We genuinely appreciate your thoughtful questions. Before we address your subsequent questions, could you confirm if our response has addressed your concerns regarding **Q1-Q7**? We want to ensure that all your concerns have been adequately addressed before we discuss any new questions. If there are any aspects you feel need further clarification or improvement, please do not hesitate to let us know.
The responses to your new questions are given as follows.
**Q8: Generalization capability of ARC**
**A8:** From a theoretical perspective, ARC is equipped with an in-context learning (ICL) technique, which enhances its generalization capability on new datasets. Specifically, the ICL mechanism enables ARC to distinguish between normal and abnormal samples by using a set of normal samples from each given testing dataset, rather than fitting to the anomaly patterns of a specific training dataset. Following the assumption that normal samples are more similar to each other than anomalies, the cross-attention-based ICL module can learn to make accurate predictions by leveraging inter-sample similarity on new datasets.
From an experimental perspective, we have made every effort to demonstrate the generalizability of ARC through experiments on datasets from a wide range of domains. Specifically, in the original manuscript, we included 8 datasets from 3 domains: **citation network, social network, and co-review network**. During the review process, we further expanded our experiments to cover 4 additional domains: **co-author network, co-purchase network, transaction record network, and work collaboration network**. It is important to note that our model had **never seen** data from these 4 new domains during training, providing strong empirical evidence of ARC’s generalization capabilities on diverse and unseen tasks. To the best of our knowledge, we have covered datasets from most of the mainstream domains in graph anomaly detection (GAD), including **4 datasets for training and 12 datasets for testing**, which exceeds the evaluation scope of the vast majority of GAD studies.
**Q9: Results discussion of experiments of pseudo normal trick**
**A9:** Regarding your question about the performance of ARC with a limited number of pseudo samples, we would like to clarify the following points:
**Optimal number of pseudo samples.** The statement “Why does the performance of ARC achieve the best performance with limited number of pseudo samples (e.g., #Pseudo Normal = 2) on some datasets, such as Amazon, Facebook and Weibo?” is not accurate. As mentioned in our response to Reviewer RwhU, ARC achieves its best performance on these datasets with #Pseudo Normal = 10. This is because, in real-world scenarios, the number of normal samples typically far exceeds the number of anomalies. Therefore, with a sample size of 10, the probability of selecting true normal samples is higher, which is why we set the default sample size to 10 in all our experiments. This demonstrates that ARC can achieve competitive performance with a small number of shot samples.
**Impact of sample size on performance.** The statement "the advantage of having a larger number of normal samples reduces the impact of anomalies within the pseudo normal samples" was intended to illustrate the intuitive feasibility of the "pseudo normal trick." However, this does not imply that increasing the number of pseudo normal samples will always enhance ARC’s performance. On the contrary, when the sample size is very small (e.g., $n_k$=2 or 10), the probability of selecting true normal samples is higher. As the sample size increases, the likelihood of including false normal samples also increases (especially when the anomaly ratio of the dataset is relevantly large), introducing more noisy psuedo samples and potentially degrading performance. To sum up, the performance of ARC with the "pseudo normal trick" is influenced by the "purity" of the sampled pseudo examples, which in turn depends heavily on the anomaly ratio of the dataset and experimental randomness. Therefore, it is reasonable that different datasets exhibit varying preferences for the pseudo sample size.
Last but not least, we'd like to clarify that **the exploration of the pseudo-label trick is merely an additional exploratory experiment**. We aimed to investigate the possibility of applying ARC to a purely unsupervised (zero-shot) scenario. Surprisingly, the results show that the performance is still acceptable, indicating ARC's potential. However, our main focus remains on the few-shot setting, where reliable content samples are available for ICL. We hope this experiment demonstrates the potential of ARC and inspires future work in this direction, but it is not the main contribution of this paper.