Bayesian network parameter learning with constraint balance

Maximum likelihood estimation (MLE) is a method for data modeling in artificial intelligence, which is intended to pursue the best fit of data and parameters. When the training data are insufficient, the model parameters that rely only on MLE for learning are unreliable. Similarly, for Bayesian networks (BNs), a crucial tool for uncertainty representation, the problem of low accuracy of parameter learning caused by insufficient data is also faced. Due to their clear structure and interpretability, BN can effectively fuse the parameter constraints transformed from prior knowledge to improve learning accuracy using the maximum a posteriori estimation (MAP). However, balancing data and constraints evolves into a new concern. With too much reliance on constraints, the data loses meaning, and the obtained prior knowledge can not be disregarded. Therefore, this paper proposes a constraint-balanced maximum a posteriori estimation method, which performs parameter learning by balancing the roles played by data and constraints in the parameter learning process. Through experiments, the proposed method can effectively improve the accuracy of parameter learning.