Data Stream: Statistics, Challenges, Concept Drift Detector Methods, Applications and Datasets

Abstract

Through real-world applications, the data stream is generated. In contrast to traditional data, data streams have different characteristics; a huge and endless amount, on-line arriving with high speed, single processing as well as the nature of being not static, in the sense that it evolves over time; this is the concept drift. Accordingly, the mining and analysis of the data stream is an arduous and attractive task. Various frameworks for data stream (analysis, mining, etc.) have been proposed over the past years. In the same context, identifying the number of classes of data streams is an important initial step when designing a model for processing the data stream. At present, deep neural networks (DNNs) are a fundamental technique in various applications. DNNs have many structures, including the multilayer perceptron (MLP). In this paper, we propose a deep neural network (DNN) model based on multilayer perceptron (MLP) to classify the streaming datasets and detect their classes as an analysis step for this data type. The proposed model tests different synthetic and real-world stream datasets. The results proved that this model detects the actual number of classes for the given stream dataset. Moreover, this paper presents a systematic review of data stream, its statistics, challenges, concept drift detector methods, data stream applications in different sectors in addition to the streaming datasets.