Table A3.

Hyperparameter values for machine learning and graph based recommender algorithms on all datasets.

Algorithm	Hyperparameter	ContentWise Impressions	Netflix Prize	Movielens 20M
P3α	topK	146	5	1000
	alpha	1.2748	0.9150	1.0915
	normalize similarity	True	False	True
RP3β	topK	5	173	245
	alpha	0.0000	0.7040	0.5775
	beta	0.3487	0.6578	0.5625
	normalize similarity	True	True	True
SLIM ElasticNet	topK	1000	503	588
	l1 ratio	1.14E-02	1.33E-02	2.17E-01
	alpha	0.0010	0.0010	0.0023
MF BPR	sgd mode	adagrad	adagrad	adagrad
	epochs	1415	1500	1065
	num factors	200	200	158
	batch size	2	1	256
	positive reg	2.32E-04	1.00E-05	1.00E-05
	negative reg	1.50E-04	1.00E-05	7.03E-05
	learning rate	5.86E-02	2.96E-02	2.16E-02
MF FunkSVD	sgd mode	adam	adam	adagrad
	epochs	500	280	160
	use bias	False	False	False
	batch size	32	1024	2
	num factors	184	186	127
	item reg	3.92E-04	5.98E-04	1.47E-05
	user reg	2.84E-03	9.44E-04	3.18E-03
	learning rate	1.62E-04	1.10E-03	7.11E-02
	negative quota	0.0133	0.1198	0.0266
PureSVD	num factors	350	47	29
NMF	num factors	324	79	79
	solver	mult. update	mult. update	mult. update
	init type	nndsvda	random	random
	beta loss	frobenius	frobenius	kullback-leibler
IALS	num factors	200	63	60
	epochs	5	65	60
	confidence scaling	linear	linear	linear
	alpha	0.7942	0.2199	0.7748
	epsilon	0.1012	0.0010	0.1358
	reg	1.00E-02	1.00E-02	1.00E-02
EASER	l2 norm	-	2.12E+06	7.52E+05
SLIM BPR	topK	-	36	685
	epochs	-	595	970
	symmetric	-	True	True
	sgd mode	-	sgd	sgd
	lambda i	-	2.07E-03	1.00E-05
	lambda j	-	9.20E-03	1.00E-05
	learning rate	-	9.22E-04	1.56E-03

Results for EASE^R and SLIM BPR are missing from the ContentWise Impressions dataset because the algorithms required more than the available 64GB of RAM.