In Partial Fulfillment of the Requirements for the Degree of Doctor of Philosophy
will defend her dissertation proposal
Feature Selection in Classification Tasks
Feature selection is one of the important pre-processing steps in classification tasks. It requires choosing a small subset of features that are the most relevant to the target class. Many classifiers do not perform well in the presence of irrelevant and redundant features. As an example, studies show that SVM has a worst-case sample complexity that grows at least linearly in the number of irrelevant features. Moreover, high-dimensional datasets make the classification task computationally expensive. Thus, application of a feature selection algorithm to detect the most relevant set of features is crucial. Many dimensional reduction methods are unsupervised feature extraction techniques, and their outcomes are not necessarily a subset of the original feature set. In this research, we introduce weight evolution waves derived from distance-based feature selection algorithms. We show how analyzing these waves along with a clustering approach sheds light on the relationship between features and the target class. Our proposed approach can be used to identify duplicates and interactive features (in the form of synergy or redundancy). Empirical results show that our method can effectively distinguish relevant features while unveiling hidden relationships between features and the target class.
Date: Friday, April 27, 2018
Time: 10:30 AM
Place: MREB 222
Advisors: Dr. Ricard Vilalta
Faculty, students, and the general public are invited.