Mengjie Zhang, Genetic Programming for Feature Selection and Construction

Genetic Programming for Feature Selection and Construction

• Overview:

  Classification tasks arise in a wide variety of practical situations. Diagnosing medical conditions from medical imaging, recognising words in streams of speech, and identifying fraudulent financial transactions are just three examples. Given the amount of data that need to be classified, computer based solutions to many of these tasks would be of immense social and economic value.

  Genetic Programming (GP) is a promising approach for automatically building reliable classification programs quickly and automatically, given only a set of example data on which a program can be evaluated. GP uses ideas analogous to biological evolution to search the space of possible programs to evolve a good program for a particular task. GP has been applied to a range of classification tasks with some success.

• Tasks and Goals of this Project

  Most classification problems use a large number of features to describe. Since the relative importance of them is usually unknown, GP has to use all of them for classification. However, not all the features are critical for that task, the features are seldom equally important, and some features are too abstract to contribute to classification, which often lead to unsatisfactory performance.

  Based on our initial work on low level features selection for speech stress detection and object recognition, the goal of this project is to investigate novel feature discovery and construction algorithms for classification that can (a) automatically discover features useful for a particular task, (b) evenly reduce redundant features, and (c) effectively construct high-level super-features from the discovered features for classification.

  We expect these algorithms to efficiently rank the features, to effectively improve the system accuracy for all classes, and to significantly improve the comprehensibility of evolved programs.

• More information can be seen from Meng's research projects . Related Publications can be viewed from here: Meng's recent publications. Contact me if you want to get a copy of those papers, or if you want to know more detail about this project.