Based on the independence of the causal influence model and the context-specific independent relations arising in the model, a parameters learning algorithm of Bayesian networks suiting for sample learning is proposed. Through decomposing and dimension-reducing the local probability model, the algorithm can synthesize the conditional probability distribution of the local structure with the probability distribution of the child nodes under the single parent node. The algorithm has low parameter-defining complexity and can better deal with the sparse structure sample set. Compared with the standard maximum likelihood estimation algorithm, the experimental results show that the proposed algorithm can fully extract the information from sample data and has higher learning accuracy.