SCSPA Algorithm Documentation

🔍Overview

The Soft CSPA (sCSPA) is a consensus clustering algorithm designed to combine multiple soft clustering results using similarity-based vector space modeling. It is the soft version of the classic CSPA algorithm.

⚙️ Class Definition

Class Name: SCSPA

Implements the soft CSPA algorithm that merges several soft membership matrices and applies KMeans in the combined space.

class SCSPA:
    def __init__(self, n_clusters: int):
        self.n_clusters = n_clusters

📋 Parameters

Parameter	Type	Default	Description
`n_clusters`	int	—	Number of consensus clusters to output

🚀Using Example

from soft_clustering._scspa._scspa import SCSPA import numpy as np

Simulate 3 soft clusterings

soft1 = np.random.dirichlet(np.ones(3), size=100)
soft2 = np.random.dirichlet(np.ones(3), size=100)
soft3 = np.random.dirichlet(np.ones(3), size=100)

model = SCSPA(n_clusters=3)
labels = model.fit_predict([soft1, soft2, soft3])
print("Consensus Labels:", labels)

📥 Input / 📤 Output

Input to fit_predict(soft_memberships):
- soft_memberships (list of np.ndarray): List of soft clustering matrices (shape N x K)
Returns:
- labels (np.ndarray): Consensus clustering labels (length N)

🔧 Methods

__init__(self, n_clusters: int): Initializes the model with number of consensus clusters.
fit_predict(self, soft_memberships: List[np.ndarray]): Concatenates soft membership matrices and applies KMeans to generate consensus clusters.

🛠️ Implementation Notes

Each soft clustering is treated as a probabilistic embedding of objects.
The vectors are concatenated and normalized before clustering.
KMeans is used for deriving final cluster assignments.
Cosine similarity is optionally computed for debugging purposes.

📚 Reference

Punera, K., & Ghosh, J. (2008). Consensus-Based Ensembles of Soft Clusterings.