Traceprop: End-to-End Provenance-Guided Data Attribution for Auditable ML
:::tip Full Breakdown This is a complete engineering breakdown of the Traceprop preprint (Zenodo DOI: 10.5281/zenodo.20036000, May 2026). The library is available at pypi.org/project/traceprop. Apache 2.0 license. :::
| Authors | Amit N. |
| Year | 2026 |
| DOI | 10.5281/zenodo.20036000 |
| Code | github.com/AmitoVrito/Traceprop |
| Install | pip install traceprop |
| License | Apache 2.0 |
Abstract
Traceprop is an open-source Python library providing the first unified system for end-to-end data provenance in machine learning pipelines, connecting raw source files through preprocessing, through model training, to individual predictions. Existing data attribution methods (Koh and Liang, 2017; Park et al., 2023; Engstrom et al., 2024) identify which training samples influenced a prediction but operate in isolation from the data pipeline. Existing computation lineage tools (MLflow, DVC, TensorFlow MLMD) track artifact-level provenance but do not descend into the computation graph or connect to gradient-level attribution. Traceprop fills this gap by introducing a computation-level lineage layer that integrates natively with gradient-based attribution.
Key results: sub-1% lineage overhead at 10^6+ array elements (1.007x macOS, 0.979x Linux); LDS 0.622 +/- 0.180 on tabular data (UCI Adult Income, logistic regression) at 0.22s CPU; Traceprop-LL achieving LDS 0.0168 on CIFAR-2/ResNet-9 vs TRAK's 0.0290 at 266x lower wall-clock cost (2.6s CPU vs 691s GPU); provenance-guided approximate unlearning exceeding the retrain-from-scratch gold standard (forget-set loss 0.425 vs gold 0.401) with only 0.5pp test accuracy drop.
Engineering Breakdown
The Problem
Existing data attribution methods (Koh and Liang, 2017; Park et al., 2023; Engstrom et al., 2024) identify which training samples influenced a prediction but operate in isolation from the data pipeline. Existing computation lineage tools (MLflow, DVC, TensorFlow MLMD) track artifact-level provenance but do not descend into the computation graph or connect to gradient-level attribution.
The Approach
Existing data attribution methods (Koh and Liang, 2017; Park et al., 2023; Engstrom et al., 2024) identify which training samples influenced a prediction but operate in isolation from the data pipeline.
Key Results
Key results: sub-1% lineage overhead at 10^6+ array elements (1.007x macOS, 0.979x Linux); LDS 0.622 +/- 0.180 on tabular data (UCI Adult Income, logistic regression) at 0.22s CPU; Traceprop-LL achieving LDS 0.0168 on CIFAR-2/ResNet-9 vs TRAK's 0.0290 at 266x lower wall-clock cost (2.6s CPU vs 691s GPU); provenance-guided approximate unlearning exceeding the retrain-from-scratch gold standard (forget-set loss 0.425 vs gold 0.401) with only 0.5pp test accuracy drop.
Research Areas
This paper contributes to the following areas of AI/ML engineering:
- Machine learning
- Deep learning
- Neural networks
- Model optimization
- AI systems
- Attribution
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::