Link prediction based on fundamental heuristic elements

Abstract

Considerable efforts have been made for link prediction by researchers from various disciplines because of its important value in a wide range of applications. Heuristics methods, which predict links based on some assumptions, can attain commendable accuracy when their assumptions are met. Otherwise, their effectiveness may be unsatisfactory. On the other hand, the methods that leverage Graph Neural Networks to learn the representations of node pairs have been confirmed to be effective for link prediction. However, they are usually very time-consuming. To circumvent these issues, we put forth the HELF method, a new link prediction technique built on fully connected neural networks (FCNNs) with fundamental heuristic elements. By investigating the formulas of a collection of heuristic methods, we extract a series of fundamental heuristic elements from them, which cover the core structural profiles of node pairs. Then, we encode target node pairs into feature vectors using these fundamental heuristic elements and feed the feature vectors to a FCNN to gauge the existence of links. Extensive experiments are conducted on several networks to evaluate the performance of our HELF method. The results demonstrate that HELF outperforms the well-known heuristic methods and state-of-the-art neural network-based methods subject to AUC and AP. Additionally, HELF runs much faster than these neural network-based methods.