biocore/hclust-benchmarks

A benchmarking repository for hierarchical clustering methods

hclust-benchmarks's Issues

Tools to benchmark

• Scipy linkage
• Hdbscan
• Scikit learn agglomerative clustering

For each function, we'll want the following wrapper

def function(otu_table : pd.DataFrame ) -> skbio.TreeNode

Where each of the rows corresponds to an OTU, and the columns correspond to sample.

Here is an example of how we'd expect to generate a tree with mean pH

from skbio import TreeNode, DistanceMatrix
def linkage(mean_ph):
    dm = DistanceMatrix.from_iterable(mean_ph.values, euclidean)
    lm = average(dm.condensed_form()) 
    ph_tree = TreeNode.from_linkage_matrix(lm, otu_table.index)
    return ph_tree

Simulation benchmarks

Possible Simulations

• Uniform distribution along line (i.e. pH gradient). Can determine if the sorted tree has the same ordering of pH as the original distribution.
• Number of clusters (variance, mean separation). Can determine if the top level clusters match the original cluster

For each simulation, we will want to investigate

1. Running time
2. Memory usage
3. Parallelizability