This repository provides supplementary materials for the paper "IDEA: Automated Design Space Exploration for Visualization Design".

.
├── design_spaces/              # Design space definitions
│   ├── schema.json             # General schema for design spaces
│   ├── data_fact.json          # Data fact design space (Section 3.2)
│   ├── narrative_composition.json  # Narrative composition design space (Section 3.2)
│   ├── standard_visualization.json # Standard visualization design space (Section 3.2)
│   └── pictorial_visualization.json # Pictorial visualization design space (Section 3.2)
├── constraints_examples/       # Example ASP constraints
│   ├── data fact.lp            # Example constraints for data fact dimensions
│   └── narrative composition.lp # Example constraints for narrative composition
├── prompt_templates.md         # LLM prompt templates for constraint generation (Section 4.2.2)
└── params.json                 # Hyperparameters for all experiments (Section 4)

Each design space is defined as a JSON file listing the dimensions and their candidate elements. See design_spaces/schema.json for the general format. The four design spaces cover:

• Data Fact: Fact type, breakdown, measure, subspace, focus, and visualization title.
• Narrative Composition: Headline, narrative intent/structure/pattern/perspective, and data-driven/context-enhancing strategies.
• Standard Visualization: Mark type and visual encodings (x, y, color, size, shape, row, column, text, stack).
• Pictorial Visualization: Chart type, icon binding/strategy/theme, and color assignments.

The constraints_examples/ folder contains example ASP (Answer Set Programming) constraints generated by the LLM for a case study (Case I in Section 5). These illustrate the three constraint types used in IDEA:

• Hard constraints: Design choices that must be avoided.
• Soft positive constraints: Recommended design choices.
• Soft negative constraints: Discouraged design choices.

prompt_templates.md documents the complete prompt templates used for LLM-based constraint generation and validation, following the five-component framework: Role Specification, Variable Injection, Constraint Semantics, Generation Rules, and Domain-anchored Examples.

params.json contains all hyperparameters used in the experiments, including:

• Reward function weights (Eq. 7) and sensitivity analysis configurations (Section 4.5.3)
• MCTS parameters: exploration weight, max iterations, convergence criteria
• Baseline parameters: GA, SA, and Beam Search configurations (Section 4.5.1)
• Random seeds for reproducibility