DiEC: Diffusion Embedded Clustering

2025-12-23

AI Summary

Existing deep clustering methods rely on a single encoder to produce static embeddings, overlooking the discriminative clustering potential embedded in the dynamic representation trajectories of pretrained diffusion models—specifically, their layer-wise and noise-step-wise internal activations. This paper proposes DiEC, the first method to directly leverage the two-dimensional (layer–time) internal activations of diffusion models for unsupervised clustering. DiEC employs weakly coupled decomposition and a two-stage search (CML + OTS) to identify clustering-favorable bottleneck layers and optimal denoising timesteps. To enhance representation separability and structural robustness, it introduces denoising consistency regularization, adaptive graph regularization, and entropy regularization. Evaluated on multiple benchmarks, DiEC achieves state-of-the-art or leading performance, significantly outperforming conventional clustering paradigms based on fixed embeddings.

Technology Category

• Machine Learning: Deep Generative Models & Autoencoders
• Computer Vision: Diffusion Models for Vision
• Search and Optimization: Learning to Search

Application Category

• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• Web Mining and Content Analysis: Normalization, clustering, classification, and summarization of Web text
• User Modeling, Personalization and Recommendation: Explainable and interpretable methods for personalization

Abstract

Deep clustering hinges on learning representations that are inherently clusterable. However, using a single encoder to produce a fixed embedding ignores the representation trajectory formed by a pretrained diffusion model across network hierarchies and noise timesteps, where clusterability varies substantially. We propose DiEC (Diffusion Embedded Clustering), which performs unsupervised clustering by directly reading internal activations from a pretrained diffusion U-Net. DiEC formulates representation selection as a two-dimensional search over layer x timestep, and exploits a weak-coupling property to decompose it into two stages. Specifically, we first fix the U-Net bottleneck layer as the Clustering-friendly Middle Layer (CML), and then use Optimal Timestep Search (OTS) to identify the clustering-optimal timestep (t*). During training, we extract bottleneck features at the fixed t* and obtain clustering representations via a lightweight residual mapping. We optimize a DEC-style KL self-training objective, augmented with adaptive graph regularization and entropy regularization to strengthen cluster structures. In parallel, we introduce a denoising-consistency branch at random timesteps to stabilize the representations and preserve generative consistency. Experiments show that DiEC achieves competitive clustering performance on multiple standard benchmarks.

Problem

• Unsupervised clustering using diffusion model activations
• Selecting optimal layer and timestep for clusterable representations
• Enhancing cluster structures with regularization and consistency branches

Innovation

• Uses pretrained diffusion U-Net internal activations for clustering
• Decomposes layer-timestep search into two-stage selection process
• Combines KL self-training with adaptive and entropy regularization