Mechanisms Under Shifts: Interpretable Clustering With Self-Improving Heterogeneous Causal Graphs.

Area of Science:

• Machine Learning
• Causal Inference
• Data Science

Background:

• Existing machine learning methods struggle with causal heterogeneity, confounding, and observational constraints.
• This leads to poor interpretability and difficulty distinguishing true causal heterogeneity from spurious associations.

Purpose of the Study:

• To propose an unsupervised framework, HCL (Interpretable Causal Mechanism-Aware Clustering with Self-Improving Adaptive Heterogeneous Causal Structure Learning), for inferring latent clusters and causal structures.
• To address limitations in existing methods by incorporating causal awareness and handling observational data without prior knowledge.

Main Methods:

• HCL jointly infers latent clusters and associated causal structures from mixed-type observational data.
• It utilizes an equivalent representation to encode structural heterogeneity and confounding, relaxing homogeneity and sufficiency assumptions.
• A bi-directional iterative strategy refines clustering and structure learning, coupled with self-supervised regularization for mechanism balancing.

Main Results:

• HCL demonstrates superior performance in both clustering and structure learning tasks.
• The framework successfully recovers biologically meaningful mechanisms in single-cell perturbation and clinical intervention data.
• Identifiability of heterogeneous causal structures is theoretically shown under mild conditions.

Conclusions:

• HCL provides a robust, interpretable solution for discovering mechanism-level causal heterogeneity.
• The framework enhances the reliability of learning systems operating under environmental shifts.
• It offers a significant advancement in unsupervised causal discovery from observational data.