Exponential kernelized feature map Theil-Sen regression-based deep belief neural learning classifier for drift detection with data stream

Data streams are potentially large and thus data stream classification tasks are not strictly stationary. In the process of data analysis, the fundamental structure may vary over time and the changes in the primary distribution of the data are known as drift. Early drift detection achieves better detection results in the evolving data stream analysis. In order to perform accurate drift detection with minimum time, a novel deep learning technique called exponential kernelized feature map Theil-Sen regression-based deep belief neural learning classifier (EKFMTR-DBNLC) was introduced. The main aim of the proposed EKFMTR-DBNLC technique is to perform multiple drift detection from the data stream using multiple layers. The proposed deep belief network comprises of various layers such as an input layer, an output layer and two hidden layers. The input layer receives the number of features and data from the dataset. The hidden layers perform the significant feature selection to reduce the drift detection time. The exponential kernelized semantic feature mapping technique is applied for identifying the significant feature for data classifications. Then, using Theil-Sen regression (TSR) function, the drifts in the data stream are detected and classified from the selected relevant features in the next hidden layer. The regression function analyzes the distribution of the data between the two-time intervals. Based on regression analysis, multiple drifts such as incremental drift, gradual drift, sudden drift and recurring drift are identified. Experimental estimation of the proposed EKFMTR-DBNLC technique and conventional methods are performed with different factors such as classification accuracy, precision, recall, F-score and drift detection time using real-world and synthetic datasets. The analyzed numerical result confirms that the proposed technique EKFMTR-DBNLC achieves 10% higher classification accuracy and also minimizes the time consumption by 13.5% than the conventional methods.