Logo Logo
Help
Contact
Switch language to German
Monitoring data streams. Classification under uncertainty and entropy-based dependency-detection on streaming data
Monitoring data streams. Classification under uncertainty and entropy-based dependency-detection on streaming data
Stream monitoring is concerned with analyzing data that is represented in the form of infinite streams. This field has gained prominence in recent years, as streaming data is generated in increasing volume and dimension in a variety of areas. It finds application in connection with monitoring industrial sensors, "smart" technology like smart houses and smart cars, wearable devices used for medical and physiological monitoring, but also in environmental surveillance or finance. However, stream monitoring is a challenging task due to the diverse and changing nature of the streaming data, its high volume and high dimensionality with thousands of sensors producing streams with millions of measurements over short time spans. Automated, scalable and efficient analysis of these streams can help to keep track of important events, highlight relevant aspects and provide better insights into the monitored system. In this thesis, we propose techniques adapted to these tasks in supervised and unsupervised settings, in particular Stream Classification and Stream Dependency Monitoring. After a motivating introduction, we introduce concepts related to streaming data and discuss technological frameworks that have emerged to deal with streaming data in the second chapter of this thesis. We introduce the notion of information theoretical entropy as a useful basis for data monitoring in the third chapter. In the second part of the thesis, we present Probabilistic Hoeffding Trees, a novel approach towards stream classification. We will show how probabilistic learning greatly improves the flexibility of decision trees and their ability to adapt to changes in data streams. The general technique is applicable to a variety of classification models and fast to compute without significantly greater memory cost compared to regular Hoeffding Trees. We show that our technique achieves better or on-par results to current state-of-the-art tree classification models on a variety of large, synthetic and real life data sets. In the third part of the thesis, we concentrate on unsupervised monitoring of data streams. We will use mutual information as entropic measure to identify the most important relationships in a monitored system. By using the powerful concept of mutual information we can, first, capture relevant aspects in a great variety of data sources with different underlying concepts and possible relationships and, second, analyze theoretical and computational complexity. We present the MID and DIMID algorithms. They perform extremely efficient on high dimensional data streams and provide accurate results, outperforming state-of-the-art algorithms for dependency monitoring. In the fourth part of this thesis, we introduce delayed relationships as a further feature in the dependency analysis. In reality, the phenomena monitored by e.g. some type of sensor might depend on another, but measurable effects can be delayed. This delay might be due to technical reasons, i.e. different stream processing speeds, or because the effects actually appear delayed over time. We present Loglag, the first algorithm that monitors dependency with respect to an optimal delay. It utilizes several approximation techniques to achieve competitive resource requirements. We demonstrate its scalability and accuracy on real world data, and also give theoretical guarantees to its accuracy.
Not available
Boidol, Jonathan
2017
English
Universitätsbibliothek der Ludwig-Maximilians-Universität München
Boidol, Jonathan (2017): Monitoring data streams: Classification under uncertainty and entropy-based dependency-detection on streaming data. Dissertation, LMU München: Faculty of Mathematics, Computer Science and Statistics
[img]
Preview
PDF
Boidol_Jonathan.pdf

1MB

Abstract

Stream monitoring is concerned with analyzing data that is represented in the form of infinite streams. This field has gained prominence in recent years, as streaming data is generated in increasing volume and dimension in a variety of areas. It finds application in connection with monitoring industrial sensors, "smart" technology like smart houses and smart cars, wearable devices used for medical and physiological monitoring, but also in environmental surveillance or finance. However, stream monitoring is a challenging task due to the diverse and changing nature of the streaming data, its high volume and high dimensionality with thousands of sensors producing streams with millions of measurements over short time spans. Automated, scalable and efficient analysis of these streams can help to keep track of important events, highlight relevant aspects and provide better insights into the monitored system. In this thesis, we propose techniques adapted to these tasks in supervised and unsupervised settings, in particular Stream Classification and Stream Dependency Monitoring. After a motivating introduction, we introduce concepts related to streaming data and discuss technological frameworks that have emerged to deal with streaming data in the second chapter of this thesis. We introduce the notion of information theoretical entropy as a useful basis for data monitoring in the third chapter. In the second part of the thesis, we present Probabilistic Hoeffding Trees, a novel approach towards stream classification. We will show how probabilistic learning greatly improves the flexibility of decision trees and their ability to adapt to changes in data streams. The general technique is applicable to a variety of classification models and fast to compute without significantly greater memory cost compared to regular Hoeffding Trees. We show that our technique achieves better or on-par results to current state-of-the-art tree classification models on a variety of large, synthetic and real life data sets. In the third part of the thesis, we concentrate on unsupervised monitoring of data streams. We will use mutual information as entropic measure to identify the most important relationships in a monitored system. By using the powerful concept of mutual information we can, first, capture relevant aspects in a great variety of data sources with different underlying concepts and possible relationships and, second, analyze theoretical and computational complexity. We present the MID and DIMID algorithms. They perform extremely efficient on high dimensional data streams and provide accurate results, outperforming state-of-the-art algorithms for dependency monitoring. In the fourth part of this thesis, we introduce delayed relationships as a further feature in the dependency analysis. In reality, the phenomena monitored by e.g. some type of sensor might depend on another, but measurable effects can be delayed. This delay might be due to technical reasons, i.e. different stream processing speeds, or because the effects actually appear delayed over time. We present Loglag, the first algorithm that monitors dependency with respect to an optimal delay. It utilizes several approximation techniques to achieve competitive resource requirements. We demonstrate its scalability and accuracy on real world data, and also give theoretical guarantees to its accuracy.