What Is a Machine Learning (ML) Model? Full Guide

Machine learning models are typically categorized based on the type of learning method — or the learning algorithm — used to train the model. There are four such algorithmic types, though sometimes just three categories are used. Let’s break them down and explain how they differ.

Supervised learning models comprise the vast majority of ML models used in the industry today because they have the most direct and obvious business use cases and benefits. Supervised learning uses labeled data for training. This is introductory data which has already been tagged with the correct output. For example: A picture of an apple may be the input, the desired output would be the textual classification of the specific type of fruit. Once trained on a vast amount of this labeled data, the goal of the supervised ML model is to then take an unlabeled piece of new data (say, a photo of a banana) and use what it has learned to successfully catalog it.

Unsupervised learning models use unlabeled training data and do not provide a preset output for the model. In the above example, the training database of photographs would have no identifying context assigned from which the model can learn. Thus, it would be up to the model to find similarities, patterns and relationships among them. This is usually a less useful technique for categorizing images, but unsupervised learning finds its greatest value when a data scientist doesn’t necessarily have a specific end result in mind. Unsupervised learning is commonly used to search for patterns or anomalies in large datasets — such as clustering medical patients based on similar symptoms and demographics to uncover the cause of an illness, information which is impossible to properly label.

A lesser known (and not always recognized) ML type is semi-supervised learning. This method uses a combination of labeled and unlabeled data to train the model and is frequently seen when it is unfeasible to label every piece of data in a very large dataset — such as billions of photographs of people. Here, a subset of the data is labeled (or captioned, in this case) and fed to the algorithm along with the unlabeled ones. The unlabeled images are labeled over the course of the training.

ML models are designed through a lifecycle that typically involves six steps. The process is cyclical in nature, defined by a repeating series of stages that broadly comprise training, testing and deployment.

You can’t (or at least you shouldn’t) train a machine learning model if you don’t have a business problem you’re trying to solve. That could be something like reducing credit card fraud by 90% in your stores within 3 months or lowering machine downtime by 15% within the next year. The problem should be specific and measurable — and achievable within a given time period.

With the problem defined, it’s time to start focusing on the data you need to solve it. Gather, consolidate and — importantly — clean and transform data into a format that is usable for model training. Dirty data that is filled with inconsistencies, duplicates, outliers and gaps will lead to a poor ML model — and poor recommendations.

At this point, it’s time to decide what algorithm you’ll use to train your model. A range of algorithms are available, each designed for a certain use case, as we explored in the previous section. A trained data scientist will be invaluable in helping to guide this decision, and in determining what the resulting model should look like to maximize its utility.

With algorithm and data in hand, it’s time for the model to begin the training process. A data scientist will monitor this process, tuning parameters as the training proceeds to minimize the required training time and the number of errors produced during the training.

Once the initial training process is complete, you can use the model in a sandbox environment to get your first look at how well it did. To do this, you can feed  the model with new, live data that the model has not yet seen. Compare the performance of the model against your expectations and the metrics you laid out in the early stages of the project. Did the model successfully categorize new images? Did it detect the desired proportion of fraudulent transactions? By deploying the model as a pilot project or proof of concept, you can get an early idea about whether it is successful enough to push live or whether it needs fine-tuning or additional training.

ML models are being used every day in business- and consumer-facing applications. These include:

When an online store suggests another product based on something you’ve added to your cart, this is ML (usually unsupervised learning) at work.

ML-trained spam filters comb through billions of legitimate and spam messages looking for patterns that help identify bogus emails.

A common use of ML models in marketing is to attempt to segment customers based on characteristics that humans can’t identify but which an ML algorithm can uncover based on a much deeper analysis of the data. Teams can use this information to target new products or promotions based on those customers’ likelihood of making a purchase.

In machine learning, an algorithm is used to train a model. Put another way, the model is the output of the algorithm once the training has finished. This model is what is ultimately used to make decisions or predictions.

The four types of learning are supervised learning, unsupervised learning, semi-supervised learning and reinforcement learning. These vary primarily based on whether the data used to train the ML model is labeled, unlabeled or some combination of the two.