Get a knn object from reduced dimensions

Search for approximate nearest neighbors between cells in the reduced dimensions (e.g. PCA), and return the k nearest neighbors (knn) for each cell. Optionally, we can find neighbors between two separate sets of cells by utilizing both data and query.

Usage

knn_hnsw(
  data,
  query = NULL,
  k = 10,
  metric = c("euclidean", "cosine"),
  verbose = TRUE,
  threads = 1,
  ef = 100
)

knn_annoy(
  data,
  query = NULL,
  k = 10,
  metric = c("euclidean", "cosine", "manhattan", "hamming"),
  n_trees = 50,
  search_k = -1
)

Arguments

data: cell x dims matrix for reference dataset
query: cell x dims matrix for query dataset (optional)
k: number of neighbors to calculate
metric: distance metric to use
verbose: whether to print progress information during search
threads: Number of threads to use. Note that result is non-deterministic if threads > 1
ef: ef parameter for RcppHNSW::hnsw_search(). Increase for slower search but improved accuracy
n_trees: Number of trees during index build time. More trees gives higher accuracy
search_k: Number of nodes to inspect during the query, or -1 for default value. Higher number gives higher accuracy

Value

Named list of two matrices of dimension (cells x k):

idx: Neighbor indices, where idx[c, n] is the index of the nth nearest neighbor to cell c.
dist: Neighbor distances, where dist[c, n] is the distance between cell c and its nth nearest neighbor.

If no query is given, nearest neighbors are found by mapping the data matrix to itself, likely including self-neighbors (i.e. idx[c,1] == c for most cells).

Details

knn_hnsw: Use RcppHNSW as knn engine

knn_annoy: Use RcppAnnoy as knn engine