Preventing Failures by Dataset Shift Detection in Safety-Critical Graph Applications.

Area of Science:

• Machine Learning
• Graph Theory
• Statistical Hypothesis Testing

Background:

• Dataset shift, a change in data distribution between training and testing, poses challenges in critical applications like healthcare and drug discovery.
• Existing dataset shift detection methods primarily focus on image/video data, neglecting graph-structured data.
• Detecting dataset shift in graph data is vital for maintaining model performance and reliability.

Purpose of the Study:

• To investigate dataset shift detection specifically for graph-structured data.
• To propose a novel, practical approach for identifying distribution changes in large-scale graph datasets.
• To address the limitations of existing methods by focusing on graph data.

Main Methods:

• A statistical hypothesis testing framework is employed for shift detection.
• A flexible two-sample test is developed, applicable to both undirected and directed graphs.
• The method does not require equal sample sizes, enhancing its practicality.

Main Results:

• Empirical studies demonstrate the proposed test's effectiveness in detecting dataset shifts.
• The approach is validated on real-world datasets featuring directed graphs and numerous nodes.
• The method proves robust and adaptable to different graph structures.

Conclusions:

• The proposed two-sample test provides an effective solution for dataset shift detection in graph-structured data.
• This research fills a gap in the literature by addressing shift detection for graph data.
• The findings have significant implications for machine learning applications in healthcare, drug discovery, and other safety-critical domains.