Supporting Knowledge Discovery in Data Stream Management Systems

Data mining represents an exciting and vibrant area of research. In particular, on-line mining has gained significant momentum in recent years. The changing data characteristics and real-time response constraints of streaming data preclude the use of existing mining algorithms that were designed for stored datasets. Therefore, researchers are proposing new fast and light algorithms for on-line mining tasks, such as classification, clustering, frequent itemsets, pattern matching, and many others. Beyond the interesting research problems posed by the design of individual algorithms for different mining methods, there remains the issue that these must be integrated into an Inductive Data Stream Mining System that supports (i) libraries of interoperable mining methods, (ii) all essential functions of data stream management systems, such as continuous query optimization, load shedding, synoptic constructs, and non-stop computing, and (iii) ease-of-use and extensibility. The issue of Inductive Data Stream Mining System has received little attention in the past, and thus offers a pristine opportunity for research contributions in my thesis work. Furthermore, there will be an opportunity for advancing the state of the art in mining algorithms, particularly in terms of extensibility, interoperability, and genericity of data representation. In this prospectus, I briefly discuss data stream management systems, data mining methods and systems, and other areas closely related to the topic of this thesis. Then, I present recently obtained preliminary results, including data representations for achieving generic implementations for many on-line mining algorithms. Furthermore, I discuss advanced techniques such as ensemble-based bagging and boosting, for which generic implementations were also devised. Finally, preliminary experiments are presented to verify the efficiency of the proposed approach. Future work will focus on more experiments and fine-tuning of mining algorithms. Furthermore, support for high-level mining languages also represents a promising topic for future research.