Scalable Graph Hashing with Feature Transformation Model and Leamning Feature Transformation 。Here,weuse an approximation2z+是≈e2 =x+ 0.5 -0.5 0.5 We assume-l≤axx≤l.It is easy to prove that p=2max{lxl3}是1 can make-1≤3xx;≤1. Then we have S P(X)TQ(X) 日24元，互)Q0 Li (http://cs.nju.edu.cn/lvj) Learning to Hash LAMDA,CS.NJU 25/43Scalable Graph Hashing with Feature Transformation Model and Learning Feature Transformation Here, we use an approximation e 2−1 2e x + e 2+1 2e ≈ e x We assume −1 ≤ 2 ρ x T i xj ≤ 1. It is easy to prove that ρ = 2 max{kxik 2 F } n i=1 can make −1 ≤ 2 ρ x T i xj ≤ 1. Then we have Se ≈ P(X) TQ(X) Li (http://cs.nju.edu.cn/lwj) Learning to Hash LAMDA, CS, NJU 25 / 43