Feature Selection Based On Class Center And Feature Weighting

With science and technology changing rapidly in the era of big data, we are faced with more and more large datasets. Massive amounts of irrelevant and redundant information that exists in the datasets makes existing machine learning algorithms face severe challenge. How to extract the most effective and reasonable data from the original data to meet the demand of storage and improve the efficiency of information processing is a problem that is in urgent need. Therefore, the feature selection problem has been one of the research hotspots in the field of pattern recognition.Maximum-margin based feature selection algorithm is an efficient feature selection algorithm for its ability to remove irrelevant features in high dimensional data and important applications in machine learning. However, the computational complexity is high. Although the local learning based feature selection algorithm is famous for its less computation, its computation complexity is logarithmical with respect to the number of features. In order to overcome the computational complexity issue, a class center and feature weighting based feature selection algorithm is proposed in this thesis. The basic idea is to take the center of a class, and look for the nearest hit and nearest miss to construct the margin. Then obtain the weighting of an instance with respect to a classification rule, which maximizes the margin in the feature weighted space.The features selected from the algorithm are used for SVM classification to verify its performance. The experiments on four UCI datasets show that the algorithm not only has less computational complexity and better classification accuracy, but also is nearly insensitive to irrelevant features.