This　paper　mainly　employed　the　random　forest　(RF),　gradient　boosting　(GB)　and　extreme　gradient　boosting　(XGB)　to　predict　the　compressive　strength　of　alkali-activated　construction　demolition　waste　geopolymers　(CDWG).　The　performances　of　three　ensemble　machine　learning　(ML)　models　were　evaluated　and　the　effects　of　eight　different　input　features　on　the　compressive　strength　of　CDWG　were　deeply　analyzed.　The　results　confirm　the　applicability　of　RF,　GB　and　XGB　algorithms　in　aspect　of　strength　prediction　for　the　CDWG　with　the　high　predictive　accuracy　(R2　＞　0.9).　Among　them,　the　performances　of　GB　and　XGB　models　are　better　than　RF　model.　The　liquid　to　solid　ratio　(L/S)　has　a　negative　correlation　with　the　compressive　strength　of　CDWG,　while　the　pretreatment　temperature,　heat　treatment　time　and　curing　age　have　a　positive　correlation　with　the　compressive　strength　of　CDWG.　The　obvious　enhancement　of　compressive　strength　of　CDWG　mainly　occurs　in　the　early　age.　Decreasing　L/S　and　raising　pretreatment　temperature　have　a　significant　positive　gain　on　the　compressive　strength　of　CDWG.　In　the　preparation　process　of　CDWG,　it　is　suggested　that　the　L/S　and　%Na2O　of　Na2SiO3-based　alkaline　activators　are　controlled　at　about　0.3　and　7　%　respectively,　and　appropriately　increasing　pretreatment　temperature　and　prolonging　heat　treatment　time.