The weight is the distance to the k-th nearest neighbour, which is a standard
non-parametric density estimate: large in sparse regions, small in dense ones.
Distances are computed on a bounded random reference subset rather than
between all pairs, which makes the cost linear in cells rather than quadratic.
Usage
density_weights(X, k = 20L, n_ref = 2000L, seed = 1L)
Arguments
- X
numeric matrix of features, one row per cell
- k
neighbour rank used for the density estimate
- n_ref
reference cells the distances are measured against
- seed
seed for the local RNG stream, which is restored on exit
Value
numeric vector of non-negative weights, one per row of X
