Function kmedoids::alternating
source · [−]pub fn alternating<M, N, L>(
mat: &M,
med: &mut [usize],
maxiter: usize
) -> (L, Vec<usize>, usize) where
N: Zero + PartialOrd + Copy,
L: AddAssign + Signed + Zero + PartialOrd + Copy + From<N>,
M: ArrayAdapter<N>,
Expand description
Run the Alternating algorithm, a k-means-style alternate optimization.
This is fairly fast (O(n²), like the FasterPAM method), but because the newly chosen medoid must cover the entire existing cluster, it tends to get stuck in worse local optima as the alternatives. Hence, it is not really recommended to use this algorithm (also known as “Alternate” in classic facility location literature, and re-invented by Park and Jun 2009)
- type
M
- matrix data type such asndarray::Array2
orkmedoids::arrayadapter::LowerTriangle
- type
N
- number data type such asu32
orf64
- type
L
- number data type such asi64
orf64
for the loss (must be signed) mat
- a pairwise distance matrixmed
- the list of medoidsmaxiter
- the maximum number of iterations allowed
returns a tuple containing:
- the final loss
- the final cluster assignment
- the number of iterations needed
Panics
- panics when the dissimilarity matrix is not square
- panics when k is 0 or larger than N
Example
Given a dissimilarity matrix of size 4 x 4, use:
let data = ndarray::arr2(&[[0,1,2,3],[1,0,4,5],[2,4,0,6],[3,5,6,0]]);
let mut meds = kmedoids::random_initialization(4, 2, &mut rand::thread_rng());
let (loss, assi, n_iter): (f64, _, _) = kmedoids::alternating(&data, &mut meds, 100);
println!("Loss is: {}", loss);