A0995
Title: A fuzzy spectral clustering model
Authors: Giorgia Zaccaria - University of Milano-Bicocca (Italy)
Cinzia Di Nuzzo - University of Catania (Italy) [presenting]
Abstract: A new fuzzy approach to the spectral clustering model is introduced. Standard spectral clustering is a technique that exploits the spectral structure of data to partition them into homogeneous groups. Unlike traditional methods such as k-means, spectral clustering does not assume a specific cluster shape and can handle non-linearly separable data. The process involves constructing a similarity matrix, computing the Laplacian matrix, identifying its eigenvectors, and using these eigenvectors to represent the data in a reduced-dimensional space where clustering is more evident. It has been demonstrated that spectral clustering is effective for complex or non-linear structures and can handle high-dimensional data. Integrating this method with a fuzzy approach for clustering has been deemed crucial for maximizing efficiency and coherence in data representation by capturing the intrinsic relationships among data. In the proposed fuzzy method, a least squares approach is used to estimate the model, resulting in a fuzzy Laplacian configuration representative of the entire dataset. The utility of this method is demonstrated through empirical evaluations of synthetic and real-world datasets. The results show the effectiveness of the approach in uncovering complex patterns and providing significant insights. Further developments in a robust framework are also outlined.
