Publication:

A review of Active Learning methods for classification problems

Within the Artificial Intelligence field, and, more specifically, in the context of Machine Learning, the need for a set of labelled data is a common need to carry out the learning process of supervised models. However, the retrieval of such labelled data can be a rather arduous, expensive, and/or time-consuming task. In order to increase this process’ efficiency by reducing the number of required labelled data, the concept of Active Learning was introduced in the literature. The main idea behind it is that, given a set of unlabeled data, the most useful instances for the learning process are selected, therefore labelling only the most important examples, rather than the whole dataset or a random subset of it. This task is dealt with by means of different metrics, which allow us to quantify the representativeness and informativeness of each individual instance, with the objective of determining whether it should be labelled. In this project, we review several Active Learning methods for classification problems, implement the most relevant approaches and test them in a common experimental framework.