A Differential Evolution Approach to Feature Selection and Instance Selection
More and more data is being collected due to constant improvements in storage hardware and data collection techniques. The incoming flow of data is so much that data mining techniques cannot keep up with. The data collected often has redundant or irreleva
- PDF / 296,428 Bytes
- 15 Pages / 439.37 x 666.142 pts Page_size
- 84 Downloads / 253 Views
Abstract. More and more data is being collected due to constant improvements in storage hardware and data collection techniques. The incoming flow of data is so much that data mining techniques cannot keep up with. The data collected often has redundant or irrelevant features/instances that limit classification performance. Feature selection and instance selection are processes that help reduce this problem by eliminating useless data. This paper develops a set of algorithms using Differential Evolution to achieve feature selection, instance selection, and combined feature and instance selection. The reduction of the data, the classification accuracy and the training time are compared with the original data and existing algorithms. Experiments on ten datasets of varying difficulty show that the newly developed algorithms can successfully reduce the size of the data, and maintain or increase the classification performance in most cases. In addition, the computational time is also substantially reduced. This work is the first time for systematically investigating a series of algorithms on feature and/or instance selection in classification and the findings show that instance selection is a much harder task to solve than feature selection, but with effective methods, it can significantly reduce the size of the data and provide great benefit. Keywords: Differential evolution tion · Classification
1
·
Feature selection
·
Instance selec-
Introduction
As hardware technology improves, more and more data is collected at a rate machine learning and data mining techniques cannot deal with. Often the data collected contains redundant or irrelevant features and instances [7,9,14,22,25], which may slow down and hindering the learning process in many tasks such as classification, reduce the learning performance, and/or learn complex models. A pre-processing step is often needed to remove some of the irrelevant or even noisy data, which can be achieved by feature selection (FS) for selecting only a small subset of informative features, instance selection (IS) for selecting only a small subset of representative examples/instances, or FS and IS for removing useless or redundant features and instances [13,17]. However, FS and/or IS is a c Springer International Publishing Switzerland 2016 R. Booth and M.-L. Zhang (Eds.): PRICAI 2016, LNAI 9810, pp. 588–602, 2016. DOI: 10.1007/978-3-319-42911-3 49
A Differential Evolution Approach to FS and IS
589
challenging problem due to two main reasons. The first is the large search space, which grows exponentially with the total number of features and instances. The second is that there are almost always interactions between features, which leads to a complex search space with many local optima and a good fitness function is often needed to guide the search in order to find a good solution. There have been a large number of works on FS, but not much work on IS, or FS and IS [13]. Different search techniques have been used for FS, but existing algorithms still suffer from the problem of stagnation in local
Data Loading...
