A Study Of Learned Hashing Index Based On Deep Learning

Nowadays,we are in an age of high lying information.The data we can create and store is increasing exponentially.It is challenging to retrieve large amounts of high-dimensional data quickly and effectively,and traditional data structures can not meet the new requirements must be optimized and improved.Recent research shows that neural networks can partially replace or even wholly replace some traditional data structures,which provides a new way of building a unique hash index,that is,using neural network to build hash indexes.Based on the previous work,a learned Local-sensitive hashing algorithm--LLSH algorithm based on deep learning is proposed according to the learned index architecture.The traditional LSH(Locality-sensitive Hashing)functions are replaced by neural networks while sacrificing some precision.Compared with the previous work,this paper proposes two solutions for the LLSH algorithm: the LLSH algorithm based on Euclidean Local-sensitive hashing and the LLSH algorithm based on ensemble learning.The main contributions of this article are in three areas:(1)The LLSH algorithm based on Euclidean Local-sensitive hashing is designed in detail,and extensive experiments are carried out to evaluate the performance of the LLSH algorithm.Using different parameters and configurations,proving the efficiency and flexibility of the LLSH algorithm.(2)This paper explores the LLSH algorithm based on ensemble learning.The algorithm integrates several unique nearest neighbor algorithms by using blended learning methods and finally uses neural networks to fit the final solution space of the integrated algorithm to achieve precision and performance improvement.(3)The traditional E2LSH(Exact Euclidean Locality-Sensitive Hashing)algorithm is reproduced in this paper and uses matrix operations to speed up the algorithm.It is compared with the LLSH algorithm based on Euclidean Local-sensitive hashing on running time and memory footprint.The paper also re-emerges several commonly used nearest neighbor algorithms.It compares them with the LLSH algorithm based on ensemble learning in various aspects,highlighting the advantages of the LLSH algorithm.The experimental results show that the LLSH algorithm based on EuclideanLocality-sensitive hashing can be more efficient in mapping,ensuring absolute accuracy.The LLSH algorithm based on ensemble learning not only has lower time and space complexity but also has a significant improvement in accuracy compared with the traditional nearest neighbor algorithm.