In　order　to　reduce　the　semantic　gap　between　low-level　visual　features　and　high-level　semantics,　a　novel　approach　for　constructing　user　preference　profile　in　personalized　image　retrieval　is　proposed.　In　proposed　approach,　the　user　interest　is　divided　into　two　parts:　the　short-tern　interest　and　the　long-term　interest.　Semantic　feature　vector　in　the　short-term　interest　is　constructed　by　building　the　correlation　between　image　low-level　visual　features　and　high-level　semantics　on　the　basis　of　SVM　after　collecting　the　visual　feature　vector　in　the　short-term　interest　with　relevance　feedback.　Moreover,　the　visual　feature　vector　in　the　long-term　interest　can　be　collected　by　the　non-linear　gradual　forgetting　interest　inference　algorithm.　Semantic　feature　vector　in　the　long-term　is　constructed　with　clustering　algorithm.　Experiments　results　show　that　the　average　recall/precision　are　significantly　improved　and　satisfied　by　personalized　user　as　well.　©　2008　IEEE.