Skip to contents

Comparison with Existing Consensus Clustering Methods

Source: vignettes/comparison-existing-methods.Rmd
comparison-existing-methods.Rmd

Background

Consensus clustering is not one method but a family of related strategies. Some packages focus on representing and combining partitions, some focus on repeated resampling across algorithms, and some focus on calibrating the existence of structure against a null baseline. uccdf was built because, in practice, analysts often need all of those pieces at once on a mixed-type dataframe.

Objective

The goal of this article is to position uccdf relative to several widely used consensus clustering toolkits:

  • clue
  • diceR
  • M3C
  • sharp

The comparison is intentionally practical. The question is not which package is “best” in the abstract, but which problem each package is actually designed to solve.

High-level comparison

The most important distinction is the unit of design.

  • clue is a general cluster ensemble framework. It provides the data structures, agreement measures, and consensus operations needed to work with sets of partitions and hierarchies.
  • diceR is an end-to-end consensus clustering workflow for generating diverse clustering solutions across algorithms, resamples, and cluster sizes.
  • M3C is aimed at null-calibrated consensus clustering, especially in genomics-style matrix settings where overestimation of K is a known risk.
  • sharp emphasizes automated calibration and stability-based consensus clustering under a weighted distance framework.
  • uccdf is designed as a typed dataframe workflow that combines schema inference, multiple representations, ensemble clustering, and null-calibrated selection in one package.

Comparison table

Package Main design target Input style Mixed-type dataframe awareness Null calibration Typical strength
clue Cluster ensemble data structures and consensus operations Partitions, hierarchies, ensemble objects No built-in typed schema layer No built-in global null test Extremely flexible ensemble tooling
diceR Diverse ensemble generation across algorithms and resamples Mostly matrix-like tabular input Limited, depends on user preprocessing Not the main focus Strong multi-algorithm consensus workflow
M3C Monte Carlo reference-based consensus clustering Continuous expression-like matrices Not built for general mixed-type tables Yes, central feature Guards against false positive cluster discovery
sharp Stability and calibration of weighted consensus clustering Distance-based or weighted clustering workflows Not a general dataframe typing system Yes, through calibrated stability procedures Hyper-parameter and weighting calibration
uccdf Typed consensus clustering for structured dataframes data.frame with mixed columns Yes, explicit schema and role inference Yes, two-stage detection and K selection Unified workflow for mixed-type tables

Where the methods overlap

There is real overlap between these tools, especially once a user manually preprocesses a table into a numeric matrix.

uccdf overlaps with:

  • clue on the idea of aggregating multiple clustering results into a consensus object
  • diceR on the idea of algorithm diversity, repeated resampling, and consensus-by-K
  • M3C on the idea that a consensus score alone is not enough and should be compared with a null baseline
  • sharp on the idea that stability and calibration matter as much as the raw clustering output

So the difference is not that uccdf invents consensus clustering from scratch. The difference is that it packages several of these ideas into a dataframe-first workflow.

Where uccdf is intentionally different

uccdf makes four choices that separate it from the existing packages above.

1. Typed schema is part of the workflow

uccdf starts from a structured dataframe and explicitly infers:

  • which columns are active
  • which are identifiers
  • which should be excluded
  • which are continuous, binary, nominal, or ordinal

That step is not the main concern of clue, diceR, M3C, or sharp, but it is a recurring practical problem in mixed-table analysis.

2. Multiple views are first-class

Instead of committing to one representation, uccdf combines:

  • a mixed-distance view
  • a mixed-latent view

This matters because mixed tables often admit more than one defensible geometry. Some existing methods can be made to use multiple representations, but that is usually the analyst’s responsibility rather than the package’s core contract.

3. Detection and selection are separated

uccdf distinguishes between:

  1. is there evidence for structure at all?
  2. if yes, which K is best supported?

This is closest in spirit to M3C, which also treats null calibration as a central concern. The main practical advantage is that uccdf can report a conservative K = 1 while still exposing an exploratory split for inspection.

4. Output is designed for downstream tabular analysis

The main row-level output of uccdf is not just a partition. It is a dataframe augmentation containing:

  • selected cluster
  • confidence
  • ambiguity
  • exploratory cluster when the global null is not rejected

This makes the package easier to drop into reporting pipelines built around tables rather than only around partition objects.

When to use each package

Use clue when

  • you already have several partitions or hierarchies
  • you want a mature ensemble object system
  • you care about partition comparison, matching, and consensus operations more than about preprocessing or null calibration

Use diceR when

  • you want a broad consensus workflow across many algorithms
  • you are comfortable preparing the input matrix yourself
  • you want diversity of cluster solutions to be the central driver

Use M3C when

  • false positive structure detection is your main concern
  • your data are naturally represented as a continuous matrix
  • you specifically want Monte Carlo reference-based calibration of consensus clustering

Use sharp when

  • weighted distance design and calibration are central to the problem
  • you want automated stability-based tuning of consensus clustering settings

Use uccdf when

  • your starting point is a mixed-type dataframe
  • you want schema handling, multiple representations, consensus, and null-aware selection in one package
  • you need row-level confidence and ambiguity outputs for downstream tabular reporting

Practical conclusion

uccdf should be read as a package that sits between classic ensemble tooling and null-calibrated stability methods.

It is not trying to replace:

  • the general ensemble machinery of clue
  • the algorithm-diversity workflow of diceR
  • the matrix-focused null calibration of M3C
  • the calibration strategy of sharp

Instead, it tries to unify the parts of those ideas that are most useful when an analyst starts from a real mixed-type dataframe rather than from a clean expression matrix or a precomputed ensemble object.

References