STRIDE: Training Data Attribution via Sparse Recovery from Subset Perturbations

arXiv 2606.05165·medium signal

Dagli, Harrasse, and Zhang propose STRIDE, a training data attribution method that traces predictions back to training examples using sparse recovery over subset perturbations rather than expensive causal interventions. The aim is to approximate the gold-standard TDA signal at lower cost. Relevant for debugging data quality, auditing memorization, and provenance work.