Link Prediction in Networked Data Using Structural Features and Machine Learning
TL;DR
It is indicated that structural information remains important for link prediction in contemporary learning-based models and future work will continue to improve the scalability, robustness, temporal adaptation and interpretability of link prediction models for large-scale, dynamic real-world networks.
Abstract
Link prediction is a problem in network analysis that seeks to find missing or future links in a graph based on observed graph structures and available node information. This paper reviews the development of link prediction methods from traditional structural heuristics to network embeddings and graph neural networks. Common Neighbors and other similarity indices are traditional methods that are simple to use and interpret but rely heavily on manually designed structural features. Representation-learning methods, such as node2vec, reduce the requirement for feature engineering by learning low-dimensional node embeddings. More recently, graph neural network (GNN)-based methods have been developed to address issues such as neighbourhood overlap, multi-hop paths, graph incompleteness and long-tailed structural patterns in link prediction, including Neighborhood Overlap-aware Graph Neural Networks (Neo-GNNs), Neural Bellman-Ford Network (NBFNet), Common Neighbor Completion with Information Entropy (IECNC) and Long-Tailed Link Prediction (LTLP). Recently, temporal link prediction has also begun to attract some attention, and EAGLE has been developed to combine both recent temporal information and global structural patterns effectively. In short, the reviewed studies indicate that structural information remains important for link prediction in contemporary learning-based models. Future work will continue to improve the scalability, robustness, temporal adaptation and interpretability of link prediction models for large-scale, dynamic real-world networks.
