A novel generative adversarial networks modelling for the class imbalance problem in high dimensional omics data

Table 2 Hyperparameter Summary

id	func_optim	beta	batch_size	iter_critic	dropout_prob	gen_structure	critic_structure	max_epochs_1	max_loss_1	n_epochs_2	lr_1	lr_2	rate_save	diff_epochs	max_loss_2	min_loss_2	instab_constant
lipid_pretraining	SGD	13	30	4	0.5	50, 100, 200	200, 100, 50	500	3	1000	0.0005, 0.0001, 0.001	1e-05, 5e-05, 0.0001	20	500	15	-10	3
lipid_retraining	SGD	13	10	2	0.7	50, 100, 200	200, 100, 50	500	3	5000	0.0005, 0.0001	1e-05, 5e-05	10	500	15	-15	3
micro_pretraining	SGD	13	30	4	0.5	50, 100, 200	200, 100, 50	500	3	1000	0.0005, 0.0001, 0.001	1e-05, 5e-05, 0.0001	20	500	15	-10	3
micro_retraining	SGD	13	10	2	0.7	50, 100, 200	200, 100, 50	500	3	5000	0.0005, 0.0001	1e-05, 5e-05	10	500	15	-15	3
sim_pretraining	SGD	13	30	4	0.5	50, 100, 200	200, 100, 50	500	3	1000	0.0005, 0.0001, 0.001	1e-05, 5e-05, 0.0001	20	500	15	-10	3
sim_retraining	SGD	13	10	2	0.7	50, 100, 200	200, 100, 50	500	3	5000	0.0005, 0.0001	1e-05, 5e-05	10	500	15	-15	3

Table showing hyperparameter values used across the experiments for each dataset. Pre-training and re-training values are shown for each experiment. Iter_critic defines the number of critic training iterations before 1 training cycle of the generator. n_epochs defines the number of epochs trained, using the determined learning rates. Instab_constant was used to define the effect of instability when determining the most effective learning rates. Furthermore, max_loss_1 and max_epochs_1 were used to define this in the first training iterations, and max_loss_2, min_loss_2 in the second. diff_epochs defines the max number of epochs to train when determining this.Stochastic gradient descent (SGD), metabolomics study (lipid), microarray study (micro), simulated microarray study (sim), optimising function (func_optim), instability (instab), probability (prob), iter (iterations), learning rate (lr)

ISSN: 1472-6947