Learning interactions among multi-channel sequences with dynamical influence models

Abstract

Many real applications involve simultaneous recording and analysis of multi-channel information sources. Learning and modeling the interactions among channels is the kernel step to analyze and recognize system characteristics. This paper presents a model that learns the dynamical influence among multi-channel sequences. The model, dynamical influence model, permits functional roles of individual channels to change and models the changing influence strength between channels. By querying the values of influence factors, we can recognize the functional role of each channel qualitatively and learn about to what extent the chains influence each other quantitatively at any time. The experimental results on synthetic data and application of multi-person interaction recognition show that our model is reliable and effective.