Classifying Evolving Data Streams

• Activity Number: 268 - Extreme Machine Learning Methods and Applications: Domestic and International
• Type: Topic Contributed
• Date/Time: Tuesday, August 4, 2020 : 1:00 PM to 2:50 PM
• Sponsor: Section on Statistical Consulting
• Author(s): Kelly Toppin* and Luca Sartore
• Companies: IMT Corporation and National Institute of Statistical Sciences
• Keywords: Classification; Data stream; Extreme machine learning

Abstract

Data stream mining is the process of extracting knowledge structures represented in models and patterns in continuous non-stop streams of data. Examples of data streams include sensor data, image data, Internet and web traffic. Unfortunately several characteristics of data streams hinder the successful extraction of knowledge. These hindrances include 1) infinite stream: this prevents we from knowing the entire dataset 2) Concept drift: the data evolves over time 3) computing and storage limitations often require that data from these streams be treated as volatile data. The infinite size and evolving nature of data streams are a challenge for traditional machine learning algorithms. This paper proposes an algorithm that deals with classification in data streaming environments i.e predicting the class of new instances in a data stream. This paper looks at the goal of predicting the class of new instances in a data stream. We propose an classification algorithm for streaming data that deals with the incompleteness of available training data. Furthermore our algorithm trains with limited amount of labeled data and is robust enough to handle the emergence of new classes in the stream.