Package 'fussclust'

Title: Fuzzy Unsupervised and Semi-Supervised Clustering
Description: Methods for distance-based fuzzy unsupervised and semi-supervised clustering, including fuzzy and possibilistic models based on alternating optimization (AO) algorithm. The package introduces a vectorized estimation framework for prototype-based fuzzy clustering algorithms, enabling modular algorithm design and extensibility. It also supports storage and retrieval of intermediate AO optimization results for downstream analysis and processing. For more details see Kmita et al. (2024) <doi:10.1109/TFUZZ.2024.3370768>.
Authors: Kamil Kmita [aut, cre, cph] (ORCID: <https://orcid.org/0000-0001-8829-2420>)
Maintainer: Kamil Kmita <[email protected]>
License: MIT + file LICENSE
Version: 0.1.0
Built: 2026-06-02 18:57:53 UTC
Source: https://github.com/cran/fussclust

Help Index

Calculates data evidence matrix E from distances matrix D.

Description

Calculates data evidence matrix E from distances matrix D.

Usage

calculate_evidence(D)

Arguments

D

Distances matrix of size N x c.

Value

Matrix of size N x c.

Creates DHE (stands for "distances horizontally exploded") and DVE (stands for "distances vertically exploded") matrices.

Description

Creates DHE (stands for "distances horizontally exploded") and DVE (stands for "distances vertically exploded") matrices.

Usage

dheve(A, vertical)

Arguments

A

Matrix of size N x c.
vertical

Boolean switch. If TRUE, create DVE (vertical explosion). If FALSE, create DHE (horizontal explosion).

Value

Matrix of size Nc x c

Estimated T matrix with typicalities in semi-supervised case.

Description

Estimated T matrix with typicalities in semi-supervised case.

Usage

estimate_super_T(D, superF, alpha, gammas, b = 1)

Arguments

D

Distances matrix of size N x c.
superF

Binary supervision matrix of size N x c.
alpha

Scaling factor, a floating point > 0 regulating the impact of partial supervision.
gammas

a c-vector of cluster-specific gamma hyperparameters.
b

a scalar weighting the contribution of possibilistic membership in SPFCM (semi-supervised possibilistic fuzzy c-means) model. It is set to 1 by default for other semi-supervised models.

Estimated T matrix with typicalities in unsupervised case.

Description

Estimated T matrix with typicalities in unsupervised case.

Usage

estimate_T(D, gammas)

Arguments

D

Distances matrix of size N x c.
gammas

a c-vector of cluster-specific gamma hyperparameters.

Estimated U matrix with memberships in semi-supervised case.

Description

Estimated U matrix with memberships in semi-supervised case.

Usage

estimate_U(D, superF, alpha)

Arguments

D

Distances matrix of size N x c.
superF

Binary supervision matrix of size N x c.
alpha

Scaling factor, a floating point > 0 regulating the impact of partial supervision.

Equation to calculate clusters' prototypes matrix V^\hat{V}.

Description

Equation to calculate clusters' prototypes matrix V^\hat{V}.

Usage

estimate_V(Phi, X)

Arguments

Phi

Matrix with weights of size N x c.
X

Matrix with predictors of size N x p.

Value

Clusters' prototypes matrix of size c x p.

Fuzzy C-Means clustering model

Description

Fits a Fuzzy C-Means (FCM) clustering model using the Alternating Optimization algorithm.

Usage

FCM(
  X,
  C,
  U = NULL,
  max_iter = 200,
  conv_criterion = 1e-04,
  function_dist = rdist::cdist,
  store_history = FALSE
)

Arguments

X

A numeric feature matrix.
C

Integer specifying the number of clusters.
U

Optional initial membership matrix. Primarily intended for reproducibility purposes. If NULL (default), the algorithm uses a random initialization.
max_iter

Maximum number of iterations. Defaults to 200.
conv_criterion

Convergence threshold used at the end of each iteration of the Alternating Optimization algorithm.
function_dist

Optional distance function. The function must accept two matrices, X and V, with the same number of columns, and return a matrix of size ⁠nrow(X) x nrow(V)⁠ containing distances between each row of X and each row of V.

For the Euclidean distance, the returned distances should not be squared. Defaults to rdist::cdist().
store_history

Logical indicating whether optimization histories should be stored. If FALSE, the returned object will contain NULL history fields. Defaults to TRUE.

Value

An object of class fcm containing:

U

An N×CN \times C membership matrix.

V

A C×pC \times p matrix of cluster prototypes.

function_dist

The distance function used by the model.

counter

Number of iterations performed until convergence.

U_history

If store_history = TRUE, a list of length counter containing membership matrices estimated at each iteration; otherwise NULL.

V_history

If store_history = TRUE, a list of length counter containing prototype matrices estimated at each iteration; otherwise NULL.

Phi_history

If store_history = TRUE, a list of length counter containing phi-weight matrices estimated at each iteration; otherwise NULL.

References

Bezdek, J. C. (1981). Pattern Recognition with Fuzzy Objective Function Algorithms. Springer US. https://doi.org/10.1007/978-1-4757-0450-1

Examples

X <- matrix(rnorm(100), ncol = 2)

model_fcm <- fussclust::FCM(
  X = X,
  C = 2
)

print(model_fcm$V)

Aggregates elements of DHE and DVE matrices in a step to build evidence matrix E.

Description

Aggregates elements of DHE and DVE matrices in a step to build evidence matrix E.

Usage

gamma_fcm(dhe, dve)

Arguments

dhe

DHE matrix of size Nc x c.
dve

DVE matrix of size Nc x c.

Value

Matrix of size Nc x 1.

Initialization procedure to calculate values of gamma hyperparameters.

Description

Initialization procedure to calculate values of gamma hyperparameters.

Usage

init_gamma(.model, .X)

Arguments

.model

estimated model of class fcm
.X

features matrix of size N x c

Possibilistic C-Means clustering model

Description

Fits a Possibilistic C-Means (PCM) clustering model using the Alternating Optimization algorithm.

Usage

PCM(
  X,
  C,
  U = NULL,
  gammas = NULL,
  initFCM = NULL,
  max_iter = 200,
  conv_criterion = 1e-04,
  function_dist = rdist::cdist,
  store_history = FALSE
)

Arguments

X

A numeric feature matrix.
C

Integer specifying the number of clusters.
U

Optional initial membership matrix. Primarily intended for reproducibility purposes. If NULL (default), the algorithm uses a random initialization.
gammas

Optional vector of cluster-specific gamma hyperparameters. If NULL (default), the initialization strategy depends on the value of initFCM.

If initFCM is NULL, a vector of ones is used. Otherwise, a Fuzzy C-Means model is first fitted, and the init_gamma() function is used to estimate the cluster-specific gamma hyperparameters.
initFCM

Optional fitted Fuzzy C-Means model used to initialize cluster-specific gamma hyperparameters via weighted averaging. If NULL (default), no preliminary Fuzzy C-Means initialization is used. If provided, this argument is effective only when gammas is NULL.
max_iter

Maximum number of iterations. Defaults to 200.
conv_criterion

Convergence threshold used at the end of each iteration of the Alternating Optimization algorithm.
function_dist

Optional distance function. The function must accept two matrices, X and V, with the same number of columns, and return a matrix of size ⁠nrow(X) x nrow(V)⁠ containing distances between each row of X and each row of V.

For the Euclidean distance, the returned distances should not be squared. Defaults to rdist::cdist().
store_history

Logical indicating whether optimization histories should be stored. If FALSE, the returned object will contain NULL history fields. Defaults to TRUE.

Value

An object of class pcm containing:

U

An N×CN \times C membership matrix.

V

A C×pC \times p matrix of cluster prototypes.

function_dist

The distance function used by the model.

counter

Number of iterations performed until convergence.

gammas

Vector of cluster-specific gamma hyperparameters.

U_history

If store_history = TRUE, a list of length counter containing membership matrices estimated at each iteration; otherwise NULL.

V_history

If store_history = TRUE, a list of length counter containing prototype matrices estimated at each iteration; otherwise NULL.

Phi_history

If store_history = TRUE, a list of length counter containing phi-weight matrices estimated at each iteration; otherwise NULL.

References

Krishnapuram, R., & Keller, J. (1993). A possibilistic approach to clustering. IEEE Transactions on Fuzzy Systems, 1(2), 98–110. https://doi.org/10.1109/91.227387

Examples

X <- matrix(rnorm(100), ncol = 2)

model_pcm <- fussclust::PCM(
  X = X,
  C = 2,
  initFCM = TRUE
)

print(model_pcm$V)

Predict method for ssfcm objects

Description

Predicts cluster memberships for new observations using a fitted Semi-Supervised Fuzzy C-Means model.

Usage

## S3 method for class 'ssfcm'
predict(object, X, ...)

Arguments

object

An object of class ssfcm.
X

A numeric matrix of new observations with pp columns.
...

Additional arguments. Currently ignored.

Value

A matrix of size N×CN \times C containing predicted cluster memberships, where CC is the number of clusters.

Examples

X <- matrix(rnorm(100), ncol = 2)

superF <- matrix(0, nrow = nrow(X), ncol = 2)

superF[1:10, 1] <- 1
superF[11:20, 2] <- 1

model_ssfcm <- SSFCM(
  X = X,
  C = 2,
  superF = superF,
  alpha = 1
)

predict(model_ssfcm, matrix(rnorm(2), ncol = 2))

Predict method for sspcm objects

Description

Predicts cluster memberships for new observations using a fitted Semi-Supervised Possibilistic C-Means model.

Usage

## S3 method for class 'sspcm'
predict(object, X, ...)

Arguments

object

An object of class sspcm.
X

A numeric matrix of new observations with pp columns.
...

Additional arguments. Currently ignored.

Value

A matrix of size N×CN \times C containing predicted cluster memberships, where CC is the number of clusters.

Examples

X <- matrix(rnorm(100), ncol = 2)

superF <- matrix(0, nrow = nrow(X), ncol = 2)

superF[1:10, 1] <- 1
superF[11:20, 2] <- 1

model_sspcm <- SSPCM(
  X = X,
  C = 2,
  superF = superF,
  initFCM = TRUE,
  alpha = 1
)

predict(model_sspcm, matrix(rnorm(2), ncol = 2))

Semi-Supervised Fuzzy C-Means clustering model

Description

Fits a Semi-Supervised Fuzzy C-Means (SSFCM) clustering model using the Alternating Optimization algorithm.

Usage

SSFCM(
  X,
  C,
  U = NULL,
  max_iter = 200,
  conv_criterion = 1e-04,
  function_dist = rdist::cdist,
  store_history = FALSE,
  alpha = NULL,
  superF = NULL
)

Arguments

X

A numeric feature matrix.
C

Integer specifying the number of clusters.
U

Optional initial membership matrix. Primarily intended for reproducibility purposes. If NULL (default), the algorithm uses a random initialization.
max_iter

Maximum number of iterations. Defaults to 200.
conv_criterion

Convergence threshold used at the end of each iteration of the Alternating Optimization algorithm.
function_dist

Optional distance function. The function must accept two matrices, X and V, with the same number of columns, and return a matrix of size ⁠nrow(X) x nrow(V)⁠ containing distances between each row of X and each row of V.

For the Euclidean distance, the returned distances should not be squared. Defaults to rdist::cdist().
store_history

Logical indicating whether optimization histories should be stored. If FALSE, the returned object will contain NULL history fields. Defaults to TRUE.
alpha

Positive scaling factor regulating the impact of partial supervision.
superF

Binary supervision matrix of the same dimensions as U, indicating the available partial supervision information.

Value

An object of class sspcm containing:

U

An N×CN \times C memberships matrix.

V

A C×pC \times p matrix of cluster prototypes.

function_dist

The distance function used by the model.

counter

Number of iterations performed until convergence.

alpha

Value of scaling factor.

U_history

If store_history = TRUE, a list of length counter containing membership matrices estimated at each iteration; otherwise NULL.

V_history

If store_history = TRUE, a list of length counter containing prototype matrices estimated at each iteration; otherwise NULL.

Phi_history

If store_history = TRUE, a list of length counter containing phi-weight matrices estimated at each iteration; otherwise NULL.

References

Kmita, K., Kaczmarek-Majer, K., & Hryniewicz, O. (2024). Explainable Impact of Partial Supervision in Semi-Supervised Fuzzy Clustering. IEEE Transactions on Fuzzy Systems, 1–10. https://doi.org/10.1109/TFUZZ.2024.3370768

Examples

X <- matrix(rnorm(100), ncol = 2)

superF <- matrix(0, nrow = nrow(X), ncol = 2)

superF[1:10, 1] <- 1
superF[11:20, 2] <- 1

model_ssfcm <- SSFCM(
  X = X,
  C = 2,
  superF = superF,
  alpha = 1
)

print(model_ssfcm$V)

Semi-Supervised Possibilistic C-Means clustering model

Description

Fits a Semi-Supervised Possibilistic C-Means (SSPCM) clustering model using the Alternating Optimization algorithm.

Usage

SSPCM(
  X,
  C,
  U = NULL,
  gammas = NULL,
  initFCM = NULL,
  max_iter = 200,
  conv_criterion = 1e-04,
  function_dist = rdist::cdist,
  store_history = FALSE,
  alpha = NULL,
  superF = NULL
)

Arguments

X

A numeric feature matrix.
C

Integer specifying the number of clusters.
U

Optional initial membership matrix. Primarily intended for reproducibility purposes. If NULL (default), the algorithm uses a random initialization.
gammas

Optional vector of cluster-specific gamma hyperparameters. If NULL (default), the initialization strategy depends on the value of initFCM.

If initFCM is NULL, a vector of ones is used. Otherwise, a Fuzzy C-Means model is first fitted, and the init_gamma() function is used to estimate the cluster-specific gamma hyperparameters.
initFCM

Optional fitted Fuzzy C-Means model used to initialize cluster-specific gamma hyperparameters via weighted averaging. If NULL (default), no preliminary Fuzzy C-Means initialization is used. If provided, this argument is effective only when gammas is NULL.
max_iter

Maximum number of iterations. Defaults to 200.
conv_criterion

Convergence threshold used at the end of each iteration of the Alternating Optimization algorithm.
function_dist

Optional distance function. The function must accept two matrices, X and V, with the same number of columns, and return a matrix of size ⁠nrow(X) x nrow(V)⁠ containing distances between each row of X and each row of V.

For the Euclidean distance, the returned distances should not be squared. Defaults to rdist::cdist().
store_history

Logical indicating whether optimization histories should be stored. If FALSE, the returned object will contain NULL history fields. Defaults to TRUE.
alpha

Positive scaling factor regulating the impact of partial supervision.
superF

Binary supervision matrix of the same dimensions as U, indicating the available partial supervision information.

Value

An object of class sspcm containing:

U

An N×CN \times C typicalities matrix.

V

A C×pC \times p matrix of cluster prototypes.

function_dist

The distance function used by the model.

counter

Number of iterations performed until convergence.

gammas

Vector of cluster-specific gamma hyperparameters.

alpha

Value of scaling factor.

U_history

If store_history = TRUE, a list of length counter containing membership matrices estimated at each iteration; otherwise NULL.

V_history

If store_history = TRUE, a list of length counter containing prototype matrices estimated at each iteration; otherwise NULL.

Phi_history

If store_history = TRUE, a list of length counter containing phi-weight matrices estimated at each iteration; otherwise NULL.

References

Kmita, K., Kaczmarek-Majer, K., & Hryniewicz, O. (2024). Explainable Impact of Partial Supervision in Semi-Supervised Fuzzy Clustering. IEEE Transactions on Fuzzy Systems, 1–10. https://doi.org/10.1109/TFUZZ.2024.3370768

Examples

X <- matrix(rnorm(100), ncol = 2)

superF <- matrix(0, nrow = nrow(X), ncol = 2)

superF[1:10, 1] <- 1
superF[11:20, 2] <- 1

model_sspcm <- SSPCM(
  X = X,
  C = 2,
  superF = superF,
  alpha = 1
)

print(model_sspcm$V)

Binary supervision structure to reconstruct the issue of underimpact of partial supervision.

Description

This dataset provides a concrete superivison structure: - 'superF' matrix of size 150 x 3 with partial supervision, - 'ind' vector with indices of unsupervised observations, - 'tind' vector with indicies of observations selected to be in the test dataset, - 'tclass' vector with class membership of the observations selected to be in the test dataset.

This supervision structure is meant to reproduce a particular realization of phenomenon of underimpact of partial supervision specific to the iris dataset.

Usage

data(superFstruct_underimpact)

Format

A list with: a matrix of size 150 x 3, and three vectors.

Initialization matrix to analyze underimpact in iris data.

Description

This dataset provides a concrete initialization of membership matrix specific to the iris data that exhibits the phenomenon of underimpact of partial supervision in semi-supervised fuzzy clustering.

Usage

data(U_underimpact)

Format

A matrix of size 150 x 3.

Rearranges elements of input matrix from a block matrix with vertical blocks (column vectors) to a block matrix with horizontal blocks (row vectors).

Description

Rearranges elements of input matrix from a block matrix with vertical blocks (column vectors) to a block matrix with horizontal blocks (row vectors).

Usage

xi_fcm(A, c)

Arguments

A

Matrix of size Nc x 1.
c

Number of columns in the wanted matrix. Associated with the number of clusters.

Value

Matrix of size N x c.