Research On Classification Method Based On Statistical Learning Theory

Massive data classification is a hot research topic of data mining,machine learning and artificial intelligence, which the training set is used to trained a classifier, and the generated classifier can be used to achieve the further classification. The working principle of naive bayes classifier and Rough Sets are analyzed in this paper, and the corresponding solutions were given according to the shortcoming of the naive bayes classifier and Rough Set reduction.An incremental bayes classification algorithm based on spatial feature vector is proposed according its disadvantange which do not possess incremental classification ability and higher time complexity and space complexity. The vector space theory and space euclidean distance are introduced to abtain feature vector and incremental classification. The simulation experiment results show that the proposed algorithm can finish classification tasks quickly and accurately aiming at large amount samples, and can give a relatively precised classification results.An attribute reduction algorithm based on concentration boolean matrix is proposed aiming at the disadvantange of attribute reduction algorithm based on Skowron distinct matrix that is limited application range, time and space wasting and bottleneck performance: the boolean algebra was used to settle the time and space wasting problem, the directly generate resolution function minimum disjunctive paradigm was proposed to approve the time and space complexity and the executed efficiency of attribute reduction, also the reduction ratio is improved too.The aboved attributed reduction algorithm is not fitable to the dynamic object set and can not support incremental attribute reduction. Aiming at this question, a incremental attribute reduction algorithm which is suitable to decision table is proposed by analyze the releationship between new object and origanal decision table object. It achieves dynamic renewal, maintainance and management to attributed reduction results. The reduction efficiency is improved greatly and the proposed algorithm is feasible from theory analysis and example verification.