Similarity Range and Approximate KNN Searches with iMinMax

Abstract

In high-dimensional databases, similarity search is computationally expensive. However, for many applications where small errors can be tolerated, determining approximate answers quickly has become an acceptable alternative. Intuitively, iMinMax can be readily used to support similarity range and nearest neighbor searching by adopting a filter-and-refine strategy: generate a range query that returns a candidate set containing all the desired nearest neighbors, and prune the candidate set to obtain the nearest neighbors. However, for KNN queries, it is almost impossible to determine the optimal range query for the candidate set, since the range query here is hyper square range query.