Robust self-adaptive curriculum scheduling under cross-set information leakage

Develop adaptation signals for self-adaptive curriculum scheduling in graph unlearning that remain robust to cross-set information leakage caused by message passing between retained and forget-set nodes.

Background

CUNO currently uses a fixed number of curriculum stages and fixed per-stage epoch allocations. The paper reports that a self-adaptive variant based on online signals such as forget-set loss and prediction confidence did not consistently outperform the fixed baseline.

The unresolved difficulty is specific to graph unlearning: message passing continuously transfers retained-set information into forget-set nodes, corrupting the loss and confidence signals that adaptive scheduling would otherwise use. The paper therefore identifies the development of adaptation signals that can withstand this cross-set leakage as an open problem.