Hyperparameter Importance for Machine Learning Algorithms
This work addresses the efficiency of hyperparameter tuning for machine learning practitioners, but it is incremental as it builds on existing subsampling methods.
The paper tackles the problem of computationally expensive hyperparameter tuning for supervised machine learning algorithms by proposing a definition of hyperparameter importance that can be estimated via subsampling, enabling more efficient tuning on entire datasets. The results show theoretical consistency under weak conditions and numerical experiments confirm this, saving significant computational resources.
Hyperparameter plays an essential role in the fitting of supervised machine learning algorithms. However, it is computationally expensive to tune all the tunable hyperparameters simultaneously especially for large data sets. In this paper, we give a definition of hyperparameter importance that can be estimated by subsampling procedures. According to the importance, hyperparameters can then be tuned on the entire data set more efficiently. We show theoretically that the proposed importance on subsets of data is consistent with the one on the population data under weak conditions. Numerical experiments show that the proposed importance is consistent and can save a lot of computational resources.