New Lasso method removes batch constraints with gradient-enhanced approach

A team of researchers (Junzhuo Gao, Ling Peng, Xu Guo, Heng Lian) introduces a renewable Lasso approach that eliminates the batch-number constraint in online high-dimensional GLM estimation. Their method uses a gradient-enhanced surrogate loss that approximates the cumulative loss using only historical summaries, improving upon existing renewable estimation techniques. This allows streaming data to be processed without restricting the number of batches, which was a key limitation of prior work. The approach achieves non-asymptotic error bounds under high-dimensional scaling, proving its statistical efficiency.

For distributed settings (master-client architecture), the method adapts to handle batches partitioned across sites, exchanging only gradient vectors rather than full data. Unlike prior work, clients do not need to compute the full surrogate loss, reducing communication overhead. Simulation results under linear and logistic models, plus a real-data application, demonstrate improved accuracy over existing renewable estimators. The paper is available on arXiv (2606.11738) as a preprint.