Research On Multi-kernel Semi-supervised Classification Of Manifold Regularization Based On Sparse Graphs

In the context of the rapid development of the information age,the amount of data has increased to the mass level.But in some areas,such as image classification,pathological detection,web recommendations and other fields,where tagged data are particularly scarce and unmarked data exists in large numbers.Based on this reality,it is an urgent need to study semi-supervised learning algorithms to uncover the hidden information in unlabeled data,and make use of them to create value.In the semi-supervised learning algorithm,the graph-based semi-supervised classification algorithm is more favored by many scholars because of its better interpretability and classification performance.So it has been continuously studied and used.The main research results are as follows:First,a semi-supervised classification algorithm based on sparse graph combined with fuzzy c-means and graph transduction is proposed.Based on the promotion effect of partial unlabeled data on classification,firstly the unlabeled data helpful to the classification are pre-selected by the fuzzy c-means algorithm to build neighbor sparse graph,which greatly reduces the data nodes and improves the efficiency.Then the pre-selected unlabeled samples are classified by the way of graph transference to obtain the class labels and more reliable labeled sample data.Finally,the classifier is obtained by training all the sample data using manifold regularization framework.The experimental results show that the classification accuracy is 95% and above,which is higher than other two-class classification algorithm.Second,this paper proposes the multi-feature fusion of terracotta warriors image,and describes the features of the texture,shape,etc.,to avoid the problem that the single feature has a low classification accuracy due to the incomplete description of the data information.Third,a semi-supervised classification algorithm of manifold regularization multi-core model is proposed.According to the limitation of single feature,the algorithm combines multiple features by using multi-kernal function.Based on the high data mapping ability and flexible feature combination ability of multi-kernal learning,it provides data for semi-supervised multi-classification algorithm and improves the generalization ability of classifier.The experimental results show that adding multi-kernel fusion polynomial features to describe the classification results of sample data is far superior to that of single-sample description data.Fourth,this paper designs and implements a terracotta warriors fragment classification system.The semi-supervised classification algorithm is applied to the classification of terra cotta warriors fragment to classify the parts of terra cotta warriors.It paves the way to splicing and repair.