The learning logic of exceptions is a considerable challenge in data mining and knowledge discovery. Exceptions are the rare data which are adhered from unusual positive behavior patterns. This is important to promote confidence to a limited number of records for effective learning of abnormality. In this study, a new synthetic approach based on statistical learning theory and information granulation theory is presented for confidence improvement of exceptional data learning. The proposed method follows under sampling approach for exceptional data detection. Information granulation theory is used for granules creation from data consecutively. Then, the support vector machine is applied to each granule. Exceptions and normal data is separated based on data point distance distribution from support vectors. The knowledge of normal and abnormal behavior has been extracted by a new method as a bottom-up learning approach. Efficiency of the proposed model has been determined by applying it to the Iran stock market data for abnormal stock selection. The superiority of the obtained results toward the outcome of applying decision tree, traditional SVM and neural network is considerable. Accuracy of proposed method was measured by g-means index. The outcomes show the capability of proposed approach in abnormality detection and exceptional behavior learning.
展开▼
