我院2025级概率论与数理统计研究生肖双玲在盖玉洁教授的指导下，于我校认定的AAA类期刊同时也是中科院一区期刊的Knowledge-Based Systems上发表了一篇关于在线迁移学习的论文。该论文提出了一种用于广义线性模型预测的机器学习框架—在线迁移学习。该框架旨在通过从源域迁移知识来完成目标域的在线学习任务。其研究两种在线迁移学习场景：（1）具有共享特征空间的同构域在线迁移学习；（2）具有不同特征空间的异构域在线迁移学习。针对每种场景，该文设计了高效的广义线性模型预测算法（MLE-IASGD与HetMLE-IASGD），以解决目标域的分类问题。最终通过在模拟数据集和真实数据集上的对比实验验证了所提方法，实证结果充分证明了该框架的优越性：在同构场景下，MLE-IASGD在准确率与AUC值上均达到最优性能；在更具挑战性的异构场景中，HetMLE-IASGD同样表现卓越，以领先的准确率和AUC值显著优于所有基线方法。这些发现有力验证了在线迁移学习方法在两种场景下的有效性与泛化能力。

论文题目：

Online transfer learning for generalized linear model prediction

论文摘要：

This paper proposes a machine learning framework for generalized linear model (GLM) prediction, termed Online Transfer Learning (OTL), which aims to accomplish online learning tasks in the target domain by transferring knowledge from the source domain. OTL faces substantial challenges, primarily because the data from the source and target domains may differ not only in distribution but also in feature representation. To address this, we investigate two settings of OTL: (i) OTL on homogeneous domains with a shared feature space, and (ii) OTL on heterogeneous domains with different feature spaces. For each case, this paper designs efficient OTL algorithms (MLE-IASGD and HetMLE-IASGD) for generalized linear model prediction to address classification problems in the target domain. Finally, comparative experiments on both simulated and real-world datasets are conducted to validate the proposed methods, and empirical results confirm the superiority of the proposed framework: under homogeneous settings, MLE-IASGD achieves top performance in accuracy (0.7289) and AUC (0.8798). Under more challenging heterogeneous conditions, HetMLE-IASGD also excels, with leading accuracy (0.8151) and AUC (0.7363), significantly outperforming all baselines. These findings strongly validate the effectiveness and generalizability of the OTL approach across both scenarios.

作者简介：

盖玉洁，中央财经大学统计与数学学院数理统计系主任，教授，博士生导师。主要研究领域包括海量数据统计建模、在线学习、高维变量选择等。在《中国科学：数学》、Statistical Sinica、Statistics and Computing等国内外期刊发表学术论文40余篇。

肖双玲，来自中央财经大学统计与数学学院2025级概率论与数理统计专业，本科毕业于湘潭大学，曾获得国家励志奖学金、全国大学生数学竞赛省三等奖等荣誉，并作为项目主持人成功完成省级创新创业训练项目《图神经网络在原油价格预测方面的应用》。

撰稿人：盖玉洁

审稿人：邓 露