Directed graphical models for root cause analysis in multimodal systems collecting heterogeneous data
提出tensorDGM框架，通过带惩罚的张量回归学习有向图模型结构，用于多模态异构数据系统的根因分析，解决现有方法难以处理混合数据类型的问题。
Nowadays, complex systems are equipped with multiple sensors that collect high-dimensional and heterogeneous data, including scalars, functional signals or profiles, and images. Such systems are prone to failures, and diagnosis is hard as the intricate interactions among the variables mask the root causes for fault diagnosis. Directed graphical models (DGMs) are a powerful tool for representing the probabilistic relationships between variables in such systems. Nonetheless, frameworks learning DGMs with mixed types of data are scarce. Existing methods focus on either scalar variables or profile data, or have stringent assumptions. In this article, we propose a framework, tensorDGM, for all data types and learn the structure of the system for root-cause analysis. We learn the structure of the DGM by fitting penalized tensor-on-tensor regressions with group Lasso penalties for variable selection and an L2 penalty to handle the group selection of variables. To solve the optimization problem, we propose a cyclic coordinate accelerated proximal gradient descent algorithm with Bayesian optimization for hyperparameter tuning. Through simulations and a case study, we illustrate the advantages of tensorDGM.