Abstract:This paper introduces a regularized projection matrix approximation framework designed to recover cluster information from the affinity matrix. The model is formulated as a projection approximation problem, incorporating an entry-wise penalty function. We investigate three distinct penalty functions, each specifically tailored to address bounded, positive, and sparse scenarios. To solve this problem, we propose direct optimization on the Stiefel manifold, utilizing the Cayley transformation along with the Alternating Direction Method of Multipliers (ADMM) algorithm. Additionally, we provide a theoretical analysis that establishes the convergence properties of ADMM, demonstrating that the convergence point satisfies the KKT conditions of the original problem. Numerical experiments conducted on both synthetic and real-world datasets reveal that our regularized projection matrix approximation approach significantly outperforms state-of-the-art methods in clustering performance.

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is to improve the existing community detection methods, especially by introducing a regularized projection matrix approximation framework to enhance the clustering performance. Specifically, the paper aims to recover cluster information from the affinity matrix and proposes a projection approximation problem combined with an entry - level penalty function. ### Problem Background In unsupervised learning, community detection is an important problem, which involves dividing data points into multiple groups based on their similarities. The traditional approach is to first extract a low - dimensional representation from the affinity matrix \(A\in\mathbb{R}^{n\times n}\), and then apply clustering algorithms such as the k - means or EM algorithm for grouping. However, the success of this method depends on the quality of the data representation and the accuracy of the calculation method. ### Proposed Method To improve this process, the paper proposes a new Regularized Projection Matrix Approximation (RPMA) framework. This framework addresses the limitations of existing methods in the following ways: 1. **Introducing three different penalty functions**: Three penalty functions are designed for bounded, non - negative, and sparse scenarios respectively to adapt to different application scenarios. 2. **Solving the optimization problem**: The Cayley transform and the Alternating Direction Method of Multipliers (ADMM) are used to directly optimize the projection matrix approximation problem on the Stiefel manifold. 3. **Theoretical analysis**: The convergence analysis of the ADMM algorithm is provided, proving that the convergence point satisfies the KKT conditions of the original problem. ### Mathematical Model The core optimization problem in the paper can be formalized as: \[ \min_{X\in P_K}\|A - X\|_F^2+\lambda\sum_{i,j}g(X_{ij}) \] where: - \(P_K\subseteq\mathbb{R}^{n\times n}\) is the set of all projection matrices with rank \(K\). - \(g\in G\) is a regularization term used to limit the element range of the projection matrix. - \(\lambda\) is a non - negative parameter used to balance the approximation error and the entry - level constraint. ### Main Contributions 1. **Proposing a new regularized projection matrix approximation framework**: By introducing three different types of penalty functions, the performance of the classical spectral clustering method is improved. 2. **Developing two algorithms**: One is a direct optimization method on the Stiefel manifold, and the other is a constraint separation method based on ADMM. 3. **Experimental verification**: Through experiments on synthetic and real - world datasets, the effectiveness and robustness of this method are demonstrated. ### Conclusion By introducing the regularized projection matrix approximation framework, this paper successfully solves the problems existing in the existing community detection methods, especially showing significant advantages when dealing with complex data.

Regularized Projection Matrix Approximation with Applications to Community Detection

Adaptive Projected Clustering with Graph Regularization.

A Random Projection Method for Large-Scale Community Detection

Robust Manifold Matrix Factorization for Joint Clustering and Feature Extraction.

Regularized spectral clustering under the mixed membership stochastic block model

Community Detection and Link Prediction Via Cluster-driven Low-rank Matrix Completion.

Double low-rank representation with projection distance penalty for clustering

Low rank approximation with sparse integration of multiple manifolds for data representation

Semi-supervised sparse subspace clustering with manifold regularization

Sparse Subspace Clustering Using Square-Root Penalty

High-order Order Proximity-Incorporated, Symmetry and Graph-Regularized Nonnegative Matrix Factorization for Community Detection

Nonconvex Multi-View Subspace Clustering Via Simultaneously Learning the Representation Tensor and Affinity Matrix*

A Symmetry and Graph Regularized Nonnegative Matrix Factorization Model for Community Detection

Generalized Low-Rank Matrix Completion Model with Overlapping Group Error Representation

Fuzzy Clustering Method with Approximate Orthogonal Regularization.

Ensemble Manifold Structured Low Rank Approximation for Data Representation

Regularized Locality Preserving Projections with Two-Dimensional Discretized Laplacian Smoothing

Sparse Approximation to the Eigensubspace for Discrimination

Affinity adaptive sparse subspace clustering via constrained Laplacian rank

Low-rank matrix recovery with non-quadratic loss: projected gradient method and regularity projection oracle

GNMFO_TW: Graph Regularized Non-negative Matrix Factorization for Overlapping Community Discovery by Using Three-Way