基于列表监督的Hash排序算法

A ranking hashing algorithm based on listwise supervision

Hash学习技术目前被广泛应用于大规模数据的相似性查找中,其通过将数据转化成二进制编码的形式,同时提高查找速度和降低存储代价。目前,大多数Hash排序算法通过比较数据在欧氏空间和海明空间的排序一致性来构造损失函数,然而,在海明空间的排序过程中,因为海明距离是离散的整数值,可能存在多个数据点共享相同的海明距离,这样就无法准确地排序。针对这一问题,将编码后的数据切分成几个长度相同的子空间,并为每个子空间设置不同的权重,比较时,再根据不同的子空间权重来计算海明距离。实验结果表明,与其他Hash学习算法相比,本文算法能够有效地对海明空间中的数据进行排序,并提高查询的准确性。

Recently,learning to hash technology has been used for the similarity search of large-scale data.It can simultaneous increase the search speed and reduce the storage cost through transforming the data into binary codes.At present,most ranking hashing algorithms compare the consistency of data in the Euclidean space and the Hamming space to construct the loss function.However,because the Hamming distance is a discrete integer value,there may be many data points sharing the same Hamming distance result in the exact ranking cannot be performed.To address this challenging issue,the encoded data was divided into several subspaces with the same length.Each subspace was set with different weights.The Hamming distance was calculated according to different subspace weights.The experimental results show that this algorithm can effectively sort the data in the Hamming space and improve the accuracy of the query compared with other learning to hash algorithms.