Explanatory Analysis of Probabilistic Graphical Models based on Discriminative Pattern Mining

在这个项目中，我们开发了一种概率模型的解释性分析方法，特别是贝叶斯网络，以增加机器学习技术的解释性。我们的分析旨在从大量可能的解释中为观察到的现象找到一个合适的解释。为了在实际应用中做到这一点，我们基于一种流行的频繁模式挖掘算法FP-Growth，建立了一些复杂的区分模式挖掘技术。最后，我们实现了对解释选择标准的提炼，并实现了一种快速的区分模式挖掘算法。尽管在优化我们的解释性分析的概率推理方面仍有进一步的工作，但我们已经获得了几个新的见解和原型工具来实现我们的解释性分析方法。

In this project, we have developed an explanatory analysis method for probabilistic models, Bayesian networks in paticular, to bring explanability to machine learning techniques. Our analysis aims to find an appropriate explanation for the observation from a huge number of possible ones. To do this in practical time, we built some sophisticated techniques for discriminative pattern mining based on a popular frequent pattern mining algorithm called FP-Growth. Finally, we have achieved to refine the selection criteria of explanations and to have a fast discriminative pattern mining algorithm. Although there remains a future work on optimizing probabilistic inference for our explanatory analysis, we have obtained a couple of new insights and prototype tools towards an implementation of our explanatory analysis method.