d10andy / sad Goto Github PK

Q1:"mooc data, rand_feat = np.zeros((max_idx + 1, 172)), 172 need change to 4",I found shape of edge_feature and node_feature will be [12344,4],[4849,4]，then in feat_process_layer，i meet RuntimeError： mat1 and mat2 shapes cannot be multiplied (4849x4 and 172x128)。Do i need to change the dim of feat_process_layer when processing mooc?

Q2: I meet NaN in loss when training，”if torch.any(torch.isnan(log_probs)): raise ValueError("Log_prob has nan!")“
I change learning rate from 5e-4 to 1e-6 solve the problem in wikipedia，but I still meet NaN in reddit dataset.
Do I need to keep turn dowm learning rate,or check logits/log_pros/sim_coeff?

Thank you!

Thank you for providing an interesting anomaly detection method. However, I noticed that when the model predicts the node anomaly state, it finds the historical neighbors but does not include the current event. This seems to ignore the current event information. However, based on my understanding of this task, the label of the node represents whether the state has changed after the current event occurs, which means that the current event should have very important information. Is this just a mistake or are there other reasons for this?

would you tell me how to find The Alipay dataset?I want to detect abnormal users in the dynamic graph，but the state_label column is changing, should I keep it constant over time?

Thank you for your outstanding open-source project. I'm a newcomer in this field and have some questions about this project. Among these four datasets, three of them are publicly available datasets, all of which are bipartite graphs, and the features are included in the edges. However, the Alipay transaction network dataset might be a large-scale homogeneous graph with a larger network size and more abundant edge and node features. Could you please provide some insights into the data structure of this network? I truly appreciate it.