Search

Sparse coding (SC) exhibits impressive performance in many practical applications. However, in the unsupervised scenario, most of the conventional SC methods fail to fully take advantage of the structure of the data. Actually, the structure of the data, especially the global structure that is an implicit prior knowledge, is vital for data analysis. In this paper, we propose a novel method called structure regularized sparse coding (SRSC) for the sparse representation of the data in the unsupervised scenario. In contrast with the other SC methods, a distinct feature of SRSC is that it takes into consideration both the local and global structure of the data and fully exploits the latent category information in the data. By using the local affinity matrix that captures the local structure, we first build a global affinity matrix to encode the global structure of the data. The global affinity matrix fully carries the latent category information that is beneficial to obtain the discriminating representation of the data, Then, we define the optimization model of SRSC and develop a two-step iterative optimization algorithm based on the alternating direction method of multipliers (ADMM) to solve it. The experimental results validate that the proposed method is effective and can achieve better performance over its counterparts.

Nonnegative matrix factorization (NMF) has received intensive attention due to producing a parts-based representation of the data. However, because of the non-convexity of NMF models, these methods easily obtain a bad local solution. To alleviate this deficiency, this paper presents a novel NMF method by gradually including data points into NMF from easy to complex, namely self-paced learning (SPL), which is shown to be beneficial in avoiding a bad local solution. Furthermore, instead of using the conventional hard weighting scheme, we adopt the soft weighting strategy of SPL to further improve the performance of our model. An iterative updating algorithm is proposed to solve the optimization problem of our method. The convergence of the updating rules is also theoretically guaranteed. Experiments on both toy data and real-world benchmark datasets demonstrate the effectiveness of the proposed method.

Data streams have become a widely-adopted data representation format in many real-world applications. This data streaming may be published for different scientific research, mining, or analysis purposes. However, such streams may contain personal-specific data that could be considered as sensitive about individuals. This makes the privacy preserving of data streams against privacy disclosure attacks, while maintaining their utilization, is a real challenge. Some studies have considered privacy-preserving publishing over data streams with only Single Sensitive Attribute, in which they do not protect the published streams from all possible privacy attacks. In this paper, we propose a novel Restricted Sensitive Attributes-based Sequential Anonymization (RSA-SA) approach for privacy-preserving data stream publishing. Besides, two new privacy restrictions are introduced to restrict the published Sensitive Attributes values: Semantic-diversity and Sensitivity-diversity. RSA-SA can protect the sensitive values of the published data streams against the related privacy attacks, including the attribute disclosure, skewness, similarity, and sensitivity attacks. In addition, RSA-SA handles data streams that have either single or multiple sensitive attributes with minimum information loss and delay time. Thus, the data utility of the published data streams is efficiently maintained to provide more accurate mining and analytical results, where robust invulnerability to privacy attacks is sustained.

Due to the high data volume and non-stationarity of time series data, it is very difficult to directly use the original data for anomaly detection. In this study, a novel framework of anomaly detection is proposed, whose intent is to capture more detailed data of time series’ shape and morphology characteristics by data representation to carry out anomaly detection. First, high-order differences and intervals are employed to realize data representation, and then such rectangles and cubes are constructed with the results of data representation for similarity measurement and anomaly detection. Compared with existing state-of-the-art methods, based on the experimental studies completed on large amount of datasets, the methods proposed in this framework are effective in detecting anomalies caused by changes in shape and amplitude. Meanwhile, it can detect anomalies with higher accuracy and better performance of data anomaly resolution.

Noise label exists widely in real-world data, resulting in the degradation of classification performance. Popular methods require a known noise distribution or additional cleaning supervision, which is usually unavailable in practical scenarios. This paper presents a theoretical statistical method and designs a label confidence inference (LISR) algorithm to handle this issue. For data distribution, we define a statistical function for label inconsistency and analyze its relationship with neighbor radius. For data representation, we define trusted-neighbor, nearest-trusted-neighbor and untrusted-neighbor. For noisy label recognition, we present three inference methods to predict the labels and their confidence. The LISR algorithm establishes a practical statistical model, queries the initial trusted instances, iteratively searches for the trusted instances and corrects labels. We conducted experiments on synthetic, UCI and classic image datasets. The results of significance test verified the effectiveness of LISR and its superiority to the state-of-the-art noise label learning algorithms.

Recently, multiple graph regularizer based methods have shown promising performances in data representation. However, the parameter choice of the regularizer is crucial to the performance of clustering and its optimal value changes for different real datasets. To deal with this problem, we propose a novel method called Parameter-less Auto-weighted Multiple Graph regularized Nonnegative Matrix Factorization (PAMGNMF) in this paper. PAMGNMF employs the linear combination of multiple simple graphs to approximate the manifold structure of data as previous methods do. Moreover, the proposed method can automatically learn an optimal weight for each graph without introducing an additive parameter. Therefore, the proposed PAMGNMF method is easily applied to practical problems. Extensive experimental results on different real-world datasets have demonstrated that the proposed method achieves better performance than the state-of-the-art approaches.

Nonnegative Matrix Factorization (NMF) and Concept Factorization (CF) have been widely used for different purposes such as feature learning, dimensionality reduction and image clustering in data representation. However, CF is a variant of NMF, which is an unsupervised learning method without making use of the available label information to guide the clustering process. In this paper, we put forward a semi-supervised discriminative concept factorization (SDCF) method, which utilizes the limited label information of the data as a discriminative constraint. This constraint forces the representation of data points within the same class should be very close together or aligned on the same axis in the new representation. Furthermore, in order to utilize the local manifold regularization, we propose a novel semi-supervised graph-based discriminative concept factorization (GDCF) method, which incorporates the local manifold regularization and the label information of the data into the CF to improve the performance of CF. GDCF not only encodes the local geometrical structure of the data space by constructing K-nearest graph, but also takes into account the available label information. Thus, the discriminative abilities of data representations are enhanced in the clustering tasks. Experimental results on several databases expose the strength of our proposed SDCF and GDCF methods compared to the state-of-the-art methods.

Matrix factorization based techniques, such as nonnegative matrix factorization (NMF) and concept factorization (CF), have attracted a great deal of attentions in recent years, mainly due to their ability of dimension reduction and sparse data representation. Both techniques are of unsupervised nature and thus do not make use of a priori knowledge to guide the clustering process. This could lead to inferior performance in some scenarios. As a remedy to this, a semi-supervised learning method called Pairwise Constrained Concept Factorization (PCCF) was introduced to incorporate some pairwise constraints into the CF framework. Despite its improved performance, PCCF uses only a priori knowledge and neglects the proximity information of the whole data distribution; this could lead to rather poor performance (although slightly improved comparing to CF) when only limited a priori information is available. To address this issue, we propose in this paper a novel method called Constrained Neighborhood Preserving Concept Factorization (CNPCF). CNPCF utilizes both a priori knowledge and local geometric structure of the dataset to guide its clustering. Experimental studies on three real-world clustering tasks demonstrate that our method yields a better data representation and achieves much improved clustering performance in terms of accuracy and mutual information comparing to the state-of-the-arts techniques.

Nonnegative matrix factorization (NMF) has been a vital data representation technique, and has demonstrated significant potential in the field of machine learning and data mining. Nonnegative matrix tri-factorization (NMTF) is an extension of NMF, and provides more degrees of freedom than NMF. In this paper, we propose the correntropy based orthogonal nonnegative matrix tri-factorization (CNMTF) algorithm, which is robust to noisy data contaminated by non-Gaussian noise and outliers. Different from previous NMF algorithms, CNMTF firstly applies correntropy to NMTF to measure the similarity, and preserves double orthogonality conditions and dual graph regularization. We adopt the half-quadratic technique to solve the optimization problem of CNMTF, and derive the multiplicative update rules. The complexity issue of CNMTF is also presented. Furthermore, the robustness of the proposed algorithm is analyzed, and the relationships between CNMTF and several previous NMF based methods are discussed. Experimental results demonstrate that the proposed CNMTF method has better performance on real world image and text datasets for clustering tasks, compared with several state-of-the-art methods. Nanyang Technological University This work was partially supported by Nanyang Technological University Research Scholarships of Singapore, National Basic Research Program (973 Program) of China (No. 2015CB351703), and National Natural Science Foundation of China (No. 91648208).

Pattern-based change detection (PBCD) describes a class of change detection algorithms for evolving data. Contrary to conventional solutions, PBCD seeks changes exhibited by the patterns over time and therefore works on an abstract form of the data, which prevents the search for changes on the raw data. Moreover, PBCD provides arguments on the validity of the results because patterns mirror changes occurred with any form of evidence. However, the existing solutions differ on data representation, pattern mining algorithm and change identification strategy, which we can deem as main modules of a general architecture, so that any PBCD task could be designed by accommodating custom implementations for those modules. This is what we propose in this paper through jKarma, a highly-modular framework written in Java for defining and performing PBCD.

Most healthcare data are available in XML format, which mainly focuses on the structure level and lacks support for data representation. Therefore, a variety of medical applications and medical semantic search engines have difficulty understanding and integrating healthcare data in a highly heterogeneous environment. OWL (Web Ontology Language) and Semantic Web technologies provide an infrastructure that can solve these problems. The aim of our study is to present a mechanism to ease the interpretation and automate the semantic transformation of XML healthcare data into the OWL ontology (S-Trans), which allows an easier and better semantic communication among hospital information systems. On the basis of the XML schemas (XSD or DTD), we extract the document structure and add more descriptions for XML elements. Moreover, to classify the semantic level of duplicate elements in an XML schema, we propose novel metrics to measure the similarity between them. Experimental results show that the proposed method reliably predicts semantic similarity of duplicates and produces a better-quality OWL ontology.