Emergence-Inspired Multi-Granularity Causal Learning

Hanwen Luo, Guoxian Yu, Jun Wang, Yanyu Xu, Yongqing Zheng, Qingzhong LiAAAI 2025 · DOI

Abstract

Causal discovery in complex systems often faces challenges when causal power resides at a collective "macro" level rather than among individual "micro" variables. We present EMCausal, a framework that leverages the concept of causal emergence to identify multi-granularity relationships. By integrating a Progressive Mapping Encoder with cross-layer consistency constraints, EMCausal effectively uncovers hidden macro-level influences and maintains structural alignment across different scales of observation.

Introduction

Most contemporary causal discovery algorithms are designed to operate at the "micro" level, identifying direct dependencies between individual variables. However, in complex systems such as biological networks or economic structures, significant causal signals may only manifest at a "macro" level—where the collective behavior of a group exerts a stronger influence than any single component.

This phenomenon presents two primary obstacles for standard discovery methods. First, diluted effects occur when causal signals are spread across numerous micro-variables, making them difficult to detect. Second, structural conflicts arise when independent modeling of micro and macro levels leads to logically inconsistent causal graphs. EMCausal addresses these issues by drawing inspiration from emergence, where a system displays properties not present in its constituent parts.

Methodology

The EMCausal framework is built upon three core pillars designed to bridge the gap between micro and macro causal structures.

Progressive Mapping Encoder (PME)

The PME module is responsible for aggregating micro-level variables into meaningful macro-level representations. It utilizes a Gumbel-Softmax operation to learn a discrete mapping of micro-variables to macro-groups. This approach allows the model to optimize the mapping through gradient-based methods while maintaining the interpretability of discrete assignments.

Cross-layer Collaborative Causal Learning

To simulate how macro-states supervene on micro-states, the model reconstructs micro-variables using a combination of micro-level inputs and learned macro-representations. This collaborative approach ensures that the macro-level information actively informs the discovery of micro-level relationships.

Micro-Macro Consistency Constraint

To resolve potential structural conflicts, a mathematical constraint, $H(C) = 0$, is enforced. This constraint ensures that the combined multi-granularity structure remains a Directed Acyclic Graph (DAG), forcing the micro and macro levels to align into a single, coherent causal flow.

Overview of the EMCausal framework, illustrating the Progressive Mapping Encoder (left) and the multi-granularity reconstruction process with consistency constraints (right).

Experimental Evaluation

The researchers evaluated EMCausal against several competitive baselines, including PC, GES, DAG-GNN, and MgCSL, using both synthetic and real-world datasets.

  • Synthetic Data: Multi-granularity datasets were generated using Erdős-Rényi and Scale-Free graph models, incorporating non-linear relationships via Multi-Layer Perceptrons (MLPs) to simulate realistic complexity.
  • Sachs Dataset: A benchmark biological dataset consisting of flow cytometry data (853 samples) measuring 11 protein and phospholipid expression levels in human T cells.

Results and Discussion

EMCausal demonstrated superior performance across all evaluated benchmarks. A key finding was the "Macro Advantage": observing the system at a macro level often made certain causal relationships easier to identify as effects manifested more strongly.

Performance comparison of EMCausal against the MgCSL baseline across Precision, Recall, SHD, and NNZ metrics as the system scale increases.

As shown in the results, EMCausal significantly improved Recall compared to score-based methods, which often miss true edges in noisy micro-level data. On the Sachs dataset, EMCausal achieved the highest F1 score (43.75%) and the lowest Structural Hamming Distance (SHD=11), confirming its effectiveness in real-world biological signaling discovery.

Conclusion

EMCausal represents a significant step toward multi-level causal discovery by formalizing the relationship between micro and macro states through emergence-inspired constraints. While the current model assumes a DAG structure and requires a pre-defined number of macro variables, it provides a robust foundation for analyzing complex systems where causal power is distributed across scales. Future work will focus on automating macro-level discovery and extending the framework to handle cyclic feedback loops.

Summary updated

BibTeX
@inproceedings{luo2025emergence,
  author = {Hanwen Luo and Guoxian Yu and Jun Wang and Yanyu Xu and Yongqing Zheng and Qingzhong Li},
  title = {{Emergence-Inspired Multi-Granularity Causal Learning}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i18.34113},
}
Figure