In location-based services, data objects often possess dual attributes, featuring both structured geographic attributes and vector attributes derived from unstructured sources (i.e., images and text). This paper studies the problem of top-k spatial-Range-constrained Approximate Nearest Neighbor Search (k-RANNS) for such data. Specifically, given a query vector q and a geospatial range R, the k-RANNS query identifies top-k objects whose spatial coordinates are located within R and whose vectors are most similar to v. This query is quite common in location-based services. For example, food delivery platforms often recommend restaurants within a user’s city that align with their previous orders. Albeit existing solutions can be applied to k-RANNS queries, they face challenges in memory efficiency and performance stability across various selective queries. This is due to their lack of effective mechanisms to manage the index memory overhead and their disregard for the impact of query selectivity on query performance. To address these challenges, we introduce Mesh, a Memory-Efficient index for the Spatial-range-constrained High-dimensional k-ANNS query. Mesh is a workload-aware index designed to optimize query efficiency for a given workload with various selective queries while ensuring memory overhead below a predefined constraint. Its construction poses a combinatorial optimization problem, proven to be NP-hard. We propose a theoretically guaranteed approximation algorithm as a practical solution for the combinatorial optimization problem. We also develop an efficient query algorithm that can adaptively decide the execution strategy for each query with specific selectivity, and present several techniques to further enhance query efficiency. Extensive experiments show that Mesh outperforms competitors by up to orders of magnitude while using less memory, and exhibits stable performance across various selective queries.
