Feature　engineering　determines　the　upper　limit　of　the　performance　of　machine　learning　algorithm.　And　feature　selection　is　the　most　critical　step　in　feature　engineering.　However,　the　dimensional　disasters　are　caused　by　high-dimensional　and　multi-granularity　feature　data,　which　makes　effective　feature　selection　very　difficult.　We　propose　a　feature　selection　based　on　the　Convolutional　Neural　Networks　and　Random　Forest　(FSCNNRF)　for　this　issue.　The　model　includes　two　parts,　Feature　Selection　Convolutional　Neural　Networks　(FSCNN)　and　Random　Forest　(RF).　It　can　select　more　effective　feature　set　by　using　FSCNN　for　dimensionality　reduction　and　RF　for　feature　selection.　Firstly,　the　high-dimensional　and　multi-granularity　feature　data　are　subjected　to　dimensionality　reduction　processing　by　FSCNN,　so　that　each　feature　becomes　a　single　granularity　feature.　Then　the　RF　is　used　to　select　valid　features.　Experiments　show　that　the　model　has　better　effect　on　feature　selection　on　high-dimensional　and　multi-granularity　dataset　and　improves　the　performance　of　machine　learning　algorithms.　©　Springer　Nature　Switzerland　AG,　2020.