AutoML analysis, optimisation and comparison

Abstract

This master’s thesis deals with the examination of the fundamental properties of Automated Machine Learning (AutoML) and its distinction from traditional methods of
machine learning (ML). The central research question is how the three AutoML libraries
Auto-sklearn, TPOT, and AutoGluon function and in what ways they can be optimised.
To answer this question, a detailed comparative analysis of the three AutoML frameworks was conducted. The methodology included a comprehensive literature review and
the experimental evaluation of the frameworks using binary classification and regression tasks, with particular focus on the optimisation potential of the individual libraries
through data preprocessing.
The attempt to optimise the AutoML frameworks through data preprocessing showed
that not every type of preprocessing is suitable. It was found that the traditional preprocessing steps LDA and kPCA led to poorer results than when the models were trained
on the original data. It was also shown that conventional preprocessing steps led to the
best results in the classification task.
Therefore, it can be said that data preprocessing offers possibilities to improve the results of AutoML, but should be applied with caution. These findings provide valuable
insights for further optimisations and developments in the field of AutoML.

Date of Award	2024
Original language	English (American)
Supervisor	Bogdan Burlacu (Supervisor)