Analysis and classification of privacy-sensitive content in social media posts

EPJ Data Science

Table 16 Classification results (macro-averaged F1-scores and percentage gain w.r.t the best bag-of-word classifier)

Dataset	Classifier	F1-score	Gain
SENS2	BoW-LR	0.68	–
	BoW-RF	0.67	–
	BoW-SVM	0.68	–
	CNN1	0.73	7.35%
	CNN2	0.73	7.35%
	CNN3	0.72	5.88%
	CNN4	0.71	4.41%
	RNN1	0.77	13.24%
	RNN2	0.77	13.24%
	BERT	0.78	14.71%
SENS3	BoW-LR	0.73	–
	BoW-RF	0.73	–
	BoW-SVM	0.78	–
	CNN1	0.81	3.85%
	CNN2	0.81	3.85%
	CNN3	0.60	−11.76%
	CNN4	0.81	3.85%
	RNN3	0.87	11.54%
	RNN4	0.86	10.26%
	BERT	0.89	14.10%
OMC	BoW-LR	0.60	–
	BoW-RF	0.59	–
	BoW-SVM	0.60	–
	CNN1	0.63	5.28%
	CNN2	0.64	5.95%
	CNN3	0.65	8.29%
	CNN4	0.65	8.07%
	RNN1	0.65	11.67%
	RNN2	0.65	11.67%
	RNN3	0.66	13.33%
	RNN4	0.67	14.39%
	BERT	0.68	13.32%
WH+TW	BoW-LR	0.80 ± 0.01	–
	BoW-RF	0.78 ± 0.01	–
	BoW-SVM	0.81 ± 0.01	–
	CNN1	0.77 ± 0.03	−3.62 ± 3.89
	CNN2	0.78 ± 0.03	−4.51 ± 4.41
	CNN3	0.75 ± 0.07	−7.32 ± 8.31
	CNN4	0.68 ± 0.15	−15.15 ± 18.92
	RNN1	0.83 ± 0.02	2.59 ± 1.78%
	RNN2	0.83 ± 0.01	3.08 ± 1.48%
	RNN3	0.83 ± 0.01	3.28 ± 2.00%
	RNN4	0.84 ± 0.02	3.85 ± 1.85%
	BERT	0.88 ± 0.01	8.80 ± 1.55%