作者:bosio mattia 13 年以前
215
Article map
In the realm of microarray classification, the challenge of dealing with few samples compared to the vast number of genes is prevalent. Various filter approaches for univariate feature selection are often employed but risk overfitting to the training data.
開啟
Computation
Time Tree construction time ? Need to include that? Feature selection 99% of time Time scalability with features IFFS SW : linear x Wsize IFFS : non linear and unfeasible SFFS : linear Is our algorhtm
better ? Common
Ground MAQC II set MCC value comparable Many samples Contemporary YES for
MCC mean
vlue MCC value
Mean value across
endpoints Metagenes
are useful? Treelet/
Euclidean ? Is the improvement
enough to compensate
for tree construcion? PROS -Resume of + genes - Expanded feat space - Interpretable comb. -Common behaviour NEW
ELEMENT? Reliability
Score useful? Score parameter Include it in article? Selection how? Classifier
Transparent Any other, only needed dist from boundary LDA Interpretable Robust Simple NEW
ELEMENT! More useful for
small sample number
After that is it totally
related with error rate? gives more info
about data distribution
wrt classifier boundary
than ERROR RATE only Microarray classification overfit to train data
Infer data distribution
from train set Many filter approaches
problem of univariate
feature selection Few samples wrt gene number