報(bào) 告 人：劉慧航 博士
報(bào)告題目：Trans-MA: Sufficiency-principled Transfer Learning via Model Averaging
報(bào)告時(shí)間：2025年03月05日（周三）下午3：00
報(bào)告地點(diǎn)：靜遠(yuǎn)樓1506學(xué)術(shù)報(bào)告廳
主辦單位：數(shù)學(xué)與統(tǒng)計(jì)學(xué)院、數(shù)學(xué)研究院、科學(xué)技術(shù)研究院
報(bào)告人簡(jiǎn)介：
劉慧航博士是中國(guó)科學(xué)技術(shù)大學(xué)國(guó)際金融研究院博士后. 2023年畢業(yè)于中國(guó)科學(xué)技術(shù)大學(xué). 研究方向?yàn)槟Ｐ推骄c遷移學(xué)習(xí). 主要的工作內(nèi)容包括針對(duì)有向和無(wú)向高斯圖模型, 遷移學(xué)習(xí), 非對(duì)稱損失的回歸模型進(jìn)行參數(shù)的模型平均.論文發(fā)表于 Biometrics 和 Journal of Business & Economic Statistics 等期刊.
報(bào)告摘要：
Domain aggregation in multi-source transfer learning faces a critical challenge: effectively integrating knowledge from heterogeneous sources while addressing statistical uncertainties. Existing methods rely on restrictive single-similarity assumptions (i.e., individual or combinatorial similarity) and often neglect practical variability, leading to suboptimal performance. To address these limitations, we propose a sufficiency-principled transfer learning framework that systematically balances model averaging and model selection during domain aggregation with unknown informative knowledge. The framework employs a sufficiency principle for quantifying transferable knowledge to eliminate the challenges of spurious correlation and perturbated evaluation. The proposed model averaging algorithms accommodate both individual and combinatorial similarity regimes, and also has privacy-preserving mechanisms. Theoretically, we establish the asymptotic optimality, estimator convergence and asymptotic normality, for multiple source domain linear regression models with diverging parameters. Especially, compared with existing results, we provide enhanced rate of converge for parameter of interest. Empirical validation through extensive simulations and an analysis of Beijing housing rental data demonstrates the statistical superiority of our framework over conventional domain aggregation methods. The proposed methodology extends beyond regression models, offering a generalizable paradigm for transfer learning in statistical decision theory.