. 2020 Jun 19;20(12):3463. doi: 10.3390/s20123463

Table 3.

Comparison of pooling techniques: hold-out setting.

Composite Activity Recognition Results
Pooling Technique	Average F1	Accuracy	Weighted F1
Rank Pooling	51.64	52.63	51.87
$R P_{F w d R e v}$	55.99	56.94	55.89
$R P_{H e l l i n g e r}$	52.78	53.11	52.9
$R P_{H e l l i n g e r_F w d R e v}$	55.14	55.5	55.08
$R P_{P o w e r}$	44.28	46.41	44.45
$R P_{P o w e r_F w d R e v}$	55.17	55.5	55.12
$R P_{H e l l i n g e r_P o w e r}$	56.25	55.74	56.38
$R P_{H e l l i n g e r F w d R e v_P o w e r F w d R e v}$	60.91	61.48	60.95
Average Pooling	50.88	53.11	50.33
Max Pooling	55.26	57.42	54.6
LSTM ( $e p o c h s = 1000, # o f l a y e r s = 4$ )	51.82	51.91	51.24
Hidden Markov Models (1000)	47.69	51.2	47.22