Home
Tags
No Tags
EC-Facilitated Cosine Classifier Optimization as Applied to Protein Solvation
| Title: | EC-Facilitated Cosine Classifier Optimization as Applied to Protein Solvation |
| Author: | Peterson, Michael R. |
| Description: | Peterson, Michael R., M.S., Department of Computer Science and Engineering, Wright State University, 2003. The performance of a pattern classification system depends heavily upon the quality of available features. Classifiers predict which of a predefined set of classes a particular object or concept belongs to, given a set of quantitative and/or qualitative measures, or features, describing the object or concept to be classified. The quality and choice of these features plays a major role in determining the overall performance of the classifier. Evolutionary Computation (EC) can be used to boost the predictive accuracy of certain types of classifiers by automating feature selection and extraction, the processes by which an optimal set of high quality features is chosen for a classifier. This thesis presents a new technique for enhancing a cosine-based k-nearest neighbor (knn) classifier using EC to search for optimal feature weights, an optimal point of reference for computing angles between feature vectors, and an optimal k-value. The use of EC to automate the processes of feature extraction and selection is shown to significantly boost the performance of this classifier. A further advantage of the techniques described here is the ability to perform data mining and analysis using the sets of weights and o®sets produced by EC. The pattern recognition system developed here is used toaddress two important problems in pharmaceutical drug development, prediction of water-binding sites on protein surfaces and distinguishing between conserved and displaced water sites upon protein-ligand binding. Over datasets developed to address these problems, this method achieves higher accuracy than both traditional pattern recognition techniques and more recently developed hybrid EC/classification techniques, while reducing the number of features considered by the classifier. The work described here also demonstrates the use of population-adaptive EC mutation to implicitly perform feature selection without allocating part of the chromsome for an explicit feature mask, as has been employed in previous related work. |
| Permanent Link: |
http://rave.ohiolink.edu/etdc/view?acc_num=wright1166729846
http://hdl.handle.net/2374.OX/19488 |
| Date: | 2003 |
Files in this item
| Files | Size | Format | View |
|---|---|---|---|
|
There are no files associated with this item. |
|||