Research On One-Class Classification Based On Structural Information

One-class classification has recently attracted many researches in machine learning and pat-tern recognition. In practical application, the one-class classification is very common, such as detection eruption in satellite image, rare case of medical diagnosis, network intrusion detection, trade fraud and malicious arrears recognition. Therefore, improving classifying performance and generalization ability is very important and significant.The paper firstly investigated the current one-class classification algorithms based on den-sity analysis, support vector and sample structure. Although density analysis and support vector domain are able to improve the performance of classifiers on some problems, both two methods need a large amount of data in training phase. Specially, when the sample space is high dimensional, the training data cannot satisfy the requirements which leads to the low classifying accuracy. Based on mining sample structure information, this thesis pro-poses three effective one-class classification algorithms with clustering method to solve the mentioned disadvantages. A large number of experiments on the artificial datasets and UCI datasets have been conducted to verify the effectiveness of the proposed methods.In this thesis, the existing algorithms are improved from three different views. Three effec-tive one-class classification algorithms have been proposed based on structure information:Firstly, we optimize the traditional one-class classification from the clustering view. We introduce the clustering evaluation and DBCV evaluation indicator, moreover construct a MST base classifier, which has high accuracy and stability. The efficiency of proposed algorithm is validated on artificial datasets and UCI datasets.Secondly, adopted ensemble learning method, this thesis proposes a fast one-class classifi-cation algorithm based on structural information. The algorithm enhances the accuracy of weak classifier and reduces the time complexity of algorithm with divided and conquered method. Considering different datasets, we give the calculating methods for the best number of clustering and provide a detailed proof. Moreover, the proposed algorithm requires less input parameter dependency of classifier.Lastly, we propose an approximate search reverse nearest neighbor algorithm combing both algorithm and data, which optimizes and accelerates traditional K-nearest neighbor. This algorithm first constructs priority search K-means tree, then proposes prototype approximate search method to solve the K-nearest problem. We use the parent node of leaf node to search approximately, reducing the searching times. The method accelerates the calculating procedure under the requirement of accuracy, thus enabling the trained classifier gain high accuracy and stability.To sum up, the problem of complex sample space is solved by introducing the clustering evaluation. The study of integrated learning further reduced the classifier dependency on input parameters, on the other hand, using reverse neighbor algorithm performs well in the classification of the characteristics of high dimensional data. And the experiments on the artificial datasets and UCI datasets verify the effectiveness of the proposed method.