Table 6 Feature selection methods along with their description, category, advantages, and limitations

Filter Methods [71, 72]

Evaluate feature relevance independently using statistical measures or scoring functions

Unsupervised

- Computationally efficient

- No reliance on the learning algorithm

- Can handle high

-dimensional data

- Ignores feature dependencies

- May not consider interactions with the target variable

Wrapper Methods [73]

Evaluate feature subsets by training and testing the model on different combinations of features

Supervised

- Considers feature interactions

- Can optimize model performance

- Provides feature importance ranking

- Computationally expensive for large feature sets

- Prone to overfitting if the evaluation criterion is not carefully chosen

Embedded Methods [74]

Incorporate feature selection within the model training process by considering feature importance or regularization

Supervised

- Simultaneously performs feature selection and model training

- Automatically assigns feature weights

- Considers feature interactions

- Limited to specific learning algorithms that support embedded feature selection

- May not handle highly correlated features effectively

Principal Component Analysis (PCA) [75]

Dimensionality reduction technique that transforms original features into a lower-dimensional space

Unsupervised

- Reduces dimensionality while preserving important information

- Removes multicollinearity among features

- Can be used for data visualization

- May lose some interpretability as transformed features are linear combinations of the original features

- Assumes linear relationship between features

Genetic Algorithms

[76]

Optimization algorithms inspired by natural selection that search for an optimal feature subset

Unsupervised

- Handles large feature spaces

- Considers feature interactions

- Can handle non-linear relationships

- Computationally expensive

- Requires careful parameter tuning

- Not guaranteed to find the global optimum