How can this training-evaluation gap be addressed? While specific techniques have been adopted to optimize certain confusion matrix based metrics, it is challenging or impossible in some cases to generalize the techniques to other metrics.

We appreciate the valuable feedback from all the reviewers and will include the following discussion into our work. We believe that this is not an elegant way to have a global view of the GNN model. "Since the explanatory motifs are not learned end-to-end, the model however may suffer from sub-optimal generalization PGExplainer is natively designed for collectively explaining multiple instances. The source code of PGExplainer can be found in GitHub with the name "PGExplainer". We follow the experimental setting in GNNExplainer. "explanation accuracy" is not formally defined. We didn't report std for baselines because they don't have sampling processes in Baselines' stds are shown in the table below. PGExplainer is a general model compatible with different GNNs and diverse learning tasks. Besides, instead of edge-level important scores, they just calculate node-level important scores. We select a method "Gradient" in the CVPR paper which doesn't require The AUC scores on BA-2motifs and MUT AG are 0.773 and 0.653, The KDD paper mentioned just showed up (June 3). Second, it only provides model-level explanations without preserving the local fidelity. That's why we call it a global method. As discussed in [38], "local fidelity" requires an explanation PGExplainer to preserve the "local fidelity", at the same time, with a global view of the GNN model. GNNExplainer is a pioneer to provide explanations for GNN's predictions. We include a parameterized network to enable explainer a global view of the GNN model. PGExplainer is much more effective and efficient than the state-of-the-art method. As discussed in Appendix D.1, "PGExplainer

Most existing point cloud detection models require large-scale, densely annotated datasets.

We would also like to discuss the limitations of the approaches including ours. As shown in Tab. 4, the PPO-Single that serves as a baseline in our paper is designed following As shown in Tab. 4, NerRewritter is most general because it can be viewed as a learning-based local It is also worth noting that there are some problems that are beyond our knowledge to tackle, e.g. the expression simplify problem, and it may requires experts with specific domain We have discussed the model details of PPO-BiHyb in Sec. 4, and in this section, we discuss the DAG. Considering the structure of DAG, we design two GCNs: the first GCN processes the original DAG, and the second GCN processes the DAG with all edges reversed. The predicted doubly-stochastic matrix by SK is processed by considering the partial matching matrix. Graph-level features are obtained via attention pooling, which are fed to the critic net.

We thank the reviewers for their very helpful comments and suggestions. The 100% recall but very poor precision (i.e., it always predicts a shift) is expected Marginal-KS is very bad because the attack model is very strong, i.e., it mimics the marginal distribution of Thus, marginal KS will naturally fail--highlighting the limitation of prior work for this adversarial attack. F or bootstrapping, does the model need to be fit multiple times? For Gaussian, this is fairly simple. For the detection stage, the FDR was controlled below 0.05 in all See also Table 6 and 7 in appendix.