Causal inference promises a revolution in machine learning, moving models from passive pattern recognition to active intervention and counterfactual reasoning. However, a formidable bottleneck stands in the way: the structural knowledge requirement.
Traditional causal frameworks assume that a perfect Directed Acyclic Graph (DAG) is handed down by an omniscient domain expert. In this ideal scenario, every causal link and hidden confounder is mapped out with absolute certainty. But in the real world—whether dealing with complex biological systems, high-dimensional image data, or dynamic reinforcement learning environments—expecting an accurate, pre-defined causal graph is highly unrealistic.
To scale causal AI, researchers are developing methods to relax, bypass, or automate this structural requirement. Understanding why this bottleneck exists, and how modern frameworks circumvent it, is key to unlocking practical causal reasoning.
Why Causal Identification Demands a Graph
To understand why the structural knowledge requirement is so restrictive, we must look at how causal effects are mathematically calculated. In Pearl’s framework, estimating the effect of an intervention (written as $P(Y \mid \text{do}(X))$) requires transforming an interventional distribution into an observational one. This process is called identification.
Identification relies entirely on the topology of the graph. To block spurious correlations between a cause $X$ and an effect $Y$, we must identify and adjust for "backdoor paths"—sequences of edges that create non-causal associations. As Tian et al. demonstrated, identifying a causal effect under confounding requires analyzing the graph's confounded components (c-components). If you do not know the graph's structure, you cannot know which variables act as confounders (which must be conditioned on) or which act as colliders (which open up spurious paths if conditioned on). Without the graph, you are effectively flying blind.
This dependency ripples across every downstream application of causality:
- Domain Adaptation: To transfer knowledge from a source domain to a target domain, von Kügelgen notes that models must distinguish between invariant causal features and shifting effect features—a task that traditionally requires prior knowledge of the causal graph.
- Transportability: When moving causal insights across different environments, Bareinboim et al. showed that an accurate graph is necessary to pinpoint the exact locations where environmental mechanisms diverge.
- Explainable AI: In reinforcement learning, the framework proposed by Madumal et al. generates human-like explanations by tracing action influences through a structural causal model. However, the authors point out that relying on an expert-defined graph severely limits the agent's ability to generalize to novel environments where the dynamics are unknown.
Strategy 1: Reasoning Over Partial or Coarse Structures
If obtaining a complete, granular DAG is impossible, can we perform valid causal inference with incomplete structural knowledge? One prominent line of research answers yes, by changing the scale or certainty of the graph.
Coarse-Graining via Cluster DAGs
Instead of mapping the exact relationship between every single pair of variables, what if we group them? Anand et al. introduced Cluster DAGs (C-DAGs) to resolve this micro-level knowledge bottleneck.
The intuition is elegant: if a set of variables is highly interconnected and its internal mechanisms are messy or unknown, we can treat the entire set as a single macro-node (a cluster). The authors proved that as long as the causal relations between the clusters are well-defined, researchers can still perform valid interventional and counterfactual reasoning on the macro-level. By isolating structural ignorance inside the clusters, C-DAGs prevent local uncertainty from breaking global inference.
Traditional DAG (Requires exact micro-edges):
[A1] -> [A2] -> [B1] -> [B2]
Cluster DAG (Blurs internal micro-edges, preserves macro-flow):
( Cluster A ) ------> ( Cluster B )
( [A1] ?? [A2] ) ( [B1] ?? [B2] )
Do-Calculus on Equivalence Classes
Another approach accepts that data alone cannot pinpoint a single true graph, but can narrow the possibilities down to a Markov equivalence class—a set of graphs that all share the same conditional independence relationships.
Hyttinen bypassed the single-graph requirement by leveraging logic-based constraint solvers to evaluate Pearl's do-calculus rules across an entire equivalence class simultaneously. If a causal effect is mathematically identifiable under every graph within the class, the exact graph does not matter; the effect is uniquely determined. This allows researchers to safely estimate causal impacts even when the true direction of certain edges remains fundamentally underdetermined.
Strategy 2: Integrating Automated Causal Discovery
Rather than treating graph specification and causal inference as two isolated steps, a growing body of work seeks to merge them, learning the structure directly from observational data to feed downstream tasks.
In root cause analysis, for example, traditional systems require a hardcoded architecture map to trace failures. To break this dependency, Ikram et al. designed an algorithm that learns the causal graph from normal system operations using observational causal discovery, and then uses that learned structure to efficiently isolate faults during a system failure.
A similar paradigm shift is happening in instrumental variable (IV) estimation. Finding a valid instrument—a variable that affects the treatment but has no direct link to the outcome—traditionally requires deep domain intuition. To automate this, Cheng et al. developed a deep generative framework (CIV.VAE) that automatically discovers conditional instrumental variables from raw data by learning latent representations of the instruments and their required conditioning sets simultaneously, removing the need for manual structural nomination.
Strategy 3: Bounding and Optimization
When a graph cannot be uniquely identified, and discovery algorithms leave lingering ambiguities, a final strategy abandons the quest for exact point estimates, opting instead for optimization and partial identification.
If a counterfactual query cannot be uniquely determined due to missing structural details, Tian et al. demonstrated that the problem can be translated into a polynomial program. By using Markov Chain Monte Carlo (MCMC) sampling, the authors showed how to compute tight, mathematically rigorous upper and lower bounds on counterfactual probabilities using an arbitrary mix of observational and experimental data, providing actionable insights even in the face of structural uncertainty.
In high-dimensional generative modeling (like GANs), trying to enforce a rigid, manually defined structural causal model over thousands of pixel dependencies is a recipe for failure. To circumvent this, Feng et al. shifted from explicit graph architectures to a sparse optimization strategy called Principal Knowledge Descent (PKD). Instead of hardcoding feature relationships, PKD uses optimization to synthesize counterfactual images (e.g., altering a specific attribute while keeping unrelated features intact), proving that causal-style reasoning can be achieved in deep networks without ever explicitly drawing the underlying graph.