Node-degree aware edge sampling mitigates inflated classification performance in biomedical random walk-based graph...

Area of Science:

• Computational biology
• Machine learning
• Graph theory

Background:

• Graph representation learning generates embeddings for nodes and graph elements, useful for tasks like predicting new relationships (edges).
• Biomedical knowledge graphs often have known positive relationships but lack explicit negative relationships, forcing models to assume most unlabeled edges are negative.
• Current methods uniformly sample negative edges, leading to imbalanced node degree distributions between positive and negative examples.

Purpose of the Study:

• To investigate the impact of uniform negative edge sampling on graph representation learning performance in biomedical applications.
• To develop and present a novel, degree-aware node sampling approach to mitigate biases in negative example selection.

Main Methods:

• Utilized a representative heterogeneous biomedical knowledge graph.
• Employed random walk-based graph machine learning techniques.
• Implemented and compared a novel degree-aware node sampling strategy against uniform sampling.

Main Results:

• Uniform negative edge sampling results in imbalanced node degree distributions, significantly affecting classification performance.
• This imbalance can artificially inflate model performance estimates during validation.
• The proposed degree-aware sampling approach effectively mitigates this bias, leading to more accurate model evaluation.

Conclusions:

• The method of selecting negative examples is critical for accurate graph representation learning, especially in biomedicine.
• Degree-aware node sampling provides a more reliable approach for training and evaluating graph-based machine learning models.
• The developed method is publicly available and easy to implement for researchers.