Search

The success of fuzzy clustering relies heavily on the features of the input data. However, traditional fuzzy clustering methods focus on primitive features that represent the physical properties of data, such as the shape and color of objects. Usually, these properties could only capture the local distribution of data, while ignoring the global structure of data. Fortunately, by performing spectral analysis on the similarity graph of data, spectral clustering methods explore more on the global structure. However, capturing a decent global structure by constructing an ideal similarity graph is not easy. In this paper, we propose a novel fuzzy clustering method, the bilevel fuzzy clustering via adaptive similarity graphs fusion (BFCS), to solve the problems mentioned above at the same time. On the one hand, the global data structure is investigated by building an adaptive similarity graph using autoweighted fusion. On the other hand, the consistent memberships for both fuzzy clustering and spectral clustering are formulated into one objective function to simultaneously retain the local physical properties and the global data structure. Notably, the proposed method shows good scalability, and the bilevel structure of the algorithm makes its application to large datasets easy. Extensive experiments demonstrate the promising clustering performance of the proposed BFCS on synthetic and real datasets. [ABSTRACT FROM AUTHOR]

The flexible paradigm of the Resource Description Framework (RDF) has accelerated the rate at which raw data is published on the web. Therefore, the volume of generated RDF data has increased impressively in the last decade, which promotes the use of compression to manage and reduce the size of RDF datasets. Furthermore, researchers have recently tried to reconstruct convolution and pooling procedures to better suit the structure of graphs and make convolutional neural networks (CNNs) more applicable to RDF graph data. In this study, we propose the Multi Kernel Inductive RDF Graph Convolution Network (MKIR-GCN), which, rather than compressing nodes/edges independently, uses similarities between nodes and the structure of graphs to reduce the size of all nodes and edges simultaneously to efficiently compress the RDF graph. By considering the topology and similarity of a network's nodes, our proposed attention based on similarity for RDF graph pooling (ASGPool) picks the most informative and representative nodes. In dynamic graphs, our proposed MKIR-GCN layer may learn more generic node representations by focusing on diverse characteristics. Through extensive experimentation, we can conclude that the proposed approach significantly improves compression over the existing graph representation learning schemes and RDF graph compression schemes. [ABSTRACT FROM AUTHOR]