Enhancing Subgraph-GNNs by way of Edge-Degree Ego-Community Encodings
Authors: Nurudin Alvarez-Gonzalez, Andreas Kaltenbrunner, Vicenç Gómez
Summary: We current a novel edge-level ego-network encoding for studying on graphs that may increase Message Passing Graph Neural Networks (MP-GNNs) by offering extra node and edge options or extending message-passing codecs. The proposed encoding is enough to differentiate Strongly Common Graphs, a household of difficult 3-WL equal graphs. We present theoretically that such encoding is extra expressive than node-based sub-graph MP-GNNs. In an empirical analysis on 4 benchmarks with 10 graph datasets, our outcomes match or enhance earlier baselines on expressivity, graph classification, graph regression, and proximity duties — whereas decreasing reminiscence utilization by 18.1x in sure real-world settings
