Overview of Concept Drifts Detection Methodology in Data Stream

Overview of Concept Drifts Detection Methodology in Data Stream

Shabina Sayed (Jodhpur National University, India), Shoeb Ahemd Ansari (Shri Jagadish Prasad Jabnormal Tibrewala University, India) and Rakesh Poonia (Bikaner Government College of Engineering, India)
DOI: 10.4018/978-1-5225-3870-7.ch018

Abstract

Real-time online applications and mobile data generate huge volume of data. There is a need to process this data into compact data structures and extract meaningful information. A number of approaches have been proposed in literature to overcome the issues of data stream mining. This chapter summarizes various issues and application techniques. The chapter is a guideline for research to identify the research issues and select the most appropriate method in order to detect and process novel class.
Chapter Preview
Top

Introduction

Consistently, enormous volumes of tactile, value-based, and web information are ceaselessly created as streams, which should be examined online as they arrive. The volumes of naturally produced information are continually expanding. As indicated by (Gantz and Reinsel, 2012) more than 2.8ZB of information were made and prepared in 2012, with an anticipated increment of 15 times by 2020. The different issues (Mark G et. al, 1999) in handling information stream are recorded in Table 1. The incredible development in the creation of computerized information comes about because of our encompassing condition being furnished with an ever-increasing number of sensors. Individuals conveying advanced mobile phones deliver information, database exchanges are being checked and put away, surges of information are removed from virtual conditions as logs or client produced. A noteworthy piece of such information is unpredictable, which implies it should be dissected progressively as it arrives. The appropriation creating the things of an information stream can change after some time. These changes, depending on the research area, are referred to as temporal evolution, covariate shift, nonstationary or concept drift (Bifet et.al, 2009). The principle qualities of the information stream display infer the accompanying limitations (Bifet, 2009):

  • It is difficult to store all the information from the information stream. Just little outlines of information streams can be processed and put away, and whatever is left of the data is discarded.

  • The incoming speed of information stream tuple compels every specific component to be handled essentially in real time and afterward disposed of.

  • The dispersion producing the things can change after some time. Along these lines, information from the past may wind up plainly insignificant or even unsafe for the current summary.

Top

Concept Drift

A normal for information stream is consistent stream of information. Information measure is to a great degree expansive and possibly unending. It's impractical to store all information. Be that as it may, significant issues identified with information stream mining are Boundless length, novel class advancement, and novel class float. Boundless length implies information stream have a vast length so require limitless length storage and training time. Concept evolution (John F et. al, 2008) implies creating novel class and Concept drift implies information changes after some time. The most prominent case to introduce the issue of concept drift is that of distinguishing and sifting through spam email. The refinement amongst undesirable and honest to goodness messages is user specific and advances with time. Novel class does not exist on the off chance that we expect no of classes are settled. Be that as it may, sometime information stream characterization issue happen like intrusion detection, content grouping, and fault detection. So this presumption is not substantial for real streaming environment, at the point when new classes might be develop whenever. Most existing information stream characterization procedure disregards this critical part of stream information, which is the landing of a Novel Class. Concept evolution takes care of the issue of endless length and novel class. As concept drift is thought to be erratic, occasional regularity is typically not considered as a novel class issue. As a special case, if regularity is not known with assurance, it may be viewed as a concept drift issue. The center supposition, when managing the concept drift issue, is vulnerability about what's to come. We expect that the wellspring of the objective occasion is not known with sureness. It can be accepted, evaluated, or anticipated, yet there is no assurance (ˇZliobait, 2010). (Mark G, 1999) displayed three routes in which concept drift may happen: A normal for information stream is nonstop stream of information. Information estimate is amazingly substantial and conceivably limitless. It's impractical to store all information. Be that as it may, significant issues identified with information stream mining are: Unending length, concept evolution, and concept drift. Unending length implies information stream have a vast length so require interminable length stockpiling and training time. Concept evolution (John F. et. al, 2008) implies creating novel class and drift may occur:

Key Terms in this Chapter

Ensemble Classifier: It consists of more than one classifier for data classification.

Single Classifier: It consists of only one classifier of any time for classifying the data.

Data Stream: It is continuous, real-time huge volume of data generated continuously from various online applications.

DSM: Is an abbreviation used for data stream mining.

Novel Class: It is defined as a newly discovered class that was previously unknown.

Complete Chapter List

Search this Book:
Reset