Data Augmentation for Visualization Design Knowledge Bases.

Area of Science:

• Computer Science
• Data Visualization

Background:

• Visualization knowledge bases aid computational reasoning in design spaces.
• Current systems use feature weights learned from limited chart pair data, lacking comprehensive trade-off assessment.

Purpose of the Study:

• To improve visualization knowledge base coverage and accuracy.
• To develop data augmentation techniques for generating and labeling chart pairs.
• To scale labeling efforts for learning updated feature weights.

Main Methods:

• Data augmentation techniques for generating novel chart pairs via design permutations and identifying under-assessed features.
• Comparison of varied methods to scale labeling efforts for annotating chart pairs.
• Evaluation within the Draco knowledge base context.

Main Results:

• An expanded corpus with thousands of new chart pairs.
• Demonstrated improvements in feature coverage.
• Enhanced chart recommendation performance.

Conclusions:

• Data augmentation significantly expands visualization knowledge base training data.
• The proposed methods improve the accuracy and coverage of learned feature weights.
• This work enhances computational reasoning and recommendation capabilities in visualization design.