Data Stream Learning / Continual Learning / Online Learning for Changing Environments

About the Project

The rise in digital device usage has caused a significant increase in data, necessitating efficient data stream learning algorithms, online learning algorithms or continual learning algorithms, which are able to learn over time and potentially adapt to changes in the environment where they are operating.

For example, in tweet topic classification, new tweet topics emerge over time, and the words that are representative of each topic may change over time, requiring adaptation. In credit card approval, new credit card customers arrive over time, and their behaviour may be affected by the current economic situation of the country. In software defect prediction, new functionalities may be developed in the software over time, affecting the likelihood of commits introducing defects in the software code.

Such changes in the underlying distribution of the problem are referred to as 'concept drifts' and pose a significant challenge to machine learning. They can cause drops in predictive performance, requiring swift adaptation of the machine learning models so that they do not become obsolete. This challenge can be further exacerbated by problem characteristics such as class imbalance / skewed distributions, missing or delayed labels, and data coming from multiple domains.

This project aims to develop novel data stream learning algorithms, online learning algorithms or continual learning algorithms to cope with challenges posed by data stream learning environments.

To apply, submit your application at https://www.birmingham.ac.uk/study/postgraduate/subjects/computer-science-and-data-science-courses/computer-science-phd listing this topic in your research proposal and Leandro Minku as a supervisor.