Neighborhood　and　sparsity　structure　preserving　projections　have　been　widely　used　in　dimensionality　reduction,　but　most　of　them　consider　single　structures.　Moreover,　existing　nonlinear　DR　methods　can　not　get　an　accurate　projection　function,　which　limits　their　applications.　To　overcome　these　problems,　we　propose　a　nonlinear　dimensionality　reduction　method　SNP-ELM　by　extending　the　extreme　learning　machine　model.　SNP-ELM　is　a　nonlinear　unsupervised　dimensionlity　reduction　method,　which　takes　both　sparsity　structure　and　neighborhood　structure　into　account.　The　experimental　results　on　toy　data,　wine　data　and　six　gene　expression　data　show　that　our　method　significantly　outperforms　the　compared　dimensionality　reduction　methods.　Copyright　©　2019　Acta　Automatica　Sinica.　All　rights　reserved.