Semantic Feature Construction

Mark Fenner (Pitt/CS)

PhD Defense

Thursday, June 14th, 2007
10:00am - SENSQ 5317

Abstract

An effective set of features is integral to the success of machine learning algorithms. Semantic feature construction is the manipulation of the propositional descriptor space of a set of examples for use in a learning algorithm. Two important sources of semantics for feature construction are the semantic type (and associated semantic properties) and the semantic class of features.  These semantics can be captured in a knowledge base and utilized to constrain search through the space of constructed features.  This dissertation presents a system that captures semantic feature construction knowledge and implements a search algorithm that respects that knowledge.  Results are presented for different combinations of features generated from different successor functions used in search.  These results are compiled over many learning problems and several learning algorithms.  Other results are also presented for different levels of detail in semantic knowledge. Generally, semantics are an effective guide in the space of constructed features.

Dissertation Adviser

Prof. Bruce Buchanan, Department of Computer Science

Committee Members

Prof. Milos Hauskrecht Department of Computer Science
Prof. Satish Iyengar Department of Statistics
Prof. Janyce Wiebe, Department of Computer Science