Table 5.

Top reported system performance for studies predicting the age of Twitter users using deep learning machine learning (ML) methods. Result metrics are reflected in this table as reported in the original publications and are not comparable to each other. Reviews are ordered by the number of classification groups.

Study	Number of age groups	Age class detail (y)	Language	ML method	Reported performance
					F1-score	Accuracy
Guimaraes et al [73], 2017	2	13-19 and >20	English	CNNa	0.94	N/Ab
Kim et al [83], 2017	2	Young (18-23) and old (25-30)	English	GRNNc	N/A	0.81
Vijayaraghavan et al [108], 2017	3	<30, 30-60, and >60	English	DMTd	0.82	N/A
Pandya et al [94], 2018	3	Dutch: <20, 20-40, and >40; English 1: 13-17, 18-40, and >40; and English 2: 13-17, 18-24, and >25	English and Dutch	CNN	0.61-0.87	N/A
Pandya et al [95], 2020	3	Dutch: <20, 20-40, and >40; English 1: 13-17, 18-40, and >40; and English 2: 13-17, 18-24, and >25	English and Dutch	CNN	0.82-0.87	N/A
Wang et al [112], 2019	4	≤18, 18-30, 30-40, and 40-99	Multilingual—28	mmDNNe	0.52	N/A
Bayot and Goncalves [55], 2017	5	18-24, 25-34, 35-49, 50-64, and ≥65	English and Spanish	CNN	N/A	0.43-0.55

^aCNN: convolutional neural network.

^bN/A: not applicable.

^cGRNN: graph recurrent neural network.

^dDMT: deep multimodal multitask.

^emmDNN: multimodal deep neural network.