Filter Based Feature Selection on Azure Machine Learning Studio supports feature selection and ranking through Pearson Correlation, Kendall Correlation, Spearman Correlation, Mutual Information, Chi Squared, Fisher Score, and Count Based.
Please could you explain the scientific methodology as a walk-through of how the feature selection is performed using a correlation method such as Pearson, apparently without ever even attempting to train a predictive model to test the predictors?
I thought it might calculate the correlation between every possible pair of columns (i.e. features/predictors) but it calculates the values so incredibly fast to believe that it is building a full correlation distance matrix. Moreover, even if it did build a matrix, how would that be used to rank the predictive power of each feature without training any models?
Microsoft don't appear to say exactly how this feature selection method works except that it is 'well known' without giving any academic references.