| dc.description.abstract | Automated Machine Learning (AutoML) has emerged as a promising solution to tackle the challenges of algorithm selection, hyperparameter optimisation, feature engineering, scalability, and interpretability in ML tasks. This master thesis aims to profoundly investigate the state-of-the-art in AutoML and suggest revolutionary approaches to enhance its capabilities.
The research begins with a thorough review and evaluation of existing AutoML frameworks and techniques. Strengths, limitations, and applicability are scrutinised, providing valuable insights into their performance and usability across different problem domains. The evaluation includes comparisons of model performance, execution time, and interpretability, enhancing the understanding of the trade-offs involved.
Based on the findings, a novel approach, AutoFlex, is proposed to integrate established algorithms with automated pre-processing techniques. This approach leverages algorithms such as Random Forest Classifier, Gradient Boosting Regressor, and Decision Tree Classifier to ensure model interpretability. Additionally, pre-processing techniques like StandardScaler, RobustScaler, and OneHotEncoder are developed to enhance the quality of input data.
Extensive experiments are conducted on diverse datasets to evaluate the performance and interpretability of the proposed approach. Visualisations and analysis provide insights into the relationship between model performance, execution time, and interpretability, helping to interpret experimental findings.
The proposed approach, AutoFlex, combines interpretable algorithms with automated pre-processing techniques, which is crucial to developing more effective and usable AutoML systems. As AutoML continues to evolve, further research and advancements are necessary to address its limitations and maximise its potential for tackling complex ML tasks. We must continuously explore and innovate AutoML to ensure it remains a reliable and safe solution for various applications. | en_US |
