When to Trust Whom: A Context-Aware Graph Routing Mechanism for Information Diffusion Prediction

Abstract

Information diffusion prediction forecasts future participants from an observed cascade prefix, enabling proactive intervention in applications such as viral marketing and misinformation mitigation. Most existing models leverage two data sources: the global social graph (exposure/trust pathways) and cascade-induced interaction relations (interest-driven co-adoption), following a ''learn-then-fuse'' pipeline that encodes both graphs with GNNs and combines them via gated fusion to condition a sequential decoder. However, we find the two views are systematically mismatched: interaction edges are largely disjoint from social links, most social neighbors never co-activate within the same cascade, and the resulting embeddings lie on near-orthogonal manifolds with negligible correspondence. With such mismatch, static fusion is ill-posed: when the views disagree, fusion enforces a compromise and can cause negative interference. We further identify three reliability mechanisms that determine when each view should be trusted: (1) behavioral consensus across views is a high-fidelity signal of influence; (2) social cues are essential in cold-start regimes where interactions are sparse and biased; and (3) social ties dominate early seeding, while interaction patterns govern the late viral stage. Motivated by these, we propose CARD, a context-aware routing framework that replaces static fusion with step-wise evidence arbitration. CARD constructs an expert pool with social and interaction experts, a consensus expert that activates when both views are confirmed to behavioral consensus, and a graph-agnostic prior expert for noisy fallback. A router hard-selects the single most reliable expert at each step, so the decoder receives a targeted signal rather than a blurred mixture. Extensive experiments on four real-world datasets show that CARD achieves state-of-the-art accuracy and stronger robustness.