Feature selection methods don't change the initial values of the variables or features; instead, they remove particular features from the source dataset that aren't relevant. Some of the feature selection methods we can use are as follows: 

• Missing value ratio: This method is based on the idea that a feature that misses many values should be eliminated from a dataset because it doesn't contain valuable information and can distort the model's performance results. So, if we have some criteria for identifying missing values, we can calculate their ratio to typical values and set a threshold that we can use to eliminate features with a high missing value ratio.
• Low variance filter: This method is used to remove features with low variance because such features don't contain enough information to improve model performance. To apply this method, we need to calculate the variance for each feature, sort them in ascending order by this value, and leave only those with the highest variance values.
• High correlation filter: This method is based on the idea that if two features have a high correlation, then they carry similar information. Also, highly correlated features can significantly reduce the performance of some machine learning models, such as linear and logistic regression. Therefore, the primary goal of this method is to leave only the features that have a high correlation with target values and don't have much correlation between each other.
• Random forest: This method can be used for feature selection effectively (although it wasn't initially designed for this kind of task). After we've built the forest, we can estimate what features are most important by estimating the impurity factor in the tree's nodes. This factor shows the measure of split distinctness in the tree's nodes, and it demonstrates how well the current feature (a random tree only uses one feature in a node to split input data) splits data into two distinct buckets. Then, this estimation can be averaged across all the trees in the forest. Features that split data better than others can be selected as the most important ones.
• Backward feature elimination and forward feature selection: These are iterative methods that are used for feature selection. In backward feature elimination, after we've trained the model with a full feature set and estimated its performance, we remove its features one by one and train the model with a reduced feature set. Then, we compare the model's performances and decide how much performance is improved by removing feature changes – in other words, we're deciding how important each feature is. In forward feature selection, the training process goes in the opposite direction. We start with one feature and then add more of them. These methods are very computationally expensive and can only be used on small datasets.