predict-email-opened

Task

Having data like: Predict whether an email will be opened or not.

Results

global_step	tp	tn	fp	fn	mcc	train_loss	eval_loss	auroc	auprc	f1
626	376	13393	225	6035	0.1151935742	0.6418074965	0.6062818189	0.6228090916	0.4420890947	0.1072447233
1252	1042	12973	645	5369	0.1934530726	0.5493763685	0.5893685939	0.6638975125	0.4921320593	0.2573474932
1878	1353	12838	780	5058	0.2325452894	0.553080976	0.573552263	0.6904007317	0.5265999733	0.3167134831
2504 (best)	1700	12579	1039	4711	0.2564462134	0.5565853715	0.5723564906	0.704683717	0.5430064945	0.3715846995
3130	1742	12599	1019	4669	0.2664377197	0.5727524757	0.5862198094	0.7079365433	0.5477412575	0.3798517226

Conclusion

Do not use the Bert classifier alone, even with additional dense features. Consider using random forests.