Construct a sample-level consensus partition

Partitions with the same k are compared through sample co-assignment. Labels are aligned to the consensus partition only after fitting. This label alignment does not align SOM nodes and must not be used to force maps with different grids into a common node geometry.

Usage

consensus_som(
  partitions,
  k,
  linkage = "average",
  method = c("auto", "coassignment", "aligned_vote"),
  max_coassignment_n = 5000L
)

Arguments

partitions

A som_partitions object.

k

Number of clusters to audit.

linkage

Linkage used to partition the sample co-assignment matrix.

method

Consensus algorithm. "auto" uses full co-assignment only when every partition assigns every sample and the sample count does not exceed max_coassignment_n. It uses medoid-aligned voting for analysis-scoped resamples with incomplete assignments or larger data.

max_coassignment_n

Largest sample count for an automatically created dense co-assignment matrix. Set this deliberately because memory grows quadratically with sample count.

Value

A som_consensus object containing aligned assignments, partition-assignment, consensus-label and replicated coverage, membership support, assignment entropy and clusterwise Jaccard values. The object also records whether the consensus itself retains all requested clusters and retains the originating partition method as provenance. For co-assignment consensus, coassignment and tree contain the dense matrix and its hierarchical clustering; both are NULL for aligned voting. Aligned voting avoids quadratic growth in sample count, but alignment and upstream partition auditing can still grow quadratically in the number of ensemble members.

Examples

data <- simulate_som_scenario("clusters", n = 45, p = 3, seed = 3)
specification <- som_spec(c(3, 2), seeds = 1:2, rlen = 10, k = 2)
ensemble <- fit_som_ensemble(data, specification, keep_models = FALSE)
consensus_som(partition_som(ensemble), k = 2)
#> <som_consensus>
#>   k                  : 2 
#>   method             : coassignment 
#>   evidence scope     : analysis 
#>   partitions         : 2 
#>   samples            : 45 
#>   assigned at least once: 45 
#>   assignment coverage: 1.000 
#>   consensus coverage : 1.000 
#>   consensus clusters : 2 / 2 
#>   replicated coverage: 1.000 
#>   median support     : 1.000 
#>   median entropy     : -0.000