Abstract: Based on the correlation fast Filtering Feature selection algorithm (FCBF),which is improved by the maximum correlation coefficient. Firstly, It calculates the correlation measure between each feature and label with the ‘maximum normalized information coefficient’ criterion and ‘measurement principle of symmetric uncertainty’ and sort these feature according to the calculated value.Finally, It filters irrelevant features and redundant features by the ‘maximum normalized information coefficient’ criterion and approximate Markov Blanket and obtain the optimal feature subset. Experimental results on machine learning repository of university of california irvine(UCI) eight open datasets show that NFCBF algorithm outperforms FCBF algorithm. The number of features selected by feature selection algorithm based on maximum information coefficient (NFCBF algorithm) is less than 3.625 of the selected feature subset of FCBF algorithm, and the classification accuracy is improved by 0.075%. NFCBF algorithm gives better performance than mutual information maximization algorithm(MIM), Least absolute shrinkage and selection operator algorithm(Lasso) and Ridge algorithm.

Key words: maximal information coefficient, fast correlation based feature selection, feature relevance, feature redundancy, classification