Two papers from NTT Laboratories have been accepted for presentation at UAI 2026.

Two papers by researchers at NTT Laboratories have been accepted for presentation at UAI 2026, the 42nd Conference on Uncertainty in Artificial Intelligence, to be held in Amsterdam from August 17 to 21, 2026. UAI is one of the world’s premier international conferences for machine learning and artificial intelligence research that addresses uncertainty in prediction and inference. The conference is particularly recognized in the field of statistical causal inference, bringing together leading researchers from around the world and featuring a wide range of high-quality presentations, from foundational theoretical research to advanced studies that pave the way for future AI technologies. This year, UAI 2026 recorded an acceptance rate of 30.5% out of 1,087 submissions, underscoring its reputation as a highly selective, top-tier conference.

Abbreviated names of the laboratories:
CS: NTT Communication Science Laboratories

■Fairness under Graph Uncertainty: Achieving Interventional Fairness with Cluster Causal Graphs

Yoichi Chikahara, Research Scientist (CS)

Machine learning is increasingly used to support decisions about individuals, such as whether to escalate an online customer inquiry to a human agent. In such settings, predictions must be not only accurate but also fair with respect to sensitive attributes such as gender and race. However, existing methods often require detailed causal graphs between variables. This requirement is difficult to satisfy in high-dimensional data with many variables, where errors in graph estimation can accumulate and weaken fairness guarantees. In this study, we propose a new machine learning technique for achieving fairness by leveraging cluster causal graphs, which represent causal relationships among groups of variables and can be estimated more easily from data than variable-level causal graphs. Experiments on synthetic and real-world datasets show that the proposed method achieves a better balance between fairness and predictive accuracy than existing approaches. Future extensions that incorporate domain knowledge about causal relationships are expected to help build decision-making systems that are both efficient and fair, without placing individuals at an unjust disadvantage.

■MetaCaDI: A Meta-Learning Framework for Causal Discovery from Multiple Observational Datasets with Unknown Interventions

Hans Jarett Ong (Nara Institute of Science and Technology), Yoichi Chikahara, Research Scientist (CS), Tomoharu Iwata, Senior Distinguished Researcher (CS)

To understand complex real-world systems, such as large-scale cloud systems and drug response mechanisms, it is important to map not only the cause-and-effect relationships among variables (the causal graph) but also to identify which parts of the system are altered by external changes, known as interventions (e.g., faults or functional suppression), in each environment. However, collecting large amounts of controlled interventional data is often difficult, so these intervention targets must be identified from only a few observations. In this study, we propose MetaCaDI, a meta-learning-based causal inference framework that learns a shared causal structure across multiple environments and leverages this shared knowledge to identify unknown intervention targets in a completely new environment using only a tiny handful of data points. Experiments on synthetic datasets, including complex gene expression simulations, demonstrate that the proposed method substantially outperforms existing approaches. In the future, this technique is expected to support decision-making in data-scarce domains such as medicine, drug discovery, and large-scale system operations.