At this point, we’ve all heard of machine learning. It’s highlighted in business news, your LinkedIn feed is riddled with posts on it, and chances are, you’ve probably had discussions in your company about it. However, as machine learning continues to gain popularity among enterprises, most companies still find themselves asking, “How do I get started?”
Before you can adopt a machine learning solution, you must learn how to identify the right machine learning use case for your business.
What Business Problems are Good Business Cases for Machine Learning?
Every business has problems, and luckily, machine learning is just another tool in your problem-solving toolkit. So, what kinds of problems are good candidates for machine learning? Typically, problems that require a series of decisions and have relevant data are best. After all, that’s what machine learning does—it automates decisions using relevant data.
To benefit from machine learning, you need a clear understanding of your business problems. These problems should be actual business problems—not just problems that are in a beginner’s machine learning tutorial, published in a blog post or written in an academic journal. And for you to justify a machine learning investment, your business problems should be impactful. For example, a machine learning approach would come in handy if you wanted to predict customer churn to improve customer retention initiatives—and in turn, your bottom line. Keep in mind, there are some problems that don’t require a machine learning approach because you can easily and inexpensively solve them by other means. For example, you wouldn’t use machine learning to determine which customers were most valuable last month; you’d just calculate sales received from each customer.
How Do I Turn my Business Case into a Machine Learning Problem?
Once you have a basic understanding of machine learning and you identify your key business problem, you can map your problem to a machine learning problem pattern—a generic problem-solving workflow used to construct a machine learning solution. By doing this, you’ll learn how to reframe your business problem as a machine learning problem.
A machine learning problem pattern is composed of building blocks just like LEGO® blocks. Two basic types of building blocks are an input data block—which can be structured data from a database, images, text, audio, etc.—and an output decision block—which can be a number, a classification, a ranking, some text, a similarity measure, etc. With these two blocks, you can get pretty far in translating simple business problems to machine learning problems. However, for more complicated business problems, there are other types of blocks to consider, like a granularity block and an algorithm block. A granularity block determines how granular you want your prediction to be, and an algorithm block determines which of the various algorithms (with different properties) you might apply to your problem. Keep in mind, machine learning problem patterns are modular in nature, so blocks can be added or removed interchangeably depending on the complexity of your business problem.
To help you visualize these concepts, here are some examples of different blocks and machine learning problem patterns.
Types of Blocks for Machine Learning Problem Patterns
Before we highlight specific examples, here are the types of blocks we’ll use for our example patterns. (Note: This isn’t an exhaustive list of possibilities, merely a sample.)
Example Pattern # 1
Let’s say we have a client who wants to use machine learning to identify and extract attributes from contracts to expedite the review process. Here are what the blocks for this pattern look like:
Input Data: Our input data is a contract (Text).
Algorithm: We’ll use an algorithm that’s good at learning time-based dependencies (LSTM – Long Short-Term Memory).
Granularity: We want to predict whether each word in the contract is important or not (Intra-Input).
Output Decision: We need to decide whether a word is important or not (Class).
Given this decision, we can automatically highlight or extract important parts of a contract to accelerate and simplify the review process for legal teams.
Example Pattern # 2
Imagine a client who wants to identify unexpected values in financial statements to streamline the auditing process. Here are what the blocks for this pattern would look like:
Input Data: Our input is historical financial statements (Structured).
Algorithm: We’ll use an algorithm for time series forecasting (ARIMA – Autoregressive Integrated Moving Average).
Granularity: Across multiple inputs (i.e. historical statement periods), we want to produce a single output specifically for the current statement period (Inter-Input).
Output Decision: Our output decision is the value we expected to be in the current period financial statement (Number).
Given this decision, we could compare the observed value in the current financial statement to our expected value to identify any discrepancies. From there, we can decide if the current value is reasonable or if we need to further investigate (i.e. for fraudulence).
Example Pattern #3
In this example, imagine a client wants to predict when an industrial container will leak to eliminate the need for 24/7 manual surveillance and to reduce costs. Here are what the blocks for this pattern would look like:
Input Data: Our input is real-time pictures of the container (Images).
Algorithm: We’ll use an algorithm that’s effective for learning spatial relationships (ConvNet – Convolutional Neural Network).
Granularity: We need a prediction for each picture taken of the container (Per Input).
Output Decision Block: The output decision is the likelihood that the container in the picture is leaking (Probability).
Given the outcome of this decision, we might send a text to our on-call maintenance technician to investigate, rather than paying him around-the-clock to monitor the container.
Now that you’ve seen some examples of machine learning use cases, what are some business problems that you think could benefit from machine learning?
To learn more insights, tips and advice on machine learning and artificial intelligence, check out our other data science blogs.