*5.4. Influence of Imbalance Ratio*

This section discusses how the performance of the analyzed models varies as the imbalance of the training set changes. Similar to [5,17,49], we varied the percentage of global minority samples removed in [10%, 20, ..., 80%, 90%] on each dataset, according to the GMD ranking (see Section 3.2). The more samples were removed, the more imbalanced the training set became. The sampling ratios for the BLOG CATALOG3, FLICKR, and YOU TUBE networks were set to 10%, 1%, and 1%, respectively. The oversampling rates were set as presented in Table 4. All the parameters were the same as those in Section 5.2. The performance variation of each method is as follows (Figure 6). For comparison, we also report the performance of the state-of-the-art approach **GraphSMOTE**.

**Figure 6.** *Cont*.

**Figure 6.** Influence of imbalance ratio: (**a**) BLOGCATALOG3, MICRO-F1; (**b**) BLOGCATALOG3, MACRO-F1; (**c**) FLICKR, MICRO-F1; (**d**) FLICKR, MACRO-F1; (**e**) YOUTUBE, MICRO-F1; (**f**) YOUTUBE, MACRO-F1.

Figure 6 demonstrates the strong robustness of **SORAG** in a variety of scenarios. As the imbalance ratio increases, the performance of **SORAG** is maintained at a high level. It can be observed that **SORAG** outperformed **GraphSMOTE** in nearly all comparisons. In particular, **SORAG***F* was the best method. It achieved the best performance in 15 out of 30 test scenarios in terms of Micro-F1 and in 17 out of 30 test scenarios in terms of Macro-F1. In contrast, **GraphSMOTE** performed best only on the YOUTUBE network.
