Causality-Driven Disentangled Representation Learning in Multiplex Graphs
New self-supervised method separates shared and private node information across network layers.
A team of researchers led by Saba Nasiri has published a new paper titled "Causality-Driven Disentangled Representation Learning in Multiplex Graphs," introducing a framework called CaDeM. The work addresses a fundamental challenge in analyzing multiplex graphs—multi-layer networks where nodes interact through multiple types of relationships. Traditional methods often entangle shared information common across all layers with private information specific to individual layers, which limits both generalization to new tasks and interpretability of the learned representations.
CaDeM tackles this by applying principles from causal inference to disentangle these components in a self-supervised way. The framework jointly performs three key operations: it aligns shared embeddings across different network layers, enforces private embeddings to capture only layer-specific signals, and critically, applies a statistical technique called backdoor adjustment. This ensures the common embeddings capture purely global information, separated from the private representations.
The researchers validated their approach through experiments on both synthetic and real-world datasets, reporting consistent improvements over existing baseline methods. The paper has been submitted to IEEE Transactions on Signal and Information Processing over Networks, indicating its technical rigor and potential impact in the signal processing and network science communities. This work represents a significant step toward more interpretable and robust AI models for complex relational data.
- Introduces CaDeM, a self-supervised framework using causal inference (backdoor adjustment) to separate common and private node embeddings in multiplex graphs.
- Designed for multi-layer networks where nodes have multiple relation types, improving over methods where information is entangled.
- Shows consistent performance improvements in experiments, leading to more interpretable and generalizable representations for networks.
Why It Matters
Enables clearer analysis of complex systems like social networks or protein interactions, leading to more trustworthy and effective AI models.