Advanced deep graph neural networks for explainable anomaly detection study

本财年共取得了三项成果：颗粒数据描述、颗粒ADD算法及其应用。第一个成果是关于数据粒度描述，我们开发了先进的数据描述粒度计算，旨在降低大数据分析的计算成本。特别是构造了结构颗粒，用于数据分布描述和数据结构挖掘。第二部分是关于异常数据检测算法。利用构造的颗粒，特别是具有描述数据结构能力的颗粒，构建了基于规则建模的初级异常数据检测方法。在带噪声的监督场景和无监督场景下，该算法的性能都得到了很好的评价。第三个成果是完成了一些与异常数据检测相关的工业应用。通过这些应用，验证了该方法在异常数据检测中考虑数据结构信息的有效性。

In this fiscal year, there are totally three achievements achieved, including granular data description, granular ADD algorithms and its application.The first achievement is about granular data description, we have developed advanced granular computing for data description, aiming to reduce the computation cost of big data analysis. Especially, structural granules are constructed for data distribution description and data structure mining in this task.The second one is about anomaly data detection algorithms. By making use of the constructed granules, especially ones having the ability of describing data structure, a primary anomaly data detection method via rule-based modeling method was built. Its performance was evaluated good on both supervised scenarios with noise and unsupervised scenarios.The third achievement is about completing some industrial applications related to anomaly data detection. Through these applications, it is verified useful by considering data structure information in anomaly data detection.