Discovering causal relationships from observational data is a fundamental challenge across science and engineering. Historically, researchers tackling this problem fell into two distinct camps: constraint-based algorithms and score-based algorithms.
Constraint-based methods (like the classic PC algorithm) use statistical conditional independence (CI) tests to systematically eliminate edges. While computationally efficient, they suffer from a major vulnerability: they rely on binary thresholds, meaning a single incorrect test early on can cause cascading errors throughout the graph. Furthermore, they can typically only resolve graphs up to a Markov Equivalence Class (MEC), leaving the direction of many edges fundamentally ambiguous.
Score-based methods, on the other hand, treat causal discovery as an optimization problem, searching the space of Directed Acyclic Graphs (DAGs) to maximize a global likelihood score. While more robust to local noise, the space of possible DAGs grows exponentially with the number of variables. This makes pure score-based searches computationally explosive and prone to getting trapped in local minima.
The core intuition behind two-stage hybrid causal discovery is simple: why choose? By decoupling the discovery process into two distinct phases—first identifying the undirected graph skeleton, and then orienting the edges—hybrid algorithms combine the scalability of constraints with the precision of score-based optimization and functional noise models.
Phase 1: High-Recall Skeleton Identification
The first stage of a hybrid framework focuses entirely on connectivity: determining if two variables are causally linked, without worrying about which way the arrow points. The primary goal here is to drastically prune the search space while maintaining high recall, ensuring that true causal links are not accidentally discarded.
To overcome the cascading errors of traditional binary CI tests, modern hybrid methods inject continuous or score-based metrics directly into the skeleton-discovery phase. For instance, Affeldt et al. (2015) integrated information-theoretic scoring into constraint-based discovery. By evaluating the statistical magnitude of 2-point and 3-point information rather than relying on hard independence thresholds, their approach prevents early noisy tests from corrupting the network structure.
In a similar vein, Ling et al. (2025) designed a local causal discovery algorithm that intentionally uses a lenient constraint-based search to maximize the recall of potential neighbors, relying on a subsequent score-based refinement phase to filter out the resulting false positives.
[All Variables]
│
▼ (Phase 1: Lenient Constraints / Adjacency Tests)
[Undirected Skeleton] <-- Search space is now drastically reduced
│
▼ (Phase 2: Functional Models / Score Optimization)
[Fully Oriented DAG]
For high-dimensional datasets where testing every pair of variables is unfeasible, researchers have turned to continuous relaxations and supervised learning to isolate the skeleton. Amin et al. (2024) introduced a Differentiable Adjacency Test (DAT) that transforms skeleton identification into a smooth, gradient-based optimization problem. By separating connectivity from directionality, they achieved precise skeleton recovery on graphs with over a thousand variables.
Taking this a step further, Ma et al. (2024) utilized a supervised model to estimate an undirected "skeleton posterior," which then acts as a probabilistic guide for a downstream differentiable orientation phase.
Phase 2: Breaking Equivalence via Edge Orientation
Once the skeleton is established, the astronomical space of all possible graphs collapses into a highly constrained neighborhood of candidate structures. This reduction in complexity unlocks a massive advantage: it allows algorithms to deploy powerful, computationally expensive directional tests locally—tests that would be completely intractable if applied to the full graph from scratch.
This is where Functional Causal Models (FCMs) and noise asymmetry analysis shine. In nonlinear systems, the true causal direction $X \rightarrow Y$ often possesses an independent additive noise residual, whereas the reverse direction $Y \rightarrow X$ does not.
Tillman et al. (2009) leveraged this asymmetry by first extracting the global graph skeleton via independence tests and then applying Weakly Additive Noise Models (WANM) locally to orient the edges. Similarly, Zhang et al. (2009) demonstrated that complex, post-nonlinear causal networks can be uniquely resolved by first establishing the Markov equivalence class and then using post-nonlinear functional analysis to orient the remaining ambiguous edges.
By focusing noise analysis purely on the pre-discovered skeleton, hybrid frameworks successfully break the Markov equivalence barrier.
- Cai et al. (2018) developed the Structural Equational Likelihood Framework (SELF), which unifies global network scoring with local structural equations, utilizing gradient boosting to evaluate noise residuals and determine exact edge directions.
- Bystrova et al. (2024) applied this paradigm to time-series data, proving that hybridizing noise-based ordering with constraint-based pruning provides robust mathematical protection against violations of the traditional "faithfulness" assumption.
- Hiremath et al. (2024) combined a top-down hierarchical sorting strategy with local functional searches, demonstrating that exploiting noise distribution properties over a constrained search space yields highly accurate linear and nonlinear DAGs.
Adapting to Real-World Complexities
The two-stage hybrid architecture is not just a trick for speed; its modularity makes it uniquely resilient when dealing with messy, real-world data constraints like latent confounders, interventional data, or noisy priors.
Latent Confounders
Unmeasured variables can introduce spurious correlations that look identical to direct causation. To counter this, Ogarrio et al. (2016) inverted the standard hybrid pipeline: they used an initial score-based phase to rapidly map out a candidate graph skeleton, and then applied rigorous constraint-based statistical independence tests to prune away connections caused by latent confounders.
Interventional Data
When data includes experimental interventions (e.g., gene knockouts), the targets of those interventions are often unknown. Cao et al. (2022) showed that constraint-based rules (such as V-structure invariance) can first narrow the search space down to an interventional MEC, after which a greedy score-and-search mechanism easily optimizes the remaining edges. For massive biological networks, Shah et al. (2023) scaled network recovery to genome-scale proportions by combining observational skeleton estimation with targeted interventional orientation.
Integrating External Priors
Because the scoring and constraint phases are distinct, hybrid models can easily ingest external domain knowledge. Chen et al. (2023) demonstrated this by using Large Language Models (LLMs) to provide structural feedback. They integrated this feedback as hard constraints to enforce known edges and soft constraints within the scoring function to guide orientation, ensuring the statistical search remained anchored to common-sense reality.
Ultimately, as early foundational work by Maier et al. (2010) and Zhalama et al. (2017) anticipated, the power of hybridization lies in mutual error correction. By allowing constraints to handle the "where" of causal connections and scoring functions or functional models to handle the "which way," hybrid causal discovery bypasses the fundamental limitations that plague either method in isolation.
Key Foundational Work
On the Identifiability of the Post-Nonlinear Causal Model (2009) — contributed 2 concept elements to this direction.
Nonlinear directed acyclic structure learning with weakly additive noise models (2009) — contributed 1 concept elements to this direction.
Learning Causal Models of Relational Domains (2010) — contributed 1 concept elements to this direction.