Internet　advertising　has　become　a　major　source　of　income　for　Internet　companies,　among　which　the　prediction　of　ad　click-through　rate　is　the　most　important　task.　The　accuracy　of　ad　click-through　rate　can　directly　generate　revenue　for　the　company.　At　present,　the　mainstream　methods　such　as　Baidu　and　Google　are　linear　models　with　a　lot　of　artificial　features,　which　are　more　and　more　unsustainable.　Because　a　lot　of　manual　features　consume　a　lot　of　manpower,　their　benefits　are　declining.　Linear　models　cannot　learn　the　nonlinear　relationship　between　features.　In　this　paper,　we　propose　a　method　for　forecasting　click　rate　of　advertisements　based　on　deep　learning,　which　can　make　full　use　of　large-scale　sparse　data　and　learn　non-linear　features,　and　further　analyze　the　role　of　different　features　in　predicting　ad　click　through　rate.　The　experimental　results　on　the　KDD　Cup　2012　Track2　validate　that　the　proposed　method　can　improve　the　predictive　performance　of　search　ads,　with　an　AUC　value　of　0.771.　©　2018　ACM.