Fuzzy measure with regularization for gene selection and cancer prediction

Dealing with high-dimensional gene expression data is a challenging issue, and it is crucial to select multiple informative subsets of genes for cancer classification. In this regard, many statistical and machine learning methods with regulations have been developed. However, these methods neglected the epistasis, i.e., some genes may cover or affect other genes. In this article, we propose a fuzzy measure with regularization, which adopts L-1 and L-1/2 norms for sparse solutions, known as FMR, to describe the interaction between genes. Regularization with L-1 and L-1/2 can obtain a series of sparse solutions which help solving fuzzy measure quicker than traditional methods, such as Genetic Algorithm. FMR obtains a subset of genes corresponding to the fewest nonzero fuzzy measure values, and consequently, selects the important gene(s) according to the frequency of appearance in the selected gene subsets. Besides, three base classifiers, including SVM, KNN and DBN, are employed as underlying models to verify the effectiveness of the selected subset(s) of genes. Experimental results indicate that the selected genes by FMR are consistent with several clinical studies. In addition, it can produce comparable results in terms of accuracy as compared with other methods reported in the literature. The codes used in this article are freely available at: https://github.com/wangphoenix/ICMLC.