Review for NeurIPS paper: Can Graph Neural Networks Count Substructures?

The authors may need to justify the concrete value of solving substructure counting problems by GNNs. Substructure counting is a problem that has been widely discussed in the domain of theory, database, and data mining. Existing combinatorial algorithms enhanced with improved system designs [1] enables accurate counting on large-scale graphs for simple substructures, such as triangle and stars discussed in this work. Compared with existing combinatorial method based solutions, what is the essential gain from using GNNs? The design of LRP may also need both theoretical and empirical justification, compared with existing ideas.