Files
Abstract
Curse of dimensionality in modeling occurs when the subject of study is analyzed in high-dimensional data, while it cannot be easily identifiable in low dimensional spaces. But migrating to higher dimensional spaces causes challenges in modeling and interpretation of the subject. Advances in machine learning, particularly in the form of neural networks, supposedly tackles the challenge of modeling, but such techniques require a plethora of input data for training. Additionally, those techniques can be opaque and brittle when they highly become performant as a result of learning in complex spaces. It is not directly understandable as to why and when they work well, and why they may fail entirely when faced with new cases not seen in the training data. In this dissertation, we tackle those issues by proposing two techniques. (1) In the case of modeling, we propose a novel method that can help unlock the power of neural networks on limited data to produce competitive results. With extensive experiments, we demonstrate that our proposed method can be effective on limited data, and we test and evaluate our method on intermediate length time-series data that may not be suitable for simple neural networks due to lack of data with high-dimensional features. (2) In the interpretation context, we propose a new framework for 2-D interpreting (features and samples) black-box machine learning models via a metamodeling technique. Our interpretable toolset can explain the behavior and verify the properties of black-box models, by which we study the output and input relationships of the underlying machine learning models. We show how our method facilitates the analysis of a black-box, aiding practitioners to demystify its behavior, and in turn, providing transparency towards learning better and more reliable models.