Not Every Node Benefits from Expensive Recovery: Risk-Aware Routing for Incomplete Multimodal Graph Learning

Nodes with missing modalities have limited information for downstream prediction in multimodal graphs, prompting the development of increasingly sophisticated and computationally expensive recovery methods, which are often applied uniformly to affected nodes. However, our analysis shows that expensive recovery can turn correct predictions into errors even when it improves overall performance, while it can still correct cheap-path errors when it degrades overall performance. These motivate selecting nodes for expensive recovery by balancing potential correction against the risk of introducing errors. We propose Risk-aware Selective Modality Recovery (R-SMR), which formulates this selection as budget-constrained routing between cheap and expensive recovery endpoints. Using only cheap-path information at inference, R-SMR estimates the probabilities of four joint correctness states and scores each switch by balancing predicted correction against harm. It selects the highest-scoring nodes with positive utility within a budget for expensive recovery calls, leaving budget unused when further switches are predicted to be unfavorable. Across six datasets and five recovery methods, R-SMR consistently outperforms the stronger endpoint up to 4.10%, with improvements persisting in both globally beneficial and harmful recovery regimes even at a 5% budget. Comparisons with random routing and ablations demonstrate the importance of balancing correction against harm and rejecting unfavorable switches.