link prediction datasets. Each number is the average performance for 10 random initialization of the experiments. Bold

Each number is the average performance for 10 random initialization of the experiments. To compare our proposed methods with additional popular heuristics methods (Jaccard (Jac.), preferential attachment (P A), Katz, PageRank (PR), and SimRank (SR)) beyond overlapped neighbors-based Neo-GNN consistently shows better performance than overlapped-based heuristic methods. Interestingly, though overlap-based heuristic methods perform worse in Power dataset, our Neo-GNN show the best performance compared to all heuristic methods. This result shows that Neo-GNN is not limited to the limitations of existing neighborhood-overlap based heuristics. The direction of generalizing these heuristic methods will be a good future work.