[Paper Review] Unified Spectral Clustering with Optimal Graph

Zhao Kang, Chong Peng+2 more|arXiv (Cornell University)|Nov 12, 2017

Face and Expression Recognition14 references3 citations

TL;DR

This paper proposes a unified spectral clustering framework that jointly learns an optimal similarity graph and discrete cluster labels in a single optimization process, avoiding the pitfalls of traditional three-step approaches. By integrating similarity learning, continuous label relaxation, and discrete label refinement—enhanced with multiple kernel learning—it achieves superior clustering performance across diverse datasets, outperforming state-of-the-art methods in accuracy, NMI, and purity.

ABSTRACT

Spectral clustering has found extensive use in many areas. Most traditional spectral clustering algorithms work in three separate steps: similarity graph construction; continuous labels learning; discretizing the learned labels by k-means clustering. Such common practice has two potential flaws, which may lead to severe information loss and performance degradation. First, predefined similarity graph might not be optimal for subsequent clustering. It is well-accepted that similarity graph highly affects the clustering results. To this end, we propose to automatically learn similarity information from data and simultaneously consider the constraint that the similarity matrix has exact c connected components if there are c clusters. Second, the discrete solution may deviate from the spectral solution since k-means method is well-known as sensitive to the initialization of cluster centers. In this work, we transform the candidate solution into a new one that better approximates the discrete one. Finally, those three subtasks are integrated into a unified framework, with each subtask iteratively boosted by using the results of the others towards an overall optimal solution. It is known that the performance of a kernel method is largely determined by the choice of kernels. To tackle this practical problem of how to select the most suitable kernel for a particular data set, we further extend our model to incorporate multiple kernel learning ability. Extensive experiments demonstrate the superiority of our proposed method as compared to existing clustering approaches.

Motivation & Objective

• To address the performance degradation caused by predefined, suboptimal similarity graphs in traditional spectral clustering.
• To overcome the sensitivity and approximation error introduced by k-means in discretizing continuous spectral labels.
• To unify the three stages of spectral clustering—similarity graph construction, continuous label learning, and discrete label assignment—into a single optimization framework.
• To develop a multiple kernel learning extension that adaptively selects optimal kernel combinations for complex, high-dimensional data.
• To provide a robust, end-to-end solution that iteratively improves similarity graph, continuous labels, and discrete cluster assignments through mutual boosting.

Proposed method

• Introduces a unified optimization framework that jointly learns the similarity graph, continuous cluster indicators, and discrete cluster labels in a single objective function.
• Uses kernel space mapping to enable flexible, data-adaptive similarity learning, replacing fixed metrics like Euclidean distance or k-NN.
• Incorporates a constraint that the learned similarity matrix must have exactly c connected components for c clusters, ensuring structural consistency.
• Applies a reassignment strategy to transform continuous spectral solutions into better-approximated discrete labels, reducing deviation from the true discrete solution.
• Employs an iterative alternating optimization scheme where each component (graph, continuous labels, discrete labels) is updated using the current state of the others.
• Extends the single-kernel model to multiple kernel learning by learning optimal kernel weights, enhancing adaptability to heterogeneous and high-dimensional data.

Experimental results

Research questions

• RQ1 Can a unified optimization framework jointly learn the optimal similarity graph and discrete cluster labels, improving upon the sequential three-step spectral clustering pipeline?
• RQ2 How does end-to-end learning of similarity graph and discrete labels reduce information loss compared to traditional methods?
• RQ3 To what extent does the proposed method outperform existing single-kernel and multiple-kernel clustering approaches on benchmark datasets?
• RQ4 How sensitive is the proposed method to hyperparameters, and does it maintain robust performance across diverse data types and noise levels?
• RQ5 Can multiple kernel learning within the unified framework further enhance clustering performance on complex, high-dimensional data?

Key findings

• The proposed method achieves superior clustering performance in terms of accuracy, NMI, and purity across nine benchmark datasets, consistently outperforming state-of-the-art methods.
• The average performance of the proposed method is significantly higher than that of traditional three-step approaches, even though the best results of the latter are sometimes close.
• The method is highly robust to parameter choices, with minimal performance degradation across wide ranges of α, β, and γ, indicating strong stability.
• The multiple kernel extension (SCMK) consistently improves performance over the single-kernel version, demonstrating the value of adaptive kernel combination.
• Random forest-based affinity methods (ClustRF-a) perform well on image data but fail on text data, highlighting the limitations of such approaches compared to the proposed unified framework.
• The method achieves state-of-the-art results with a single run, unlike k-means-based methods that require multiple restarts to achieve best results.