Supplementing Markov Chains with Additional Features for Behavioural Analysis

The combination of decision structures has been proposed by numerous researchers in the behavioural analysis domain, and been shown to improve accuracy over network structures tested in isolation; however, the vast majority of researchers use a simplistic combination strategy, amounting to little more than bridging of the network structures. This paper introduces a fusion mechanism between Bayesian and Markovian networks, which provides the Markov states with additional low-level features. This hybrid approach affords users a simplified network structure and provides the basis for an automatic technique, which allows the transitional probabilities in the network to be learned during online running of the system without the need for a training phase. The hybrid technique is validated using two very different datasets and is shown to outperform a standard network approach tested.