学术报告：孙毅 新疆大学

主讲人：孙毅 新疆大学

题 目： Model Decomposition for Bayesian Networks with its Applications in Local and Parallized Inference

时间：2025年10月18日 14:00-15:30

地点：VSport体育官网新校园 B514

摘要：In high-dimensional Bayesian networks, computing the joint distribution for inference poses a significant challenge. To address this issue, we propose a decomposition framework based on an optimal d-decomposition tree, which decompose the network into parallelizable subgraphs. Through this transformation, the estimation of joint distributions in high-dimensional networks is simplified into parameter learning and inference tasks on lower-dimensional subnetworks.

Within this framework, we design two algorithms for parallel parameter estimation and probabilistic inference. Experimental results demonstrate that, compared to variable elimination and belief propagation, our approach achieves substantial improvements in computational efficiency while maintaining inference accuracy.

As the network size increases, the acceleration effect of the decomposition becomes more significant. Concurrently, the distributional distance between the estimated network and the true network diminishes rapidly as the sample size increases. when the sample size reaches 10,000, the difference between them becomes negligible. In probabilistic inference, our method achieves up to an 8× speed-up over conventional techniques.