Attention‐Guided Causal Structure Learning With Gated Feature Fusion

ABSTRACT Causal discovery, also referred to as causal structure learning, aims to identify causal relationships among variables from observational data and represent them as a directed acyclic graph (DAG). Existing differentiable causal discovery methods alleviate the difficulty of structural search through continuous optimization. However, their learning process is typically driven by the ELBO‐based reconstruction accuracy term, which may be insufficient for explicitly characterizing dependencies among variables. In complex scenarios, this limitation is prone to yielding spurious edges that are inconsistent with the real causal structure. To address this issue, this article proposes AGA‐DAG, an attention‐guided causal structure learning method. The proposed method incorporates a self‐attention mechanism into the structure learning process to capture dependencies among variables. Meanwhile, a gated feature fusion strategy is designed to adaptively balance variables' representation and structure representations, thereby reducing the adverse impact of excessive dependencies on causal structure learning. Furthermore, an attention‐DAG structural consistency loss is constructed by penalizing the discrepancy between the attention weight matrix and the currently learned adjacency matrix, encouraging consistency between them and improving the accuracy of causal structure estimation. Experimental results on synthetic data sets show that AGA‐DAG generally outperforms existing baseline methods under different node settings and functional data‐generating mechanisms. Experiments on real‐world data further demonstrate the structural learning performance of the proposed method.