Back to search

University of Minnesota

Graph-Based Causal Topology Learning for Complex Dynamical Systems under Observational Non-Idealities

Abstract

dc:description.abstract

Complex dynamical systems are often understood through their topology: the pattern of connections, dependencies, or influences among interacting variables. These graph-based representations can support process monitoring, fault diagnosis, control, and biological interpretation. However, topology learned from observational data depends on how the system is measured. Temporal aggregation, subsampling, unresolved fast interactions, delayed propagation, feedback, self-dependence, latent variables, and finite data can alter both the accuracy and the meaning of inferred edges. This dissertation develops a framework for learning, refining, and interpreting graph-based topology under these realistic non-idealities, using industrial chemical processes as mechanistically inspectable testbeds and biological systems as applications with limited ground truth. First, established causal discovery methods are evaluated on chemical-process systems containing contemporaneous and lagged effects. Analyzing these effects separately can produce spurious relations, while temporal aggregation and subsampling change the graph that is identifiable from the data. Sufficiently aggregated dynamics can resemble a contemporaneous structural-equation model, whereas subsampled observations may encode ancestral rather than direct causal relations. Latent disturbances can also generate misleading observed-variable connections; Wiener-separation-based conditional independence combined with Fast Causal Inference provides a way to represent signatures of unobserved confounding. These findings motivate the main methodological contribution of the dissertation. A lagged directed mixed graph formalizes observed-variable mixed-delay structure after latent variables are marginalized, and a lagged directed partial ancestral graph (LDPAG) provides its estimable partial representation. Its adjacencies encode lag-aligned inducing-path relations rather than automatically direct mechanisms. The Pruning-Sanitation-Recovery (PSR) algorithm refines candidate LDPAGs produced by existing discovery methods: pruning removes unsupported adjacencies, sanitation weakens endpoint claims no longer supported by the refined skeleton, and recovery restores endpoint information justified by collider and orientation evidence. For candidate graphs formed by pairing separate contemporaneous and temporal methods, PSR increases combined skeleton F1 from 0.473 to 0.922 in a synthetic benchmark and from 0.720 to 0.913 in a chemical-process benchmark. It also increases correct definite endpoint and source marks and improves recovery of latent-confounding signatures. Hybrid methods generally begin closer to the reference graph and receive smaller but consistent improvements. Finally, the graph-based perspective is extended to biological systems. Sparse-feedback control of structural and functional brain networks shows that, under high feedback cost, control is concentrated in a small subset of highly connected regions. In resting-state functional magnetic resonance imaging from 260 analyzed participants, a restricted conditional-pruning and partial-orientation procedure removes an average of 55.2% of candidate edges, particularly at positive lags, while preferentially preserving within-network and homotopic relationships and increasing modularity. Spatially thinned causal discovery applied to gene-centered regulatory features in Calu-3 and A549 identifies a reproducible feature-class backbone, including persistent enhancer--transposable-element and associated-promoter--gene adjacencies, while exact endpoint orientations remain less stable than adjacencies across replicates, genomic windows, and datasets. Collectively, these results show that useful topological information can be extracted from observational data only when the graph representation, discovery procedure, validation strategy, and interpretation are matched to the observation process and the available evidence. Across domains, adjacency structure is generally more consistently supported than exact endpoint direction, making partial graphs meaningful scientific outputs rather than incomplete substitutes for fully directed mechanisms.

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Dewantoro, Harman

Subjects

dc:subject × 5

Rights

Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Citation
Dewantoro, Harman. (2026). Graph-Based Causal Topology Learning for Complex Dynamical Systems under Observational Non-Idealities. Retrieved from the University Digital Conservancy, https://hdl.handle.net/11299/283077.

Chain of custody

source
Harvested from
University of Minnesota
Base URL
conservancy.umn.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Dewantoro, Harman. Graph-Based Causal Topology Learning for Complex Dynamical Systems under Observational Non-Idealities. 2026. https://hdl.handle.net/11299/283077