We live in a spectacular era where we can teach computers to think and make decisions for us. Gartner predicts that by 2025, 38% of companies will make decisions based on algorithms. As a business leader, it is up to you to set your team and organization up for success when building and deploying machine learning-based solutions.
To lead a team to success with machine learning, you first need to understand what machine learning is, how and where to apply it, and how your company can benefit from it. One major benefit of machine learning is that it can help you answer questions about your business and clients.
There are two main types of predictive models that you can apply in your work. The type of model you choose to use should depend on the model outputs and how they can feed into and guide decision-making processes in your business. This article will define these two types of models, go through how to design and set up a successful model, and talk about how to prepare your employees to embrace the benefits of automated decision-making.
The two main types of predictive models are active models and passive models:
1. Active Models
In active models, predictions made by the model are acted upon automatically, with no need for human involvement. For instance, in debt collection, prediction models prioritize debtors based on those who are most likely to repay what they owe. The created, prioritized list is then loaded into an automatic dialer that calls down the list. When someone picks up, they are put through to a debt collector.
2. Passive Models
In passive models, predictions made by the model are passed on to people who consider the model outputs and make the final decision. For example, in tax collection, a central system identifies potential fraudulent tax returns. A local tax inspector will then investigate and reject cases at will. Active models are generally less likely to be overridden by human decision makers than passive models.
What Are the Necessary Preconditions for Success with Machine Learning?
One of the mistakes that business leaders make when investing in machine learning is believing that it will bring them rapid financial returns. With this belief comes the temptation to invest too much money into machine learning-based tools before preparing their company for the change. Preparation for a data science project should begin by making sure a company has everything they need in place to successfully deploy machine learning-based decision-making into their existing business processes. The following is a list of six characteristics of a company that is ready to leverage their data for success using machine learning:
1. You are open to changing the event you want to predict with machine learning.
Machine learning provides the most benefit for users and for the bottom line when a company has the ability to pivot their current operations to implement model outputs into the actual business processes that the model is targeting. If you ask a model to predict an event that your company can’t (or won’t) change, creating and running the model won’t provide any value. For instance, if your customers are always offered the same discount no matter what your pricing or intent model says, there is no opportunity for the model prediction to be implemented and demonstrate success.
2. You have a specific problem that is worth solving.
For example, many people overpay their taxes by small amounts. However, if you’re the IRS, the relatively small total dollar value for these errors means that it's not likely a major priority to identify these cases. This would not be an ideal application for machine learning. Machine learning should be used when a company has an issue that is causing a significant problem for their employees or their business.
3. Your company has a data-favorable culture and an appetite for change.
Your company’s values are important for the success of machine learning in your business. So before you invest in machine learning-based tools, first make sure you have a data-driven culture in place on your team. Many experts in their fields believe that predictive models can’t be as good as they are at predicting outcomes, which can lead to constant override of model predictions. If the experts at your company don’t trust the model outputs, consider using active models where possible, to help minimize the possibility of human override of decisions that could be automated.
4. Your company has the ability to integrate your machine learning models within existing systems.
Your organization may already be developing other machine learning projects, or already have an automated decision-making system in place. If that sounds like you or your team, congratulations! Your new machine learning project will likely benefit from your existing infrastructure, as well as from an internal peer group with experience in machine learning. While planning your new machine learning model, make sure that your project is compatible with the existing systems and that it can easily be integrated within the existing infrastructure.
5. You have available, stable, and usable data.
When implementing your model, make sure your data is available and will remain available. If you already know that certain data will not be available for use in the future, simply don’t use it to train your model. Also, if relationships within your data change, your model’s predictions will be affected. This phenomenon is known as model rot. If you know that the relationships within your data are unstable or changing, you’ll need to build a new model to reflect those changes.
6. You are aware of your stakeholders’ expectations.
Machine learning models should be designed with the expectations of business stakeholders in mind. In the banking industry, for instance, models need to be explicable and transparent because banks are constantly scrutinized by regulators. In contrast, in target marketing, dating sites, or language translation, the end user is not as interested in the underlying decision-making logic of the product as they are in the outcome, giving more freedom to model creators.
What Defines a Viable Machine Learning Project?
The following is a list of four characteristics of a successful machine learning project, which produces a model that can be considered a minimum viable product, or MVP. A minimum viable product for a machine learning project is an early version of a model that is usable as-is, yet is also buildable upon as it is refined and developed further.
1. Your project solves a problem worth solving.
Working on a solution for a problem that no one cares about is a waste of time, no matter how much fun you have doing it or how cool the solution is that you come up with. In contrast, a problem that’s worth solving may be something that you or your employees are encountering often and working on frequently. Make sure you know who a potential solution would benefit and how.
2. Your project focuses on a well-defined problem.
When you identify a problem that's worth solving, make sure that you also precisely define it. Defining the problem for your data science project is a large portion of the initial work. State exactly what you want to accomplish and break down the problem into smaller and simpler components. You may want to start by solving the problem manually by yourself. You should also consider speaking with customers as you define your problem to ensure your project is compatible with their needs and wishes.
3. Your project will have access to plenty of high-quality and labeled training data.
I can’t stress enough how important access to high-quality data is for a successful venture into machine learning. Ask yourself before you start your project if you can build data labeling into normal user activities within your product or service (for example, flagging spam emails). Having labeled data ready for your machine learning system is a great place to start your data science project.
4. Your project has optimal accuracy levels and reasonable error rates, and has communicated them to users and management.
You don’t need to obsess about perfection in your model, but you should remember that errors can be more costly in some projects than in others. Think about your end result when you consider acceptable levels of error in your project. Errors made in a model that produces product recommendations are relatively risk-free, but think about errors in a machine learning tool for self-driving cars—those could be catastrophic.
Calculating Return on Investment (ROI) and Allocating Budget
A successful machine learning project will generate value for your business, and can also decrease the likelihood of inefficient use of capital. However, a machine learning project will only be worth it for your business if you can prove your return on investment. Consider if your project can be associated with one of the following observations:
1. Increased Revenue
Ask yourself if the project you’re considering will increase revenue directly in a customer-facing department, and whether or not it will allow you to offer new products and services. To calculate revenue gains, make predictions for your project about resulting price changes and sales volumes. Consider if your service quality will increase and allow you to adjust prices based on your results.
2. Decreased Costs
Consider whether or not your project can achieve increased output and decreased human capital costs, alongside an increase in efficiency. Include in your analysis potential cost reductions resulting from potentially improved compliance or potentially decreased legal risks that may be happy side effects of automated decision making. To identify potential areas to decrease costs in your business, it may be useful to map your current business strategy. Identify what each of your employees do, and consider if they could be retrained to complete higher-value work.
Encouraging innovation within your corporate culture may also contribute to decreased costs, as automation can oftentimes enhance creative thinking and boost your employees’ productivity. However, the ultimate cost-saving strategy is to avoid the opportunity costs of inaction. Invest in artificial intelligence before it’s too late to catch up to those competitors who may be investing today.
3. Measured ROI
When calculating a project’s return on investment, the total time scale you use may vary depending on your project’s complexity and whether you build the model in-house or if you have an external provider. Some external providers have a faster roll-out time than others, which can lead to a quicker ROI.
4. A Portfolio Approach
Your approach and mindset towards your project needs to be the right one. Remember that if you expect a large return on investment, then you should also expect a longer roll-out time and higher level of risk. Assume that you’ll have some failures, and view your investments as research and development. And it never hurts to prioritize projects that are sure wins in order to boost your organization's confidence in machine learning applications.
Make Machine Learning Operational in Your Organization
Accepting the use of automated decision-making can be a big challenge for any organization. A successful and useful model needs to be fully embraced by your team in order to be operational. So, how do you solve this challenge?
To get started, draw parallels with your team to a Mergers and Acquisitions initiative, in which machine learning is the new entity to merge with or acquire. Investigate all of the alternatives to using machine learning. If your investigation indicates that machine learning can indeed add value to your work, go for it.
Similar to a mergers and acquisitions process, you’ll need strong technical talent and “translators” to bridge the gaps between data, machine learning, and decision makers at your company, so that generalist managers can execute based on actionable insight. Also make sure you have a good infrastructure in place before you begin—that means strong data storage, data movement, and model deployment protocols, as well as MLOps (What’s MLOps, you ask? See our thorough introduction to MLOps here).
Finally, make everyone comfortable with making decisions based on the model’s predictions. Make sure your employees don’t override these decisions by encouraging them to understand that a successful machine learning project is not only about the model, but also about the team of employees that surround the model. Consider machine learning as an engine in a car: if the rest of the car is missing, the engine alone is useless. A successful machine learning project includes the people and the business.
Hopefully this article gives you an idea of what it may take to carry out a successful machine learning project within your company, both before and after you build your model. Once you choose your project, be sure to check out our article about how to manage a data science project. As machine learning is used more and more to automate and speed up more mundane and time-consuming decisions and processes, it can also flatten the hierarchies within your organization and enhance collaboration. As a leader, you’ll need to evaluate the risks for your team associated with a machine learning project and mitigate them carefully. If you do, you'll open your company up to the possibility of real value generated for your business.