Evaluating Feature Selection and Classification Methods for Unbalanced Multiclass Data with Small Sample Size

Feature selection (FS) and sample classification are important data mining techniques that can be applied to biological data. FS is used to identify biomarkers associated with specific sample type, and classification can classify a new sample using biomarkers identified in FS. The challenges in this research area lie in the factor that many biological data sets have high dimensions in feature space and are multiclass with small numbers of samples. In this project, we manipulated Iris data set from UCI machine learning repository by introducing different levels of noise and imbalance. The simulated Iris data sets were used to evaluate the performance of FS methods combined with classification methods. The same procedure was then applied to a soybean rhizosphere bacterial metagenomics data. Results show that appropriate FS method, together with classification method, is able to accurately identify biomarkers from multiclass biological data even when the sample size is small and data is unbalanced.