Validate whether GNN edge weights reflect ground-truth causal relations

Determine whether the graph neural network edge weights that gate object-to-object message passing in HCLSM correspond to ground-truth causal relationships between objects in environments with known causal structure, ideally evaluating with intervention-based metrics to establish correspondence.

Background

HCLSM learns interaction structure between object slots through graph neural network (GNN) edge weights. Although an explicit causal adjacency matrix with NOTEARS-style regularization is included, it collapsed to zero edges during training, so the primary causal signal comes from the GNN edge weights.

The authors explicitly note that they have not verified whether these learned GNN edges align with true causal relationships in the environment and suggest that a proper evaluation would require benchmarks with known causal graphs and intervention-based assessments.