Label Leakage
A problem in which the target variable indirectly enters the information visible to the model during training or validation, through features, edges, or the manner of data splitting. On graph data it is especially covert because nodes are interconnected.
Timeline (last 90 days)
In GNN training, the labels of validation nodes (such as the final exam scores of second-year students) should be hidden during validation to prevent label leakage.
Label leakage is especially insidious on graph data, because nodes are interconnected through edges, and the neighbors of validation nodes may include training nodes, causing validation metrics to be overestimated.
All Facts (2)
In GNN training, the labels of validation nodes (such as the final exam scores of second-year students) should be hidden during validation to prevent label leakage.
50%UnverifiedLabel leakage is especially insidious on graph data, because nodes are interconnected through edges, and the neighbors of validation nodes may include training nodes, causing validation metrics to be overestimated.
50%