Authors : Nabeel Al-A’araji, Eman Al-Shamery and Alyaa Abdul-Hussein
Abstract: Data that continuously arrive at time-varying and possibly unbound known as data streams. Connected, communicating devices that is called the Internet of Things (IoT) requires stream processing to deal with data coming from the sensors of connected devices which are the basic source of data stream now and in the future. These data streams are actually huge such as (e.g., sensor readings, call records, web page visits). This study aims to detect the drift and identify type of drift of stream. In addition to the pre-processing this consists of applying multi polynomial regression to solve the missing values problem, windowing and features extraction. An Adaptive Regression Neural Network (ARNN) for error prediction is suggested as new model to facilitate the distinction between real and virtual drift. A Modified Page Hinckley (MPH) Model is proposed to detect the drift and to determine the type of this drift. Common performance measures are used to evaluate detected drift by assuming a change point can be considered the “positive” class while no change can be considered the “negative” class. These models are formed the learning operation for concept drift. The system is applied on two real dataset (PAMAP2) for physical activity and (ELEC2) for electricity. Artificial datasets, also are used to evaluate the system performance. These dataset are SEA, SIN, moving hyper plan, circle and wave generation dataset. Accordingly, the results show that the performance of proposed system is better with time accuracy has reached to 96% when compared with previous studies and the MPH method can detect all drift types.
Nabeel Al-A’araji, Eman Al-Shamery and Alyaa Abdul-Hussein, 2018. ARNN for Enhancing Drift Detection of Data Stream Based on Modified Page Hinckley Model. Journal of Engineering and Applied Sciences, 13: 8281-8291.