PRISM: Partial-label Relational Inference with Spatial and Spectral Cues

In many real-world scenarios, acquiring precise labels for graph-structured data is expensive or even infeasible, as reliable annotation often requires substantial expert knowledge or computational resources. As a result, graph labels are often noisy and ambiguous. This challenge motivates partial-label graph learning, where each graph is weakly annotated with a candidate label set containing the true label. However, such ambiguous supervision makes it hard to extract reliable graph semantics and increases the risk of overfitting to noisy candidate labels. To address these challenges, we propose a unified framework named PRISM that performs relational inference with spatial and spectral cues to alleviate the impact of label ambiguity. On the one hand, PRISM captures discriminative spatial cues by aligning prototype-guided substructures across graphs. On the other hand, it decomposes graph signals into multiple frequency bands and extracts global spectral cues with an attention mechanism, which preserve frequency-specific semantics. We integrate these complementary views into a hybrid relational graph and perform an iterative label propagation under candidate constraints. Extensive experiments on a range of well-known datasets demonstrate that PRISM consistently outperforms strong baselines under various noise settings.