Using slisemap to interpret physical data.

Area of Science:

• Data visualization
• Machine learning
• Physical sciences

Background:

• High-dimensional datasets are common in physical sciences.
• Manifold visualization techniques are widely used for data exploration.
• Explainable artificial intelligence (XAI) is crucial for understanding complex models.

Purpose of the Study:

• To apply and evaluate the slisemap manifold visualization technique on physics and chemistry datasets.
• To demonstrate how slisemap integrates manifold visualization with XAI.
• To showcase slisemap's utility in uncovering patterns and behaviors in scientific data.

Main Methods:

• Application of slisemap, a novel manifold visualization method.
• Integration of slisemap with explainable artificial intelligence (XAI) principles.
• Analysis of datasets from physics and chemistry domains.

Main Results:

• slisemap successfully creates embeddings where data items with similar local explanations are clustered.
• The patterns in the slisemap embedding reflect target properties of the data.
• Meaningful insights were found in classification and regression models trained on physical data.

Conclusions:

• slisemap provides a valuable overview of black box model behaviors.
• The technique is effective for analyzing and interpreting scientific datasets.
• slisemap aids in extracting meaningful information from machine learning models in physical sciences.