|
| 摘要: |
| 本文在非一致抽样分布下，研究与高斯核有关的在线分位数算法的收敛阶.本文引入阈值ε到Pinball损失函数产生算法的稀疏性，用Hölder对偶空间刻画抽样分布的非一致性，通过误差分解和迭代方法推导算法的收敛速度.并且以中位数回归为例，得到算法的具体收敛速度，同时也指明本文的背景和数学方法适用于一般分位数回归. |
| 关键词: 在线算法 分位数回归 ε-pinball损失函数 再生核Hilbert空间 非一致分布 |
| DOI： |
| 分类号:O29 |
| 基金项目:国家自然科学基金资助（11671307）. |
|
| QUANTILE REGRESSION WITH SAMPLES DRAWN FROM NON-IDENTICAL DISTRIBUTIONS |
|
YANG Peng-wei
|
| Abstract: |
| This paper considers the online quantile regression with Gaussain kernel and non-identical sampling process. Under the sparsity condition and non-identical marginal distributions, we derive the convergence rate of the the online quantile regression algorithm. Mathematical analysis depends on the error decomposition and the iteration method. Specially, we get the explicit learning rate of online median regression. Finally, we illustrate that our main result can extend to general quantile regression problem. |
| Key words: online learning quantile regression ε-pinball loss reproducing kernel Hilbert space non-identical distribution |
