Curriculum negative mining for temporal networks.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Network Science

Background:

• Temporal networks model dynamic interactions, crucial for social and e-commerce applications.
• Existing Temporal Graph Neural Networks (TGNNs) focus on model architecture, neglecting negative sample quality.
• Negative sampling in TGNNs faces challenges like positive sparsity and positive shift.

Purpose of the Study:

• To introduce Curriculum Negative Mining (CurNM), a framework to enhance TGNN training by improving negative sample quality.
• To address positive sparsity and positive shift challenges inherent in temporal network data.
• To develop a robust and adaptive method for selecting informative negative samples.

Main Methods:

• Curriculum Negative Mining (CurNM) framework employing model-aware curriculum learning.
• Dynamically updated negative pool balancing random, historical, and hard negatives.
• Temporal-aware negative selection module and annealing random negatives for stable training.

Main Results:

• CurNM significantly outperforms baseline methods across 12 datasets and 3 TGNN architectures.
• Ablation studies and parameter sensitivity analyses confirm the method's effectiveness and robustness.
• The approach successfully tackles positive sparsity and positive shift in temporal network training.

Conclusions:

• Curriculum Negative Mining (CurNM) provides a robust solution for negative sampling in TGNNs.
• The proposed framework enhances representation quality and training stability for temporal networks.
• This work offers a significant advancement in the field of Temporal Graph Neural Networks.