报告时间：2026.9.14日10：00

报告地点：曲江校区教9楼320

题目：Heterogeneous Collaborative Learning for Data Silos: Qualified Distributed Local Average Regression

摘要：Collaborative prediction promises to overcome data scarcity by uniting distributed data silos. Its practical deployment, however, rests on a premise that existing schemes largely ignore: multi-agent collaboration is inherently heterogeneous, as practitioners differ in how they predict and how much data they invest, while interpretability and privacy remain largely unaddressed. We propose qualified distributed local average regression (QDLAR), a framework that accommodates algorithmic, parameter, and data-investment heterogeneity, provides built-in resistance to membership inference attacks, and delivers inherently interpretable predictions, all without sacrificing prediction accuracy. QDLAR adopts a similarity-based modeling paradigm while imposing no restriction on practitioners' choice of similarity metrics, thereby structurally embedding privacy, interpretability, and algorithmic diversity into the framework itself. Three integrated mechanisms, namely logarithmic calibration, qualification, and contribution-aware weighted aggregation, accommodate heterogeneity while ensuring near-optimal global performance. Within the framework of statistical learning theory, we prove that QDLAR attains near-optimal generalization error bounds. Experiments on synthetic and real-world data corroborate the theoretical results and demonstrate the feasibility and effectiveness of QDLAR.

个人简介：刘小彤，华中农业大学信息学院副教授，博士毕业于西安交通大学。主要研究领域为机器学习理论、隐私保护与可解释性。致力于将数学与机器学习理论同信息安全领域关注的隐私、安全等可信要素相结合，开展交叉研究。近年来在JMLR、SIIMS 等国际顶级期刊发表多篇学术论文。