**4. Experimental Settings**

*4.1. Datasets*

We use three popular multi-label networks: BLOGCATALOG3, FLICKR, and YOUTUBE as benchmark datasets. In Table 1, we list the statistical information of all datasets used, including the number of nodes, number of edges, number of node classes, and the tuned optimal value of key parameters of **SORAG***F*: {learning rate, weight decay, dropout rate, *k* (Section 3.2), *ρ* (Section 3.2), *α* (Section 3.3), *β* (Section 3.3), *λ* (Section 3.6), *μ* (Section 3.6)} (see the detailed parameter tuning discussion in Section 5.5). For each dataset, we assume that a majority class is one with more samples than the average class size, while a minority class is one with less samples. Below is a brief description of each dataset used.


For all datasets, we attribute each node with a 64-dim embedding vector obtained by performing dimensionality reduction on the adjacency matrix using PCA [52], similar to [5,17]. All of the above datasets are available at http://zhang18f.myweb.cs.uwindsor.ca/ datasets/ (accessed on 25 June 2022).


**Table 1.** Dataset statistics.
