Articles

Types of learning used by Machine Learning algorithms

4 min read

Machine Learning models are those capable of learning to perform a task without being explicitly programmed for it. These algorithms can find patterns within large volumes of data and make predictions that are used, for example, in intelligent search engines, diagnostic systems, image, speech, and language recognition, and more.

But how is it that they “learn on their own”? We say that an algorithm learns when its performance improves with experience and through the use of data. Machine Learning algorithms can learn in four different ways: supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning. In this article, we discuss the different types of learning used by these models.

Supervised learning

Supervised learning can be used when labeled datasets exist, in other words, when prior knowledge is available. For example, a clothing store may have daily records of sold products over the past years, along with daily weather data. Here, we might want to train a model that, based on the weather data of a given day, predicts how many items will be sold. That is, using historical data (training data), it learns a mapping function between inputs and output.

Within supervised learning, depending on whether the variable to predict is quantitative or qualitative, there are two types of models: regression or classification models.

Classification models output a discrete label (one label from a finite set), such as clothing image classification (a set of possible apparel labels) or sentiment analysis (a finite set of sentiments).

Regression model outputs, however, are real values. For example, the number of garments that will be sold on a specific day, as in the earlier example.

Unsupervised learning

Unlike supervised learning, unsupervised learning works with unlabeled data, that is, there is no prior knowledge or target label to predict. These models start from large volumes of data, analyzing their structure and distribution to extract new insights and group entities by similarity (detecting patterns). They can also be used to reduce dimensionality and simplify datasets.

Clustering algorithms define similarity or distance metrics to compare data. They are used, for example, to identify customers with similar behaviors to target them with specific products or services, or to detect suspicious fraud patterns.

Dimensionality reduction techniques reduce model training time, improve performance, or speed up data visualization. They rely on mathematical and statistical methods to reduce data dimensionality while losing only a small amount of information.

Semi-supervised learning

Semi-supervised learning is a mix of supervised and unsupervised learning. Since fully labeled datasets are often difficult to obtain, this approach combines labeled and unlabeled data to build an improved supervised model.

For example, imagine analyzing tweets about a company to determine if they are positive or negative. First, we collect the tweets, but since they are not labeled, we do not know the sentiment. Semi-supervised learning would label a small subset manually, train a supervised model with those labels, use it to label the remaining unlabeled tweets, and finally retrain a new supervised model using all labeled data.

Reinforcement learning

Reinforcement learning differs from the others not by the type of data but by the learning process itself. Here, we deal with unsupervised problems where feedback or rewards (wins or losses) are given. The idea is simple: reward desired behaviors and penalize undesired ones. This allows an agent to learn through trial and error in a dynamic and uncertain environment, mapping states to actions that maximize a reward function.

Examples include chess models or AlphaGo. The agent knows the rules of the game and learns to play by itself. Initially with random moves, eventually using increasingly sophisticated strategies.

Machine Learning is undoubtedly a powerful technology that transforms data into knowledge and value, improving decision-making. At Linkroad we specialize in developing and implementing such models, with successful projects across multiple industries. A key success factor is clearly defining from the beginning the type of learning required based on the characteristics of the dataset.

If you’re considering applying Machine Learning to your business, feel free to contact us so we can conduct an initial assessment and discover the true potential of applying ML to your results.

CONTACT US

Our experts can help you

Tell us the constraint you are working against: the regulator, the core system, the data, and we will tell you where we would start.