mrmr

Why we included this project

Most feature selection methods return every feature that has any relationship with the target, which leaves you sorting through a pile of columns. mRMR works the other way: it hunts for the smallest subset that still carries the predictive signal, ranking each feature by how much unique value it adds beyond what the others already provide. That minimal-optimal behavior is what makes it useful for teams that need to shrink wide datasets before training, since fewer columns means less memory, faster runs, and a model that is easier to explain. The same two functions, one for categorical targets and one for numeric ones, run on Pandas, Polars, Spark, and Google BigQuery, so the selection logic stays identical whether your data sits in a local DataFrame or a distributed warehouse. Uber engineers described using it in their marketing ML platform in a 2019 paper. For data scientists who rebuild models on messy, wide tables, it is a straightforward way to automate a step that is usually done by hand.