MDL-based segmentation of multi-attribute sequences

Many real-life multi-attribute sequences (multi-sequences) have a segmental structure, with segments of differing structures of attribute dependencies, that reflect an evolving nature of the dependencies over time and space. We propose a new approach for discovering a segmental structure of such evolving dependencies in probabilistic terms as a sequence of Dynamic Bayesian Networks (DBN). We use the Minimum Description Length (MDL) Principle to partition the multi-sequence into non-overlapping and homogeneous segments by fitting an optimal sequence of DBNs to the multi-sequence. In experiments, conducted on daily rainfall data we showed the applicability of the method for discovering interesting spatio-temporal evolving dependencies between rainfall occurrences in south-western Australia.