Table 1.
Comparison of our proposed SuRe and existing methods with respect to various aspects in ranking and link prediction tasks.
SuRe outperforms all learning methods in terms of accuracy, speed, scalability, and memory usage.
Table 2.
Table of symbols.
Fig 1.
Example of RWR and our proposed approaches RWER & SuRe on a political blog network.
RWR uses the fixed restart probability 0.15 or 0.5 while our proposed RWER uses distinct restart probabilities on nodes.
Fig 2.
Each node has its own restart probability.
Fig 3.
Flowchart of RWER (Algorithms 1 and 2) and SuRe (Algorithm 3).
SuRe learns restart probability vector c, and RWER computes our node relevance score vector r for a given seed node s.
Table 3.
The query nodes are used for the ranking and the link prediction tasks.
Table 4.
Ranking results of our proposed method SuRe and other methods w.r.t. a query node obsidianwings, a liberal blog.
Bold nodes are conservative blogs, and the non-bold ones are liberal. Our ranking result from SuRe contains only liberal nodes, indicating the best result, while other ranking results wrongly contain conservative nodes.
Fig 4.
Ranking performance on Polblogs.
Our method SuRe provides the best ranking performance compared to existing methods in terms of MAP and Precision@20. Note that the maximal values are 1 for each plot.
Fig 5.
Hop count plot in the datasets used in the link prediction task.
Note that the majority of nodes are within 2 hops in both real-world networks.
Fig 6.
Link prediction performance on HepPh dataset, where the degree of each query node is greater than or equal to 30.
SuRe shows the highest accuracies: 10.8% higher MAP, and 5.7% higher Precision@20 compared to the best existing method.
Fig 7.
Link prediction performance on the HepTh dataset, where the degree of each query node is greater than or equal to 30.
SuRe shows the highest accuracies: 14.7% higher MAP, and 10.1% higher Precision@20 compared to the best existing method.
Table 5.
AUC result, where the degree of each query node is greater than or equal to 30.
We compare SuRe with other baselines. SuRe provides the best link prediction accuracy.
Table 6.
Link prediction performance on the HepPh dataset, where the degree of each query node is less than 30.
Bold and italic fonts indicate the best and the second best methods, respectively.
Table 7.
Link prediction performance on the HepTh dataset, where the degree of each query node is less than 30.
Bold and italic fonts indicate the best and the second best methods, respectively.
Table 8.
Number of parameters for each supervised method.
Fig 8.
Sensitivity of parameter o of our method SuRe in the HepPh and HepTh datasets.
We report the link prediction accuracy using MAP measure, changing the values of the elements in the origin vector o, where all the elements of o are set to a same value. Note that the performance of SuRe is improved by introducing the origin parameter o, compared to not using o which corresponds to setting o = 0.
Fig 9.
Sensitivity of parameter λ of our method SuRe in the HepPh and HepTh datasets.
We report the MAP scores in the link prediction task, varying the value of λ. Note that SuRe avoids overfitting problem and shows improvement by introducing the regularization parameter λ. When λ is 100 = 1, SuRe exhibits the best link prediction performance in both datasets.
Fig 10.
Link prediction performance of SuRe varying the number of labeled nodes in the HepPh and HepTh datasets.
We sample labeled nodes with a sampling rate. Note that a higher sampling rate leads to more training examples which in turn improve the performance.
Fig 11.
Scalability of SuRe compared to baselines in the Wikipedia dataset.
SuRe has the smallest slope 0.76 of the fitted line, while the slopes for RWR, SRW, and QUINT are 0.83, 0.88, and 2.57, respectively. Note that SuRe is the fastest among the supervised methods SRW and QUINT. Although SuRe shows slower running time compared to RWR which is an unsupervised method, SuRe provides higher accuracy than RWR in most cases in both ranking and link prediction tasks as shown in Figs 7, 4 and 6.