aiqbench analysis
CRISP addresses the computational bottleneck of training gradient-boosted trees on massive, imbalanced tabular datasets by introducing a linear-time coreset reduction method. By strategically allocating sampling budgets across quantile strata and applying inverse-propensity weighting, it achieves significant data reduction while maintaining near-full model accuracy on production fraud detection tasks. This approach offers a practical pathway to scaling tree-based models without sacrificing performance on critical minority classes.
CRISP: Scalable Importance-Stratified Coresets for Imbalanced Tabular Learning
arXiv:2609.26962v1 Announce Type: new Abstract: Large imbalanced tabular datasets make repeated gradient-boosted tree training expensive. Existing coreset methods often lose accuracy when most majority examples are removed. We present CRISP (Coreset Reduction via Importance-Stratified Pruning), a linear-time method that allocates a negative-class budget across quantile strata of a proxy-model score. Sample weights account for unequal inclusion probabilities. At 95% negative-class reduction on a production fraud dataset, CRISP trains on approximately 1.70M of 25M rows and retains 99.7% of full-data Average Precision. This is a 93.2% reduction in total training rows. On public CriteoPrivateAds, CRISP has the highest mean Average Precision at each tested rate from 90% to 99.4% majority reduction. Sparkov results are mixed at lower rates, but CRISP has the highest mean at 99.2% and 99.4%. Ablations identify budget allocation and inverse-propensity weighting as the main sources of the production-dataset gain.
Abstract:Large imbalanced tabular datasets make repeated gradient-boosted tree training expensive. Existing coreset methods often lose accuracy when most majority examples are removed. We present CRISP (Coreset Reduction via Importance-Stratified Pruning), a linear-time method that allocates a negative-class budget across quantile strata of a proxy-model score. Sample weights account for unequal inclusion probabilities. At 95% negative-class reduction on a production fraud dataset, CRISP trains on approximately 1.70M of 25M rows and retains 99.7% of full-data Average Precision. This is a 93.2% reduction in total training rows. On public CriteoPrivateAds, CRISP has the highest mean Average Precision at each tested rate from 90% to 99.4% majority reduction. Sparkov results are mixed at lower rates, but CRISP has the highest mean at 99.2% and 99.4%. Ablations identify budget allocation and inverse-propensity weighting as the main sources of the production-dataset gain.
Machine Learning (cs.LG)
arXiv:2609.26962 [cs.LG]
(or arXiv:2609.26962v1 [cs.LG] for this version)
https://doi.org/10.48550/arXiv.2609.26962 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 231 words
Related coverage
Handling Imbalanced Classification: What Works Better Than SMOTE
Analytics Vidhya · Jul 12, 2026
Tabular Imbalanced Learning: A Survey, Benchmark, and Practical Guide
arXiv · Sep 28, 2026
K-IPO: Kendall-constrained Importance Preserving Oversampling for Imbalanced Tabular Data
arXiv · Jul 21, 2026
Quantum Circuits Cut Data Imbalance Recall Gap by 10–25%
Quantum Zeitgeist · Jul 13, 2026