TABLE 2.

Average RMSE over five replications. We compare SeER to only explicit feedback models (MF, NeuMF, and MLP) because we cannot guarantee to have enough items for all the users in the test data to compute ranking metrics. SeER achieves a lower RMSE compared to the other approaches, for increasing item cold start levels, which means that it is more robust in dealing with unseen items. All differences are significant (Tukey test $p-values<0.05$).

	% Item cold start
	0% (no cold start)	5%	10%	15%	20%
MF	2.4977	2.5696	2.5344	2.5100	2.5487
NeuMF	1.2765	1.3273	1.3166	1.2889	1.3237
MLP	1.2750	1.3123	1.3046	1.2750	1.3017
SeER	1.2433	1.3055	1.2914	1.2652	1.2940