Naïve Bayes Classification Ensembles to Support Modeling Decisions in Data Stream Mining

Data stream mining is the process of applying data mining methods to a data stream in real-time in order to create descriptive or predictive models. Due to the dynamic nature of data streams, new classes may emerge as a data stream evolves, and the concept being modeled may change with time. This gives rise to the need to continuously make revisions to the predictive model. Revising the predictive model requires that labeled training data should be available. Manual labeling of training data may not be able to cope with the speed at which data needs to be labeled. This paper proposes a predictive modeling framework which supports two of the common decisions that need to be made in stream mining. These decisions are: (1) determining when model revision should be performed and (2) deciding which newly arrived instances should be used as training data. The framework consists of an online component and an offline component. The online component uses Naïve Bayes ensemble base models to make predictions for newly arrived data stream instances. The offline component consists of algorithms to combine base model predictions, determine the reliability of the ensemble predictions, select training data for new base models, create new base models, and determine whether the current online base models need to be replaced.