Abstract. Evolutionary algorithms have received much attention in extracting knowledge on high-dimensional micro-array data, being crucial to their success a suitable definition of the search space of the potential solutions. In this paper, we present an evolutionary approach for selecting informative genes (features) to predict and diagnose cancer. We propose a procedure that combines results of filter methods, which are commonly used in the field of data mining, to reduce the search space where a genetic algorithm looks for solutions (i.e. gene subsets) with better classification performance, being the quality (fitness) of each solution evaluated by a classification method. The methodology is quite general because any classification algorithm could be incorporated as well a variety of filter methods. Extensive experiments on a public micro-array dataset are presented using four popular filter methods and SVM.

Recent research efforts attempt to combine multiple feature selection techniques instead of using a single one. However, this combination is often made on an “ad hoc” basis, depending on the specific problem at hand, without considering the degree of diversity/similarity of the involved methods. Moreover, though it is recognized that different techniques may return quite dissimilar outputs, especially in high dimensional/small sample size domains, few direct comparisons exist that quantify these differences and their implications on classification performance. This paper aims to provide a contribution in this direction by proposing a general methodology for assessing the similarity between the outputs of different feature selection methods in high dimensional classification problems. Using as benchmark the genomics domain, an empirical study has been conducted to compare some of the most popular feature selection methods, and useful insight has been obtained about their pattern of agreement

No abstract

As data mining develops and expands to new application areas, feature selection also reveals various aspects to be considered. This paper underlines two aspects that seem to categorize the large body of available feature selection algorithms: the effectiveness and the representation level. The effectiveness deals with selecting the minimum set of variables that maximize the accuracy of a classifier and the representation level concerns discovering how relevant the variables are for the domain of interest. For balancing the above aspects, the paper proposes an evolutionary framework for feature selection that expresses a hybrid method, organized in layers, each of them exploits a specific model of search strategy. Extensive experiments on gene selection from DNA-microarray datasets are presented and discussed. Results indicate that the framework compares well with different hybrid methods proposed in literature as it has the capability of finding well suited subsets of informative features while improving classification accuracy. Accuracy (%) 98.4 97.0 99.4 95.2 93.6 Subset size 4 7 10 6 4 ENDNOTES 1 Experiments were performed with a CPU of 2.8 GHz and 4GB of RAM.

Methods currently used for micro-array data classification aim to select a minimum subset of features, namely a predictor, that is necessary to construct a classifier of best accuracy. Although effective, they lack in facing the primary goal of domain experts that are interested in detecting different groups of biologically relevant markers. In this paper, we present and test a framework which aims to provide different subsets of relevant genes. It considers initial gene filtering to define a set of feature spaces each of ones is further refined by taking advantage from a genetic algorithm. Experiments show that the overall process results in a certain number of predictors with high classification accuracy. Compared to state-of-art feature selection algorithms, the proposed framework consistently generates better feature subsets and keeps improving the quality of selected subsets in terms of accuracy and size