SGER: A Framework for Explanation in Bayesian Networks

现有的基于贝叶斯网络的诊断方法通常要么产生不充分的解释，不能完全解释给定的观察结果，要么产生过度的解释，其中包含解释观察结果时不必要的变量。本项目通过阐述一个新的框架——最相关解释（MRE）来解决这些限制，该框架用于贝叶斯网络中的解释，可以自动识别给定观测值中最相关的故障。特别是，该项目包括基于MRE框架的多故障近似、高效诊断算法的开发和评估。由于诊断很可能是贝叶斯网络最成功的应用，基于mre的方法将通过提高医学和制造业等领域的可理解性和效率，显著改善当前计算机辅助诊断的实践。

Existing Bayesian network-based diagnostic approaches often produce either underspecified explanations that do not fully account for given observations, or overspecified explanations that contain variables unnecessary in explaining the observations. This project addresses these limitations by elaborating a new framework, Most Relevant Explanation (MRE), for explanation in Bayesian networks that can automatically identify the most relevant fault(s) for given observations. In particular, the project includes development and evaluation of approximate, efficient diagnostic algorithms for multiple faults based on the MRE framework. Because diagnosis is quite likely the most successful application of Bayesian networks, MRE-based methodologies will significantly improve the current practice of computer-aided diagnosis by improving the comprehensibility and efficiency in areas such as medicine and manufacturing.