Add filter feature selection using Pearson correlation

[mandjevant]

In short:

Adding a filter feature selection method for applications that require efficient computation over the detection of complex relationships.

Implementation

1. Transpose the data: self.dataX contains the data as a matrix. This data must be transposed to calculated Pearson correlation.
2. Calculate Pearson correlation: Since scipy is already a requirement for this library, we can simply use scipy.stats.pearsonr. This will also return the p-value.
3. Remove the features where the absolute correlation and p-value do not obey the set minimum and maximum.
4. Return the indices of the selected features and the names of the selected features.

Function arguments:

Args:
    min_corr: Minimum correlation value for feature to be selected. Standard: 0.2
    max_corr: Maximum correlation value for feature to be selected. Standard: 1.0
    max_pvalue: Maximum p-value to determine statistical significance. Standard: 0.05

Results

Wrapper method:

• The following features were selected: ['RM', 'TAX', 'LSTAT', 'PTRATIO', 'DIS', 'AGE']
• The estimated error of the developed model is: 2.7131430936707424
• Method took 73.3180787563324 seconds to complete.

fst-pso method:

• The following features have been selected: ['CRIM', 'ZN', 'NOX', 'RM', 'DIS', 'RAD', 'TAX', 'PTRATIO', 'LSTAT'] with a MAE of 2.79
• The estimated error of the developed model is: 2.907917419119525
• Method took 3338.643133163452 seconds to complete.

Filter method:

• The following features were selected: ['CRIM', 'ZN', 'INDUS', 'NOX', 'RM', 'AGE', 'DIS', 'RAD', 'TAX', 'PTRATIO', 'LSTAT']
• The estimated error of the developed model is: 2.6634497163494317
• Method took 2.928159236907959 seconds to complete.