Causal Discovery as a Supervised Classification Task

This direction is no longer present in the current build. Its last version stays readable and citable here.

Version 1 of 118 papers28 conceptsfrozen 2026-10-04

Traditionally, causal inference has been treated as a problem of deductive logic. Researchers would define a set of rigid statistical assumptions—such as "the noise must be additive and Gaussian"—and then use mathematical tests to see if a dataset fits a specific directed graph. However, a growing body of research suggests that we can bypass these restrictive assumptions by reframing causal discovery as a supervised pattern recognition problem.

Instead of asking "Which mathematical model generated this data?", these researchers ask: "What does the statistical signature of a cause-effect relationship look like?"

From Deductive Logic to Inductive Patterns

The core intuition behind this shift is that the joint probability distribution of a cause ($X$) and its effect ($Y$) contains subtle asymmetries. In a perfectly linear, noise-free world, $X \to Y$ might look identical to $Y \to X$. But in the real world, the process of generating an effect usually leaves a "fingerprint" on the data distribution that is not present in the reverse direction.

David Lopez-Paz, Bernhard Schölkopf, and their colleagues pioneered the idea of training a classifier to recognize these fingerprints in Towards a Learning Theory of Cause-Effect Inference. By using kernel mean embeddings, they transformed entire probability distributions into fixed-length feature vectors. They then trained a model on thousands of synthetic pairs where the ground truth (e.g., $X$ causes $Y$) was known. Once trained, this model could look at a brand-new pair of variables and "see" the causal direction based on learned statistical signatures, rather than pre-defined rules.

Visualizing Causality

If causal directionality is a pattern, can we see it? Ye Yuan and Ziv Bar-Joseph took this literally in Causal inference using deep neural networks. They transformed observational data into 2D "Normalized Empirical Probability Density Function" (NEPDF) matrices—essentially turning a scatter plot of data into an image.

By treating causal inference as an image classification task, they allowed Convolutional Neural Networks (CNNs) to detect geometric features in the distribution that signify the direction of influence. This approach works because CNNs are exceptionally good at finding non-linear, local invariants that human-designed statistical tests often miss.

Learning from Simulations and Interventions

One of the biggest hurdles in causal discovery is the lack of "labels" in the real world; we rarely know the true causal graph of a complex system. However, we can simulate them.

Konrad Kording and Xinyue Wang argue in Learning domain-specific causal discovery from time series that we should use supervised learning to capture domain-specific dynamics. By training Transformers on data generated from known physical or biological simulators, the models learn to identify causal links that are specific to that field (e.g., neuroscience or genomics), significantly outperforming "one-size-fits-all" methods like Granger Causality.

Similarly, Nan Rosemary Ke and colleagues developed DiscoGen, which uses supervised learning to recognize how interventions (like CRISPR gene edits) manifest in high-dimensional data. By learning from simulated interventions, the model becomes a highly accurate "discovery engine" for real-world biological networks.

Scaling to High Dimensions

Traditional causal algorithms often struggle with "the curse of dimensionality"—as the number of variables grows, the number of possible graphs grows exponentially. Reframing the problem as link prediction helps solve this.

In Deep Learning of Causal Structures in High Dimensions, Kai Lagemann and Sach Mukherjee introduced Deep Discriminative Causal Learning (D2CL). Instead of trying to solve the entire graph at once, D2CL uses a neural network to predict the existence of an edge between any two variables. This discriminative approach allows the model to scale to systems with thousands of variables, such as genome-wide networks, by focusing on local causal regularities.

Propagating Knowledge via Semi-Supervised Learning

In many scientific fields, we aren't starting from zero; we might know a few confirmed causal links from decades of literature. Chris J. Oates and Sach Mukherjee proposed treating causal discovery as a semi-supervised task in Causal Discovery as Semi-Supervised Learning.

By using a small set of "labeled" causal links to anchor the model, they use manifold regularization to propagate that knowledge to "unlabeled" pairs. If the joint distribution of pair A (unknown) looks statistically similar to pair B (known to be causal), the model can confidently label pair A as causal too. This turns causal discovery into an iterative process where every new experimental discovery makes the overall inference engine stronger.

Papers in this version

History & citationVersion 1 · 2026-10-04

No longer present in the current build · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_causal_discovery_as_a_supervised_classif_f15e1d79,
  title = {Causal Discovery as a Supervised Classification Task},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-discovery-as-a-supervised-classification-task/rdr_f15e1d79a2f846298226fe07c57fc2d6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-18},
}

This exact version (version 1, frozen text):

@misc{lacuna_causal_discovery_as_a_supervised_classif_f15e1d79_v1,
  title = {Causal Discovery as a Supervised Classification Task},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/causal-discovery-as-a-supervised-classification-task/rdr_f15e1d79a2f846298226fe07c57fc2d6/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-18},
}