Traditional cloud similarity measures are only suitable for single granularity, and the multi-granularity cloud similarity measure is insufficient in research. In this paper, a knowledge distance framework is proposed and its relative properties are proved. The relationships between knowledge distance and information measure and information granularity are established. Moreover, two valuable conclusions are drawn in a hierarchical granular structure. The granularity difference of two granular spaces in a hierarchical granular structure is positive correlation to the knowledge distance framework between them. The granular spaces change continuously with the granularity and they can be mapped into the one-dimension coordinate. Finally, a similarity measure of cloud model based on the knowledge distance framework(KDFCM) is proposed. The experiments verify that the properties of KDFCM are consistent with the above conclusions.

Community detection based on edge clustering is capable of detecting overlapping communities naturally. However, it engenders the problems of obscure belongingness of the nodes on community borders and the excessive overlap of communities. In this paper, an overlapping community detection based on edge density clustering(OCDEDC) algorithm is proposed. Firstly, density clustering based on edges is employed to extract core edge communities. Next, a partitioning strategy is designed to dispatch border edges to its closest core edge community. In addition, a strategy based on the degrees and community belongingness of edges is designed to handle the isolated edges, and thus the excessive overlap of communities is avoided. Finally, edge communities are transformed back into node communities as the output. Experiments on artificial and real datasets show that the proposed algorithm detects overlapping communities efficiently and effectively.

Firstly, the OE-concept information system is defined, and the distance between OE-concepts is produced on the basis of characteristics of the OE-concepts. Then, the OE-concepts are clustered by K-Modes clustering, and the subcontext is obtained through class centers. K-deletion transformation is defined to study the relationship between the OE-concept lattices of the original context and the subcontext, and it is employed to realize the compression of OE-concept lattice. The relationship between the compressed lattice and the original OE-concept lattice is discussed. Finally, experiments are conducted to prove the effectiveness and superiority of the proposed method.

Some discriminative text features are lost due to the lack of guidance of label information during dimensionality reduction in unsupervised topic models. Thus, the classification result is unsatisfactory. A supervised topic model with deep learning, namely supervised latent Dirichlet allocation with deep learning(DL-sLDA) is proposed in this paper. A deep neural network is exploited to establish the mapping between the topic of a document and its label. Both variational expectation-maximization and deep learning method are adopted to update the defined parameters under Bayesian framework in DL-sLDA. When the structure of the deep network and the type of the activation function are changed properly, the proposed model can be utilized for both classification and regression tasks. The experimental result demonstrates that DL-sLDA maintains the ability of topic extraction and gains a better predictive ability.

Aiming at the increasingly serious haze pollution, a haze prediction method combining co-evolution artificial fish swarm algorithm(CEAFSA) and support vector machine(SVM) is proposed. Firstly, an improved artificial fish swarm algorithm is proposed by initializing evenly distributed population using the good point set, and introducing adaptive strategies for visual scope and step and co-evolution strategies among subpopulations. Then, the main parameters of SVM are optimized by co-evolution artificial fish swarm algorithm. Finally, haze prediction model is established by SVM. Experimental results on 10 Benchmark testing functions verify the validity of CEAFSA and the results on 6 UCI datasets demonstrate its high stability and effectiveness.

The correlation between labels are studied through the related information about the marked labels. However, the influence of the relationship between unmarked and marked labels on the quality of the multi-label set is not taken into account. Inspired by k-nearest neighbors(KNN), a non-equilibrium labels completion of neighboring labels space(NeLC-NLS) is proposed to improve the quality of the neighboring label space and the performance of the multi-label classification. Firstly, the information entropy between labels is utilized to measure the strength of the relationship between labels, and the confidence matrix of the basic label is obtained. Then, the confidence matrix of non-equilibrium labels containing more information is obtained via the proposed non-equilibrium label confidence matrix. Secondly, the similarity of samples is measured in the feature space and the k-nearest neighbors are obtained. Then, the non-equilibrium labels completion matrix is employed to calculate the label completion matrix of the neighboring labels space. Finally, the extreme learning machine is adopted as a linear classifier. The experimental results of the proposed algorithm on 8 public multi-label datasets show that NeLC-NLS is superior to other multi-label learning algorithms. The effectiveness of NeLC-NLS is further illustrated by using hypothesis testing and stability analysis.

Extracting keywords directly from a single document cannot satisfy the precision requirements of keyphrase extraction, and the existing methods for keyphrase extraction based on neighbor information are time-consuming. In this paper, common author relations in research papers are utilized to build a neighbor network, and neighbor network information as well as document content is used to extract keyphrases. Based on those, high frequency pairs of phrase co-occurrence in domain are incorporated to further acquire high-quality keyphrases. Experimental results demonstrate that the proposed method is more computationally efficient than the existing methods.