Research Of Classification Algorithm On Data Stream

Data stream is a new type of data with the characteristics of continuity, unlimitation, variations, orderliness and high speed affluxing.And these data streams can only be scanned once, and because of the limitations in the hard disk resources these records can not be stored in medium permanently. Classification is important in data stream mining, also is an important data analysis format.This thesis has found the existed correlative classification algorithm and the characteristics among them in current data stream classification mining research through analyzing and researching the literature of China and abroad. But due to the limitation in technology and theory, research in dynamic classification algorithm is relatively limited. Plus, the conception of data stream pre-processing is unclear and the application is not so prevalent. And all these are due to how to design a better pre-process with a lighter scale so that it can face the data steam easily and improve the classification analysis and it's data stream mining outcomes.This thesis brings forward a Bayesian dynamic classification algorithm trough researching data stream mining technology and the existed classification algorithms. This algorithm pre-processes the data stream through combining the methods of Landmark Model Concise Sampling with discretization conception, which simplizes the redundant data stream, and brings the conveniences to the classification algorithm. Data stream classification mechanism bases on Bayesian and sets two types of threshold values to perform the process of elimination, classification, and finding new classes. Also min critical value of useful info is setted to define pre-new class. And the eliminated info is stored for the future usage.