Table 3.

Performance of EINMF compared with other algorithms (embedding size = 64).

Dataset	MovieLens-100k						MovieLens-1m
	HR@N			NDCG@N			HR@N			NDCG@N
Model	N = 5	N = 10	N = 20	N = 5	N = 10	N = 20	N = 5	N = 10	N = 20	N = 5	N = 10	N = 20
Pop [2]	0.2031	0.3712	0.4761	0.0718	0.0786	0.0863	0.2015	0.2983	0.4228	0.0718	0.0786	0.0863
Item-KNN [5]	0.3160	0.4730	0.5758	0.0976	0.1067	0.1185	0.2237	0.3371	0.4874	0.0677	0.0714	0.0817
BPR-MF [10]	0.2874	0.4380	0.5822	0.0971	0.1088	0.1277	0.3340	0.4804	0.6267	0.1135	0.1206	0.1409
NCF [13]	0.6002	0.7540	0.8367	0.2662	0.2641	0.2890	0.603	0.7278	0.8268	0.2608	0.2497	0.2525
DMF [20]	0.6458	0.7667	0.8738	0.2672	0.2692	0.2835	0.5892	0.7197	0.8257	0.2401	0.2354	0.2413
EINMF	0.6978	0.8038	0.8887	0.3163	0.3092	0.3179	0.6540	0.7781	0.8618	0.2880	0.2728	0.2825
MI (%)	8.05	4.84	1.71	18.38	14.86	10.00	8.46	6.91	4.23	7.46	9.25	11.88

“MI” indicates the smallest improvements of our EINMF over the corresponding baseline. The optimal value of each metric of the baseline top-N task is underlined in the table.