Feature Importance Calculator

The recommended feature count is generated entirely from fixed, model-specific heuristics — sample-to-feature ratios for linear models, square-root or log2 rules for tree ensembles, a looser ratio for neural networks — rather than being solved backward from your target variance figure. Compare your Target Variance against the reported Est. Variance Retained to see whether the heuristic recommendation happens to meet your target, rather than expecting the target itself to drive the recommendation.

Linear models have no built-in mechanism to ignore irrelevant features, so keeping too many relative to your sample size directly destabilizes the coefficient estimates, which is why the sample-to-feature ratio rule applies there. Tree ensembles instead select a random subset of features at each split, so the relevant heuristic is about how many features to consider per split for good diversity across trees, which is why it follows the square-root or log2 convention popularized by random forest research instead of a strict sample-ratio rule.

It means your current dataset size falls short of what's generally recommended to reliably support the number of features you have for your chosen model type, which raises the risk of an unstable or overfit model regardless of which specific features you keep. Options include collecting more data, more aggressively reducing your feature count, or choosing a model type with lower sample requirements per feature for your data volume.

It estimates what fraction of the dataset's total useful information the recommended feature subset captures, based on a diminishing-returns model where early features contribute more and later features contribute progressively less. It's a heuristic estimate rather than a measurement of your actual data's variance structure, so use it as a planning signal for how aggressive your feature reduction is, and verify against real cross-validated model performance before finalizing a feature set.