Introduction
Discovering causal relationships from observational data is a cornerstone of scientific inquiry. While traditional methods focus on scalar variables, real-world data often exhibits a grouped structure—where multiple measurements represent a single entity or process. This paper introduces GroupRESIT, a framework designed to identify causal links between these high-dimensional random vectors.
The Challenge of Grouped Data
Standard causal discovery algorithms typically treat variables as independent scalars. In complex systems like neuroscience (brain regions) or manufacturing (sensor clusters), this approach either loses information through averaging or ignores the inherent group structure. GroupRESIT addresses this by modeling interactions directly between vectors, ensuring mathematical identifiability in nonlinear settings.
Methodology: The GroupRESIT Framework
GroupRESIT operates in two distinct phases:
- Phase I: Causal Ordering. The algorithm determines the sequence of causal influence. It employs multi-output Deep Neural Networks (MLPs) to regress each group against others, using the Hilbert-Schmidt Independence Criterion (HSIC) to evaluate the independence of residuals. By iteratively identifying the "least dependent" group, it establishes a causal order.
- Phase II: DAG Pruning. To refine the causal graph, the authors introduce MURGS (Multi-Response Group Sparse Additive Models). This novel feature selection method uses a specialized regularization functional to enforce group-level sparsity, effectively removing spurious connections. A major contribution is the derivation of a closed-form backfitting update for MURGS, making the optimization efficient for practical use.
Experimental Results
The framework was rigorously tested against synthetic and real-world datasets. In synthetic benchmarks, GRESIT-MURGS demonstrated superior accuracy, particularly in the Ancestor Adjustment Identification Distance (AAID) metric, which measures how well the algorithm captures the overall causal flow.
The method's robustness was further validated using manufacturing data from Robert Bosch GmbH. GroupRESIT successfully reconstructed a causal graph that closely mirrors the physical layout of the assembly line, a feat where competing methods like GGraN-DAG and GroupPC struggled. While other methods produced too many undirected edges or failed to find connections, GroupRESIT produced a meaningful graph matching the factory's physical process cells.
Conclusion
GroupRESIT provides a robust, mathematically grounded approach to causal discovery in grouped data. By leveraging deep learning for ordering and MURGS for pruning, it captures complex nonlinear interactions without the information loss typical of dimensionality reduction. Future research directions include extending the framework to handle cycles (feedback loops) and latent confounders, which are common in complex biological and social systems.
Summary updated