Maximum Variance Hashing via Column Generation

Algorithm 2

MVH-A: Maximum variance hashing with anchors.

Input: Training data 𝒳={x_𝑖}^𝑛_𝑖=1; balancing parameter  𝛾>0; length of hash codes  ℓ;

   number of anchors 𝑐≪𝑛.

(1) Generate the anchor set 𝒜 by a clustering method (e.g. K-means) from 𝒳;

(2) Find ℋ via running Algorithm 1 on 𝒜 with parameters 𝛾 and ℓ.

(3) Map 𝒳 to 𝒴={𝐲_𝑖}^𝑛_𝑖=1 using ℋ.

Output: The learnt hash functions ℋ and the binary codes 𝒴.