CDCGS (Community Detection Clustering via Gumbel Softmax) Documentation

🔍 Overview

CDCGS is a graph clustering algorithm that applies the Gumbel-Softmax trick to learn soft assignments of nodes to communities. It computes a community relationship matrix R using these assignments and applies softmax to normalize interactions between clusters.

---

⚙️ Class Definition

Class Name: CDCGS This class implements the Gumbel-Softmax based community clustering model.

class CDCGS(nn.Module):
    def __init__(self, num_nodes: int, n_clusters: int, tau: float = 1.0):
        ...

---

📋 Parameters

Parameter	Type	Default	Description
num_nodes	int	—	Number of graph nodes
n_clusters	int	—	Number of target communities
tau	float	1.0	Gumbel-Softmax temperature

---

🚀 Usage Examples

from soft_clustering._cdcgs._cdcgs import CDCGS
import torch

adj = torch.tensor([
    [0, 1, 0, 0, 0, 0],
    [1, 0, 1, 0, 0, 0],
    [0, 1, 0, 1, 0, 0],
    [0, 0, 1, 0, 1, 0],
    [0, 0, 0, 1, 0, 1],
    [0, 0, 0, 0, 1, 0]
], dtype=torch.float)

model = CDCGS(num_nodes=6, n_clusters=2, tau=1.0)
R, soft_assign = model(adj)
loss = model.loss(R)

print("Loss:", loss.item())

---

🛠️ Methods

• forward(adj): Computes community assignments and relation matrix

• loss(output): Encourages the relation matrix to be close to identity

---

📝 Implementation Notes

• Learns soft cluster assignments via Gumbel-Softmax

• Encourages orthogonal (independent) clusters via identity loss

• Simple and fully unsupervised

---

📚 Reference

This implementation is based on:
“Community Detection Clustering via Gumbel Softmax”
by H. Zhang, J. Bu, Y. Wang, and C. Chen.