Table 4:

Table of hyperparameters for prediction on affective computing dataset.

Component	Model	Parameter	Value
GRU Encoder	GRU	Input sizes	[5,20,35,74,300,704]
		Hidden sizes	[32,32,64,128,512,1024]
		Num of layers	1 or 2
		Dropout	0:0 or 0:1
Transformer Encoder [158]	Transformer [158]	Input sizes	[5,20,35,74,300,704]
		Hidden sizes	[32,32,64,128,512,1024]
		Num of layers	2 or 3
		Dropout	0.2
Head	MLP	Input sizes	[5,20,32,64,128,256]
		Hidden sizes	[5,20,32,64,128,256]
		Num layers	[2
		Dropout	0.2
MCTN [123] Encoder	GRU	Input sizes	300
		Hidden sizes	[32, 64]
		Num of layers	1 or 2
		Dropout	0.0 or 0.1
MCTN [123] Decoder	GRU	Input sizes	[32, 64]
		Hidden sizes	300
		Num of layers	1 or 2
		Dropout	0.0 or 0.1
MCTN [123] Seq2Seq	GRU+GRU	teaching ratio	0.5
		Embed sizes	32
		${\mu }_{t_1}$,${\mu }_c$ ,${\mu }_{t_2}$	0.01
Fusion	LRTF [106]	Num ranks	64
		Output sizes	128
	MI-Matrix [77]	Hidden size	128
	MulT [	Hidden size	40
		Num heads	8 or 10
Training		Loss	MAE or Cross Entropy
		Batch size	32
		Seq Length	50 or 20
		Num epochs	100 or 300
		Early stop	True
		Patience	[8,20]
		Activation	ReLU
		Optimizer	AdamW
		Weight Decay	1×10−4
		Learning rate	1×10−4