Patch seriation to visualize data and model parameters.

Area of Science:

• Data Science
• Computational Chemistry
• Materials Science
• Bioinformatics

Background:

• Visualizing high-dimensional data matrices is challenging.
• Identifying clusters within complex datasets requires robust methods.
• Existing techniques may struggle with missing data or multi-dimensional arrays.

Purpose of the Study:

• To introduce a new seriation merit function for improved data matrix visualization.
• To develop a method for powerful data cluster identification.
• To enhance the interpretability of complex data models, including artificial neural networks.

Main Methods:

• Calculation of a local similarity matrix based on neighboring objects.
• Construction of a global function to maximize local similarities.
• Application of row and column ordering to form data clusters.
• Demonstration on diverse datasets (QSAR, chemical, material science, food science, cheminformatics, environmental) in 2D and 3D.

Main Results:

• The seriation method effectively identifies data clusters when similarity is driven by distinct variable sets.
• The approach handles missing data and multi-dimensional arrays.
• Feasibility demonstrated across multiple scientific domains.

Conclusions:

• The developed seriation merit function enhances visual information in data matrices.
• The method provides a powerful tool for cluster elucidation in diverse scientific data.
• It aids in the development and interpretation of artificial neural network models by revealing interpretable clusters.