Concept drift adaptation for learning with streaming data

Publication Type:
Thesis
Issue Date:
2018
Full metadata record
Files in This Item:
Filename Description Size
01front.pdf112.55 kB
Adobe PDF
02whole.pdf2.92 MB
Adobe PDF
The term concept drift refers to the change of distribution underlying the data. It is an inherent property of evolving data streams. Concept drift detection and adaptation has been considered an important component of learning under evolving data streams and has attracted increasing attention in recent years. According to the existing literature, the most commonly used definition of concept drift is constrained to discrete feature space. The categorization of concept drift is complicated and has limited contribution to solving concept drift problems. As a result, there is a gap to uniformly describe concept drift for both discrete and continuous feature space, and to be a guideline to addressing the root causes of concept drift. The objective of existing concept drift handling methods mainly focuses on identifying when is the best time to intercept training samples from data streams to construct the cleanest concept. Most only consider concept drift as a time-related distribution change, and are disinterested in the spatial information related to the drift. As a result, if a drift detection or adaptation method does not have spatial information regarding the drift regions, it can only update learning models or their training dataset in terms of time-related information, which may result in an incomplete model update or unnecessary training data reduction. In particular, if a false alarm is raised, updating the entire training set is costly and may degrade the overall performance of the learners. For the same reason, any regional drifts, before becoming globally significant, will not trigger the adaptation process and will result in a delay in the drift detection process. These disadvantages limit the accuracy of machine learning under evolving data streams. To better address concept drift problems, this thesis proposes a novel Regional Drift Adaptation (RDA) framework that introduces spatial-related information into concept drift detection and adaptation. In other words, RDA-based algorithms consider both time-related and spatial information for concept drift handling (concept drift handling includes both drift detection and adaptation). In this thesis, a formal definition of regional drift is given which has theoretically proved that any types of concept drift can be represented as a set of regional drifts. According to these findings, a series of regional drift-oriented drift adaptation algorithms have been developed, including the Nearest Neighbor-based Density Variation Identification (NN-DVI) algorithm which focuses on improving concept drift detection accuracy, the Local Drift Degree-based Density Synchronization Drift Adaptation (LDD-DSDA) algorithm which focuses on boosting the performance of learners with concept drift adaptation, and the online Regional Drift Adaptation (online-RDA) algorithm which incrementally solves concept drift problems quickly and with limited storage requirements. Finally, an extensive evaluation on various benchmarks, consisting of both synthetic and real-world data streams, was conducted. The competitive results underline the effectiveness of RDA in relation to concept drift handling. To conclude, this thesis targets an urgent issue in modern machine learning research. The approach taken in the thesis of building regional concept drift detection and adaptation system is novel. There has previously been no systematic study on handling concept drift from spatial prespective. The findings of this thesis contribute to both scientific research and practical applications.
Please use this identifier to cite or link to this item: