Discovering causal relationships in mixed-type non-Euclidean data with applications to fault diagnosis

There has been a growing interest in developing machine learning algorithms that can handle non-Euclidean data. We introduce a causal generating process between parent nodes and child nodes based on multivariate tensor regression. Additionally, we propose a two-stage causal discovery approach involving regularized generalized canonical correlation analysis and greedy hill-climbing search. By utilizing numerical representation in the shared Euclidean subspace, we are able to more accurately discover causal relationships between heterogeneous non-Euclidean variables. The effectiveness of the algorithm is demonstrated using a dataset of mixed functional and compositional data, as well as empirical research conducted on real-world industrial sensor data.