Machine learning models allow data science teams to help businesses solve problems and deliver solutions accordingly. Regardless of where you stand in your data science journey, you should be able to identify different machine learning models.
The algorithms used for machine learning typically fall under several categories: Supervised Learning, Unsupervised Learning, Reinforcement Learning, and Semi-Supervised Learning. Let’s explore these machine learning models…
Supervised Learning
Supervised learning models use labelled input and output data. Think of it as a teacher providing you with all the answers before you take a test. You know exactly what to expect ahead of time because you have been provided with all the correct information.
For instance, an input dataset contains a list of emails labelled ‘spam’ or ‘not spam’. This data informs the algorithm what emails are safe to filter into your inbox. Because the algorithm has labelled data to refer to over time, it should correctly filter new, non-labelled input data.
Supervised learning is a task-driven approach. It lets you tackle two potential business problems:
- Classification. You need to segment your output data into specific categories. The algorithm will attempt to categorise the output data based on the input data.
Examples of classification could be ‘Spam’, ‘Not Spam’ / ‘True’, ‘False’, / ‘Red’, ‘Blue’, ‘Green’, ‘Yellow’.
Some examples of classification algorithms are:
- Decision Trees
- Naïve Bayes
- Random Forest
- SVM
- KNN
These algorithms allow you to model the relationship between variables and subsequently classify them.
- Regression. You want to determine the relationship between a dependent and independent variable(s). The algorithm will form predictions on the input data as a continuous numerical value.
Examples of regression can include weight, time, or the price of something.
Some examples of regression algorithms are:
- Linear Regression
- Logistical Regression
- Polynomial Regression
- Decision Trees
- Random Forest
- SVM
You may notice that some algorithms can be used interchangeably for classification and regression. Bear this in mind when you are using supervised learning models.
Unsupervised learning
Unsupervised learning uses non-labelled input data and has no known output/outcome. The overall aim is to find hidden patterns and meaning in large amounts of non-labelled data.
Unsupervised learning is a data-driven approach that allows you to do:
- Clustering. You are finding and grouping identifiable patterns in your non-labelled dataset.
Examples of clustering include grouping consumers by purchase behaviour, small/large distances, or fraudulent behaviours/activities.
Examples of clustering algorithms are:
- K-Means
- Hierarchical Clustering
- Mixture of Gaussians
- Association. You want to define certain rules as a result of the relationships in your non-labelled data.
A common association rule is a recommendation system. For example, if a consumer purchases one product, we can then assume they will likely buy a second closely related product.
Examples of association algorithms are:
- Apriori
- AIS
Semi-Supervised Learning
Semi-supervised learning algorithms use a blend of learning methods and data. They tend to use small amounts of labelled data and large amounts of non-labelled data.
Firstly, unsupervised learning clusters the non-labelled data. Following this, supervised learning models label any non-labelled data. Over time, semi-supervised learning algorithms are able to learn from the labelled data provided. As a result, the algorithm is able to create pseudo-labels for non-labelled data.
Applications of semi-supervised learning include classifying text documents or analysing images and audio.
Examples of semi-supervised algorithms are:
- Naïve Bayes
- GANs
Reinforcement Learning
Reinforcement learning is unique from its counterparts in that it does not know anything about the input data. Instead, a “maximum reward” or end goal is set to encourage a series of actions to be taken in a particular scenario. It’s all about finding the best possible path to take.
Reinforcement learning uses one of two methods:
- Model-based. Interacting with its environment, model-based reasoning creates a ‘model’ of its physical surroundings. It is able to assess these surroundings to make informed decisions so that it can plan ahead of time.
Model-based learning is commonly associated with robotics and gaming.
Examples of reinforcement learning algorithms with model-based reasoning are:
- Policy Iteration
- Value Iteration
- Model-free. Gathers information about its environment in real-time. It operates on a feedback mechanism that is akin to a series of rewards and punishments.
Examples of model-free reinforcement learning in action include video games and self-driving cars. Whilst the end goal is in sight (i.e. the destination), it will likely face a series of obstacles (pedestrians, cars, debris, etc) that it must assess at that moment and know to avoid.
Examples of reinforcement learning algorithms with model-free reasoning are:
- Q-Learning
- Policy Gradient
