Deep　learning　well　demonstrates　its　potential　in　learning　latent　feature　representations,　and　has　been　applied　in　the　fields　of　speech　recognition,　image　identification　and　information　retrieval.　Deep　learning　architecture　is　composed　of　multilayer　non-linear　units,　each　low　layer＇s　output　as　a　input　of　higher　layer,　and　can　learn　high-order　feature　representations　which　contain　many　structural　information　from　a　large　number　of　data.　Deep　learning　is　a　good　way　to　extract　features　from　original　data.　Web　page　is　an　important　human-machine　interface.　Identify　users＇　visual　behavior　on　Web　pages　will　promote　human-machine　interaction.　This　paper　apply　stacked　auto-encoders　(SAE)　with　logistic　regression　to　build　classification　model.　This　model　effectively　solves　the　problem　of　identifying　users＇　working　state　of　visual　search　and　visual　browse　on　Web　pages.　The　experiment　shows　that　this　model　outperforms　other　Single　model　such　as　SVM　and　logistic　regression　and　achieves　the　accuracy　of　90.32%.　Further,　we　embed　adaboost　algorithm　to　improve　recognition　accuracy　and　precision.　©　Copyright　owned　by　the　author(s)　under　the　terms　of　the　Creative　Commons　Attribution-NonCommercial-NoDerivatives　4.0　International　License　(CC　BY-NC-ND　4.0).