Abstract: Feedback Locality Preserving Projections (FLPP) incorporates user’s feedbacks into LPP. By properly disposing user’s feedbacks, FLPP can update the eigenvectors which span the image representation subspace, so we can obtain a semantic subspace which can better reflect intrinsic property of image data. FLPP can use user’s feedbacks to optimize rapidly image representation, so gain capability of long-term learning. Experimental results show that FLPP can effectively improve retrieval accuracy, and after long-term learning, an approximate optimal subspace can be obtained.