Active Learning with Low-Rank Structure for Data Selection

Researchers have introduced a new data selection framework for machine learning, leveraging low-rank approximation and residual-based sampling. This approach aims to improve model training efficiency by selecting a representative subset of data.

The data selection problem involves choosing a small subset of data to train a machine learning model effectively. Previous methods relied on clustering assumptions, but modern datasets often possess global algebraic structure. The new framework uses low-rank approximation and residual-based sampling to select a weighted subset of data points, achieving improved performance over prior strategies based on uniform sampling or clustering-based sensitivity sampling^[1]. Additionally, a cost-aware online active learning framework called QueryMarket has been developed, which queries incoming data points based on their estimated utility and price. QueryMarket unifies pricing, information gain, and rolling budget constraints under concept drift, and its variant OVBAL (online variance-based active learning) integrates data pricing with information-driven selection, adapting to nonstationary streams and heterogeneous label costs. Experiments show OVBAL is effective under seller-centric pricing, yielding a more favorable long-run error-cost trade-off in a real-world task^[2].