Table 1 Performance of different methods in terms of RMSE.

Methods	Flixster	Douban	Yahoo	ML-100K	ML-1M	ML-10M
PMF	1.370 ± 0.011	0.893 ± 0.002	26.7 ± 0.529	0.970 ± 0.005	0.885 ± 0.007	0.855 ± 0.0006
SVD++	1.150 ± 0.008	0.865 ± 0.002	24.8 ± 0.498	0.948 ± 0.004	0.866 ± 0.004	0.833 ± 0.0004
GRALS	1.296 ± 0.009	0.840 ± 0.002	37.9 ± 0.786	0.933 ± 0.002	0.846 ± 0.005	0.811 ± 0.0002
sRGCNN	1.170 ± 0.007	0.805 ± 0.002	22.8 ± 0.482	0.921 ± 0.003	0.839 ± 0.003	0.785 ± 0.0003
GC-MC	0.943 ± 0.006	0.736 ± 0.001	20.4 ± 0.361	0.906 ± 0.001	0.830 ± 0.001	0.778 ± 0.0001
PinSage	0.945 ± 0.005	0.732 ± 0.001	21.0 ± 0.332	0.914 ± 0.002	0.840 ± 0.002	0.790 ± 0.0002
NGCF	0.954 ± 0.004	0.742 ± 0.001	20.9 ± 0.370	0.916 ± 0.002	0.833 ± 0.002	0.779 ± 0.0003
GAT	0.952 ± 0.005	0.737 ± 0.001	21.2 ± 0.334	0.913 ± 0.001	0.835 ± 0.001	0.784 ± 0.0004
FCF	1.064 ± 0.008	0.823 ± 0.002	22.9 ± 0.389	0.957 ± 0.002	0.874 ± 0.005	0.847 ± 0.0007
FedMF	1.059 ± 0.006	0.817 ± 0.002	22.2 ± 0.349	0.948 ± 0.002	0.872 ± 0.004	0.841 ± 0.0005
FedPerGNN	0.980 ± 0.006	0.775 ± 0.001	20.7 ± 0.325	0.910 ± 0.001	0.839 ± 0.003	0.793 ± 0.0002

Results of FedPerGNN and the best-performed baseline are in bold. The advantage of FedPerGNN over other SOTA privacy-preserving personalization methods FCF and FedMF is significant (p < 0.1). FedPerGNN also achieves comparable performance with other centralized GNN-based personalization methods, and there is no significant difference between FedPerGNN -based personalization methods, and there is no significant difference between FedPerGNN and the best-performed method on Yahoo (p > 0.1).

Quick links

Search