Traffic　congestion　is　a　problem　faced　by　cities,　and　it　is　urgent　for　solving　this　issue.　Accurate　short-term　traffic　state　prediction　is　benefit　for　citizens　to　know　the　traffic　information　in　advance,　and　take　the　measures　in　time　to　avoid　the　congestion.　In　this　paper,　a　short-term　traffic　state　prediction　approach　is　proposed　based　on　Fuzzy　C-Means　(FCM)　clustering　and　Random　Forest.　Firstly,　a　novel　Adaptive　Multi-kernel　Support　Vector　Machine　(AMSVM)　which　incorporates　the　spatial-temporal　information　is　used　to　predict　the　short-term　traffic　parameters,　including　the　volume,　the　speed　and　the　occupancy.　Secondly,　the　historical　traffic　data　are　analyzed　based　on　FCM　algorithm,　and　the　historical　traffic　state　information　is　got.　Lastly,　the　Random　Forest　(RF)　algorithm　is　utilized　to　analyze　the　predicted　short-term　traffic　parameters,　then　the　final　predicted　short-term　traffic　state　is　obtained.　This　method　incorporates　the　spatial-temporal　information　as　well　as　applying　the　Random　Forest　to　a　new　research　field　of　short-term　traffic　state　prediction.　The　experimental　results　demonstrate　that　the　evaluation　method　of　historical　traffic　state　based　on　FCM　is　suitable　for　both　freeway　and　urban　road　scenarios.　Besides,　the　Random　Forest　has　higher　prediction　accuracy　than　other　common　machine　learning　methods,　thus　providing　the　short-term　traffic　information　timely　and　reliably.　©　2018,　Science　Press.　All　right　reserved.