How to improve the effect of Web service discovery is a key issue in service-oriented computing field. Aiming at this issue, a Web service discovery method based on bi-directional semantic and multiple relationships between concepts(BSMRC) is proposed in this paper. The bi-directional semantic and the multiple relationships between concepts are taken into account in the similarity measurement. Different from the traditional Web service discovery method, not only nouns and verbs but also adjectives are computed. In the similarity measurement of nouns and verbs, the multiple relationships, such as ISA, HASA and ANT, are covered. Moreover, both similar-to and ANT relationships are involved in the similarity measurement of adjectives. The semantic similarity between service and request is computed by integrating I/O interface similarity and description similarity to improve the accuracy of service discovery. By means of the experiments based on synthetic data, the effectiveness of the proposed method is validated. The method has better precision, recall and F-measure.

Most subspace clustering methods can not deal with nonlinear data satisfactorily, and the data in different subspaces possess higher similarity and clustering error can not be verified in time. Aiming at these problems, an overlapping subspace clustering algorithm based on local weighted least squares regression(LWLSR) is proposed. The k-nearest neighbor(KNN) is introduced to highlight the local information of data and replace the nonlinear data structure. The nearest neighbor data points are selected by the Gaussian weighting method to obtain the optimal representation coefficients. Then, an overlapping probability model is employed to determine the overlap of the data in the subspace, and the clustering results are rechecked to improve the clustering accuracy. The experimental results on both artificial datasets and real-world datasets show that the proposed algorithm achieves better clustering results.

Feature selection is to select a subset of features from the original feature space to yield similar or better learning performance compared with the original feature set in the task of classification. In this paper, an information granulation based multi-label feature selection is firstly proposed. Then, the label weight and sample average margin is fused. Finally, the improved neighborhood information entropy is applied to multi-label feature selection. Experiments are conducted on six datasets and five evaluation metrics, and experimental results show that the proposed algorithm is effective.

The domain adaptability of link prediction method based on matrix factorization is fine. However, in the existing link prediction method based on matrix factorization, the network data representation of 0-1 matrix has a strong assumption of unknown edge in the network, while the importance of the known edges in the network is indistinguishable. The network data representation hypothesis of 0-1 matrix is relaxed in this paper, and no assumption to the edges of the unknown node-pairs is made. The measure method of importance of edges is put forward. Finally, the link prediction model based on the network weight matrix factorization is established by measuring the importance of the known edges in the network. The model is compared with the prediction algorithms based on metric and matrix factorization. Experimental results on eight public network datasets show the proposed algorithm is more effective.

Aiming at the detection of textile defects with complex periodic patterns, an unsupervised fabric defect detection method based on modified Markov random field model is proposed. The defects of periodic textile images are detected and the areas of defect are judged via the Markov neighborhood feature. The minimum image block computing unit of Markov random field is determined by combining the segmentation of periodic image, and the computational complexity of the algorithm is reduced. In the definition of the random field potential function, the difference of adjacent image blocks is comprehensively taken into account. The location of defect area is judged by the global characteristics of the Markov random field. The concept of fuzzy similarity relation matrix is introduced to solve the parameters of the improved Markov random field model, and the local energy of all image blocks is optimized.Experiments show that the proposed defect detection method gains high recall.

Compared with the traditional single kernel, multiple kernel learning is more flexible and interpretable while dealing with heterogeneous, irregular and non-flat distribution samples. It is difficult for the existing accurate regularization path algorithms to solve large-scale problems. Therefore, an approximation algorithm for multiple kernel learning regularization path is proposed in this paper. The kernel matrix is randomly sampled according to the sampling distribution function, and the corresponding lines are extracted from the Lagrange multiplier vector. Then, the product of matrix is approximately computed, and the efficiency of the regularization path of multiple kernel learning is improved. Finally, the approximation error bounds and the computational complexity of the algorithm are analyzed. Experimental results on standard datasets verify the validity and efficiency of the proposed algorithm.