Personalized　recommendation　of　tourist　attractions　for　visitors　is　useful　based　on　the　vast　amount　of　tourism　information　and　data.　In　this　paper,　we　use　Flickr＇s　geotagged　photos　from　2013　to　2018　in　Hong　Kong　to　identify　tourist　hot　spots　and　reconstruct　the　tourism　trajectory　according　to　the　tourist　visiting　order.　On　this　basis,　we　propose　a　personalized　recommendation　method　based　on　Latent　Dirichlet　Allocation　(LDA)　model　and　User＇s　Long-term　and　Short-term　Preference　(L-ULSP)　to　address　the　problem　that　existing　methods　do　not　take　into　account　the　dynamic　changes　in　visitor　preferences　during　the　travel　process.　In　this　method,　the　LDA　model　is　used　to　obtain　the　feature　information　of　attractions,　and　the　correlation　between　attractions　is　explored.　Then,　attention　mechanism　is　used　to　focus　on　the　important　information　in　the　long-term　sequence　to　capture　the　long-term　preference　of　tourists,　and　LSTM　is　used　to　model　the　short-term　sequence　information　to　learn　the　short-term　preference　of　tourists.　Finally,　the　long-term　and　short-term　preferences　are　weighted　to　obtain　the　final　preferences　of　tourists　to　capture　the　dynamic　changes　of　user　preferences.　The　algorithm　has　the　following　advantages:　(1)　By　mining　the　topic　feature　information　of　Geotagged　photo　text,　the　description　information　of　attractions　is　added,　which　can　capture　users＇　travel　preference　more　accurately;　(2)　The　algorithm　considers　both　the　long-term　and　short-term　preferences　of　users,　and　can　learn　the　dynamic　changes　of　users＇　preferences　in　the　process　of　travel　while　modeling　the　sequence　information　of　attractions.　The　experimental　results　show　that:　(1)　The　attractions　recommended　by　the　L-ULSP　method　outperform　other　existing　methods　in　both　Hit　Rate　and　Mean　Reciprocal　Rank,　two　common　evaluation　metrics　for　recommendation　algorithms,　proving　that　the　proposed　method　can　effectively　learn　visitor　preferences　from　a　sequence　of　attractions　and　recommend　the　next　attraction　to　visitors.　It　is　demonstrated　that　the　method　can　achieve　good　recommendation　results　in　travel　recommendation　scenarios;　(2)　The　comparison　experiments　between　the　model　using　long-term　preference　as　the　user＇s　final　preference　and　the　model　combining　user＇s　long-　term　and　short-　term　preferences　as　the　final　preference　further　validate　that　considering　both　the　user＇s　long-term　and　short-term　preferences　can　better　learn　the　user＇s　preference　changes　and　thus　improve　the　accuracy　of　recommendations;　(3)　This　paper　further　compares　the　calculation　efficiency　of　L-ULSP　with　different　deep　learning　recommendation　models　based　on　RNN,　and　counts　the　running　time　of　each　model.　The　results　show　that　this　method　is　better　than　most　methods　in　efficiency.　©　2021,　Science　Press.　All　right　reserved.