Use this URL to cite or link to this record in EThOS:
Title: Dynamic data mining : methodology and algorithms
Author: Deng, Xiong
ISNI:       0000 0004 2705 0008
Awarding Body: Imperial College London
Current Institution: Imperial College London
Date of Award: 2010
Availability of Full Text:
Access from EThOS:
Access from Institution:
Supervised data stream mining has become an important and challenging data mining task in modern organizations. The key challenges are threefold: (1) a possibly infinite number of streaming examples and time-critical analysis constraints; (2) concept drift; and (3) skewed data distributions. To address these three challenges, this thesis proposes the novel dynamic data mining (DDM) methodology by effectively applying supervised ensemble models to data stream mining. DDM can be loosely defined as categorization-organization-selection of supervised ensemble models. It is inspired by the idea that although the underlying concepts in a data stream are time-varying, their distinctions can be identified. Therefore, the models trained on the distinct concepts can be dynamically selected in order to classify incoming examples of similar concepts. First, following the general paradigm of DDM, we examine the different concept-drifting stream mining scenarios and propose corresponding effective and efficient data mining algorithms. • To address concept drift caused merely by changes of variable distributions, which we term pseudo concept drift, base models built on categorized streaming data are organized and selected in line with their corresponding variable distribution characteristics. • To address concept drift caused by changes of variable and class joint distributions, which we term true concept drift, an effective data categorization scheme is introduced. A group of working models is dynamically organized and selected for reacting to the drifting concept. Secondly, we introduce an integration stream mining framework, enabling the paradigm advocated by DDM to be widely applicable for other stream mining problems. Therefore, we are able to introduce easily six effective algorithms for mining data streams with skewed class distributions. In addition, we also introduce a new ensemble model approach for batch learning, following the same methodology. Both theoretical and empirical studies demonstrate its effectiveness. Future work would be targeted at improving the effectiveness and efficiency of the proposed algorithms. Meantime, we would explore the possibilities of using the integration framework to solve other open stream mining research problems.
Supervisor: Guo, Yike ; Ghanem, Moustafa Sponsor: UK/China Scholarships for Excellence Programme
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral