HypoML: Visual Analysis for Hypothesis-based Evaluation of Machine Learning Models.

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Visualization

Background:

• Evaluating machine learning (ML) models often relies on statistical hypothesis testing.
• Reasoning about multiple hypotheses in ML model evaluation can be complex and non-intuitive.
• Existing methods may not fully integrate logical inference with statistical testing for comprehensive model assessment.

Purpose of the Study:

• To present HypoML, a visual analytics tool designed for hypothesis-based evaluation of ML models.
• To introduce a novel ML-testing framework integrating statistical hypothesis testing with logical reasoning.
• To facilitate the understanding of how specific information impacts ML model performance.

Main Methods:

• Developed a novel ML-testing framework combining statistical hypothesis testing and logical reasoning.
• Created HypoML, a visual analytics tool to process and visualize multi-thread testing results.
• Transformed statistical and logical inferences into intuitive visual representations for rapid observation.

Main Results:

• HypoML enables hypothesis-based evaluation of ML models through a novel framework.
• The tool visualizes analytical results derived from statistical and logical inferences.
• Demonstrated the intuitive and explainable nature of the visual analysis with hypothesized concepts.

Conclusions:

• HypoML provides an intuitive and explainable visual analysis for ML model evaluation.
• The framework supports rigorous testing of hypotheses regarding feature impact on ML models.
• Visual analytics significantly aids in understanding complex hypothesis testing outcomes in ML.