Generally, based on the way algorithms learn from data, machine learning can be divided into three paradigms: supervised learning, unsupervised learning, and reinforcement learning.
The primary difference between them lies in the results they produce, and the way they validate and calibrate those results.
Using Pecan, you can create different types of models that fall within the domain of supervised learning. The type of model that best fits your data – whether binary, multiclass or regression – will automatically be detected by Pecan’s platform once data is fed into it.
In supervised learning, a model learns to predict the future based on what happened in the past.
Here, historical data with “ground-truth labels” (where there is an actual observed outcome for each input) is fed into supervised learning algorithms. So you might classify users who performed a certain behavior with a 1, and others with a 0. The algorithm will identify correlations, patterns and trends that are historically correlated with these ground-truth outcomes, and then use them to predict the future.
Generally, you want your predictions to match ground-truth outcomes as frequently as possible – that is, you want your model to make as many correct predictions as it can.
In unsupervised learning, a model attempts to understand what happened in the past so it can be used to classify future outcomes.
Unsupervised learning makes it possible to uncover patterns in datasets where ground-truth labels are not available – that is, when records cannot be labeled with a certain target behavior (i.e. classified with a 0 or 1). The objective is to understand past outcomes, how they can be grouped or labeled, and what hidden patterns they hold – so that future behavior can be labeled and predicted.
Here, you would run an unsupervised model where the algorithm doesn’t have a target (e.g. 1 or 0), but rather, looks at the behavior of different customers and tells you which ones are behaving similarly to each other. It’s unsupervised because the data scientist never tells the AI what classifications to perform. But you can manually assess the results and identify certain qualities that exist among groups. So while you may not be able to identify, say, people who are likely to commit fraudulent activity, you may identify classifications related to income level, shopping habits, etc.
In reinforcement learning, a model uses trial and error to iteratively improve its approach to making decisions. The algorithm will continually try new things in order to learn how to behave better in the future.
Here, raw data (whether in real time or in batches) serves as the input. The reinforcement learning algorithm learns by constantly seeking new approaches to a problem, and by improving upon approaches that maximize the defined “reward”. This type of learning is frequently used in robotics and gaming.