Data Centric MLedit

Data Centric ML, or data-centric machine learning, studies how changes to data can improve model behavior, treating data alongside model architecture as a central focus of improvement. The relevant operations include selection, pruning, weighting, deletion, synthesis, and cross-party evaluation.

Introductionedit

Data-centric ML studies interventions made through data operations. Some operations happen after training, such as deletion and reweighting; others happen before or during training, such as pruning, synthetic-data filtering, and cross-silo evaluation.

Research contextedit

Data-centric ML connects Qiao's earlier machine-unlearning work with his current AI and networks direction. Research on data selection, sample selection bias, synthetic data, machine unlearning, and collaborative evaluation examines how a model changes when the data process changes.

Publicationsedit

Paper	Venue/status
Hessian-Free Online Certified Unlearning	ICLR 2025, 24-28 April 2025, Singapore.
DynFrs: An Efficient Framework for Machine Unlearning in Random Forest	ICLR 2025, 24-28 April 2025, Singapore.
Beyond Binary Erasure: Soft-Weighted Unlearning for Fairness and Robustness	AAAI 2026, 20-27 January 2026, Singapore.
When Sample Selection Bias Precipitates Model Collapse	ICML 2026, 6-11 July 2026, Seoul.

Connection to Qiao's workedit

In Qiao's publication record, data-centric ML appears in several forms. In unlearning, the data operation is removal or reweighting after training. In model-collapse work, the operation is selection of real or synthetic examples before recursive training, with low-resource verification exposing how local filters can mistake rare valid modes for low-quality samples. The common question is whether a learning system can identify which data matter, which data harm reliability, and which data can be safely ignored under realistic cost constraints.

See alsoedit

• Data Selection
• Sample Selection Bias
• Synthetic Data
• Machine Unlearning
• AI and Networks