As　a　kind　of　behavioral　characteristics,　the　analysis　of　keystroke　behavior　and　the　selection　of　keystroke　features　are　crucial　operations　to　improve　the　accuracy　of　user　identification　using　shallow　machine　learning　algorithms.　In　this　paper,　we　discuss　the　typing　behavior　phenomena　and　put　forward　a　targeted　feature　optimization　strategy　in　order　to　meet　the　need　of　improving　the　accuracy　of　user　identification.　Three　types　of　keystroke　features　are　analyzed　including　duration　time　features　(Hold　Time)　and　two　types　of　latency　time　features　(UD　Time　and　DD　Time)　from　three　aspects　of　features　distribution,　features　correlation　and　features　contribution.　The　differential　evolution　(DE)　algorithm　is　used　to　optimize　keystroke　features　and　the　new　proposed　fitness　function　of　DE　algorithm　is　defined　based　on　the　analysis　of　keystroke　features.　Finally,　random　forest　(RF)　algorithm　is　devoted　to　evaluate　the　performance　of　feature　optimization.　Feature　analysis　results　show　that　latency　time　features　distribution　among　users　is　more　diverse　than　duration　time;　the　two　types　of　latency　time　features　have　a　strong　correlation　of　which　is　as　high　as　0.9766;　the　combination　of　duration　and　latency　time　features　have　the　best　classification　accuracy.　Final　experimental　results　of　an　open　dataset　show　that　DE　algorithm　based　on　features　correlation　analysis　selects　features　which　have　better　contribution　to　user　classification,　reduces　the　correlation　between　UD　Time　and　DD　Time　features,　and　improves　the　classification　accuracy　by　an　average　of　2.6206%.　©　2019　IEEE.