With　the　rapid　development　of　Internet　technology　and　the　continuous　explosive　growth　of　network　traffic,　Traffic　Engineering　(TE),　as　a　viable　method　for　optimizing　network　traffic　distribution　and　improving　network　performance,　attracts　widespread　attention　from　both　industry　and　academia.　Software　Defined　Networks　(SDN),　which　decouples　the　data　plane　and　the　control　plane,　realizes　a　flexible　routing　and　improves　the　TE　performance.　Existing　TE　approaches　in　SDN　mainly　utilize　Reinforcement　Learning　(RL)　methods　to　learn　the　mapping　relationship　between　network　traffic　and　routing　policies.　However,　due　to　the　continuous　expansion　of　network　size　and　dynamic　changes　in　traffic,　the　enlargement　of　traffic　state　space　hinders　RL　from　converging　to　the　optimal　routing　policy,　leading　to　a　decline　in　network　performance.　To　address　these　issues,　this　paper　presents　a　TE　method　based　on　unsupervised　contrastive　representation　and　RL.　This　method　first　shrinks　the　original　traffic　state　space　by　efficiently　extracting　traffic　features　through　Contrastive　Learning　(CL),　aiding　quick　convergence　of　RL.　It　then　uses　RL　to　directly　learn　the　mapping　from　traffic　features　to　traffic　splitting　policies.　Finally,　through　numerous　experiments　on　real　network　traffic　and　topology,　it　demonstrates　that　the　proposed　TE　method　can　effectively　achieve　load　balancing　of　network　traffic　under　complex　and　volatile　dynamic　traffic　demands,　thereby　enhancing　network　performance.　©　2024　IEEE.