoughts, when I make "that face" in the kitchen, she has learned I'm unhappy that the kids left their dishes in the sink. She has observed a pattern and therefore makes accurate predictions.

1. Model

In the field of ML, "model" doesn't yet have a standard definition. Think of the model as the system that embodies the techniques, programs, and data. It's the model that will ultimately yield results. You might inquire about a certain model's performance on a new dataset. If you were to ask a data scientist if the model was ready to solve your problem, the reply might be, "The model isn't accurate enough yet; it still needs more tuning."

2. Algorithm

An algorithm is a procedure for completing a task, like a recipe. In ML, it's a mathematical approach to finding the patterns in the data. ML software includes libraries of these generic programs that can be used in your models. There are many algorithms written by mathematicians, each with its own strengths and weaknesses for different situations. They often have cryptic names like empirical Bayesian kriging (EBK) or ordinary least squares (OLS), and some testing is required to find the best algorithm for each application. Data science training is usually necessary to select the best algorithm for a particular task.

3. Training

Training helps set ML apart from other types of problem solving. The model is trained using a set of data with known outcomes. It applies the algorithm to analyze the training data, finding the correlations and patterns in that data. Once the model is sufficiently trained, it can then be used to make predictions on new data where the outcomes are yet unknown.

Solving a Real Problem

Electric utilities turn service on and off as part of normal business. For most utility customers, this doesn't happen very often. However, for some, this action is required as frequently as several times a year. This could be due to people moving in and out, or to unpaid utility bills. Some utilities install meters that can be turned on and off remotely from the office, without the need to send an employee and a vehicle to the location. These meters save time for the customer while reducing transaction costs and environmental impact. How could a utility analyze data to predict which customers would benefit from these more expensive meters?

To address this problem, we could assemble data from past orders and make predictions about future turn-on-and-off activity. The data could include account type, rate, payment history, bill amounts, address, etc. We would choose an algorithm, gather historic data, and train the model. This model could then be used to examine current accounts and predict which of those are most likely to have a high demand for this service in the future.

In this example, as with most utility concerns, information about location would be priceless in helping to solve this problem. We understand that university students occupy apartments with the school cycle and then often move out for the summer. Student moves directly affect electric service orders. If we could also use location information, such as proximity to universities, the model could perform even better.

Machine learning techniques will continue to evolve. As they do, it's important to include location intelligence every step of the way. Location is a natural part of utility data preparation, model training, and meaningful predictions. There is no better way to incorporate location intelligence in ML than with GIS.

The job of exploiting big data is far too great for manual analysis. Utilities need the best tools and techniques to find actionable insights in data, visualize the results, and clearly communicate the impact to stakeholders. The ArcGIS platform includes these tools along with powerful location-based machine learning techniques.

The next time you come out of the kitchen and someone reads your mind based on your facial expression, realize that they are using the same technique as machine learning. Find out how location intelligence and machine learning can address your business challenges.

[Hussam_AlJabri]

@Pat_Hohl, I like how you simplify complicated topics like AI/ML.

Thanks for your insightful blog.