Web　page　is　an　important　human-computer　interface　and　the　classification　of　visual　behavior　on　web　page　has　drawn　widely　attention　recently.　People＇s　visual　behavior　can　be　reflected　by　recording　users＇　eye　movement　data　and　analyzing　eye　movement　features.　This　paper　studies　the　problem　of　web　page　visual　behavior　classification　based　on　the　deep　learning　approach.　Compared　with　most　existing　works　with　traditional　deep　neural　network　architecture,　where　either　supervised　learning　or　unsupervised　learning　is　adopted,　this　paper　propose　a　comprehensive　deep　neural　network　architecture,　considering　both　denoising　auto-encoders　based　unsupervised　learning　and　error　back-propagation　based　supervised　learning.　The　experimental　results　show　that　the　proposed　comprehensive　neural　network　architecture　achieves　better　performance　than　most　existing　classical　pure　supervised　learning　or　unsupervised　learning　architecture.