基于深度学习的可学习索引研究

A Study of Learned Index Based on Deep Learning

如何对数据进行高效的检索一直是个热门话题。传统的索引方法在大数据环境下进行最近邻查找时,面临着查找速度慢、准确率不高等问题。为了保证检索效率,人们往往会牺牲一定的准确度来换取更高的查询效率。随着机器学习和神经网络的发展,采用基于深度学习的可学习索引模型,将检索过程使用神经网络的查找进行代替成为一种可行的方法。实验结果表明,在解决最近邻查找问题时,使用包含输入层、神经网络层、索引层和输出层等四个层次的深度学习模型,能够在保持一定查找准确率的基础上,在查找时间上取得优势。

How to efficiently retrieve data has always been a hot topic.When traditional indexing methods solve the nearest neighbor search problem in a big data environment,they face problems such as slow search speed and low accuracy.In order to ensure retrieval efficiency,people often sacrifice certain accuracy to exchange higher query efficiency.With the development of machine learning and neural networks,it has become a feasible method to use a learned index model based on deep learning to replace neural network search in the retrieval process.The experimental results show that in solving the nearest neighbor search problem,using a four-level deep learning model including input layer,neural network layer,index layer and output layer can gain an advantage in search time while maintaining a certain search accuracy.