The Optimization Of AUC Assessment Methods Based On Weighed

The realization of classifier is one of the important research projects in the area of data mining, and the assessment to the classification performance is also an important basis to judge whether it is good or bad. Accuracy is a common method of assessment, but it is not perfect enough when it is used to the train set and test set are different with class distribution and misclassification costs. The assessment method of AUC(the area under the ROC) based on ROC can make up these deficiencies. The value of AUC got by calculation of ROC curves has more distinctions than accuracy. Its measure is better than the standard of accuracy and its judgment to classifier is more accurate.Today, many of the classification algorithms usually regard all the condition attributes as having the same importance to decision attribute, but this assuming can not be widely used in realistic world, because it ignores some sort of connection between attributes and thinks that attributes are independent existence from each other. In response to this deficiency, this article introduces a new weighted assessment method called AWI-AUC(Attribute Weighted with Information -AUC). Through analyzing the correlation between attributes, the higher the correlation between condition attribute and decision attribute is, the more important the condition attribute to decision attribute is, and then the higher weight given is. And vice versa. Obtaining the correlation coefficient,then, finding the formula of weight combined with the concept of mutual information.Finally,combine with AUC values to assess.To realize this method under the MBNC experiment platform and do simulation with UCI data set. And then compare the result with the result of the method of AUC. From this we can see the feasibility and effectiveness of weighting algorithm, which can assess the classification performance more accurately.