서동혁/김세인’s paper has been accepted in

Title: SCOUT: Structure-Aware Aspect and Anchor-Count Selection for Node Attribute Augmentation via Positional Information
Author: Dong-Hyuk Seo, Sein Kim, Taeri Kim, Won-Yong Shin, and Sang-Wook Kim
Abstract
When node attributes are absent, graph neural networks often fail to distinguish locally isomorphic nodes and underperform on graph downstream tasks. Positional information (PI) augmentation mitigates this failure by selecting representative nodes as anchors and encoding anchor–node distances as augmented attributes. However, the performance of PI-based methods hinges on two graph dependent choices: the structural measures used for anchor selection and anchor-node distance metrics, and the anchor-count 𝐾. To address these challenges, we propose SCOUT, a model-agnostic augmentation that learns which structural aspect is used and how many anchors are selected for each graph. To this end, first, we regard each centrality–similarity pair as a positional aspect; encode diverse positional aspect attributes, and learn a graph-level attention that selects the aspect that best fits the underlying graph and task while sharing the selection across nodes. Second, leveraging the longstanding observation that centrality distributions are heavy-tailed and frequently approximate power laws, we apply an elbow detector to the ranked centrality curve to identify the most central nodes and thereby determine the anchor-count 𝐾. SCOUT is model-agnostic and thus can be integrated into standard GNNs for downstream tasks such as node classification and link prediction. Without original attributes, SCOUT yields +26.88 Hits@20 points on ogbl-ddi and +4.52 accuracy points on ogbn-mag; with original attributes, it brings additional gains of +6.15 AUC points on Cora and +11.69 accuracy points on ogbn-mag. The source code of SCOUT is available at https://anonymous.4open.science/r/SCOUT-DB7B/.