Local Skeleton Discovery for Incremental Bayesian Network Structure Learning

Nowadays there are a huge number of applications produce the immense amount of data in the form of a data stream, which needs real time analysis. Sensor networks, real-time surveillance and telecommunication systems are the examples of such applications. The real time analysis of the data stream leads to a number of computational and mining challenges. In this scenario new data arrives continuously and an efficient learning algorithm must be able to improve its learning accuracy by incorporating the time and memory constraints. This paper addresses the problem of incremental Bayesian network structure learning for high dimensional domains. The local skeleton discovery methods for Bayesian network structure learning are outperforming to deal with such domains. Here we transformed the local discovery algorithm Max-Min Parents and Childrens (MMPC) into an incremental fashion. We learned a set of candidate-parent-children for each variable by using incremental hill-climbing. The reduced search space saves a lot of computations and reduces the complexity. Our algorithm is then illustrated with a toy example.