Hyperparameter ranges for all considered baseline models

Algorithms topK shrink Similarity Normalize α β ℓ2 norm
kNN algorithms [5 to 1000] [0 to 1000] [cosine, jaccard, asymmetric, dice, tversky] [True, False] — — —
Graph based algorithms [5 to 1000] — — — [0 to 2], strategy: uniform distribution [0 to 2], strategy: uniform distribution —
EASER — — — — — — [1 to 1e7], strategy: log-uniform distribution


Optimal hyperparameter values

Algorithms topK shrink Similarity Normalize α β ℓ2 norm
ItemkNN 144 1000 cosine True — — —
UserkNN 146 1000 cosine True — — —
P3α 496 — — False 0.7681732734954694 — —
RP3β 496 — — True 0.44477903655656115 0.5968193614337285 —
EASER — — — — — — 1e3


Statistics of the provided train-test splits

Users Items Interactions Density Data leakage Interactions per usermin Interactions per itemmin Interactions per usermax Interactions per itemmax Testing users
52,643 91,599 2,984,108 0.00062 No 16 10 13313 1907 52,639


Full result table

Methods Amazon-Book Time
R@20 N@20 T-Time (s) P-Time (s) Average P-Time Time Per User (ms)
Random 0.00022 0.00018 0 150.7472472 2.863
TopPop 0.0024 0.0044 0.046885014 128.3760633 2.438
ItemkNN 0.056 0.045 426.9121618 107.231256 2.036
UserkNN 0.051 0.041 283.7361674 104.065511 1.976
P3α 0.056 0.043 197.1278133 114.8666947 2.181
RP3β 0.068 0.056 224.2510846 126.1138134 2.395
EASER 0.071 0.056 — — —
DGCF (Paper) 0.039 0.030
DGCF (Code) 0.011 0.017 — — —

Optimal hyperparameter values

Algorithms topK shrink Similarity Normalize α β ℓ2 norm
ItemkNN 144 1000 cosine True — — —
UserkNN 146 1000 cosine True — — —
P3α 496 — — False 0.7681732734954694 — —
RP3β 350 — — True 0.7681732734954694 0.4181395996963926 —
EASER — — — — — — 1e3


Statistics of the provided train-test splits

Users Items Interactions Density Data leakage Interactions per usermin Interactions per itemmin Interactions per usermax Interactions per itemmax Testing users
31,668 38,048 1,561,406 0.00130 No 18 10 2311 1492 31,668


Full result table

Methods Yelp '18 Time
R@20 N@20 T-Time (s) P-Time (s) Average P-Time Per User (ms)
Random 0.0003 0.0004 0 53.14148092 1.678
TopPop 0.0125 0.0101 0.006148577 45.95453548 1.451
ItemkNN 0.064 0.054 81.77051783 47.61694789 1.503
UserkNN 0.061 0.050 72.28791833 43.42841244 1.371
P3α 0.064 0.052 64.42336011 46.88240695 1.480
RP3β 0.067 0.056 54.6901195 48.22654772 1.522
EASER 0.064 0.052 555.2657838 84.11382508 2.656
DGCF (Paper) 0.064 0.052
DGCF (Code) 0.037 0.028 — — —

Optimal hyperparameter values

Algorithms topK shrink Similarity Normalize α β ℓ2 norm
ItemkNN 508 1000 cosine True — — —
UserkNN 146 1000 cosine True — — —
P3α 496 — — False 1.087096950563704 — —
RP3β 350 — — True 0.5663562161452378 0.001085447926739258 —
EASER — — — — — — 1e3


Statistics of the provided train-test splits

Users Items Interactions Density Data leakage Interactions per usermin Interactions per itemmin Interactions per usermax Interactions per itemmax Testing users
29,858 40,981 1,027,370 0.00084 No 9 10 1,014 2,310 29,858


Full result table

Methods Yelp '18 Time
R@20 N@20 T-Time (s) P-Time (s) Average Prediction Time Per User (ms)
Random 0.00013 0.00022 0 51.75596523 1.733
TopPop 0.0139 0.0317 0 45.00996447 1.507
ItemkNN 0.156 0.123 89.02916408 49.55570626 1.659
UserkNN 0.155 0.128 58.89179349 42.49520731 1.423
P3α 0.175 0.146 90.30393124 48.97286344 1.640
RP3β 0.183 0.152 91.19598675 49.15153646 1.646
EASER 0.163 0.136 724.8426113 76.08247638 2.548
DGCF (Paper) 0.179 0.152
DGCF (Code) 0.133 0.109 — — —