Thesis available

[Martin]

>>> 	FEATURE SELECTION
>>>  CLASSIFICATION PROBLEMS


My the PhD thesis: 

       "Distinction Senstivie Learning Vector Quantization (DSLVQ)"   
       (TU-Graz, 120 pages)

is now available for anonymous ftp from: 

       FTP-host:            fdpmial03.tu-graz.ac.at
       FTP-filename:     /pub/outgoing/dslvq.ps.Z (900 kbyte)
                                   /pub/outgoing/dslvq.ps	(4.4 mbyte)

Martin Pregnezer

---------------------------

Abstract

This thesis introduces a new feature selection method:
Distinction Sensitive Learning Vector Quantization (DSLVQ). 
DSLVQ is not based on the individual testing of different candidate 
feature subsets; the relevance of the features is deduced from the 
implicit problem representation through an exemplar based 
classification method. While most of the common feature selection 
methods require repeated training of the target classifier on 
selected feature subsets, only a single learning process is necessary 
with DSLVQ. This makes the new method exceptionally quick.   

The DSLVQ algorithm is motivated theoretically and evaluated
empirically. On a very complex and high dimensional artificial data
set it is shown that DSLVQ can separate relevant and irrelevant
features reliably. A real world application of DSLVQ is the selection
of optimal frequency bands for an EEG-based Brain Computer Interface
(BCI). DSLVQ has been used to individually adapt the filter settings
for each subject. This can improve the performance of the BCI. 

LVQ classifier: conditions under which stability problems with 
different training algorithms can occur are outlined in chapter 4 of 
this thesis or in a 15 pages draft paper which can be downloaded from 
the same site (lvq_stab.ps.Z /  lvq_stab.ps).