🤖 machine learning

Enes Causal Discovery

The paper proposes a mixture-of-experts architecture for causal discovery on observational data, though the provided abstract contains significant grammatical incoherence and contradictory claims regarding the necessity of neural networks versus simple linear models.

Original authors: Alexis Kafantaris

Published 2026-03-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Alexis Kafantaris

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 solve a mystery: Who caused what?

In the real world, we often see things happening together (correlation), but we don't know if one actually caused the other. For example, ice cream sales and shark attacks both go up in the summer. Did eating ice cream cause shark attacks? No, the hot weather caused both. Figuring out the true "cause-and-effect" chain is called Causal Discovery.

This paper introduces a new detective tool called Enes (Edge node edge similarity). Here is how it works, explained simply:

1. The Problem: A Messy Crime Scene

Usually, to solve a mystery, you need a clear list of clues. But in science and business, we often only have a pile of random observations (like a blurry photo of a crime scene). We don't know the rules of the game, and we can't run experiments to test our theories easily.

Most existing tools are like old-school detectives:

Some use simple math (like drawing straight lines between points). They are fast but miss complex clues.
Others use strict logic rules. They are smart but get confused when the data is messy or huge.

2. The Solution: A Team of Specialized Experts (MoE)

The Enes model is different. Instead of one detective trying to do everything, it uses a Mixture of Experts (MoE). Think of this as a detective agency with two specialists working together:

Expert A (The Linear Detective): Good at spotting simple, straight-line relationships (like "more rain = more mud").
Expert B (The Non-Linear Detective): Good at spotting complex, twisty relationships (like "a little bit of poison helps, but too much kills").

A Gating Mechanism acts like a Manager. When a new clue comes in, the Manager decides: "Is this a simple clue? Send it to Expert A. Is this a weird, complex clue? Send it to Expert B." They combine their answers to make the final call.

3. The Training: Learning from "Fake" Crime Scenes

Here is the tricky part: The model didn't learn from real protein data (the "Sachs dataset") because that would be like cheating—it would just memorize the answers.

Instead, the team created thousands of fake, random crime scenes using math simulations (called SEMs and Michaelis-Menten dynamics).

The Analogy: Imagine training a pilot by putting them in a flight simulator with random storms, rather than letting them fly a real plane immediately.
The model learned to recognize patterns in these fake storms. The goal was to see if it could learn the rules of the road well enough to drive a real car later.

4. The Rules of the Game (The "Physics")

To keep the model from making wild guesses, the researchers gave it three strict rules (penalties) to follow, like a coach correcting a player:

No Time Travel (DAG Penalty): You can't say the future caused the past. The model must build a logical timeline.
No Coincidences (Pearson Penalty): If two things move together just by chance, the model shouldn't think they are connected.
Consistency Check (Cosine Similarity): The model checks if its explanation makes sense geometrically.

5. The Results: The New Detective Wins

The researchers tested Enes against the old-school detectives (the "Linear Pearson" method and the "PC Algorithm") using two types of data:

Real Data: The famous Sachs protein dataset (a complex biological map).
Scalability Test: They made the maps bigger and bigger (from 11 nodes to 50 nodes) to see if the model got confused.

The Verdict:

Accuracy: Enes was the most accurate. It made fewer mistakes about who caused what.
Robustness: Even when the maps got huge and messy, Enes didn't break down.
The Catch: Enes is very careful. It only claims a connection if it is very sure. This means it misses a few true connections (low "Recall"), but the ones it does find are almost always correct (high "Precision"). It's better to be a detective who finds one solid lead than one who accuses everyone in town.

Why Does This Matter?

The authors say that while they tested this on proteins (biology), the real goal is Service Design (like figuring out why a customer service process fails).

If this model can figure out the complex cause-and-effect chains in a living cell (which is incredibly hard), it can definitely help businesses figure out why their processes are broken. It's a new, smart, and adaptable way to turn a pile of confusing data into a clear map of cause and effect.

Based on the provided paper "Enes Causal Discovery" by Alexis Kafantaris (March 2026), here is a detailed technical summary:

1. Problem Statement

The paper addresses the challenge of observational causal discovery, specifically the task of inferring the Directed Acyclic Graph (DAG) structure of causal relationships from observational data without interventional labels.

Core Difficulty: The field suffers from the "Garbage In, Garbage Out" (GIGO) problem, where models struggle to generalize when trained on biased or insufficient data. Traditional statistical methods (e.g., PC algorithm) often lack scalability or struggle with non-linear relationships, while standard neural networks typically fail due to the lack of formal training data for causal patterns.
Target Application: The primary benchmark is the Sachs dataset (a classic protein signaling network), with secondary testing on Michaelis-Menten (MM) kinetics to evaluate scalability and efficiency. The ultimate goal is to apply this logic to service design process ground truth retrieval.

2. Methodology: The Enes Architecture

The authors propose Enes (Edge node edge similarity), a novel neural evolutionary method utilizing a Mixture of Experts (MoE) architecture combined with physics-informed constraints.

Architecture Design:
- Input: The model processes random non-linear Structural Equation Model (SEM) data and graph triplets ($Edge-Node-Edge$).
- Dual Expert System:
  - Expert A: A Non-linear expert designed to capture complex, non-linear causal dependencies.
  - Expert B: A Linear expert to handle linear relationships (similar to Pearson correlations).
  - Gating Mechanism: A soft-gating mechanism ( $Score = w \cdot A + (1-w) \cdot B$ ) dynamically weights the contributions of both experts to classify the causal motif of the triplet.
- Output: A classification of the causal pattern (edge existence and direction), effectively reconstructing the causal map.
Training & Optimization:
- Physics-Informed Constraints: The network is trained using a "physics-informed" approach where physical constraints are merged into the training matrix.
- Loss Functions & Penalties: The model minimizes a fusion loss incorporating:
  - DAG Enforcement Penalty: Ensures the output graph is acyclic.
  - Pearson Correlation Penalty: $P(edge) \cdot (1 - |r|)$ , penalizing edges that do not correlate linearly.
  - Cosine Similarity: Used to measure consistency between edge-node-edge probes.
  - Adjacency Matrix Constraint: Enforces structural validity.
- Stabilization: To handle chaotic gradient descent and ensure stability, the authors employ Langevin dynamics and Simulated Annealing optimization techniques.
- Data Strategy: The model is trained on randomly generated non-linear SEMs and MM dynamics rather than real-world data to prevent "cheating" (overfitting to specific dataset motifs), forcing the model to learn generalizable causal patterns.

3. Key Contributions

Hybrid MoE Approach: Successfully integrates linear and non-linear experts within a single neural framework to handle the diverse nature of causal relationships, a task where standard neural nets often fail.
Physics-Informed Training: Introduces a novel method of embedding DAG constraints and correlation penalties directly into the neural network's learning mechanism, bridging the gap between statistical constraints and deep learning.
Generalization via Random Synthesis: Demonstrates that training on random, non-biased synthetic data (SEM/MM) allows the model to generalize effectively to real-world datasets (Sachs) without prior knowledge of the ground truth.
Stability Mechanisms: The application of Langevin dynamics and Simulated Annealing to stabilize the training of causal discovery models, addressing the volatility often seen in neural causal inference.

4. Results and Performance

The Enes model was evaluated against two baselines: a Linear Fast Pearson coefficient and the PC Algorithm (a constrained statistical baseline).

Sachs Dataset (11 Nodes):
- Accuracy: Enes achieved 0.8818, outperforming Linear Pearson (0.8694) and PC (0.7025).
- Structural Hamming Distance (SHD): Enes achieved the lowest error rate (14.30), significantly better than PC (36.00) and Linear Pearson (15.80).
- Precision: Enes showed high precision (0.7214), indicating it rarely predicts false edges, though it had lower recall (0.2706), suggesting a conservative "cherry-picking" strategy.
Scalability (MM Kinetics - 25 & 50 Nodes):
- Enes maintained high accuracy (0.9614 and 0.9802 respectively) and low SHD compared to baselines as the graph size increased.
- While the F1 score dropped for larger graphs (a common issue in sparse causal graphs), Enes consistently outperformed the PC algorithm in SHD and accuracy across all node counts.

5. Significance and Conclusion

Robustness: The study proves that a neural network can outperform state-of-the-art statistical baselines (PC) and simple linear methods in observational causal discovery, provided it is trained on diverse synthetic data with strict physical constraints.
Scalability: The MoE architecture demonstrates superior scalability compared to the PC algorithm, which struggles with larger node counts (e.g., SHD of 74.00 at 50 nodes vs. Enes' 49.60).
Future Implications: While the current focus is biological (Sachs), the authors argue the system is a "hybrid" capable of being retuned for service design processes. The ability to retrieve ground truth from complex, non-linear observational data without interventional labels represents a significant step forward in automated causal inference.

Critique noted in the paper: The model tends to have lower recall, meaning it is conservative in identifying edges. However, the high precision and low SHD suggest it is highly reliable for the edges it does predict, making it a stable tool for causal discovery in complex systems.