Learning temporal causal graphs for relational time-series analysis

Learning temporal causal graph structures from multivariate time-series data reveals important dependency relationships between current observations and histories, and provides a better understanding of complex systems. In this paper, we examine learning tasks where one is presented with multiple multivariate time-series, as well as a relational graph among the different time-series. We propose an L1 regularized hidden Markov random field regression framework to leverage the information provided by the relational graph and jointly infer more accurate temporal causal structures for all time-series. We test the proposed model on climate modeling and cross-species microarray data analysis applications. Copyright 2010 by the author(s)/owner(s).