Related Experiment Videos
Systematic benchmarking of microarray data classification: assessing the role of non-linearity and dimensionality
Nathalie Pochet1, Frank De Smet, Johan A K Suykens
1ESAT-SCD (SISTA), K.U. Leuven, Kasteelpark Arenberg 10, 3001 Leuven-Heverlee, Belgium. Nathalie.Pochet@esat.kuleuven.ac.be
Bioinformatics (Oxford, England)
|July 3, 2004
Summary
Non-linear kernel methods, particularly with radial basis function (RBF) kernels, can improve cancer classification accuracy from microarray data. However, linear methods with regularization or linear kernel principal component analysis (kernel PCA) are less prone to overfitting.
Area of Science:
- Bioinformatics
- Computational Biology
- Genomics
Background:
- Microarray technology enables simultaneous measurement of thousands of gene expression levels.
- Gene expression profiling combined with classification methods aids clinical decisions, especially in oncology.
- Benchmarking non-linear versus linear techniques and dimensionality reduction is crucial for optimizing these methods.
Purpose of the Study:
- To systematically compare the performance of linear and non-linear classification and dimensionality reduction techniques using microarray data.
- To evaluate the impact of radial basis function (RBF) kernels against linear kernels in classification tasks.
- To assess the effectiveness of regularization and kernel principal component analysis (kernel PCA) in gene expression-based cancer classification.
Main Methods:
- A systematic benchmarking study comparing linear and non-linear classification techniques, including least squares support vector machines (LS-SVMs).
- Evaluation of dimensionality reduction methods, specifically linear and kernel principal component analysis (kernel PCA) with linear and RBF kernels.
- Utilized 9 binary cancer classification problems from 7 public microarray datasets with 20 randomizations each.
Main Results:
- Well-tuned RBF kernels in LS-SVMs (without dimensionality reduction) performed comparably or better than linear kernels, with minimal overfitting risk.
- Regularization proved essential for linear classifiers, including LS-SVM with a linear kernel.
- Kernel PCA using an RBF kernel showed a tendency towards overfitting, especially with supervised feature selection, whereas linear kernel PCA yielded superior results.
Conclusions:
- Non-linear RBF kernels can enhance LS-SVM classification performance on microarray data without significant overfitting.
- Regularization is critical for achieving robust performance with linear classifiers.
- Linear kernel PCA is preferable to RBF kernel PCA for dimensionality reduction prior to classification to avoid overfitting.