Question on MLPAE

[Jing-DS]

Hello,

We noticed in your paper that MLPAE has the best AUC score in identifying contextual outliers (Table 5). However, according to the code in pygod/pygod/models/mlpae.py, MLPAE seems to be only using the nodes features, and not the adjacency matrix as input.

If we understood correctly from your paper, contextual outliers are generated by replacing the feature matrix of node i by one that deviates most from it among q nodes. If MLPAE is not using any structural information, how is it able to distinguish the outlier nodes from the nodes whose features it is using to replace its original features?

Thank you!

[kayzliu]

Thanks for your questions. We found that, based on the contextual outlier injection methods proposed by previous works, the selected substitute node attribute is likely to be global feature outliers'. We use t-SNE to roughly visualize the node attributes in inj_amazon without considering of any graph structure. The red nodes are contextual outliers, while green nodes are other nodes. As we can see, The contextual outliers are primarily located around the periphery, indicating that they are also global feature outliers.