Research On Hierarchical Classification Based On Conceptual Semantic Hierarchy

With the development of computer and Internet technology,the social media on the internet grow rapidly.The scale of the classification problem is becoming larger and larger,which includes the rapid growth of the number of samples,feature dimension and the number of categories.This will brings a great challenge to the efficiency and accuracy of classification.How to manage these large amounts of data and how to classify them in a large number of classes are the problems that needs to be solved at present.In the face of these problems,researchers exploit the hierarchical structure,which can be used to manage data,such as taxonomy or the hierarchy of text.There are some researchers exploiting hierarchical classification method to solve large classification problem and accurate classification.Based on the current research,this paper develops a hierarchical classification framework,which includes the construction of category hierarchy and hierarchical classification model based on structured support vector machine.The main research work includes the following aspects:(1)This paper proposes a new method,which fuses multiple semantic similarities from multiple views to build the category hierarchy.The algorithm fuses multiple similarities to obtain a new similarity measure by learning the weights of every similarity.After obtaining the fused measure,we develop an algorithm to construct the hierarchy automatically,which is based the Self-Tuning Spectral Clustering.The results on two images dataset show the effectiveness of the constructed hierarchies.(2)For the hierarchical classification model,we translate the classification task with a given hierarchy into a structured support vector machine(SVM)learning framework.We also develop a new semantic loss function,which is based the fused similar measure of classes.The experiments verify the hierarchical classification framework and our semantic loss function.(3)We apply the hierarchical classification framework based on structured support vector machine to the protein folding prediction problem.As to the feature selection,we extract several popular features following the former work,and combine all the features.The experimental results on two baseline dataset show that the hierarchical classification framework get good performance in protein folds prediction.