Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
Télécom Paris, Institut Polytechnique de Paris
4
0
4
Edge-girth outperforms traditional structural descriptors in GNNs, achieving a significant reduction in error while exposing inherent limitations in graph discrimination.
Targeted sampling of informative pairs can slash computational costs in pairwise loss estimation while maintaining performance on par with full evaluations.
Link prediction models amplify structural biases beyond immediate connections, but a new $k$-hop fairness approach can mitigate these disparities with favorable performance trade-offs.
Algorithmic recommendations subtly warp network evolution, and this new measure reveals how link prediction strategies amplify existing biases in dynamic graphs.