Research Of Hypernetwork Based On Multi-class Data Classification

The classification problem is one of the main fields in data mining, many learning algorithms were originally just developed for two-class classification problems. Data with multiple classes introduce intricate interaction between attributes and classes than two-class data, and it challenges the traditional learning methods. Hypernetwork(HN) is a novel machine learning method with simplicity, readable learning outcomes and so on. As a result, it has been widely used in disease diagnosis, text mining and other fields, and has achieved good results. But hypernetwork was designed to be a classifier for two-class data, has own drawback and shortcoming in dealing with multi-class data. This thesis improves the traditional hypernetwork for better solving multi-category dataset classification. The main research work of the thesis is as follows:1. The traditional evolutionary hypernetwork model has drawbacks of unreadable calculation result for hyperedge fitness value and unstable classification result for each category data. The thesis designs multi-classification evolutionary hypernetwork by researching the relationship between hyperedges and categories, improving calculation method of hyperedge fitness. Besides, the setting value of hypernetwork order seriously affects the classification performance. The traditional hypernetwork determines order value in the way of exhaustive search method before the evolutionary learning. The thesis discusses the possibility of setting hypernetwork order values by associating with C4.5 decision tree after analyzing the decision rule of C4.5 decision tree. The experiment shows that this method can determine the hypernetwork order value independently and enhance the classification precision effectively.2. For the poor generalization of hypernetwork model causes by the hyperedge generated randomly, the thesis designs a subspace integration multi-classification hypernetwork with feature cluster. Firstly, this thesis measures relationship between attributes and categories by using symmetric uncertainty, then the relevant features are used for clustering operation. Put redundancy attributes to the same cluster, and extract attribute to construct a subspace from every attribute cluster. Then a hyperedge library is generated by using one subspace, finally the thesis integrates all hyperedge libraries into one hypernetwork classifier. Through subspace integration, hypernetwork can search more decisive hyperedges and makes hyperdge library better fitting the probability distribution of training dataset, thus enhances the predictable capability of hypernetwork. The experiment shows that subspace integration multi-classification hypernetwork has higher classification accuracy and generalization performance than traditional hypernetwork and multi-classification hypernetwork.