Clustering and Pruning in Causal Data Fusion

2026-06-18 · Source: stat.ML updates on arXiv.org · Field: Science & Research — Mathematics & Computational Sciences, Research Methodology & Innovation · Depth: Expert, medium

Summary

Otto Tabell, Santtu Tikka, and Juha Karvanen propose pruning and clustering as preprocessing techniques to enhance causal data fusion, particularly when dealing with multiple data sources and complex causal graphs. Data fusion combines observational and experimental data to identify causal effects, but general-purpose tools like do-calculus face computational challenges with increasing variables. The authors generalize existing results for single data sources, establishing sufficient conditions for applying pruning (removing irrelevant variables) and clustering (combining related variables) in multi-source scenarios. Their work demonstrates how to infer identifiability or non-identifiability from a reduced graph and derive corresponding identifying functionals for the original, larger graph. This approach aims to reduce computational load for identification algorithms and present identifying functionals more concisely, with examples from epidemiology and social science.

Key takeaway

For research scientists working with complex causal graphs and multiple data sources, you should consider implementing pruning and clustering as preprocessing steps. These methods reduce graph size, mitigating the computational burden of do-calculus-based identification algorithms. This allows you to more efficiently determine causal effect identifiability and derive identifying functionals, accelerating your analysis in fields like epidemiology or social science.

Key insights

Pruning and clustering reduce causal graph complexity in multi-source data fusion while preserving causal effect identifiability.

Principles

• Graph reduction can preserve identifiability.
• Do-calculus scalability improves with smaller graphs.
• Non-identifiability in clustered graphs implies non-identifiability in original graphs.

Method

The paper proposes applying pruning (removing irrelevant variables) and clustering (combining related variables) as preprocessing steps. It derives sufficient conditions for these operations in multiple data source contexts and shows how to transfer identifying functionals.

In practice

• Apply pruning to remove descendants of response variables.
• Use clustering to group related vertices.
• Utilize identifying functionals from reduced graphs.

Topics

• Causal Data Fusion
• Graph Pruning
• Variable Clustering
• Causal Identifiability
• Do-calculus
• Multi-source Causal Inference

Code references

• ottotabell/clustering-pruning

Best for: AI Scientist, Research Scientist

Related on AIssential

• Identifiability of Causal Graphs under Non-Additive Conditionally Parametric Causal Models — CPCMs offer flexible causal discovery by modeling conditional distributions with diverse parametric families.
• A causal fused lasso for interpretable heterogeneous treatment effects estimation — The causal fused lasso adaptively estimates interpretable, piecewise constant heterogeneous treatment effects.
• Optimal Experiments for Partial Causal Effect Identification — Optimal experimental design for partially identifiable causal queries is NP-hard but can be significantly pruned using graphical criteria.
• Estimate Collapsibility of Causal Effects in Completed Partial DAGs via Strong d-Convex Hulls — A collapsible method estimates causal effects in CPDAGs, maintaining consistency before and after marginalization.
• Moment Matters: Mean and Variance Causal Graph Discovery from Heteroscedastic Observational Data — Separating mean and variance causal influences improves interpretability and intervention design in heteroscedastic systems.
• Multi-Level Causal Embeddings — Causal embeddings generalize abstraction, mapping multiple detailed models into a coarser causal model's sub-systems.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by stat.ML updates on arXiv.org.