A Novel HOSFS Algorithm for Online Streaming Feature Selection

Abstract

In recent days, Data stream mining is important for many of the real time and IOT based applications. Online feature selection is the one big topic of data stream mining which attracted researchers with intensive interest. This technique reduces the dimensionality of the streaming features by excluding inappropriate and redundant features. The researchers have proposed many online feature selection algorithm for streaming features like Grafting, Alpha-investing, OSFS, OGFS and SAOLA. Based on above studies the exiting algorithm has limitation over prediction accuracy and the large number of selected features. To overcome the limitations of above mentioned approaches, we propose an online feature selection algorithm for streaming features called Heuristic Online Streaming Feature Selection (HOSFS) which has advantages on choosing features from streaming features and omits the irrelevant and redundant features in real-time by using self-adaption sliding window protocol, and Heuristic function. The HOSFS algorithm assigns heuristic value to the features using the trained heuristic function and selects features with higher heuristic value where other features are considered as irrelevant features. This proposed technique results reduced number of strongly related features and obtains greater prediction accuracy with optimal features. HOSFS algorithm efficiency was tested with three different Health care datasets using MOA tools. Through the experimental outcomes, HOSFS has greater prediction accuracy and reduced number of selected features than alpha - investing, OSFS, and SAOLA.