Synthetic Bottleneck Graph Dataset for "Over-squashing as Transport Congestion: A Sandpile Dynamics Perspective"

This dataset accompanies the paper “Over-squashing as Transport Congestion: A Sandpile Dynamics Perspective”. The uploaded file `synthetic_bottleneck_graph.npz` contains a controlled 1,140-node synthetic bottleneck graph used for the NearColl-based congestion-collapse diagnostic reported in Fig. 4(d) of the paper. The dataset includes node features, labels, graph edges, train/validation/test masks, bottleneck-region annotations, and trained GCN hidden representations used to compute NearColl over label-discordant node pairs. The diagnostic is intended as a controlled mechanism dataset rather than a natural real-world benchmark. It supports the paper’s analysis of the association between transport congestion and representation collapse under bottlenecked graph structure. Associated paper:Over-squashing as Transport Congestion: A Sandpile Dynamics Perspective