Learning Bayesian Network Structure from Distributed Homogeneous Data

In this paper, we propose an algorithm: parallel three-phase dependency analysis (P-TPDA), for learning the structure of Bayesian network from distributed homogenous datasets: each of which has same variables. The algorithm has two steps: local learning and global learning. In local learning, we first obtain local Bayesian networks on each dataset independently using Bayesian network power constructor system. Then in global learning, we combine those local structures into the final structure with conditional independency (CI) test. The simulated experimental results for alarm networks indicate: when the number of records in dataset is more than 10000, the final structure obtained with P-TPDA algorithm is consistent with the structure obtained with centralized solution. But the running time in P-TPDA algorithm is shorter than the running time in centralized solution.