Hyperparameters

The hyperparameters were tuned for 3 different setups for BPR-Max and Cross-Entropy respectively. We select the best performing setup with the best hyperparameters (1-1 for both BPR-Max and Cross-Entropy), and use it in our experiments.

Fixed Hyperparameters

The setups are defined by setting the constrained_embedding and embedding parameters to one of the following combinations:

Setups

Setup	constrained_embedding	embedding
setup 1	True	0
setup 2	False	0
setup 3	False	matching layers

And by fixing the following hyperparameters:

BPR-Max

Fixed hyperparameters

parameter	value
loss	bpr-max
logq	0
n_sample	2048
adapt	adagrad
n_epochs	10

Cross-Entropy

Fixed hyperparameters

parameter	value
loss	cross-entropy
final_act	softmax
logq	1
bpreg	0
n_sample	2048
adapt	adagrad
n_epochs	10

Optimized Hyperparameters

For each setup the following hyperparameters are optimized:

BPR-Max

Optimized hyperparameters

parameter	type	range	step	scale
layers	int	[64..512]	32	Uniform
batch_size	int	[32..256]	16	Uniform
learning_rate	float	[0.01..0.25]	0.005	Uniform
dropout_p_embed	float	[0.0..0.5]	0.05	Uniform
dropout_p_hidden	float	[0.0..0.7]	0.05	Uniform
momentum	float	[0.0..0.9]	0.05	Uniform
sample_alpha	float	[0.0..1.0]	0.1	Uniform
bpreg	float	[0.0..2.0]	0.05	Uniform
final_act	categorical	options: [elu-0.5,elu-1,linear]

Cross-Entropy

Optimized hyperparameters

parameter	type	range	step	scale
layers	int	[64..512]	32	Uniform
batch_size	int	[32..256]	16	Uniform
learning_rate	float	[0.01..0.25]	0.005	Uniform
dropout_p_embed	float	[0.0..0.5]	0.05	Uniform
dropout_p_hidden	float	[0.0..0.7]	0.05	Uniform
momentum	float	[0.0..0.9]	0.05	Uniform
sample_alpha	float	[0.0..1.0]	0.1	Uniform