期刊名称：Electronic Journal of Applied Statistical Analysis
电子版ISSN：2070-5948
出版年度：2019
卷号：12
期号：1
页码：69-84
DOI：10.1285/i20705948v12n1p69
出版社：University of Salento
摘要：Lasso regression methods are widely used for a number of scientific applications.
Many practitioners of statistics were not aware that a small change
in the data would result in an unstable Lasso solution path. For instance,
in the presence of outlying observations, Lasso perhaps leads to an increase
in the percentage of the false selection rate of predictors. The discussions
on determining an optimal shrinkage parameter of Lasso are still ongoing.
Therefore, this paper proposed a robust algorithm to tackle the instability
of Lasso in the presence of outliers. The new weight function is proposed to
overcome the problem of outlying observations. The weighted observations
are for a certain number of subsamples to control the false Lasso selection.
The simulation study has been carried out and uses real data to assess the
performance of our proposed algorithm. Consequently, the proposed method
shows more efficiency than LAD-Lasso and weighted LAD-Lasso and more
reliable results.
其他摘要：Lasso regression methods are widely used for a number of scientic applications.Many practitioners of statistics were not aware that a small changein the data would results in unstable Lasso solution path. For instance, inthe presence of outlying observations, Lasso perhaps leads the increase inthe percentage of the false selection rate of predictors. On the other hand,the discussions on determining an optimal shrinkage parameter of Lasso isstill ongoing. Therefore, this paper proposed a robust algorithm to tacklethe instability of Lasso in the presence of outliers. A new weight function isproposed to overcome the problem of outlying observations. The weightedobservations are subsamples for a certain number of subsamples to controlthe false Lasso selection. The simulation study has been carried out and usesreal data to assess the performance of our proposed algorithm. Consequently,the proposed method shows more eciency than LAD-Lasso and weightedLAD-Lasso and more reliable results.
