CS 68191 Masters Seminar / CS 89191 Doctoral Seminar
Spring 2008
A Hybrid Random Subspace Classifier Fusion Approach
for Protein Mass Spectra Classification
Amin Assareh
(Doctoral Student Presentation)
Classifier fusion strategies have shown great potential to enhance the
performance of pattern recognition systems. There is an agreement
among researchers in classifier combination that the major factor for
producing better accuracy is the diversity in the classifier
team. Re-sampling based approaches like bagging, boosting and random
subspace generate multiple models by training a single learning
algorithm on multiple random replicates or sub-samples, in either
feature space or the sample domain. In the present study we proposed a
hybrid random subspace fusion scheme that simultaneously utilizes both
the feature space and the sample domain to improve the diversity of
the classifier ensemble. Experimental results using two protein mass
spectra datasets of ovarian cancer demonstrate the usefulness of this
approach for six learning algorithms (LDA, 1-NN, Decision Tree,
Logistic Regression, Linear SVMs and MLP). The results also show that
the proposed strategy outperforms three conventional re-sampling based
ensemble algorithms on these datasets.