Research On Feature Selection Based On Maximum-weight Independent Set

With the fast development of the science and technology, high-capacity store technique hasbeen widely applied in many aspects. The data with abundant information appear in manyfields. Meanwhile, dealing with these data meets more and more pressure. The data makes themachine learning algorithms face serious ‘Curse of Dimensionality’. Developing thealgorithms to process these data becomes an anxious requirement. Dimensionality reductiontechnique effectively overcomes this problem. Feature selection technique quickly becomesthe focus of the pattern recognition and the data mining fields, as its simple, speediness andvalidity.Feature selection is a particularly important step in analyzing the data such as image,text and video. It aims to select the most representative features from the original data so thatthe performances of classification and recognition will be improved. Many feature selectionmethods have been proposed in the past decades. However, these methods always focus onthe importance of the features or the correlation among the features independently. Moreover,how to determine the optimal number of selected features is also a hard task. In order toovercome these limitations, we propose a novel feature selection framework termed featureselection as maximum-weight independent set (FS-MWIS). In our approach, we use thetheory of the maximum-weight independent set to select the subset of features with themaximum importance and minimum redundancy. Meanwhile, the optimal number of featuresselected by our framework can be automatically determined. This framework can embedmany algorithms to get the input parameters, providing the space for better experiment resultsand further improvement of our algorithm.In the experiment part, we test our algorithm on Extended YaleB database and ORLdatabase to compare our algorithm with other algorithms. The results of the experimentsshow the better recognition performances of our proposed algorithm and confirm thesuperiority and effectiveness of the algorithm.