An anomaly intrusion detection algorithm based on graph partitioning method was presented. It could solve the problems of the low detection rate of the intrusion detection algorithms based on unsupervised learning, and the insufficiency of training samples of the intrusion detection algorithms based on supervised learning. It firstly gets the threshold for clustering using the labeled data, then clusters lots of unlabelled data using graph partitioning method, and then classifies the clusters using graph partitioning method and minimal labeled data. It not only can improve the classification accuracy effectively but also can utilize the detection of novelty attack. The experimental results manifest that the detection results of the algorithm outperforms some methods such as SVM, NB and KNN and so on. And the algorithm our proposed can find new types of attacks under the premise of ensuring the detection rate. Copyright © 2013 Binary Information Press.

The paper proposes a feature selection method by using the hierarchical correlation between categories. It firstly presents the discriminative ability of a feature by combining probabilistic statistical method with category information. And it coins a concept of category hierarchical correlation based on the hierarchical relationship of categories in hierarchical document classification(HDC). Based on the idea above a feature selection method based on category hierarchical correlation(CHC) is proposed according to the contribution of training documents with different level to the discriminative ability of the feature. Experiments on 20NewsGroups and CCT2002 using the libSVM classifier show that the performance of CHC and IG is similar and is better than CHI in HDC generally. Especially the performance of CHC is better than IG in the process of classification in the internal node of the category tree. Copyright © 2013 Binary Information Press.

An approach of feature selection for hierarchical classification is proposed. Firstly, the concept of category hierarchical correlation degree is introduced and it is calculated according to the category tree and the probability distribution of training data on different levels. Then, the importance degrees of categories are computed according to hierarchical correlation degree. Finally, the discriminative abilities of features are calculated based on the previous computation and the features with the greater discriminative ability are chosen as the feature set for classification. Experimental results show that the proposed approach outperforms the traditional feature selection methods on both quality of the features selected and standard classification metrics in terms of accuracy, F1 and micro-precision.

One common approach in hierarchical text classification (HTC) involves associating classifiers with nodes in the category tree T and classifying documents in a top-down manner. Training a hierarchical classifier is to choose the suitable classifier and its parameters at each node of T because the features of data in different classes are usually distinct. Until now all known approaches suffer from so-called blocking problem and ignore the hierarchical semantic relationships among subcategories in training classifier. In this paper we propose a novel methodology to train a hierarchical classifier that takes both above two problems into account. The approach will punish the classifier that wrongly classified documents in the last run and update the weights of the wrongly classified samples according to the levels they locate at T. The aim of the training process is to let the defined whole loss function of the classification be the minimized. Experiments on two collections show that the trained classifier outperforms the traditional AdaBoost in terms of accuracy, precision, recall, F1 and microPrecision. © 2011 by Binary Information Press.

Most of the previous researches on sentiment analysis concentrate on the binary distinction of positive vs. negative. This paper presents the multi-class sentiment classification problem that attempt to mine the implied rating information from reviews. We use four machine learning methods and two feature selection methods to find out whether or not the multi-class sentiment classification problem is the same to the binary sentiment classification problem, and whether it is equal to the traditional multi-class classification problem. Experiments show that multi-class sentiment classification problem is difficult than that of only determining the polarity of a review and that it is different from traditional multi-class classification problem, thus traditional multi-class classification method can not be directly used to deal with this problem.

In this paper we focus on fine-grained sentiment classification of Chinese reviews rather than the binary distinction of positive vs. negative. We attempt to mine the author's implied rating information through machine learning method. After giving a brief review of sentiment classification methods, the methodology used in this paper is described. Experiments show that fine-grained sentiment classification problem is difficult than that of only determining the polarity of a review and it is different from traditional multi-classes classification problem, thus traditional multi-classes classification method can not be directly used to deal with this problem. At last, some difficulties of this kind of fine-grained sentiment classification are discussed, together with the potential future directions. 1548-7741/ Copyright © 2009 Binary Information Press.

Decision tree is a forecasting model with a tree shape. The way from the root node to the leaf node forms the rule that forecasts a class label to the object. But it is often used in the condition of having not too much attributes. And FP-tree is a fast and efficient structure to discover frequent pattern. The paper proposes a new fast method for categorization through FP-tree. It discovers frequent feature terms of documents and categories using FP-tree and then forms a class decision tree by using the top frequent feature term to be the test attributes. So the decision tree is used in text categorization by the frequent feature terms. Finally the paper gives the experiment and the analysis of the method.

Hierarchical document classification refers to assigning one or more suitable categories from a hierarchical category space to a document. This paper proposes a new hierarchical document classification method based on a backtracking algorithm. Utilizing the relationships between categories in category tree, a suitable threshold for every category is found to determine whether a document could be classified into the category. And the backtracking algorithm in our hierarchical classification approach effectively solves the problem that a misclassification at higher level directly leads to the misclassification at a lower level. Moreover, feature set is selected by integrating information gain with hierarchy information, which accords with the characteristic of a category tree. Experiments show that the method performs well when enough training documents are given.

Ontology is a conceptualization of a domain into a human understandable, machine-readable format consisting of entities, attributes, relationships, and axioms. OWL is the standard Web ontology language proposed by W3C recently that is on the basis of the description logic. But there are so much uncertain and incomplete knowledge in real world. So it is necessary to import the fuzzy ability to extend the OWL. The paper proposed the ontology model based on fuzzy description logics (FDLs).It extended the OWL by importing the conditional probability on the basis of FDLs. And then the model extracts fuzzy formal concept and proposes a method to compute the similarity of the concepts. In addition it comes up with a method of fuzzy conceptual clustering and then generates the fuzzy ontology. Finally the paper gives the analysis of the model.