Research And Application Of Feature Selection Algorithm Based On Dynamic Weights Using Redundancy

Feature selection is an important means of dimension reduction.By retaining effective features,eliminating irrelevant and redundant features,a part of features are selected from the original feature set to form a feature subset,thus achieving the purpose of data dimensionality reduction.However,some existing feature selection algorithms are difficult to ensure that the relevance,redundancy and interaction of features are considered at the same time in the selection process,resulting in limited feature selection performance.In addition,with the development of deep learning technology,convolutional neural networks are widely used in various fields.It is often used to extract features of images.Most feature selection algorithms are aimed at traditional data sets,but they lack research on features extracted by networks.For the above two issues,the main work of this thesis is as follows.Aiming at the problem that some existing feature selection algorithms have insufficient consideration of redundancy between features in feature selection process,based on mutual information theory and the algorithm framework of dynamically updating candidate feature weights,this thesis proposes a feature selection algorithm of dynamic weights using redundancy.In the feature selection process,the algorithm can consider the redundancy between features while considering the feature relevance and interaction information,so as to ensure better feature selection result.Aiming at the problem that the existing feature selection algorithms lack research on features extracted by networks,this thesis takes fine-grained image classification as the target,and analyzes the process of feature extraction of convolutional neural network to obtain image features from the network.Then utilizing traditional feature selection algorithms to select image features.Thereby achieving the purpose of reducing the number of network neurons,reducing the amount of network parameters,and improving the classification accuracy.