Five keys to a successful machine learning initiative 

Key 1: Pick a good project

While this seems like a no-brainer, it’s important to remember that machine learning is effectively applied statistics. This means that the projects with the highest chances of success are going to be ones for which there are known mathematical or statistical relationships between the thing you want to predict and your data. Think mortgage defaults, fraud detection, or anything where there are known patterns that can be used as a starting point for building your ML models. This doesn’t mean that you can’t apply ML to new problems or areas where you haven’t identified patterns or mathematical relationships. Instead, it means that you’ll likely need to perform a deep dive analysis of your data to find those patterns before you start the ML work. Something to keep in mind is that this is an opportunity for a data analyst to potentially deliver significant value before you even start the ML work.  

Key 2: Identify cause and effect relationships

This is an extension of item number one. You need to identify the specific things that lead to the outcomes you’re asking the data scientists to model or predict. Let’s say you’re trying to predict arrival times for commuter trains, on the surface it sounds like a case of knowing things like velocity, location and distance to the next stop, likely departure time, etc. Instead, it’s more an exercise of predicting what could cause a train to be delayed, followed by predicting how long it would take to resolve those issues. I.E., getting ahead of potential breakdowns, weather issues, problems caused by an unexpected spike in ridership, etc. If there aren’t any significant issues on the tracks, you can predict when a train will arrive with speed, location, and distance to the next station with simple math. Where ML can provide value is predicting the things that can go wrong and the impact of those adverse events on arrival times.  

Key 3: Define how you will measure business value or success

ML models are statistical models and come with a variety of accuracy metrics: such as “precision”, “recall”, “f1-scores”, “area under curve” and the like. While meaningful to data scientists to measure the quality of their models, these aren’t necessarily useful to business stakeholders. Instead, there needs to be a discussion of how to measure the value this model delivers to the business. For example, reduction in fraud costs or default rates if the model had been applied to a set of historical credit card transactions or mortgage loans. Another measure could be the reduction in labor costs if the model was used to automate a business process. Having a discussion around how the model’s outputs will be translated into something relevant to your business and what business metrics are the most important to you will help ensure everyone is on the same page about what success looks like. 

Key 4: Don’t be surprised if most of the effort revolves around building the infrastructure for acquiring and studying additional data

This means that removing roadblocks to gathering data from across your organization and identifying data sources that can be used to enrich your current data, etc., are all critical success factors. Think back to the commuter train example: gathering data in real–time related to weather conditions, breakdowns, ridership, and the conditions of the train cars will likely be your critical success factor. For example, one of the reasons Dutch passenger trains are #3 in the world with respect to on-time arrival % is because of the data they gather and analyze in real–time. I.e. Don’t be surprised if it turns out that your project’s critical path requires building new data ingestion pipelines and/or deploying IoT devices to collect the data your machine learning models need to be effective. Be prepared to spend a significant amount of time clearing roadblocks and assisting your data science team in getting access to the data they need. 

Key 5: It’s a science project

No, this isn’t me attempting to use a non-clever pun. It’s the understanding that a data science project is an iterative process where you spend a lot of time learning more about your data, your business, discovering additional data needs, etc. Furthermore, the data might not tell the story you think it does, which may require rethinking how to solve the problem or pivoting to work on other problems that you weren’t previously aware of or considering. I.e., the path to solving your problem with ML is unlikely to be a linear one. However, you shouldn’t view this as wasted effort. Learning what doesn’t work or doesn’t work the way you think, is just as important as learning what does work. It’s also quite likely that insights gathered during the building of your ML models can be used elsewhere in your business. E.g., developing a stronger understanding of customer spending habits while creating a fraud detection model, could potentially enhance marketing efforts, retail partnerships, and customer satisfaction. 

Adopting machine learning for the first time can seem daunting and the process can be frustrating when it doesn’t quite play out like the typically exciting headlines pertaining to businesses and AI/ML adoption. However, with a bit of planning, foresight, and understanding of the process you can greatly improve your chances of success and ensure that your machine–learning projects deliver long–term value to your business.