Agentic data science with the Data Agent Kit
cloud.google.com·7h ago
AI Research
arXiv CS.AI·Jul 27·5 min read
FrED: External Data Influence Estimation via Domain Knowledge Graph Grounding
TL;DR
FrED introduces a novel probabilistic framework for Training Data Attribution that combines continuous feature similarities with domain-specific Knowledge Graphs, enabling effective influence estimation without model access. It outperforms traditional methods in artistic image synthesis and enhances flood forecasting accuracy.
✦ Why It Matters
Implement FrED to enhance data attribution transparency in your AI models without needing internal access.
Key Takeaways
- Utilize FrED for effective data influence estimation in AI models.
- Incorporate domain-specific Knowledge Graphs for improved context.
- Benchmark against traditional methods to validate performance gains.
- Apply findings to enhance transparency in generative AI applications.
How It Works
FrED operates by merging continuous feature similarities—numerical representations of data points—with discrete Knowledge Graphs that encapsulate domain-specific relationships. This fusion allows the framework to reward highly relevant historical samples while suppressing the influence of less pertinent background data, ensuring that the attributions reflect the true structural context of the data.