Search

Convolutional neural networks (CNNs) have achieved unprecedented competitiveness in text and two-dimensional image data processing because of its good accuracy performance and high detection speed. Graph convolutional networks (GCNs), as an extension of classical CNNs in graph data processing, have attracted wide attention. At present, GCNs often use domain knowledge (such as citation recommendation system, biological cell networks) or artificial created fixed graph to achieve various semi-supervised classication tasks. Poor quality graph will lead to suboptimal results of semi-supervised classification tasks. We propose a more general GCN of reconstructed graph structure with constrained Laplacian rank. First, we use hypergraph to establish multivariate relationships between data. On the basis of the hypergraph, In virtue of Laplacian rank constraint to the graph matrix, we learn a new graph structure which has c connected components (where c is the number of classification), and then we construct an ideal graph matrix which is more suitable for the task of semi-supervised classification on GCNs. Finally, the data and the new graph are input GCNs model to get the results of classification. Experiments on 10 different datasets demonstrate that this method is more competitive than the comparison method. [ABSTRACT FROM AUTHOR]

Convolutional neural networks (CNNs) can learn the features of samples by supervised manner, and obtain outstanding achievements in many application fields. In order to improve the performance and generalization of CNNs, we propose a self-learning hybrid dilated convolution neural network (SPHDCNN), which can choose relatively reliable samples according to the current learning ability during training. In order to avoid the loss of useful feature map information caused by pooling, we introduce hybrid dilated convolution. In the proposed SPHDCNN, weight is applied to each sample to reflect the easiness of the sample. SPHDCNN employs easier samples for training first, and then adds more difficulty samples gradually according to the current learning ability. It gradually improves its performance by this learning mechanism. Experimental results show SPHDCNN has strong generalization ability, and it achieves more advanced performance compared to the baseline method. [ABSTRACT FROM AUTHOR]

Graph convolutional networks (GCNs), as an extension of classic convolutional neural networks (CNNs) in graph processing, have achieved good results in completing semi-supervised learning tasks. Traditional GCNs usually use fixed graph to complete various semi-supervised classification tasks, such as chemical molecules and social networks. Graph is an important basis for the classification of GCNs model, and its quality has a large impact on the performance of the model. For low-quality input graph, the classification results of the GCNs model are often not ideal. In order to improve the classification effect of GCNs model, we propose a graph learning method to generate high-quality topological graph, which is more suitable for GCNs model classification. We use the correlation between the data to generate a data similarity matrix, and apply Laplacian rank constraint to similarity matrix, so that the number of connected components of the topological graph is consistent with the number of categories of the original data. Experimental results on 10 real datasets show that our method is better than the comparison method in classification effect. [ABSTRACT FROM AUTHOR]

Many clustering methods may have poor performance when the data structure is complex (i.e., the data has an aspheric shape or non-linear relationship). Inspired by this view, we proposed a clustering model which combines kernel function and Locality Preserving Projections (LPP) together. Specifically, we map original data into the high-dimensional feature space according to the idea of kernel function. Secondly, it is feasible to explore the local structure of data in clustering tasks. LPP is used to preserve the original local structure information of data to improve the validity of the clustering model. Finally, some outliers are often included in real data, so we embedded sparse regularization items in the model to adjust feature weights and remove outliers. In addition, we design a simple iterative optimization method to solve the final objective function and show the convergence of the optimization method in the experimental part. The experimental analysis of ten public data sets showed that our proposed method has better efficiency and performance in clustering tasks than existing clustering methods. [ABSTRACT FROM AUTHOR]

In this paper, we propose a new nonlinear clustering method to preserve local structure of the features. Specifically, our method applies the gaussian kernel function to achieve high dimensional projection so as to make the original data linearly separable. Our method establishes the similarity matrix of data features in low-dimensional space to conduct local structure learning, as a result, it can avoid the divergence of sample sets and retain the original nearest neighbor structural relations. Furthermore, our method uses the sparse learning to remove the redundant features to make the model more robust in the process of learning. Experimental results on eight benchmark datasets show that our proposed method was superior to the state-of-the-art clustering methods in terms of clustering performance. [ABSTRACT FROM AUTHOR]

Text classification is an important task in machine learning. Specifically, deep neural network has been shown strong capability to improve performance in different fields, for example speech recognition, objects recognition and natural language processing. However, in most previous work, the extracted feature models do not achieve the relative text tasks well. To address this issue, we introduce a novel multi-task learning approach called a hybrid representation-learning network for text classification tasks. Our method consists of two network components: a bidirectional gated recurrent unit with attention network module and a convolutional neural network module. In particular, the attention module allows for the task learning private feature representation in local dependence from training texts and that the convolutional neural network module can learn the global representation on sharing. Experiments on 16 subsets of Amazon review data show that our method outperforms several baselines and also proves the effectiveness of joint learning multi-relative tasks. [ABSTRACT FROM AUTHOR]

Spectral clustering has been demonstrated to often outperform K-means clustering in real applications because it improves the similarity measurement of K-means clustering. However, previous spectral clustering method still suffers from the following issues: 1) easily being affected by outliers; 2) constructing the affinity matrix from original data which often contains redundant features and outliers; and 3) unable to automatically specify the cluster number. This paper focuses on address these issues by proposing a new clustering algorithm along with the technique of half-quadratic optimization. Specifically, the proposed method learns the affinity matrix from low-dimensional space of original data, which is obtained by using a robust estimator to remove the influence of outliers as well as a sparsity regularization to remove redundant features. Moreover, the proposed method employs the ℓ2,1-norm regularization to automatically learn the cluster number according to the data distribution. Experimental results on both synthetic and real data sets demonstrated that the proposed method outperforms the state-of-the-art methods in terms of clustering performance. [ABSTRACT FROM AUTHOR]

Feature selection is one of the most important machine learning procedure, and it has been successfully applied to make a preprocessing before using classification and clustering methods. High-dimensional features often appear in big data, and it’s characters block data processing. So spectral feature selection algorithms have been increasing attention by researchers. However, most feature selection methods, they consider these tasks as two steps, learn similarity matrix from original feature space (may be include redundancy for all features), and then conduct data clustering. Due to these limitations, they do not get good performance on classification and clustering tasks in big data processing applications. To address this problem, we propose an Unsupervised Feature Selection method with graph learning framework, which can reduce the redundancy features influence and utilize a low-rank constraint on the weight matrix simultaneously. More importantly, we design a new objective function to handle this problem. We evaluate our approach by six benchmark datasets. And all empirical classification results show that our new approach outperforms state-of-the-art feature selection approaches. [ABSTRACT FROM AUTHOR]