FrED: Black-Box Data Attribution via Knowledge Graph Grounding

FrED provides a probabilistic framework for training data attribution in black-box settings by fusing continuous feature similarities with discrete Knowledge Graphs. This method enables estimating external data influence without model weight access, specifically targeting complex domains like artistic synthesis and weather forecasting.