Gkioulekas, Ioannis; (2020) Optimisation based approaches for machine learning. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text PhD_Thesis_Gkioulekas_Ioannis.pdf - Accepted Version Download (5MB) | Preview

Abstract

Machine learning has attracted a lot of attention in recent years and it has become an integral part of many commercial and research projects, with a wide range of applications. With current developments in technology, more data is generated and stored than ever before. Identifying patterns, trends and anomalies in these datasets and summarising them with simple quantitative models is a vital task. This thesis focuses on the development of machine learning algorithms based on mathematical programming for datasets that are relatively small in size. The first topic of this doctoral thesis is piecewise regression, where a dataset is partitioned into multiple regions and a regression model is fitted to each one. This work uses an existing algorithm from the literature and extends the mathematical formulation in order to include information criteria. The inclusion of such criteria targets to deal with overfitting, which is a common problem in supervised learning tasks, by finding a balance between predictive performance and model complexity. The improvement in overall performance is demonstrated by testing and comparing the proposed method with various algorithms from the literature on various regression datasets. Extending the topic of regression, a decision tree regressor is also proposed. Decision trees are powerful and easy to understand structures that can be used both for regression and classification. In this work, an optimisation model is used for the binary splitting of nodes. A statistical test is introduced to check whether the partitioning of nodes is statistically meaningful and as a result control the tree generation process. Additionally, a novel mathematical formulation is proposed to perform feature selection and ultimately identify the appropriate variable to be selected for the splitting of nodes. The performance of the proposed algorithm is once again compared with a number of literature algorithms and it is shown that the introduction of the variable selection model is useful for reducing the training time of the algorithm without major sacrifices in performance. Lastly, a novel decision tree classifier is proposed. This algorithm is based on a mathematical formulation that identifies the optimal splitting variable and break value, applies a linear transformation to the data and then assigns them to a class while minimising the number of misclassified samples. The introduction of the linear transformation step reduces the dimensionality of the examined dataset down to a single variable, aiding the classification accuracy of the algorithm for more complex datasets. Popular classifiers from the literature have been used to compare the accuracy of the proposed algorithm on both synthetic and publicly available classification datasets.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as JSONCSVXML

Downloads by country - last 12 months

Export as JSONXMLCSV

Archive Staff Only

View Item