Committee:

Dr. David Anderson, ECE, Chair , Advisor

Dr. Justin Romberg, ECE

Dr. Christopher Rozell, ECE

Dr. Wing Li, Math

Dr. Gari Clifford, Emory

Abstract:

The objective of this research is to explore the use of sparse coding as a tool for unsupervised feature learning to help in the classification of imbalanced datasets. Traditional sparse coding dictionaries are learned by minimizing the average approximation error between a vector and its sparse decomposition. As such, these dictionaries may overlook important features that are only present in the minority class. Without these features, it may be difficult to accurately classify between the two classes. To overcome this problem, this work explores novel modifications to the dictionary learning framework that encourage dictionaries to include anomalous features. Sparse coding also inherently assumes that a vector can be represented as a sparse linear combination of a feature set. This work addresses the ability of sparse coding to learn a representative dictionary when the underlying data has a nonlinear sparse structure. Finally, this work applies the intuition gained from exploring dictionary learning in this manner to the classification of three imbalanced datasets.