Discovery of Information Flow in Multivariate Dynamic Systems

Y. Tanokura and G. Kitagawa (Japan)


Knowledge Discovery in Databases, Detection of NoiseSource, Akaike's Power Contribution, Multivariate TimeSeries


Akaike's power contribution, a useful concept in detecting influential noise sources of multivariate dynamic systems with feedback, is not applicable to systems with significant correlations of the noise because of its assumption of inde pendence of the noise. To relax this assumption, we present a decomposition of the variance covariance matrix of the noise, modeling cross correlations, and define an extended power contribution that can be applied to general dynamic systems. The extended power contribution succeeded in discovering mutual information flows among variables. By applying this method to the ship and the stock index data, the newly discovered information flows were presented.

Important Links:

Go Back