Overviev and Practical Examples: Which Different Types of ML-Alghorithms Are There?

The Most Important Categories of Machine Learning

Blog Entry EP-KI-Blog / November 23, 2022

After explaining in our previous blog post how the two terms »Artificial Intelligence« (AI) and »Machine learning« (ML) differ from each other, this text is specifically about ML and different categories. We introduce the four main types of ML below and show real-world examples. As an overview, they are shown in the graphic. The first three categories discussed differ in terms of what data is used. The fourth ML category is about learning through interaction with the environment.

Author: Axel Preis

Supervised Learning – an AI Learns With the Support of a Human

»Supervised Learning« is the term used by experts to describe methods that learn using a labeled data set. A labeled data set consists of labels and features. Labels are the manifestations of a variable that is of primary interest and should be predicted, for example, fraud cases in a booking system. Features, in turn, are variables that are used to explain the variable of interest. Supervised Learning methods learn the relationship between features and labels from the data. Often, experts are used to create labels. So human intelligence comes into play first to make the learning possible at all. These people assign a label to each data point.

Providing the data is thus often very time-consuming – especially in times of Big Data, where data sets can consist of millions of data points. As we can see, supervised ML methods require labels during the learning process that are generated under human supervision, which is why we speak of Supervised Learning.

Application Example: Artificial Intelligence in Detecting Fraud in Booking Systems

An example of the use of Supervised Learning, which we in the EP-KI team often deal with, is the detection of fraud in a booking system. A corresponding AI system decides whether a booking is considered suspicious or not based on factors such as »amount of money« or »date of booking«. These factors form the features of a possible ML model. For such decisions, it needs booking data for learning that has already been labeled as »fraud« or »no fraud«. These markings represent the labels. An ML model thus learns whether new bookings are potential fraud based on the features and labels of the dataset.

© Fraunhofer ITWM
The Four Main Categories or Types of Machine Learning

Unsupervised Learning – Recognizing Without Human Clues

»Unsupervised Learning« in German refers to methods that learn patterns and relationships in data sets without labels. Such ML methods are mainly used in the following areas:

for grouping similar data points
for finding anomalies in data
for dimensionality reduction of a data set (i.e., to reduce the number of features).

This kind of learning has the advantage that no labels are needed, which have to be generated with high effort as described. On the other hand, labels contain helpful information that is not available in Unsupervised Learning.

Application Example: Automated Fraud Detection
The example of the AI system that detects fraud in submitted invoices introduced under »Unsupervised Learning« can also be used to illustrate »Supervised Learning«. Now, however, it is not known whether a booking is a case of fraud or not. That is, no labels are available. An Unsupervised Learning algorithm determines whether a booking deviates strongly in its features from other bookings and is therefore a potential case of fraud.

Semi-Supervised-Learning – A Balanced Learning Mix

»Semi-Supervised Learning« in German is used when there is a data set in which only a small portion of the data points have labels and the majority are unlabeled. Thus, in a way, it represents a combination of Supervised and Unsupervised Learning. The particular advantage of this approach is that labels are used without investing too much effort in creating them manually.

Application Example: Accounting Advanced
One application of Semi-Supervised Learning in the context of the accounting system, which we have already shown to illustrate Supervised Learning and Unsupervised Learning, could be to label a small amount of accounting data as »fraud« or »no fraud« by experts. Subsequently, unlabeled data points could be classified as »fraud« or »no fraud« based on their similarity to the labeled data points.

Reinforcement Learning – Learning Through Rewarding

»Reinforcement Learning« refers to methods that aim to learn an optimal strategy for performing a specific task. The so-called »agent« learns the optimal strategy via rewards and punishments. An agent can be, for example, a robot that makes decisions and acts autonomously. The learning process takes place in the agent's environment. This environment contains:

all possible states between the start and completion of the task
the amount of reward or punishment for an executed action in a given state
the possible transitions from one state to another by means of an action

The agent adapts its strategy during the learning process in such a way that it maximizes its total reward and it finds the optimal path. If this explanation is too abstract, it can be illustrated with »Super Mario«:

Application Example: Super Mario and AI Practice in Video Gaming

A well-known example of the use of Reinforcement Learning is the learning of a game by a computer. During the game, the computer repeatedly adapts its solution strategy to win the game in a wide variety of situations. In Super Mario, for example, the character can perform three actions: run to the right, run to the left and jump.

To successfully complete a level, it is necessary to move the character quickly and frequently to the right. The level together with the rewards (mushrooms, coins) and obstacles (enemies, carnivorous plants, fire, holes in the ground) represents the environment. At the beginning of a level, moving to the left is penalized as the character moves away from the target, while moving to the right is rewarded and jumping is a neutral action. With an action performed, the state and the set of selectable actions change. Situations arise in which a move to the right is bad and thus results in a punishment. This is the case, for example, when contact is made with an obstacle. By repetition the computer learns which partial strategies lead to success. After successfully mastering the level, the learned strategy is further optimized in repeated runs. The effort for the development of an optimal strategy can be reduced by the computer using the recording of the mastering of the level by a human player as a starting strategy.

© freepik
Reward or punish? Collecting a heart, star, coin or life potion leads to positive effects.

The Super Mario example is used to illustrate the concept of »Reinforcement Learning«. It is described from the perspective of a human. For us humans, Super Mario is a sequence of images with mushrooms, enemies, etc. The ML algorithm (called a computer in the example), on the other hand, doesn't know who or what Mario is, what mushrooms are, etc. For him, Super Mario is a sequence of data of pixels.

In general, »Reinforcement Learning« is the closest approach (for a given application) to »Strong Artificial Intelligence« (AI). If you're not familiar with the term »Strong AI«, you can read about what it means in our previous blog post here.

In this blog post, an overview of four important types of Machine Learning was given with examples. In addition, there are a variety of other categories of ML. Examples are Deep Learning, Federated Learning, Active Learning and Transfer Learning.

zur Blog-Übersicht

Blog Entry