📊 statistics

Differentiable Cyclic Causal Discovery Under Unmeasured Confounders

The paper proposes DCCD-CONF, a novel differentiable framework that effectively recovers nonlinear cyclic causal graphs and identifies unmeasured confounders using interventional data, outperforming existing methods while offering theoretical consistency guarantees.

Original authors: Muralikrishnna G. Sethuraman, Faramarz Fekri

Published 2026-01-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Muralikrishnna G. Sethuraman, Faramarz Fekri

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to figure out how a complex machine works. You have a bunch of dials (variables) that move around, and you want to know: Does turning Dial A cause Dial B to move, or does Dial B cause Dial A?

Usually, detectives have two big rules to make their job easier:

No Time Travel: They assume nothing can loop back on itself (A causes B, which causes C, which never causes A again).
No Hidden Suspects: They assume they can see every single dial in the room.

But in the real world—especially in biology—these rules often break. Machines have feedback loops (time travel!), and there are often hidden dials (confounders) that the detective can't see but are secretly pushing both Dial A and Dial B at the same time.

This paper introduces a new detective tool called DCCD-CONF. Here is how it works, using simple analogies:

1. The Problem: The "Invisible Puppeteer"

In many real systems (like genes in a cell), two things might move together not because one pushes the other, but because an invisible puppeteer (an unmeasured confounder) is pulling both strings.

Old tools either assumed the puppeteer didn't exist or got confused when the system had loops (like a snake eating its own tail).
The new tool is designed specifically to spot these invisible puppeteers and handle the loops.

2. The Solution: A "Smart Simulator"

Instead of trying to solve a giant, impossible math puzzle all at once, the authors built a smart simulator (a neural network).

The Setup: Imagine you have a black box. You put in some noise (random static), and it spits out the behavior of your dials.
The Trick: The simulator is built so that if you know the dials, it can work backward to guess the noise. If the noise for Dial A and Dial B is "correlated" (they wiggle in sync), the simulator knows there is a hidden puppeteer connecting them.
The Learning Process: The tool plays a game of "guess and check." It tries to build a map of who causes whom. It keeps adjusting the map until the simulator can perfectly predict the real-world data. If the map is wrong, the prediction fails. If the map is right, the prediction works.

3. The "Intervention" Test

To make sure the detective isn't just guessing, the tool uses interventions.

Imagine you forcibly turn Dial A to a specific setting (like a surgeon cutting a wire).
If Dial B still moves, it's not because of Dial A; it's because of the hidden puppeteer.
The paper shows that by testing the system under different "forced" settings, the tool can untangle the true causes from the hidden puppeteers, even when the system has loops.

4. What They Found

The authors tested their tool on two types of challenges:

Fake Data: They built computer simulations with known loops and hidden puppeteers. DCCD-CONF was better at finding the true map and identifying the hidden puppeteers than any other existing tool.
Real Data: They used real data from a gene study (melanoma cells). Since no one knows the "true" map of these genes, they tested if the tool could predict what would happen if they changed a gene. DCCD-CONF was the best at predicting the outcome, proving it understood the system better than the others.

The Bottom Line

This paper presents a new method that can figure out cause-and-effect relationships in messy, real-world systems where:

Things can loop back on themselves.
There are invisible factors influencing the results.

It does this by using a flexible, learning-based simulator that gets better the more data (and forced experiments) it sees, outperforming older methods that get stuck when things get complicated.

Technical Summary: Differentiable Cyclic Causal Discovery Under Unmeasured Confounders

Problem Statement
Causal discovery aims to infer the directed graph structure representing cause-effect relationships between variables. Existing algorithms typically rely on two restrictive assumptions: (i) the underlying causal graph is a Directed Acyclic Graph (DAG), and (ii) all variables are observed (no unmeasured confounders). These assumptions are frequently violated in real-world systems, particularly in biological networks, which exhibit feedback loops (cycles) and hidden confounders. While some methods address cycles and others address confounders, few handle both simultaneously, especially in nonlinear settings. Furthermore, existing approaches for cyclic graphs with confounders often assume linearity or struggle with scalability. This paper addresses the challenge of learning nonlinear cyclic causal graphs in the presence of unmeasured confounders using interventional data.

Methodology: DCCD-CONF
The authors propose DCCD-CONF, a differentiable framework for learning causal structures under these complex conditions. The methodology is built upon the following components:

Structural Equation Model (SEM) with Latent Confounders:
The system is modeled using a Directed Mixed Graph (DMG) containing both directed edges ( $i \to j$ ) and bidirectional edges ( $i \leftrightarrow j$ ). The bidirectional edges represent hidden confounders. The authors model confounding not as a separate latent variable node, but as correlations in the exogenous noise terms. Specifically, the noise vector $Z$ follows a Gaussian distribution $Z \sim \mathcal{N}(0, \Sigma_Z)$ , where non-zero off-diagonal entries in the covariance matrix $\Sigma_Z$ indicate confounding between variables.
Differentiable Causal Mechanism via Implicit Flows:
To handle cycles and nonlinearity, the causal mechanisms are modeled using implicit flows. The structural equations are defined as $X = F(X, Z)$ , where $F$ is a function parameterized by neural networks. To ensure the existence of a unique fixed point (equilibrium) and invertibility, the authors enforce a contractive mapping using a specific form: $F(x, z) = -g_x(x) + g_z(z) + z$ , where $g_x$ and $g_z$ are contractive functions (enforced via spectral normalization). This allows the use of root-finding algorithms (e.g., Broyden's method) to compute the forward map $z = f_x(x)$ and the reverse map.
Interventional Data and Mutilated Graphs:
The framework utilizes "surgical" (hard) interventions. When a node is intervened upon, all incoming edges (directed and bidirectional) are removed, and the node is set to a value sampled from a known distribution. The likelihood of the data under these interventions is derived based on the Jacobian of the forward map.
Optimization and Score Function:
The learning objective is to maximize the regularized log-likelihood of the observed data across multiple interventional settings.
- Graph Structure: The adjacency matrix is treated as a Bernoulli random variable parameterized by a matrix $B$ . The $L_1$ -norm is used as a proxy for sparsity regularization.
- Jacobian Determinant: Computing the log-determinant of the Jacobian is computationally expensive ( $O(d^2)$ ). The authors employ an unbiased estimator based on the power series expansion of $\log(1+x)$ , truncated stochastically to reduce bias, combined with the Hutchinson trace estimator.
- Confounder Distribution: The covariance matrix $\Sigma_Z$ is updated by solving a sparsity-regularized optimization problem (formulated as a Lasso regression) to ensure positive definiteness and stability.

Key Contributions

Novel Framework: Introduction of DCCD-CONF, the first differentiable framework capable of learning nonlinear cyclic causal graphs while explicitly modeling unmeasured confounders via correlated Gaussian noise.
Theoretical Guarantees: The authors prove that maximizing the proposed score function identifies the interventional equivalence class of the ground truth graph under specific assumptions (including the data lying within the model class and $\sigma$ -separation properties).
Algorithmic Efficiency: The development of a scalable optimization procedure that handles the computational challenges of implicit flows and Jacobian determinants through stochastic estimation and structured updates for the noise covariance.

Experimental Results
The framework was evaluated on both synthetic and real-world datasets:

Synthetic Data: DCCD-CONF was compared against state-of-the-art baselines (NODAGS-Flow, LLC, DAGMA, ADMG) across varying numbers of confounders, cycles, degrees of nonlinearity, and intervention counts.
- Performance: DCCD-CONF consistently achieved lower Structural Hamming Distance (SHD) for graph recovery and higher F1 scores for confounder identification compared to baselines, particularly in nonlinear settings.
- Robustness: The method demonstrated robustness as the number of cycles and confounders increased. It outperformed baselines as the number of interventions increased, whereas some baselines (like LLC) failed in purely observational settings.
- Scalability: Performance remained superior as the number of nodes increased from 10 to 80.
Real-World Data: The model was applied to the Perturb-CITE-seq gene perturbation dataset (melanoma cells under control, co-culture, and IFN- $\gamma$ $γ$ conditions).
- Metric: Since ground truth graphs were unavailable, performance was measured by Negative Log-Likelihood (NLL) on held-out test data.
- Outcome: DCCD-CONF achieved the lowest NLL across all three conditions, outperforming NODAGS, LLC, and DCDI, suggesting superior predictive accuracy and effective modeling of confounding in biological systems.

Significance and Claims
The paper claims that DCCD-CONF provides a unified solution for a previously fragmented problem space, addressing nonlinearity, cycles, and latent confounders simultaneously. The authors assert that their approach offers theoretical consistency guarantees and practical superiority over existing methods, which are often limited to linear assumptions, acyclic structures, or complete observability. The work highlights the prevalence of confounders in real-world systems and demonstrates that accounting for them via differentiable learning significantly improves causal discovery accuracy. The authors note that while the current work assumes Gaussian exogenous noise, future directions include exploring other noise distributions, handling missing data, and relaxing assumptions regarding intervention targets.