The original feature preprocess on datasets

[khan-yin]

hello author， I am a student now focused on GNN. I am curious about the original feature preprocess details on datasets for node classification. I want to know whether the origin feature has Heterogeneity or not. for example, the origin feature is generated by metapath2vec/transE ..etc, or maybe randomwalk? because I noticed that on some dataset，the original features (feats-type 0) can not even work better than only target with others zero features (feats-type 1). Thanks a lot. looking forward to your reply.😆

[1049451037]

Hi. Thank you for your attention. The original features depends on the datasets. For example, the paper node in ACM and DBLP features are paper keyword n-gram. The author nodes are aggregated features from papers as suggested in HAN and MAGNN. Maybe the early aggregation causes the worse performance. For other information, you can refer to the dataset preprocessing scripts:

• ACM：2c6535d
• IMDB：4e05b43
• DBLP：https://github.com/cynricfu/MAGNN/blob/master/preprocess_DBLP.ipynb