Really simple 'trick' for feature selection: before producing a feature importance analysis, temporarily inject a number of random noise features into ones dataset. Any feature that ranks BELOW the most important of these noise features, by extension, is also essentially noise and can be dropped. Indeed in small datasets one can be shockingly surprised how spuriously important these noise features can be! This technique is adapted from the 2003 paper "Ranking a Random Feature for Variable and Feature Selection" (https://lnkd.in/dNRC8s44) ---------------------------------------------------------- For more on classical ML get my book "The Orange Book of Machine Learning - Green edition" via https://lnkd.in/dZVnK67t #datascience #machinelearning | 45 comments on LinkedIn