Representation learning on continuous-time dynamic graphs has garnered substantial attention for its capacity to model evolving entity relationships. However, existing methods exhibit pronounced overfitting, particularly in complex and sparse data scenarios. We empirically substantiate this overfitting through multiple indicators: (1) a significant performance discrepancy between training and validation/test sets, especially for long-term interaction predictions; (2) an inverse correlation between model complexity and generalization performance; (3) a widening temporal generalization gap as the prediction horizons extend; and (4) rapid performance deterioration under data-sparse conditions. These phenomena collectively demonstrate the overfitting issue, limiting the applicability of current approaches in cold-start scenarios and dynamic environments. To address this, we propose Continuous-Time Dynamic Graph Learning via Uncertainty Masked MIx-UP (ConUMIP), a novel data augmentation method operating in the representation space of continuous-time dynamic graphs. Unlike conventional techniques that perturb raw graph data, ConUMIP adaptively captures temporal evolution patterns and generates diverse augmented samples. This approach effectively mitigates overfitting while enhancing long-term dependency modeling. By eschewing predefined time windows and integrating both local and global structures, ConUMIP demonstrates superior adaptation to complex dynamic evolution patterns. Comprehensive evaluations across five real-world datasets validate ConUMIP's efficacy in substantially improving both the performance and generalizability of existing continuous-time dynamic graph models, particularly in long-term predictions and data-sparse scenarios, without incurring additional computational complexity, thus offering a robust solution to the overfitting challenge in this domain.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号