Transactions on Machine Learning and Data Mining (ISSN: 1865-6781)
Volume 3 - Number 1 - July 2010
Incremental Learning of Statistical Models from a Temporal Data Stream – An MML-Based Approach
J. Feng, A. Attig, M. Schwarz and P. Perner
Institute of Computer Vision and Applied Computer Sciences, Leipzig, Germany
Classifying data into groups is an important task in Data Mining. The Bayesian classifier is one method of choice as a classification model. If all data are available at the outset we can use that data and develop the classifier. The number of classes might be know a-priori or might be derived from the data by clustering. Another method of choice is the estimation of the right number of classes based on the minimum message length principle (MML). If the data stream arrives in temporal sequence, we must incrementally update the classifier model based on the newly arriving data and the MML principle. In this paper, we describe the development of the classification model and how we can update the model incrementally based on the new data and the MML principle.