Hybrid Chain Graphs for Causal and Symmetric Dependencies

This direction is split into other directions: hybrid dependency modeling. Its last version stays readable and citable here.

Version 1 of 135 papers45 conceptsfrozen 2026-10-04

In traditional probabilistic modeling, researchers usually have to choose between two worlds. If they want to model causal "flow"—where $A$ causes $B$—they use Directed Acyclic Graphs (DAGs). If they want to model symmetric relationships or equilibrium states—where $A$ and $B$ are simply associated or influence each other simultaneously—they use undirected Markov Random Fields (MRFs).

However, real-world systems rarely fit neatly into one of these boxes. In a social network, an external policy might cause a change in behavior (directed), but individuals also influence their peers in a reciprocal loop (undirected). Wray Buntine argues in Chain Graphs for Learning that forcing these complex systems into purely directed or undirected models results in a loss of information. The solution is the Chain Graph, a hybrid structure that allows both directed and undirected edges to coexist.

The Intuition of Mixed Dependencies

The core power of a chain graph lies in its ability to represent "blocks" of variables. Within a block, variables are connected by undirected edges, representing symmetric dependencies or mutual influence. Between these blocks, directed edges represent a clear hierarchical or causal progression.

This hybrid approach is essential because, as Jose M. Peña demonstrates in Approximate Counting of Graphical Models Via MCMC, the vast majority of independence structures in complex networks cannot be simplified into purely directed or undirected formats. By using chain graphs, we can model systems where causal drivers (like drug dosage) coexist with complex, non-causal correlations among effects (like various physiological symptoms).

Causal Inference Under Interference

One of the most significant applications of this research is in causal inference, particularly when the "no interference" assumption fails. Standard causal models assume that treating one person doesn't affect another. In social networks or epidemiology, this is rarely true.

Elizabeth L. Ogburn and Youjin Lee show in Causal Inference, Social Networks and Chain Graphs that chain graphs provide a mathematically rigorous way to model "social contagion." By using directed edges for external exposures and undirected edges for reciprocal peer interactions, researchers can estimate causal effects even when outcomes "spill over" between connected individuals.

This direction has been further refined into Segregated Graphs. As Rohit Bhattacharya and Ilya Shpitser explain in Causal Inference Under Interference And Network Uncertainty, these models add a third edge type—bidirected edges—to represent latent confounding (unobserved common causes). This triple-threat representation (directed, undirected, and bidirected) allows for the identification of causal effects in systems where units are both interconnected and influenced by hidden variables.

Learning the Structure of Hybrid Systems

While chain graphs are more expressive, they are also harder to learn from data. In a DAG, the "Markov blanket" of a node is its parents, children, and children's parents. In a chain graph, this boundary is more complex. Mohammad Ali Javidian and Marco Valtorta address this in Learning LWF Chain Graphs: A Markov Blanket Discovery Approach, defining the blanket as a union of the node's boundary, its children, and its "complex-spouses."

The difficulty lies in the fact that many different graph structures can represent the same set of statistical independencies. To solve this, An inclusion optimal algorithm for chain graph structure learning introduced the CKES algorithm. This was a breakthrough because it proved that we could find an "inclusion optimal" structure—the simplest graph that fits the data—without needing the overly restrictive "faithfulness" assumptions required by earlier methods.

Why It Works: Expressivity and Robustness

The reason this research direction is gaining traction is its physical and mathematical robustness. José M. Peña proved in Faithfulness in Chain Graphs: The Gaussian Case that for almost all Gaussian distributions that fit a chain graph, the graph is a "faithful" representation of the underlying reality. This means the unique independence models captured by these hybrid graphs are not just mathematical curiosities; they are robust features of the data.

By bridging the gap between directed causal flow and undirected relational constraints, hybrid chain graphs allow us to model the world as it is: a messy, interconnected system where things cause other things, but everything also influences its neighbors.

Papers in this version

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_hybrid_chain_graphs_for_causal_and_symme_acb66339,
  title = {Hybrid Chain Graphs for Causal and Symmetric Dependencies},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hybrid-chain-graphs-for-causal-and-symmetric-dependencies/rdr_acb66339ebd145f38946663005ba80cb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_hybrid_chain_graphs_for_causal_and_symme_acb66339_v1,
  title = {Hybrid Chain Graphs for Causal and Symmetric Dependencies},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hybrid-chain-graphs-for-causal-and-symmetric-dependencies/rdr_acb66339ebd145f38946663005ba80cb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}