Optimal Feature Selection using Fuzzy Combination of Feature Subset for Transcriptome Data

Applying machine learning algorithms directly on high dimensional datasets, like those encountered in transcriptome analysis, may lead to high time complexity and low performance of learning models, especially when the number of samples is small compared to the dimensionality. Selecting the optimal set of features then becomes an essential task for such datasets. Filter methods are one of the main class of techniques used for feature selection wherein a score is assigned to features based on criteria such as information gain, statistical measures or similarity based measures and then selects the best scored features. Using filter methods on the complete dataset results in features that have good performance over the dataset but might perform poorly in certain regions of the data, which affects accuracy for data points of those regions. To overcome this degradation in performance, we propose two novel methods to assign a robust score by using the fuzzy combination of the region-specific optimal feature subsets obtained using a standard feature selection algorithm (we use mRMR for this paper).We compare the result with state-of-the-art feature selection algorithm, mRMR (Minimum Redundancy Maximum Relevance) in the terms of accuracy on certain standard datasets.