Approximate Nearest Neighbor Search on High Dimensional Data — Experiments, Analyses, and Improvement

Abstract

Approximate Nearest neighbor search (ANNS) is fundamental and essential operation in applications from many domains, such as databases, machine learning, multimedia, and computer vision. Although many algorithms have been continuously proposed in the literature in the above domains each year, there is no comprehensive evaluation and analysis of their performances.In this paper, we conduct a comprehensive experimental evaluation of many state-of-the-art methods for approximate nearest neighbor search. Our study (1) is cross-disciplinary (i.e., including 16 algorithms in different domains, and from practitioners) and ( 2) has evaluated a diverse range of settings, including 20 datasets, several evaluation metrics, and different query workloads. The experimental results are carefully reported and analyzed to understand the performance results. Furthermore, we propose a new method that achieves both high query efficiency and high recall empirically on majority of the datasets under a wide range of settings.

Orchestrating Single-Cell Analysis with Bioconductor

“…To address these challenges, Bioconductor has developed software packages that incorporate recent advance in nearest-neighbors and clustering algorithms that improve computational efficiency through approaches such as using approximate methods instead of exact methods, thereby trading an acceptable amount of accuracy for vastly improved runtimes. For example, the BiocNeighbors package [97][98][99][100] can be used to search for nearest neighbors and then a shared nearest neighbor graph using cells as nodes can be built using the scran [59]. Further, approximate methods have the advantage of smoothing over noise and sparsity, and thus potentially providing a better fit to the data [101].…”

“…BiocNeighbors [97][98][99][100] Exact and approximate methods for nearest neighbor detection that uses the BiocParallel [91] framework to parallelize operations SC3 [102], clusterExperiment [103], SIMLR [104], mbkmeans [106], BEARscc [107], clustree [108] Unsupervised clustering frameworks for single-cell data edgeR [3,62], DESeq2 [7], limma [115] Methods developed for bulk RNA-seq differential expression that can be used in combination with methods such as zinbwave [31,116] to account for the zero-inflation MAST [28], scDD [117], BASiCS [64,65], SCDE [118] Methods to identify differentially expressed features using statistical models that directly model zero-inflation slingshot [126], TSCAN [30], monocle [123,124,127], cellTree [128] Methods for trajectory analysis or pseudotime inference MAST [28], AUCell [141], scmap [77], PADOG [139], fgsea [137], goseq [138], slalom [142], scCoGAPS [143,144], EnrichmentBrowser [140] Methods for gene set / signature enrichment analysis iSEE [148] Interactive data exploration and visualization countsimQC [153], batchQC…”

Accurate and Fast Retrieval for Complex Non-metric Data via Neighborhood Graphs

“…An approximate neighborhood graph can be constructed substantially more efficiently [22,11]. To improve performance, one can use various graph pruning methods [20,23,13]: In particular, it is not useful to keep neighbors that are close to each other [20,13].…”

“…For a recent experimental comparison of several retrieval approaches see [32]. Although, HNSW is possibly the best retrieval method for generic distances [23,20], in our work we use a modified variant of SW-graph, where retrieval starts from a single point (which is considerably more efficient compared to multiple starting points). The main advantage of HNSW over the older version of SW-graph is due to (1) introduction of pruning heuristics, (2) using a single starting point during retrieval.…”

The Role of Local Intrinsic Dimensionality in Benchmarking Nearest Neighbor Search

“…We use the benchmarking system described in [4] as the starting point for our study. Different approaches to benchmarking nearest neighbor search are described in [9,10,20]. We refer to [4] for a detailed comparison between the frameworks.…”