2026
Priatmojo, Widyasmoro; Bleze, Mariia; Milisavljevic, Matilda; Samandari, Ahmad; Mongelli, Maurizio; Wicaksono, Hendro
In: Expert Systems with Applications, pp. 134767, 2026.
Abstract | Links | BibTeX | Tags: artificial intelligence, causal AI, causal inference, federated learning, machine learning
@article{priatmojo2026reproducible,
title = {Reproducible benchmarking of Aggregation Strategies for Federated Constraint-Based Causal Discovery: a systematic study of a FedPC-style pipeline under controlled synthetic data scenarios},
author = {Widyasmoro Priatmojo and Mariia Bleze and Matilda Milisavljevic and Ahmad Samandari and Maurizio Mongelli and Hendro Wicaksono},
url = {https://www.sciencedirect.com/science/article/pii/S0957417426036717},
doi = {https://doi.org/10.1016/j.eswa.2026.134767},
year = {2026},
date = {2026-10-17},
urldate = {2026-10-17},
journal = {Expert Systems with Applications},
pages = {134767},
publisher = {Pergamon},
abstract = {Federated causal discovery aims to infer causal structures from distributed data without exchanging raw observations, but fragmented local samples and client level distributional or structural variation can make reliable global structure recovery challenging. Existing constraint-based federated approaches provide limited mechanisms for estimating client reliability and aggregating structural information, especially for edge orientation, while systematic and reproducible evaluation frameworks remain scarce. This paper proposes a FedPC-style benchmark framework that integrates consensus-based client reliability estimation, skeleton aggregation, Sepset aggregation, and CPDAG recovery within a unified federated causal discovery pipeline positioned as a controlled aggregation-ablation study. The framework enables controlled evaluation across client scales, sample sizes, graph densities, aggregation parameters, and sample-level noise-scaling and client-level mechanism-shift heterogeneity regimes, and is validated on both synthetic datasets and the Sachs protein-signaling benchmark with the client-side PC implementation additionally compared against an established causal discovery package. Experimental results show that cross client Sepset aggregation substantially improves edge orientation recovery over naive aggregation under the tested conditions. Under non-severe corruption, uniform and consensus weighting are statistically indistinguishable, while consensus weighting provides a significant benefit when corrupted clients form the majority. Federated variants with Sepset aggregation also maintain skeleton recovery comparable to centralized learning under moderate federated conditions. Further, orientation quality is more sensitive than skeleton recovery to cross client Sepset aggregation in the tested settings, indicating that Sepset aggregation contributes more to orientation recovery than the weighting scheme under non corrupted conditions. Overall, this work provides a transparent, reproducible, and extensible benchmark for analyzing federated constraint-based causal discovery and offers practical insights into how consensus-based aggregation influences causal structure recovery in distributed environments.
},
keywords = {artificial intelligence, causal AI, causal inference, federated learning, machine learning},
pubstate = {published},
tppubtype = {article}
}
Federated causal discovery aims to infer causal structures from distributed data without exchanging raw observations, but fragmented local samples and client level distributional or structural variation can make reliable global structure recovery challenging. Existing constraint-based federated approaches provide limited mechanisms for estimating client reliability and aggregating structural information, especially for edge orientation, while systematic and reproducible evaluation frameworks remain scarce. This paper proposes a FedPC-style benchmark framework that integrates consensus-based client reliability estimation, skeleton aggregation, Sepset aggregation, and CPDAG recovery within a unified federated causal discovery pipeline positioned as a controlled aggregation-ablation study. The framework enables controlled evaluation across client scales, sample sizes, graph densities, aggregation parameters, and sample-level noise-scaling and client-level mechanism-shift heterogeneity regimes, and is validated on both synthetic datasets and the Sachs protein-signaling benchmark with the client-side PC implementation additionally compared against an established causal discovery package. Experimental results show that cross client Sepset aggregation substantially improves edge orientation recovery over naive aggregation under the tested conditions. Under non-severe corruption, uniform and consensus weighting are statistically indistinguishable, while consensus weighting provides a significant benefit when corrupted clients form the majority. Federated variants with Sepset aggregation also maintain skeleton recovery comparable to centralized learning under moderate federated conditions. Further, orientation quality is more sensitive than skeleton recovery to cross client Sepset aggregation in the tested settings, indicating that Sepset aggregation contributes more to orientation recovery than the weighting scheme under non corrupted conditions. Overall, this work provides a transparent, reproducible, and extensible benchmark for analyzing federated constraint-based causal discovery and offers practical insights into how consensus-based aggregation influences causal structure recovery in distributed environments.