Detecting and Revamping of X-Outliers in Time Series Database

Dataset with Outliers causes poor accuracy in future analysis of data mining tasks. To improve the performance of mining task, it is necessary to detect and revamp of outliers which are there in the dataset. Existing techniques like ARMA (Auto-Regressive Moving Average), ARIMA (AutoRegressive Integrated Moving Average) and Multivariate Linear Gaussian state space model don't consider the periodicity for outlier detection. The above methods are used to find out only Y Outliers which are present in Y axis. These methods are not applicable to detect the time at which the peculiar data occurs (so called X-Outliers). This paper focuses different methods for detecting and revamping of X-Outliers that have abnormal data according to a known periodicity. These are practically applied in fraud detection, Market-basket analysis and medical applications to detect certain abnormal diseases. First the data is modeled to get the trend of the data and to remove noises by means of kernel smoothing. Next the outliers are detected by similarity measurements. If the dataset has outliers it can be replaced by considering periodic indices from the historical dataset. The performance of system is measured by precision, recall and F Score. The proposed method is tested with three different time series datasets namely, Electricity power consumption dataset, Weather dataset and