Developing an online hate classifier for multiple social media platforms

Table 11 Generalizability of our best models (XGBoost with All and BERT features) across social media platforms

	YouTube	Reddit	Twitter	Wikipedia
F1_{xgboost_all}	0.911	0.776	*0.980*	*0.861*
F1_{xgboost_BERT}	0.907	*0.778*	0.975	0.846
F1 in original paper^a	*0.960* [5]	0.749 [73]	0.900 [16]	–
ROC-AUC_{xgboost_all}	*0.968*	*0.967*	*0.994*	*0.993*
ROC-AUC_{xgboost_BERT}	0.964	*0.967*	0.991	*0.991*
ROC-AUC in original paper^a	–	0.957 [73]	–	0.972 [7]

We also present the results from previous research—when not reported, the cell contains (−). To facilitate reading, our results are given in italic. The results between the XGBoost (All vs BERT) in Wikipedia and Twitter platforms are significantly different with p < 0.01 (McNemar’s test). Note that the results from previous studies are not directly comparable with the current study findings
^aBrackets refer to sources