Graph Neural Networks (GNNs) are prominent in graph machine learning and have
shown state-of-the-art performance in Link Prediction (LP) tasks. Nonetheless,
recent studies show that GNNs struggle to produce good results on low-degree
nodes despite their overall strong performance. In practical applications of
LP, like recommendation systems, improving performance on low-degree nodes is
critical, as it amounts to tackling the cold-start problem of improving the
experiences of users with few observed interactions. In this paper, we
investigate improving GNNs' LP performance on low-degree nodes while preserving
their performance on high-degree nodes and propose a simple yet surprisingly
effective augmentation technique called NodeDup. Specifically, NodeDup
duplicates low-degree nodes and creates links between nodes and their own
duplicates before following the standard supervised LP training scheme. By
leveraging a ''multi-view'' perspective for low-degree nodes, NodeDup shows
significant LP performance improvements on low-degree nodes without
compromising any performance on high-degree nodes. Additionally, as a
plug-and-play augmentation module, NodeDup can be easily applied to existing
GNNs with very light computational cost. Extensive experiments show that
NodeDup achieves 38.49%, 13.34%, and 6.76% improvements on isolated,
low-degree, and warm nodes, respectively, on average across all datasets
compared to GNNs and state-of-the-art cold-start methods.

通过使用名为 NodeDup 的简单但有效的增强技术，研究论文提出了一种改进 GNN 对低度节点的链接预测性能的方法，从而在不损害高度节点性能的同时，在低度节点上显著改善性能。

节点复制提高冷启动链路预测

Node Duplication Improves Cold-start Link Prediction

To evaluate the performance of prediction of missing links, the known data
are randomly divided into two parts, the training set and the probe set. We
argue that this straightforward and standard method may lead to terrible bias,
since in real biological and information networks, missing links are more
likely to be links connecting low-degree nodes. We therefore study how to
uncover missing links with low-degree nodes, namely links in the probe set are
of lower degree products than a random sampling. Experimental analysis on ten
local similarity indices and four disparate real networks reveals a surprising
result that the Leicht-Holme-Newman index [E. A. Leicht, P. Holme, and M. E. J.
Newman, Phys. Rev. E 73, 026120 (2006)] performs the best, although it was
known to be one of the worst indices if the probe set is a random sampling of
all links. We further propose an parameter-dependent index, which considerably
improves the prediction accuracy. Finally, we show the relevance of the
proposed index on three real sampling methods.

本文研究了如何预测链接缺失的低度节点，实验表明，Leicht-Holme-Newman 指数表现最佳，提出的一种与参数相关的指数显著提高了预测准确性，并在三种实际抽样方法上验证了所提出的指数的相关性。