Table 2. Precision values of different methods on 15 networks.
Precision | C. elegant | Karate | Word | Jazz | USAir | Yeast | PB | NS | Router | Power | Baydry | School | SmaGri | SW | |
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
NMF1* | 0.142 | 0.111 | 0.201 | 0.042 | 0.548 | 0.320 | 0.139 | 0.143 | 0.265 | 0.025 | 0.022 | 0.477 | 0.195 | 0.053 | 0.123 |
NMF-D1*(2) | 0.171 | 0.143 | 0.234 | 0.042 | 0.615 | 0.362 | 0.165 | 0.170 | 0.298 | 0.016 | 0.023 | 0.539 | 0.216 | 0.068 | 0.153 |
NMF-A1*(2) | 0.175 | 0.143 | 0.214 | 0.042 | 0.605 | 0.386 | 0.168 | 0.170 | 0.311 | 0.021 | 0.026 | 0.541 | 0.218 | 0.068 | 0.153 |
NMF2* | 0.122 | 0.075 | 0.183 | 0.055 | 0.512 | 0.350 | 0.079 | 0.158 | 0.202 | 0.103 | 0.018 | 0.430 | 0.168 | 0.049 | 0.165 |
NMF-D2*(2) | 0.185 | 0.144 | 0.201 | 0.079 | 0.593 | 0.445 | 0.169 | 0.234 | 0.303 | 0.067 | 0.029 | 0.520 | 0.182 | 0.118 | 0.246 |
NMF-A2*(2) | 0.191 | 0.149 | 0.201 | 0.080 | 0.600 | 0.470 | 0.186 | 0.248 | 0.310 | 0.235 | 0.036 | 0.544 | 0.218 | 0.130 | 0.276 |
SPM*(2) | 0.171 | 0.144 | 0.210 | 0.101 | 0.650 | 0.449 | 0.160 | 0.238 | 0.420 | 0.224 | 0.057 | 0.552 | 0.227 | 0.118 | 0.211 |
Katz(1) | 0.102 | 0.131 | 0.169 | 0.072 | 0.449 | 0.365 | 0.108 | 0.175 | 0.299 | 0.060 | 0.058 | 0.085 | 0.142 | 0.099 | 0.151 |
LHNII(2) | 0.000 | 0.000 | 0.000 | 0.001 | 0.047 | 0.003 | 0.000 | 0.000 | 0.008 | 0.000 | 0.010 | 0.005 | 0.062 | 0.000 | 0.001 |
ACT | 0.053 | 0.024 | 0.128 | 0.087 | 0.169 | 0.332 | 0.000 | 0.077 | 0.193 | 0.160 | 0.034 | 0.118 | 0.142 | 0.035 | 0.101 |
TSCN(1) | 0.018 | 0.014 | 0.145 | 0.002 | 0.024 | 0.133 | 0.032 | 0.027 | 0.087 | 0.096 | 0.056 | 0.036 | 0.197 | 0.028 | 0.039 |
Salton | 0.024 | 0.050 | 0.001 | 0.001 | 0.535 | 0.046 | 0.000 | 0.013 | 0.253 | 0.000 | 0.015 | 0.011 | 0.175 | 0.000 | 0.001 |
Jaccard | 0.028 | 0.071 | 0.001 | 0.002 | 0.521 | 0.064 | 0.000 | 0.017 | 0.252 | 0.000 | 0.007 | 0.010 | 0.180 | 0.000 | 0.001 |
Sorenson | 0.028 | 0.065 | 0.001 | 0.002 | 0.521 | 0.064 | 0.000 | 0.017 | 0.252 | 0.000 | 0.009 | 0.010 | 0.180 | 0.000 | 0.001 |
HPI | 0.015 | 0.007 | 0.091 | 0.005 | 0.255 | 0.016 | 0.012 | 0.003 | 0.146 | 0.000 | 0.005 | 0.055 | 0.105 | 0.002 | 0.000 |
HDI | 0.029 | 0.069 | 0.004 | 0.004 | 0.465 | 0.083 | 0.000 | 0.025 | 0.264 | 0.000 | 0.007 | 0.009 | 0.173 | 0.000 | 0.001 |
LHN | 0.000 | 0.003 | 0.004 | 0.000 | 0.093 | 0.004 | 0.000 | 0.000 | 0.084 | 0.000 | 0.010 | 0.014 | 0.082 | 0.000 | 0.001 |
CN | 0.095 | 0.139 | 0.164 | 0.064 | 0.509 | 0.372 | 0.104 | 0.174 | 0.379 | 0.057 | 0.051 | 0.065 | 0.162 | 0.090 | 0.112 |
AA | 0.112 | 0.151 | 0.163 | 0.067 | 0.524 | 0.396 | 0.104 | 0.172 | 0.563 | 0.038 | 0.030 | 0.063 | 0.148 | 0.103 | 0.131 |
RA | 0.112 | 0.138 | 0.165 | 0.056 | 0.545 | 0.473 | 0.083 | 0.151 | 0.586 | 0.020 | 0.030 | 0.065 | 0.187 | 0.102 | 0.139 |
PA | 0.060 | 0.014 | 0.096 | 0.089 | 0.130 | 0.318 | 0.012 | 0.069 | 0.012 | 0.025 | 0.001 | 0.167 | 0.025 | 0.051 | 0.099 |
LP(1) | 0.100 | 0.131 | 0.169 | 0.072 | 0.495 | 0.370 | 0.107 | 0.175 | 0.299 | 0.059 | 0.054 | 0.071 | 0.113 | 0.095 | 0.128 |
CRA | 0.116 | 0.157 | 0.199 | 0.038 | 0.557 | 0.391 | 0.123 | 0.177 | 0.481 | 0.062 | 0.033 | 0.085 | 0.210 | 0.118 | 0.147 |
Random | 0.005 | 8.5e-4 | 0.016 | 0.007 | 0.016 | 0.004 | 2.3e-4 | 0.002 | 0.001 | 5e-5 | 5.4e-5 | 0.034 | 0.010 | 0.001 | 0.004 |
We compared our methods with other methods on the 15 network data sets and the precisions are returned with the average over 100 runs. The last row is the precision value of a real random predictor which is obtained by providing a ranking list that is ordered according to a random permutation of the links. For every data set, the presented links are partitioned into training set (90%) and test set (10%). The local methods are in standard character while the global methods are in italic. Number in bracket closed to a method denotes the number of tuning parameters. The best result achieved by global methods and the best result achieved by local methods on each network are boldface. Methods with an asterisk like * denote methods based on inference and methods without an asterisk denote methods based on a paradigm (in the sense that are model-based). We tune the parameters to optimize the performance of baseline methods for comparison. In our experiments, we set α = 0.0001 for LP, parameter α = 0.01 for Katz, ϕ = 0.99 and φ = 1 for LHNII, η = 0.1 for SPM, NMF − D1, NMF − A1, NMF − D2 and NMF − D2.