RENT: A Python Package for Repeated Elastic Net Feature Selection
| Main Authors: | Jenul, Anna, Schrunner, Stefan, Huynh, Bao Ngoc, Tomic, Oliver |
|---|---|
| Format: | info software Journal |
| Bahasa: | eng |
| Terbitan: |
, 2021
|
| Subjects: | |
| Online Access: |
https://zenodo.org/record/5140349 |
Daftar Isi:
- Due to modern data acquisition techniques, the number of generated features in measurement data keeps increasing. This increase can make the analysis with standard machine learning methods difficult because of underdetermined systems where the dimensionality of the feature space (number of features) exceeds the dimensionality of the object space (number of observations). A concrete example of such a situation is data acquisition in the healthcare domain, where the number of patients (observations) suffering from a specific condition may be relatively low, but a lot of measurements (number of features) are generated for each patient to acquire a good understanding of the patient's health. A very common challenge is that not all features in a high dimensional space are equally important for predictive tasks — many might even be redundant. Feature selection deals with finding the most relevant features of a dataset. With help of appropriate methodology, feature selection can reduce (a) the complexity of and (b) noise in the dataset. More importantly, data interpretation of the model becomes easier with fewer features, which is of great importance within domains such as healthcare. Even though feature selection is a well-established research topic, relatively few approaches are focusing on the stability of the selection. The important question at hand is: can we trust that the selected features are really valid or is their selection very dependent on which observations are included in the data? Providing information on the stability of feature selection is vital, especially in wide data sets where the number of features can be many times higher than the number of observations. Here, the inclusion or exclusion of a few observations can have a high impact on which features may be selected.