Home About Login Current Archives Announcements Editorial Board
Submit Now For Authors Call for Submissions Statistics Contact
Home > Archives > Volume 20, No 8 (2022) > Article

DOI: 10.14704/nq.2022.20.8.NQ44641

A Novel Hybrid Concept Drift Detector Ensemble for Handling Pure and Hybrid Drift in Non-Stationary Data Streams

Muralikrishnan Ramane, Gnanasekar J M

Abstract

Cognitive analytics deals with complex problems that arises due to dynamically evolvingenvironments. Online machine learning methodologies learn continuously from nonstationary and evolving data streams and can handle dynamism in near real time. Evolving data can lead to change or drift in learned concepts and data distributions due to population variations. The literature presents several drift detection techniques which identify drift as real or virtual and are not sophisticated to detect pure and hybrid drift types. Past methods monitor drift in a disconnected fashion and this disconnection leads to imprecise and redundant model update requests. The main objective of this work is to detect concept drift and differentiate its types precisely, thus improving the quality and robustness of the drift handling process. This work proposes a novel hybrid concept drift detector ensemble, that utilizes a connected drift detection method for detecting pure and hybrid drifts. Further, a parallel framework is introduced to monitor input data and learner output concurrently. Experiments were conducted separately to evaluate individual and connected techniques using synthetic datasets of one million samples containing abrupt and gradual drift. This work utilizes three leaner error-ratebased methods along with the 2-Sample Kolmogorov-Smirnov test for finding concept dissimilarities. The proposed hybrid technique utilizes parallel processing to reduce computation time and performs better in aspects of drift detection, drift type differentiation, and precise model update requests.

Keywords

Concept Drift Detection, Ensemble Technique, Machine Learning, Cognitive Systems

Full Text

PDF

References

?>