Abstract
Dynamic node classification is critical for modeling evolving systems likefinancial transactions and academic collaborations. In such systems,dynamically capturing node information changes is critical for dynamic nodeclassification, which usually requires all labels at every timestamp. However,it is difficult to collect all dynamic labels in real-world scenarios due tohigh annotation costs and label uncertainty (e.g., ambiguous or delayed labelsin fraud detection). In contrast, final timestamp labels are easier to obtainas they rely on complete temporal patterns and are usually maintained as aunique label for each user in many open platforms, without tracking the historydata. To bridge this gap, we propose PTCL(Pseudo-label Temporal CurriculumLearning), a pioneering method addressing label-limited dynamic nodeclassification where only final labels are available. PTCL introduces: (1) atemporal decoupling architecture separating the backbone (learning time-awarerepresentations) and decoder (strictly aligned with final labels), whichgenerate pseudo-labels, and (2) a Temporal Curriculum Learning strategy thatprioritizes pseudo-labels closer to the final timestamp by assigning themhigher weights using an exponentially decaying function. We contribute a newacademic dataset (CoOAG), capturing long-range research interest in dynamicgraph. Experiments across real-world scenarios demonstrate PTCL's consistentsuperiority over other methods adapted to this task. Beyond methodology, wepropose a unified framework FLiD (Framework for Label-Limited Dynamic NodeClassification), consisting of a complete preparation workflow, trainingpipeline, and evaluation standards, and supporting various models and datasets.The code can be found at https://github.com/3205914485/FLiD.