Research On Feature Selection Algorithm Based On Maximum Weight And Minimum Redundancy

Feature selection is the classical method of statistical. It attaches the attention ofscholars in various fields by its huge potential and broad application prospects.According to how they combine the optimal feature subset search with theconstruction of learning models, feature selection methods can be divided into Filter,wrapper and embedded methods. Filter methods assess the importance of each featureby investigating only the intrinsic properties of the data. Wrapper methods combinethe feature subset selection with the model training and testing process. Differentfrom the wrapper methods, embedded methods built the feature subset search intolearning model construction. Though the wrapper and embedded methods alwaysoutperform filter methods in terms of accuracy, filter methods is widely used for itscomputationally simple and fast. In this paper, we mainly research filter methods.Filter methods according to the selection strategy can be divided into rankingbased filter methods and space search based filter methods. The filter methods regardthe feature selection as a ranking problem. Although the ranking based filter methodshave been applied to some real world tasks successfully, a criticism of these methodsis that the feature subset selected by them may contain redundancy. Some spacesearch based filter methods have been proposed to reduce the redundancy duringfeature selection. In these methods, the features are selected by optimizing a particularcost function which is often defined as a trade-off between the most informative andleast redundancy inside the feature subset.By analyzing the existing feature selection methods, a novel filter framework ispresented to select optimal feature set based on the maximum weight and minimumredundancy (MWMR) criterion. Comparing with ranking based filter methods,WMWR considers both the weight of features and the redundancy among features;Comparing with other space search based filter methods, MWMR doesn’t specifymeasurement to estimate the importance of features and the redundancy amongfeatures. In the experiments, we firstly verify the accuracy of MWMR on artificialdata sets. Secondly, we test the ability of MWMR to solve cluster problem andclassify problem on four databases (CMU PIE, Extended YaleB, Colon and DLBCL).At last, we use one tailed t-test confirmed our MWMR framework significantly betterthan other feature selection algorithm.