Table 2.

Hyperparameter ranges used for the optimization of all the reported algorithms, both baselines and CQFS.

Algorithm	Hyperparameter	Range	Distribution
ItemKNN cosine	topK	5–1000	uniform
	shrink	0–1000	uniform
	normalize ${}^a$	True, False	categorical
	weighting	none, TF-IDF, BM25	categorical
PureSVD	num factors	1–350	uniform
RP${}^3\beta$	topK	5–1000	uniform
	alpha	0–2	uniform
	beta	0–2	uniform
	normalize ${}^b$	True, False	categorical
CFeCBF	epochs	1–300	early-stopping
	learning rate	${10}^{-5}$–${10}^{-2}$	log-uniform
	sgd mode	Adam	categorical
	$l1$ reg	${10}^{-2}$–${10}^{+3}$	log-uniform
	$l2$ reg	${10}^{-1}$–${10}^{+3}$	log-uniform
	dropout	30–80%	uniform
	initial weight	1.0—random	categorical
	positive only	True, False	categorical
	add zero quota ${}^c$	50–100%	uniform
CQFS	$\alpha$	1	categorical
	$\beta$	${10}^0$–${10}^{-4}$	log-uniform
	s	${10}^0$–${10}^{+4}$	log-uniform
	p	$40\%,60\%,80\%,95\%$	categorical

${}^a$ The normalize hyperparameter in KNNs controls whether to use the denominator of the cosine similarity. If False, the similarity becomes the dot product alone. ${}^b$ The normalize hyperparameter in RP${}^3\beta$ refers to applying $l1$ regularization on the rows of the similarity matrix after the selection of the neighbors, to ensure it still represents a probability distribution. ${}^c$ Percentage of item similarities of value zero added as negative samples to improve its ranking performance.