| In the context of the data mining, text classification is an important area of research,KNN algorithm which is one of the best methods of text classifying in the vector spacemodel (VSM)is a simple, example based method. The main steps are: text segmentation,feature extraction (feature weight calculation and characteristics of the word choice),building the feature model, training classifier. The feature extraction which is the coreof the text classification system, the feature extraction method has a major impact on theresult of text classification.Feature selection method has a direct impact on the result of text categorization.Many researches show that mutual information is a good feature selection method. TheMutual Information has one property that distinguish it from other dependencymeasures: it is the capacity of measuring any kind of relationship between variables;But the traditional mutual information approach still has the followinglimitations:(l)Mutual information method only considers the document frequency ofterm in the corpus，without taking into account the term frequency in each category ofthe corpus.(2)Mutual information method focuses on the correlation between terms andcategories, without considering the connections between terms.(3)The number of textsin each category in corpus also has an influence on the value of mutual information.To address the limitations of traditional mutual information method，this thesisimproved the measure of mutual information by introducing the feature frequency inclass and the dispersion of feature in class. Further more, we limit the minimum termfrequency，and introduce a minimum feature redundancy measure method.In addition, K-Nearest Neighbor (KNN) classification algorithm whether can selectthe appropriate K value will also affect the quality of classification results, the fixed Kvalue ignores the influence of the category and the document number of training text. Ifthe K value is too large, the text tends to belong to the class which contains more texts,classification performance is poor; If K value is too small, text has too few neighbors,this will reduce the classification accuracy.In order to solve the defect of the fixed K value, this paper proposes a kind ofdynamically obtained k-valued for KNN classification algorithm, experimental resultsshow that the dynamically obtained k-valued KNN classification algorithm has highperformance. A new feature extraction technology which based on the improved measure of mutual information is also proposed and we use it to improve KNN textcategorization algorithm. Experimental results on data sets indicate that the combinationof the improved feature extraction algorithm with the dynamically obtained K valuescan effectively obtain high quality classification results. |
