Uncorrelated feature selection via intra-group competition and inter-group cooperation

Feature selection is an important topic in pattern recognition research, which is supposed to find the most informative subset of features and remove the redundant features as well. By doing this, feature selection not only reduces the size of data, but also improves the performance of pattern recognition algorithms. However, previous feature selection methods focus on identifying the most important features and ignore the redundancy in important features, i.e., the important features maybe very similar with each other. To address this problem, we propose a novel and efficient approach to find a subset of important and uncorrelated features. An example of the proposed approach can be summarized as follows: firstly, we evaluate the importance of each feature and, meanwhile, group the features based on their pair-wise similarity. Then, the features are ranked in each group and a new score for each feature is computed by referring to its ranks in the groups. Finally, the features are re-ranked altogether using their updated new scores. In this way, our method is able to select the important and uncorrelated features rather than the most important but similar features. Experimental results on benchmark image data sets and a UCI data set are demonstrated to show the effectiveness of the proposed method.