一种基于混合索引的最近邻查找方法

A Nearest Neighbor Search Method Based on Hybrid Index

针对某些场景下可学习KD树模型在最近邻查找中准确率较低的问题,提出了一种基于可学习索引模型和传统KD树的混合索引结构。该结构将待查找数据同时输入已经训练好的可学习KD树模型和KD树中得到若干个候选的k近邻点,从而将可学习索引模型在查找效率和传统索引方法在查找准确率上的优点相结合。试验结果证明,使用基于可学习索引模型的可学习KD树和树形结构KD树的混合索引,综合了两者在最近邻查找中的优点,实现了查找效率和查找精度的平衡,满足了多种条件下的查找需求。

Aiming at the problem that the learned KD tree model has low accuracy in nearest neighbor lookup in some scenarios,a hybrid index structure based on learned index model and traditional KD tree is proposed.The structure simultaneously inputs the data to be found into the trained learned KD tree model and several candidate k neighbors in the KD tree,so as to combine the advantages of the learned index model in the search efficiency and the traditional index method in the search accuracy.The experimental results show that the hybrid index of learned KD tree and tree structure KD tree based on learned index model combines the advantages of the two in nearest neighbor lookup,realizes the balance of search efficiency and search accuracy,and meets the search needs under various conditions.