REPOSITORY > RESULTS

Abstract

This thesis introduces a new optimization method based on deep learning, designed for data
influenced by random processes. The main contribution of this method is the combination
of advanced noise reduction techniques with recurrent neural network models, which helps
to prevent the common problem of overfitting seen when there is not much training data
available. This approach greatly improves how the model optimizes data representation
and target outcomes, and enhances the model’s performance by managing noise effectively.
The success of this method is thoroughly tested on two different types of data, financial
and textual, demonstrating its flexibility and effectiveness. For financial data, it is applied
to a large dataset containing equity prices over twenty years, which forms a large complex
process where each equity follows unique random process. For textual data, the method
is used on a dataset about required job skills from different languages spanning over five
years, to show how it can be applied in more predictable situations.
This thesis makes important contributions to the field of deep learning in particular in
recurrent neural networks, by creating a new method that effectively uses a sophisticated
approach to noise management in neural networks. This method not only deals with the
challenges of modeling random processes but also opens up new possibilities for making
accurate predictions across different types of data.
The evaluation of a proposed approach includes three main parts: a formal definition
where each part of the model is defined in detail, a statistical analysis using the Wilcoxon
page rank test to show that our method performs better than existing models, and an
in-depth insight into how the model’s parameters can be adjusted and visualized. This
detailed assessment shows that our approach performs better than current models and
significantly improves the use of recurrent neural networks in handling random processes.