Work out correlation matrix threshold instantly with clear inputs, formula shown and shareable results.
A correlation matrix over p features contains p(p-1)/2 unique pairs. Pruning collinearity means dropping one member of each highly correlated pair, and because a cluster of m mutually correlated features produces m(m-1)/2 pairs, the number of features to drop is recovered by inverting that relationship. Removing collinear inputs stabilises linear coefficients and shrinks tree models without losing signal.
Pair count and pruning
pairs = p(p-1)/2; features implicated by k pairs ~ ceil((-1 + sqrt(1 + 8k)) / 2); drop all but one per cluster
It rarely hurts accuracy but it splits importance between duplicated features, which makes explanations misleading and feature selection unstable.
Keep the one that is cheaper to collect, more stable over time, or more interpretable. If both are equal, keep the one with higher correlation to the target.