Abstract
Causal discovery in real-world systems often faces two significant hurdles: the presence of feedback loops (cycles) and incomplete data. Standard methods typically assume Directed Acyclic Graphs (DAGs) and complete observations, failing when data is Missing Not At Random (MNAR). We present MissNODAG, a differentiable framework that addresses these challenges by combining an Expectation-Maximization (EM) procedure with contractive residual normalizing flows. MissNODAG explicitly models the missingness mechanism as part of a larger "m-graph," allowing for the recovery of cyclic causal structures even from biased, incomplete datasets. Our evaluations on synthetic and real-world biological data demonstrate that MissNODAG outperforms existing imputation-based baselines in structure recovery and predictive accuracy.
Introduction
Identifying causal relationships from observational data is a cornerstone of scientific inquiry. However, traditional causal discovery algorithms are often limited by two restrictive assumptions: the absence of feedback loops and the availability of complete datasets. In domains like genomics or economics, feedback loops are ubiquitous, and data is frequently missing due to systematic biases—a phenomenon known as Missing Not At Random (MNAR).
MissNODAG bridges this gap by treating causal discovery as a continuous optimization problem. It leverages neural networks to handle cycles and an EM framework to account for missing values, providing a robust solution for complex, real-world causal inference tasks.
Methodology
MissNODAG's architecture is built on three core pillars:
1. Differentiable Cyclic Modeling
To accommodate feedback loops without computational instability, MissNODAG utilizes contractive residual flows. This approach ensures that the mapping between latent noise and observed variables remains invertible and differentiable. By relaxing the acyclicity constraint common in methods like NOTEARS, MissNODAG can efficiently search the space of cyclic graphs using gradient-based optimization.
2. Modeling the Missingness Mechanism
A key innovation of MissNODAG is its treatment of missingness. Rather than simply imputing missing values, the model explicitly incorporates missingness indicators ($R$) into a comprehensive "m-graph." This allows the framework to handle MNAR scenarios where the probability of a value being missing depends on the value itself.
3. The EM Framework and Rejection Sampling
The model iterates through an Expectation-Maximization (EM) procedure. In the E-step, missing values are imputed using a rejection sampling strategy to generate realistic samples from the conditional distribution of missing data given observed values and the current graph. In the M-step, the parameters of the causal graph and the missingness mechanism are updated to maximize the likelihood of the observed data.
Experimental Evaluation
The performance of MissNODAG was evaluated against several baselines, including MissForest and optimal transport-based imputation, across synthetic and real-world datasets.
Synthetic Benchmarks
Using Erdős-Rényi (ER) graphs with 10 nodes, the researchers tested both linear and nonlinear Structural Equation Models (SEMs) under varying levels of missingness (10% to 50%). MissNODAG consistently achieved the lowest Structural Hamming Distance (SHD), indicating superior accuracy in edge recovery.
Real-World Biological Data
The model was also applied to the Perturb-CITE-seq dataset, which captures gene expression in melanoma cells. MissNODAG successfully predicted the effects of unseen genetic interventions, achieving a lower Negative Log-Likelihood (NLL) than competing methods. This suggests that the model captures the underlying biological "wiring" more effectively than methods that ignore cycles or missingness mechanisms.
Discussion and Future Work
MissNODAG demonstrates that cycles and MNAR data are not insurmountable obstacles for causal discovery. By integrating structural modeling with missingness mechanisms, it provides a more realistic framework for analyzing complex systems.
However, several challenges remain. The current implementation is computationally intensive, limiting its scalability to larger networks (e.g., thousands of genes). Additionally, the model assumes causal sufficiency—the absence of unobserved confounders—and perfect measurements for non-missing data. Future research will focus on improving scalability, incorporating latent variables, and extending the framework to handle more diverse classes of missingness patterns.
Summary updated