The GC3 framework grid density based clustering for classification of streaming data with concept drift

Data mining is the process of discovering patterns in large sets of data. In recent years there has been a paradigm shift in how the data is viewed. Instead of considering the data as static and available in databases, data is now regarded as a stream as it continuously flows into the system. One of the challenges posed by the stream is its dynamic nature, which leads to a phenomenon known as Concept Drift. This causes a need for stream mining algorithms which are adaptive incremental learners capable of evolving and adjusting to the changes in the stream.;Several models have been developed to deal with Concept Drift. These systems are discussed in this thesis and a new system, the GC3 framework is proposed. The GC3 framework leverages the advantages of the Gris Density based Clustering and the Ensemble based classifiers for streaming data, to be able to detect the cause of the drift and deal with it accordingly. In order to demonstrate the functionality and performance of the framework a synthetic data stream called the TJSS stream is developed, which embodies a variety of drift scenarios, and the model's behavior is analyzed over time.;Experimental evaluation with the synthetic stream and two real world datasets demonstrated high prediction capability of the proposed system with a small ensemble size and labeling ratio. Comparison of the methodology with a traditional static model with no drifts detection capability and with existing ensemble techniques for stream classification, showed promising results. Also, the analysis of data structures maintained by the framework provided interpretability into the dynamics of the drift over time. The experimentation analysis of the GC3 framework shows it to be promising for use in dynamic drifting environments where concepts can be incrementally learned in the presence of only partially labeled data.