Description K-means with k-nearest neighbors (KMKNN) Vantage point (VP) trees Exhaustive search Approximate nearest neighbors Oh Yeah (Annoy) Hierarchical navigable small worlds (HNSW) Distance metrics Author(s) References
This page provides an overview of the neighbor search algorithms available in BiocNeighbors.
In the KMKNN algorithm (Wang, 2012), k-means clustering is first applied to the data points using the square root of the number of points as the number of cluster centers. The cluster assignment and distance to the assigned cluster center for each point represent the KMKNN indexing information. This speeds up the nearest neighbor search by exploiting the triangle inequality between cluster centers, the query point and each point in the cluster to narrow the search space. The advantage of the KMKNN approach is its simplicity and minimal overhead, resulting in performance improvements over conventional tree-based methods for high-dimensional data where most points need to be searched anyway. It is also trivially extended to find all neighbors within a threshold distance from a query point.
In a VP tree (Yianilos, 1993), each node contains a subset of points that is split into two further partitions. The split is determined by picking an arbitrary point inside that subset as the node center, computing the distance to all other points from the center, and taking the median as the “radius”. The left child of this node contains all points within the median distance from the radius, while the right child contains the remaining points. This is applied recursively until all points resolve to individual nodes. The nearest neighbor search traverses the tree and exploits the triangle inequality between query points, node centers and thresholds to narrow the search space. VP trees are often faster than more conventional KD-trees or ball trees as the former uses the points themselves as the nodes of the tree, avoiding the need to create many intermediate nodes and reducing the total number of distance calculations. Like KMKNN, it is also trivially extended to find all neighbors within a threshold distance from a query point.
The exhaustive search computes all pairwise distances between data and query points to identify nearest neighbors of the latter. It has quadratic complexity and is theoretically the worst-performing method; however, it has effectively no overhead from constructing or querying indexing structures, making it faster for in situations where indexing provides little benefit. This includes queries against datasets with few data points or very high dimensionality.
The Annoy algorithm was developed by Erik Bernhardsson to identify approximate k-nearest neighbors in high-dimensional data. Briefly, a tree is constructed where a random hyperplane splits the points into two subsets at each internal node. Leaf nodes are defined when the number of points in a subset falls below a threshold (close to twice the number of dimensions for the settings used here). Multiple trees are constructed in this manner, each of which is different due to the random choice of hyperplanes. For a given query point, each tree is searched to identify the subset of all points in the same leaf node as the query point. The union of these subsets across all trees is exhaustively searched to identify the actual nearest neighbors to the query.
In the HNSW algorithm (Malkov and Yashunin, 2016), each point is a node in a “nagivable small world” graph. The nearest neighbor search proceeds by starting at a node and walking through the graph to obtain closer neighbors to a given query point. Nagivable small world graphs are used to maintain connectivity across the data set by creating links between distant points. This speeds up the search by ensuring that the algorithm does not need to take many small steps to move from one cluster to another. The HNSW algorithm extends this idea by using a hierarchy of such graphs containing links of different lengths, which avoids wasting time on small steps in the early stages of the search where the current node position is far from the query.
All algorithms support neighbor searching by Euclidean, Manhattan and cosine distances. Cosine distances are implemented as the Euclidean distance between L2-normalized vectors. Note that KMKNN operates much more naturally with Euclidean distances, so your mileage may vary when using it with Manhattan distances.
Aaron Lun, using code from the cydar package for the KMKNN implementation; from Steve Hanov, for the VP tree implementation; RcppAnnoy, for the Annoy implementation; and RcppHNSW, for the HNSW implementation.
Wang X (2012). A fast exact k-nearest neighbors algorithm for high dimensional search using k-means clustering and triangle inequality. Proc Int Jt Conf Neural Netw, 43, 6:2351-2358.
Hanov S (2011). VP trees: A data structure for finding stuff fast. http://stevehanov.ca/blog/index.php?id=130
Yianilos PN (1993). Data structures and algorithms for nearest neighbor search in general metric spaces. Proceedings of the Fourth Annual ACM-SIAM Symposium on Discrete Algorithms, 311-321.
Bernhardsson E (2018). Annoy. https://github.com/spotify/annoy
Malkov YA, Yashunin DA (2016). Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs. arXiv. https://arxiv.org/abs/1603.09320
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.