Abstract |
: |
An anomaly or outlier is a deviation point which varies so much from other monitored data to indicate that it is different from others. Outliers are also mentioned as discordant, or anomalies. The chore of detecting outliers is to find data objects that are marked different from or incompatible with the original data. Data are imperfectly labeled due to data corruption or data of uncertainty. A data may be imperfectly labeled as outlier, although the data is not an outlier. These occur mainly due to negative examples or corrupted data. Typically, most outlier detection algorithms such as Index-based algorithm, Cell-based algorithm, Statistical based method use some distance measure of outlier or a statistical model. These methods cannot detect imperfectly labeled data and normal data which behave like an outlier. To find imperfectly labeled data, a membership value towards the normal and abnormal classes is proposed. The proposed approach works as follows: the Cuckoo k-means clustering method is used to detect the outliers and kernel LOF-based method will be used to reckon the membership values. The originated membership values and few negative examples are incorporated into Cuckoo-SVDD learning frame to develop a classifier for global outlier detection. The proposed method extensively deals with the imperfectly labeled data and attains the accuracy of detecting outliers. |