Hierarchical Multi Granularity Learning Method For Hierarchical Data

With the maturity of information technology,the scale of data is expanding.It is mainly reflected in the rapid growth of data samples and features.Traditional machine learning and deep learning methods are facing great challenges.On the one hand,the growth of feature dimension makes redundant features greatly increased,machine learning more difficult and inefficient.On the other hand,the increase in sample dimension makes the number of data categories increase with the increase of sample number,and the accuracy of data classification decreases suddenly.For the problem of feature dimension,we improve the efficiency of the traditional fuzzy rough set algorithm by optimizing the distance calculation formula between sets and realize the efficient feature selection of high-dimensional datasets.For the problem of class dimension,this thesis realizes the hierarchical classification of multi-class datasets using the semantic hierarchy between classes and adding branch networks to the convolutional neural network.The main research contents are as follows:(1)Hierarchical feature selection algorithm based on fuzzy rough set,To solve the problem of low complexity and high efficiency of the traditional fuzzy rough set algorithm,an improved Hausdorff distance is used instead of Gaussian distance to calculate the upper and lower approximation of the set.The complexity of the algorithm is reduced,and the features with importance higher than the threshold are selected by introducing the threshold.It significantly improves the efficiency of the algorithm and the accuracy of feature selection to a certain extent.(2)Hierarchical classification algorithm based on convolutional neural network,Convolutional neural network can only realize the end-to-end one-time classification problem.We use VGG16 as the backbone network and add a fully connected network as a branch network at different depths of the convolution layer of the backbone network.Multiple classifications can be realized on each branch network.In addition,in the classification process,the classification results of the previous branch network are transmitted to the next branch network through feature splicing.The classification results of the previous network directly affect the next network.It realizes the hierarchical classification of hierarchical category datasets.