Zhengtao Yu

Low-rank matrix factorization with multiple Hypergraph regularizer

This paper presents a novel low-rank matrix factorization method, named MultiHMMF, which incorporates multiple Hypergraph manifold regularization to the low-rank matrix factorization. In order to effectively exploit high order information among the data samples, the Hypergraph is introduced to model the local structure of the intrinsic manifold. Specifically, multiple Hypergraph regularization terms are separately constructed to consider the local invariance; the optimal intrinsic manifold is constructed by linearly combining multiple Hypergraph manifolds. Then, the regularization term is incorporated into a truncated singular value decomposition framework resulting in a unified objective function so that matrix factorization is changed into an optimization problem. Alternating optimization is used to solve the optimization problem, with the result that the low dimensional representation of data space is obtained. The experimental results of image clustering demonstrate that the proposed method outperforms state-of-the-art data representation methods. We propose a novel low rank matrix factorization method.We incorporate multiple Hypergraph manifold regularization to the matrix factorization.We adopt alternating optimization to solve the optimization problem.

Fractional differential and variational method for image fusion and super-resolution

This paper introduces a novel fractional differential and variational model that includes the terms of fusion and super-resolution, edge enhancement and noise suppression. In image fusion and super-resolution term, the structure tensor is employed to describe the geometry of all the input images. According to the fact that the fused image and the source inputs should have the same or similar structure tensor, the energy functional of the image fusion and super-resolution is established combining with the down-sampling operator. For edge enhancement, the bidirectional diffusion term is incorporated into the image fusion and super-resolution model to enhance the visualization of the fused image. In the noise suppression term, a new variational model is developed based on the fractional differential and fractional total variation. Thanks to the above three terms, the proposed model can realize the image fusion, super-resolution, and the edge information enhancement simultaneously. To search for the optimal solution, a gradient descent iteration scheme derived from the Euler-Lagrange equation of the proposed model is employed. The numerical results indicate that the proposed method is feasible and effective.

Ontology representation and mapping of common fuzzy knowledge

Abstract Fuzzy knowledge is prevalent in real life, and ontology is the main way of knowledge representation in semantic web. In this paper, two main kinds of common fuzzy knowledge are sorted out, and there is a great significance for ontology representation of common fuzzy knowledge in semantic web. In addition, searching knowledge in ontology is the most common operation of semantic web, but heterogeneous ontologies seriously affect accuracy of information retrieval, and ontology mapping is the key to solve the problem. Therefore, in this paper, firstly an ontology representation method of common fuzzy knowledge is presented. Common fuzzy knowledge is polytypic and includes most commonly used fuzzy knowledge in reality. So fuzzy sets and Cloud Model are used to reflect and represent these 2 types of common fuzzy knowledge in ontology. Then, based on the ontology representation method of common fuzzy knowledge mentioned above, a corresponding algorithm of ontology mapping is presented, which is based on similarity calculation of concepts and Support Vector Machine. The proposed methods extend the scope of ontology application and are significant for information retrieval of fuzzy knowledge in the semantic web. The experiments show that the proposed methods are practicable and effective.

Locality Preserving Collaborative Representation for Face Recognition

Face recognition has many applications in pattern recognition and computer vision, and many face recognition methods have been proposed. Among them, the recently proposed collaborative representation based face recognition has attracted the attention of researchers. Many variants and extensions of collaborative representation based classification (CRC) have been presented. However, most of CRC methods do not consider data locality, which is crucial for classification task. In this article, a novel collaborative representation based face recognition method, LP-CRC, is proposed, which balances data locality and collaborative representation. The proposed method incorporates a locality adaptor term into the robust collaborative representation based classification framework, leading to a novel unified objective function. The Augmented Lagrange Multiplier is used to optimize the objective function. Tests on standard benchmarks demonstrate that the proposed face recognition method is superior to existing methods and robust to noise and outliers.

Reuse of Chinese Domain Ontology for the restricted domain Question Answering System

Expressing knowledge by ontology is favorable for reuse and reasoning of knowledge. The structure of domain ontology for question answering system (QA) was formally defined, and the relations between ontology elements and semantic analysis of question and answer extraction were illustrated. The method of reusing Chinese ontology was proposed in the same domain, and the method can construct new initial ontology after extracting and classifying new instances based on the existing ontology using support vector machine (SVM) and reuse the other elements of the existing ontology. By the experiment of ontology reuse in certain medical domain, the average F-value of extraction and classification of instances reached 82.8%, and it verified validity of the method. The proposed method has significance for fast constructing Chinese ontology for QA in same domain.

Combining paper cooperative network and topic model for expert topic analysis and extraction

Abstract Paper cooperation network embodies expert topic similarity in an extent, thus, a novel method is proposed for expert topic analysis and extraction by combining paper cooperation network and topic model. In the method, we extract each paper’ author information and construct an expert cooperation network. At the same time, by means of LDA model, a probabilistic topic model is also built to analyze papers’ latent topics. Then, by making full use of the feature that adjacent nodes in the expert cooperation network share similar themes distribution, we makes a constraint on expert topic distribution in Gibbs sampling process of solving the probabilistic topic model. Experimental results on NIPS dataset show that the proposed method can effectively extract expert topics, and the expert paper cooperation network plays a very good supporting role on the extracting task.

Expert list-wise ranking method based on sparse learning

Expert ranking is the core issue of expert retrieval. Taking into consideration the complexity of feature redundancy in traditional dense listwise Learning to Rank method and local optimum in parameter learning, the article proposed the expert listwise Learning to Rank method based on sparse learning. The objective function was defined through the optimization process of experts listwise ranking performance index. Then the Learning to Rank loss function was solved by the objective function. Thus feature dimension reduction was achieved by the feature threshold from the loss-control function of sparse learning algorithm and the steps above. In order to verify whether the feature threshold is optimal, the article made cross validation with the feature threshold and the objective function of model parameter vector to get the optimal model parameters vector and to verify the feature threshold. Meanwhile the article realized expert ranking via the expert listwise ranking model based on sparse learning, which depends on feature dimension reduction and parameter tuning. At last, the contrast experiments of expert ranking proved the effectiveness of the proposed method, which supported expert listwise ranking strongly. This thesis represents an expert listwise Learning to Rank method based on sparse learning.The article realized expert ranking via the expert listwise ranking model based on sparse learning.The contrast experiments of expert ranking proved the effectiveness of the proposed method.

News Events Elements Extraction Based on Undirected Graph

News event elements extraction is a main task of information extraction. For news event correlation between sub-events, this paper proposes a kind of undirected graph model of news event element extraction merging associations of event elements. Firstly, splitting the news to multiple sub-event and extracting event elements. Then, the correlation between event elements and news events was analyzed, a undirected graph by extracting the correlation based on news event elements as node was established, and we transferred news event element extraction into a weighted undirected graph node calculation problem. At last, We conducted event elements extraction experiments. And comparing the experimental results show that the proposed method has good effect, correlation of sub-event can effectively improve the effect of extracting elements of news events.

Extraction of Expert Relations Integrated with Expert Topic and Associated Relationship Features

In order to utilize the topic features and the associated relationship features of experts to identify expert relations effectively, a novel extraction method of expert relations is proposed with the integration of expert topic and associated relationship features in this article. Firstly, the expert topics are extracted according to the idea that cooperative experts share the same topic distribution by integrating the expert cooperation network with Probabilistic Topic Models based on LDA Model. Secondly, the associated relationship features are extracted with the utilization of the attributes characteristics, such as the links among homepages, the mutual following on the Blogs. Finally, Markov Network is used to construct the extraction model of expert relations by integrating expert topics and associated relationship features. The experimental results have demonstrated that the proposed method that integrated with expert topic and the associated relationship features of experts supports the extraction of expert relations and shows promising performance.

Approaches to Detect Micro-Blog User Interest Communities Through the Integration of Explicit User Relationship and Implicit Topic Relations

In order to utilize effectively explicit user relationship and implicit topic relations for the detection of micro-blog user interest communities, a micro-blog user interest community detection approach is proposed. First, we analyze the follow relationship between the users to construct the user follow-ship network. Second, we construct the user interest feature vectors based on the concept of feature mapping to build a user-tag based interest relationship network. Third, we propose to build a guided user interest topic model and construct a topic-based interest relationship network. Finally, we integrate the above-mentioned three kinds of relationship network to construct a micro-blog user interest relationship network. Meanwhile, we propose a micro-blog user interest community detection algorithm based on the contribution of the neighboring nodes. The experiment result turns out that good effect has been achieved through our approach.