Kernel-based Fuzzy Competitive Learning (K-FCCL)

An adaptive soft-to-hard clustering algorithm using kernel-induced similarity and competitive learning dynamics.

🔍 Overview

Kernel-based Fuzzy Competitive Learning (K-FCCL) is a soft clustering algorithm that leverages kernel methods and competitive learning principles to uncover complex, non-linear data structures. By mapping input data into a higher-dimensional space using a Gaussian kernel, K-FCCL effectively separates clusters that are inseparable in the original feature space.

Unlike traditional fuzzy clustering methods, K-FCCL incorporates a competitive learning mechanism that drives the model toward optimal fuzzy partitioning using kernel-induced similarities and probabilistic softmax dynamics. This results in improved handling of overlapping clusters and more interpretable assignments.

⚙️ Class Definition

class soft_clustering.KFCCL(
    n_clusters: int = 2,
    lambda_: float = 10.0,
    gamma: float = 1.0,
    epsilon: float = 1e-4,
    max_iter: int = 100
)

📋 Parameters

Parameter	Type	Default	Description
`n_clusters`	`int`	`2`	Number of clusters to form.
`lambda_`	`float`	`10.0`	Fuzziness control parameter for softmax membership assignment. Larger values lead to crisper partitions.
`gamma`	`float`	`1.0`	Parameter for the Gaussian (RBF) kernel controlling how tightly the similarity decays. Must be tuned for performance.
`epsilon`	`float`	`1e-4`	Convergence tolerance. Iterations stop when inner products stabilize below this threshold.
`max_iter`	`int`	`100`	Maximum number of training iterations.

🛠️ Methods

`fit(X)`

Trains the K-FCCL model on the input dataset using kernel similarity and iterative soft competition.

Parameters:

X (np.ndarray): Data matrix of shape (n_samples, n_features).

Returns:

labels (np.ndarray): Final hard cluster assignments computed from the soft membership matrix U.

📌 Key Features

✅ Kernelized Learning: Uses Gaussian RBF kernel to model non-linear cluster boundaries.
✅ Softmax-based Fuzziness: Learns fuzzy memberships using a temperature-scaled softmax driven by inner product similarities.
✅ Online Competitive Update Rule: Incorporates competitive learning with adaptive learning rate for more robust convergence.
✅ Supports Non-Convex Clusters: Especially effective when clusters are non-spherical or not linearly separable.

📝 Notes

The gamma parameter has a major impact on clustering quality. Smaller values produce more global similarity, while larger values focus on local structure.
Fuzziness (lambda_) affects how strongly each data point is assigned to its clusters. Higher values encourage hard assignments.
The method returns hard labels, but the full fuzzy partitioning is available via the U matrix if needed.
Use visual inspection and metrics like Adjusted Rand Index (ARI) for evaluating clustering quality, especially on synthetic or labeled data.

📚 Reference

Mizutani, K., & Miyamoto, S. (2005). Kernel-Based Fuzzy Competitive Learning Clustering. IEEE Xplore 1452468.