Deep Representation Learning for Analyzing Financial and Cybersecurity Data

Abstract

Recent years have shown a surge in the capability of deep learning models for various domains of interest and different types of data. Machine learning models are now capable of understanding, categorizing and generating unstructured data on par with or even surpassing human performance. While many tasks requiring human intelligence and expertise are still out of reach for machine learning models, these systems have the potential to automate tasks that require a large amount of manual work by human experts. In this thesis, we will investigate the application of machine learning models in the financial auditing process and in the field of cybersecurity, with data primarily consisting of unstructured text data and numerical tabular data. <br /> When working with textual data, one of the most important aspects of any machine learning model is the representation of the data. Text data is inherently unstructured and difficult to process for machine learning models due to its high dimensionality and discrete nature. To address this issue, we transform the data into a lower-dimensional continuous vector space, a process that requires a model to learn a useful representation of the data. We call this process representation learning. Having learned a useful representation of the data, we can apply various machine learning models to the data for a variety of tasks in a flexible manner, meaning we can use the same learned representation for different tasks. <br /> In this thesis, we will show how neural networks are able to learn a useful representation of data and how this representation can be used to reduce the need for manual work by human experts significantly. We will further show how representations with meaningful geometric properties allow for the generation of new data beyond the capabilities of classical generation algorithms.