| Summary: | There are two main approaches to tackle the challenge of finding the best filter or embedded feature selection (FS) algorithm: searching for the one best FS algorithm and creating an ensemble of all available FS algorithms. However, in practice, these two processes usually occur as part of a larger machine learning pipeline and not separately. We posit that, due to the influence of the filter FS on the embedded FS, one should aim to optimize both of them as a single FS pipeline rather than separately. We propose a meta-learning approach that automatically finds the best filter and embedded FS pipeline for a given dataset called FSPL. We demonstrate the performance of FSPL on n = 90 datasets, obtaining 0.496 accuracy for the optimal FS pipeline, revealing an improvement of up to 5.98 percent in the model’s accuracy compared to the second-best meta-learning method.
|
|---|