Machine learning is a branch of artificial intelligence that can be defined as the ability of a program to recognise patterns in large volumes of data with which to make predictions.
Each of the types of learning has a series of specific applications and algorithms, adapting to different needs or data sets in the field of AI.
Let’s look at the main characteristics of the types of machine learning: supervised learning, unsupervised learning, semi-supervised learning and reinforcement learning.
Supervised learning
As explained on the Telefónica Tech blog, supervised learning is when algorithms work with labelled data, so that they are trained with a ‘history’ of data from which to learn in order to assign labels to a new value.
This labelled data used to train algorithms is used to make decisions and predictions and is especially useful for assessing risk, recognising images or detecting fraud.
Supervised learning uses several algorithms, some of the most prominent being the following:
- Decision trees to produce logic construction diagrams and solve problems thanks to these predictions.
- Naïve Bayes to calculate the probability of event A occurring if event B has happened, with special incidence in industry.
- Logistic regression to predict conclusions of a categorical variable according to dependent and independent variables, with use in social and health sciences.
- Set of classifiers that design a set of these to catalogue and weigh the data that is being taken.
- Support Vector Machines (SVM) to solve classification and regression problems, through the creation of a hyperplane in which the distance between two points is the maximum.
Unsupervised learning
On the other hand, in unsupervised learning, unlike supervised learning – as we have analysed previously – it lacks ‘labelled’ data for training.
By trying to describe the structure of the data, this type of learning has an exploratory character, allowing for pattern recognition and predictive modelling.
This type of learning is useful when working with complex and unstructured data, as it makes it possible to obtain results without having to manually label large volumes of data.
Its applications include customer segmentation, data comprehension and exploratory data analysis, product recommendation and pattern recognition.
Semi-supervised learning
As you might guess from the name, semi-supervised learning combines the two models we have already analysed: supervised and unsupervised learning.
In this case, some of the comments are labelled manually and when a small group is already labelled, one or more supervised learning algorithms are trained on that sample and the resulting models are used to label the rest of the comments.
For this reason, it is effective in cases where the collection of labelled data may be difficult or costly.
In short, in this machine learning model, the small volume of labelled data guides the learning process for the larger set of unlabelled data.
Thus, unsupervised learning is combined to identify data groups and then label the groups thanks to supervised learning.
Reinforcement learning
In this case, the algorithms learn through their own experience by trial and error, rewarding the former and punishing the latter.
As a result of having more practice, the algorithms can correctly predict the events under study.
Based on behaviourist psychology, it is based on the interaction of the agent with the environment, learning from situations or actions in the past without the need for known or previous results, with the exception of the final objective of the problem.
