Xianchao Zhang

Multi-view clustering via multi-manifold regularized non-negative matrix factorization

Non-negative matrix factorization based multi-view clustering algorithms have shown their competitiveness among different multi-view clustering algorithms. However, non-negative matrix factorization fails to preserve the locally geometrical structure of the data space. In this paper, we propose a multi-manifold regularized non-negative matrix factorization framework (MMNMF) which can preserve the locally geometrical structure of the manifolds for multi-view clustering. MMNMF incorporates consensus manifold and consensus coefficient matrix with multi-manifold regularization to preserve the locally geometrical structure of the multi-view data space. We use two methods to construct the consensus manifold and two methods to find the consensus coefficient matrix, which leads to four instances of the framework. Experimental results show that the proposed algorithms outperform existing non-negative matrix factorization based algorithms for multi-view clustering.

Spectroscopic and TEM studies on poly vinyl carbazole/terbium complex and fabrication of organic electroluminescent device

Abstract Terbium acetyl salicylate complex containing 1,10-phenanthroline was synthesized. Element analysis showed that the composition of the complex was Tb(C 9 O 4 H 7 ) 3 (C 12 N 2 H 8 ). A single layer electroluminescent device using polyvinyl carbazole doped with the terbium complex as an emitting layer was fabricated. The device showed bright green EL emission. The dispersion of the terbium complex in PVK matrix was investigated using transmission electron microscopy.

Supervised ranking framework for relationship prediction in heterogeneous information networks

In recent years, relationship prediction in heterogeneous information networks (HINs) has become an active topic. The most essential part of this task is how to effectively represent and utilize the important three kinds of information hidden in connections of the network, namely local structure information (Local-info), global structure information (Global-info) and attribute information (Attr-info). Although all the information indicates different features of the network and influence relationship creation in a complementary way, existing approaches utilize them separately or in a partially combined way. In this article, a novel framework named Supervised Ranking framework (S-Rank) is proposed to tackle this issue. To avoid the class imbalance problem, in S-Rank framework we treat the relationship prediction problem as a ranking task and divide it into three phases. Firstly, a Supervised PageRank strategy (SPR) is proposed to rank the candidate nodes according to Global-info and Attr-info. Secondly, a Meta Path-based Ranking method (MPR) utilizing Local-info is proposed to rank the candidate nodes based on their meta path-based features. Finally, the two ranking scores are linearly integrated into the final ranking result which combines all the Attr-info, Global-info and Local-info together. Experiments on DBLP data demonstrate that the proposed S-Rank framework can effectively take advantage of all the three kinds of information for relationship prediction over HINs and outperforms other well-known baseline approaches.

Constrained Clustering With Nonnegative Matrix Factorization

Nonnegative matrix factorization (NMF) and symmetric NMF (SymNMF) have been shown to be effective for clustering linearly separable data and nonlinearly separable data, respectively. Nevertheless, many practical applications demand constrained algorithms in which a small number of constraints in the form of must-link and cannot-link are available. In this paper, we propose an NMF-based constrained clustering framework in which the similarity between two points on a must-link is enforced to approximate 1 and the similarity between two points on a cannot-link is enforced to approximate 0. We then formulate the framework using NMF and SymNMF to deal with clustering of linearly separable data and nonlinearly separable data, respectively. Furthermore, we present multiplicative update rules to solve them and show the correctness and convergence. Experimental results on various text data sets, University of California, Irvine (UCI) data sets, and gene expression data sets demonstrate the superiority of our algorithms over existing constrained clustering algorithms.

Multi-type Co-clustering of General Heterogeneous Information Networks via Nonnegative Matrix Tri-Factorization

Many kinds of real world data can be modeled by a heterogeneous information network (HIN) which consists of multiple types of objects. Clustering plays an important role in mining knowledge from HIN. Several HIN clustering algorithms have been proposed in recent years. However, these algorithms suffer from one or moreof the following problems: (1) inability to model general HINs, (2) inability to simultaneously generate clusters for all types of objects, (3) inability to use similarity information of the objects with the same type. In this paper, we propose a powerful HIN clustering algorithm which can handle general HINs, simultaneously generate clusters for all types of objects, and use the similarity information of the same type of objects. First, we transform a general HIN into a meta-path-encoded relationship set. Second, we propose a nonnegative matrix tri-factorization multi-type co-clustering method, HMFClus, to cluster all types of objects in HIN simultaneously. Third, we integrate the information between the objects with the same type into HMFClus by using a similarity regularization. Extensive experiments on real world datasets show that the proposed algorithm outperforms the state-of-the-art methods.

Local linear neighbor reconstruction for multi-view data

We propose to learn a unified similarity matrix by using multi-view information.The unified similarity matrix interacts with multi-view original local data.The learned similarity matrix enhances both spectral clustering and label propagation.Results on several datasets show the effectiveness of the learned similarity matrix. Graph based multi-view data analysis has become a hot topic in the past decade, and multi-view similarity matrix is fundamental for such tasks. Existing multi-view similarity matrix construction methods cannot learn local geometrical information in the original data space from multiple views simultaneously. Considering the fact that an appropriate similarity matrix is block-wise with intra-class similarity, it is more reasonable to learn a similarity matrix by using local geometrical information in multiple original data space. In this paper, we propose to construct a unified similarity matrix by using local linear neighbors in multiple views. In each view, the similarity matrix can be reconstructed with the weights of the neighbors of each data point in the original space. In multiple views, we seek for a unified similarity matrix which consists of the similarity matrix in each view. The unified similarity matrix can be used for spectral clustering, label propagation and other graph based learning algorithms. Experimental results show that spectral clustering and label propagation algorithms using the unified similarity matrix outperform those using other multi-view similarity matrices, they also outperform typical multi-view spectral clustering algorithms and typical multi-view label propagation algorithms.

Multi-view clustering via graph regularized symmetric nonnegative matrix factorization

Multi-view clustering has become a hot topic since the past decade and nonnegative matrix factorization (NMF) based multi-view clustering algorithms have shown their superiorities. Nevertheless, two drawbacks prevent NMF based multi-view algorithms from being a better algorithm: (1) The solution of NMF based multi-view algorithms is not unique. (2) Standard orthogonal basis matrix is not obtained for each view. Orthogonality is utilized to settle these above problems in our framework and high computational complexity caused by orthogonality is avoided. Moreover, to preserve the locally geometrical structure between views, graph regularization is utilized. Finally, we offer an update rule for the parameter of the graph regularization to balance the reconstruct error and regularization and make the objective function converge faster. Experimental results and theoretical proof show the validity and efficiency of our algorithm.

Dynamically Maintaining Duplicate-Insensitive and Time-Decayed Sum Using Time-Decaying Bloom Filter

The duplicate-insensitive and time-decayed sum of an arbitrary subset in a stream is an important aggregation for various analyses in many distributed stream scenarios. In general, precisely providing this sum in an unbounded and high-rate stream is infeasible. Therefore, we target at this problem and introduce a sketch, namely, time-decaying Bloom Filter (TDBF). The TDBF can detect duplicates in a stream and meanwhile dynamically maintain decayed-weight of all distinct elements in the stream according to a user-specified decay function. For a query for the current decayed sum of a subset in the stream, TDBF provides an effective estimation. In our theoretical analysis, a provably approximate guarantee has been given for the error of the estimation. In addition, the experimental results on synthetic stream validate our theoretical analysis.

Possible world based consistency learning model for clustering and classifying uncertain data

Possible world has shown to be effective for handling various types of data uncertainty in uncertain data management. However, few uncertain data clustering and classification algorithms are proposed based on possible world. Moreover, existing possible world based algorithms suffer from the following issues: (1) they deal with each possible world independently and ignore the consistency principle across different possible worlds; (2) they require the extra post-processing procedure to obtain the final result, which causes that the effectiveness highly relies on the post-processing method and the efficiency is also not very good. In this paper, we propose a novel possible world based consistency learning model for uncertain data, which can be extended both for clustering and classifying uncertain data. This model utilizes the consistency principle to learn a consensus affinity matrix for uncertain data, which can make full use of the information across different possible worlds and then improve the clustering and classification performance. Meanwhile, this model imposes a new rank constraint on the Laplacian matrix of the consensus affinity matrix, thereby ensuring that the number of connected components in the consensus affinity matrix is exactly equal to the number of classes. This also means that the clustering and classification results can be directly obtained without any post-processing procedure. Furthermore, for the clustering and classification tasks, we respectively derive the efficient optimization methods to solve the proposed model. Experimental results on real benchmark datasets and real world uncertain datasets show that the proposed model outperforms the state-of-the-art uncertain data clustering and classification algorithms in effectiveness and performs competitively in efficiency.

Multi-view clustering on unmapped data via constrained non-negative matrix factorization

Existing multi-view clustering algorithms require that the data is completely or partially mapped between each pair of views. However, this requirement could not be satisfied in many practical settings. In this paper, we tackle the problem of multi-view clustering on unmapped data in the framework of NMF based clustering. With the help of inter-view constraints, we define the disagreement between each pair of views by the fact that the indicator vectors of two samples from two different views should be similar if they belong to the same cluster and dissimilar otherwise. The overall objective of our algorithm is to minimize the loss function of NMF in each view as well as the disagreement between each pair of views. Furthermore, we provide an active inter-view constraints selection strategy which tries to query the relationships between samples that are the most influential and samples that are the farthest from the existing constraint set. Experimental results show that, with a small number of (either randomly selected or actively selected) constraints, the proposed algorithm performs well on unmapped data, and outperforms the baseline algorithms on partially mapped data and completely mapped data.