Use of AI in Modern Dataviz - Øystein Moseng - NDC Oslo 2024

Hybrid approaches combining AI with traditional deterministic methods often yield better results than pure AI solutions - use AI for what it’s good at (like NLP) and traditional approaches for data processing

When working with LLMs for data visualization:

• Keep prompts concise and unambiguous
• Provide clear examples of input/output
• Send metadata about data rather than full datasets
• Use one-shot API calls when possible
• Implement proper error handling and validation

Key challenges to address when using AI for data visualization:

• Accuracy and data consistency cannot be compromised
• Performance and response times
• Privacy concerns with third-party services
• Cost management when using external APIs
• Handling hallucinations and unpredictable outputs

For data processing queries:

• Have AI generate a recipe/stack of actions rather than processing directly
• Route different queries to specialized models
• Include relevant metadata like column types, statistics, and sample values
• Validate outputs before execution

The data visualization process typically involves:

• Data gathering/processing stage
• Insights discovery/analysis stage
• Design phase for effective representation
• Making it accessible and interactive for end users

AI can assist in:

• Automatic anomaly detection
• Format translation and adapter creation
• Accessibility features like audio charts
• Providing design suggestions and best practices
• Natural language interfaces for visualization tools

Focus on using AI to augment rather than replace existing tools and processes - combining AI capabilities with deterministic approaches often provides better results while maintaining control over critical aspects