Date of Award
Fall 2017
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Mathematics, Statistics and Computer Science
First Advisor
Ahamed, Sheikh Iqbal
Second Advisor
VanHecke, Amy
Third Advisor
Chu, William
Abstract
The prediction accuracy in many classical problems across multiple domains has seen a rise since computational tools such as multi-layer neural nets and complex machine learning algorithms have become widely accessible to the research community. In this research, we take a step back and examine the feature space in two problems from very different domains. We show that novel augmentation to the feature space yields higher performance. Emotion Recognition in Adults from a Control Group: The objective is to quantify the emotional state of an individual at any time using data collected by wearable sensors. We define emotional state as a mixture of amusement, anger, disgust, fear, sadness, anxiety and neutral and their respective levels at any time. The generated model predicts an individual’s dominant state and generates an emotional spectrum, 1x7 vector indicating levels of each emotional state and anxiety. We present an iterative learning framework that alters the feature space uniquely to an individual’s emotion perception, and predicts the emotional state using the individual specific feature space. Hybrid Feature Space for Image Classification: The objective is to improve the accuracy of existing image recognition by leveraging text features from the images. As humans, we perceive objects using colors, dimensions, geometry and any textual information we can gather. Current image recognition algorithms rely exclusively on the first 3 and do not use the textual information. This study develops and tests an approach that trains a classifier on a hybrid text based feature space that has comparable accuracy to the state of the art CNN’s while being significantly inexpensive computationally. Moreover, when combined with CNN’S the approach yields a statistically significant boost in accuracy. Both models are validated using cross validation and holdout validation, and are evaluated against the state of the art.