Domain structure-based transfer learning for cross-domain word representation

Cross-domain word representation aims to learn high-quality semantic representations in an under-resourced domain by leveraging information in a resourceful domain. However, most existing methods mainly transfer the semantics of common words across domains, ignoring the semantic relations among domain-specific words. In this paper, we propose a domain structure-based transfer learning method to learn cross-domain representations by leveraging the relations among domain-specific words. To accomplish this, we first construct a semantic graph to capture the latent domain structure using domain-specific co-occurrence information. Then, in the domain adaptation process, beyond domain alignment, we employ Laplacian Eigenmaps to ensure the domain structure is consistently distributed in the learned embedding space. As such, the learned cross-domain word representations not only capture shared semantics across domains, but also maintain the latent domain structure. We performed extensive experiments on two tasks, namely sentiment analysis and query expansion. The experiment results show the effectiveness of our method for tasks in under-resourced domains.     

Weakly-Supervised Cross-Domain Dictionary Learning for Visual Recognition

We address the visual categorization problem and present a method that utilizes weakly labeled data from other visual domains as the auxiliary source data for enhancing the original learning system. The proposed method aims to expand the intra-class diversity of original training data through the collaboration with the source data. In order to bring the original target domain data and the auxiliary source domain data into the same feature space, we introduce a weakly-supervised cross-domain dictionary learning method, which learns a reconstructive, discriminative and domain-adaptive dictionary pair and the corresponding classifier parameters without using any prior information. Such a method operates at a high level, and it can be applied to different cross-domain applications. To build up the auxiliary domain data, we manually collect images from Web pages, and select human actions of specific categories from a different dataset. The proposed method is evaluated for human action recognition, image classification and event recognition tasks on the UCF YouTube dataset, the Caltech101/256 datasets and the Kodak dataset, respectively, achieving outstanding results.     

Graph regularized discriminative non-negative matrix factorization for face recognition

Non-negative matrix factorization (NMF) has been widely employed in computer vision and pattern recognition fields since the learned bases can be interpreted as a natural parts-based representation of the input space, which is consistent with the psychological intuition of combining parts to form a whole. In this paper, we propose a novel constrained nonnegative matrix factorization algorithm, called the graph regularized discriminative non-negative matrix factorization (GDNMF), to incorporate into the NMF model both intrinsic geometrical structure and discriminative information which have been essentially ignored in prior works. Specifically, both the graph Laplacian and supervised label information are jointly utilized to learn the projection matrix in the new model. Further we provide the corresponding multiplicative update solutions for the optimization framework, together with the convergence proof. A series of experiments are conducted over several benchmark face datasets to demonstrate the efficacy of our proposed GDNMF.     

Cross-view action recognition by cross-domain learning

This paper proposes a novel cross-view human action recognition method by discovering and sharing common knowledge among different video sets captured in multiple viewpoints. We treat a specific view as target domain and the others as source domains and consequently formulate the cross-view action recognition into the cross-domain learning framework. First, the classic bag-of-visual word framework is implemented for visual feature extraction in individual viewpoints. Then, we add two transformation matrices in order to transform original action feature from different views into one common feature space, and also combine the original feature and the transformation feature to proposed the new feature mapping function for target and auxiliary domains respectively. Finally, we proposed a new method to learn the two transformation matrices in model training step based on the standard SVM solver and generate the final classifier for each human action. Extensive experiments are implemented on IXMAS, and TJU. The experimental results demonstrate that the proposed method can consistently outperform the state-of-the-arts.     

跨领域文本情感分类研究进展

作为社会媒体文本情感分析的重要研究课题之一，跨领域文本情感分类旨在利用源领域资源或模型迁移地服务于目标领域的文本情感分类任务，其可以有效缓解目标领域中带标签数据不足问题.本文从三个角度对跨领域文本情感分类方法行了归纳总结：（1）按照目标领域中是否有带标签数据，可分为直推式和归纳式情感迁移方法；（2）按照不同情感适应性策略，可分为实例迁移方法、特征迁移方法、模型迁移方法、基于词典的方法、联合情感主题方法以及图模型方法等；（3）按照可用源领域个数，可分为单源和多源跨领域文本情感分类方法.此外，论文还介绍了深度迁移学习方法及其在跨领域文本情感分类的最新应用成果.最后，论文围绕跨领域文本情感分类面临的关键技术问题，对可能的突破方向进行了展望.     

A sparse neighborhood preserving non-negative tensor factorization algorithm for facial expression recognition

In this paper, a novel sparse neighborhood preserving non-negative tensor factorization (SNPNTF) algorithm is proposed for facial expression recognition. It is derived from non-negative tensor factorization (NTF), and it works in the rank-one tensor space. A sparse constraint is adopted into the objective function, which takes the optimization step in the direction of the negative gradient, and then projects onto the sparse constrained space. To consider the spatial neighborhood structure and the class-based discriminant information, a neighborhood preserving constraint is adopted based on the manifold learning and graph preserving theory. The Laplacian graph which encodes the spatial information in the face samples and the penalty graph which considers the pre-defined class information are considered in this constraint. By using it, the obtained parts-based representations of SNPNTF vary smoothly along the geodesics of the data manifold and they are more discriminant for recognition. SNPNTF is a quadratic convex function in the tensor space, and it could converge to the optimal solution. The gradient descent method is used for the optimization of SNPNTF to ensure the convergence property. Experiments are conducted on the JAFFE database, the Cohn–Kanade database and the AR database. The results demonstrate that SNPNTF provides effective facial representations and achieves better recognition performance, compared with non-negative matrix factorization, NTF and some variant algorithms. Also, the convergence property of SNPNTF is well guaranteed.     

基于迁移学习的无监督跨域人脸表情识别

本文主要研究了基于迁移学习的无监督跨域人脸表情识别.在过去的几年里,提出的许多方法在人脸表情识别方面取得了令人满意的识别效果.但这些方法通常认为训练和测试数据来自同一个数据集,因此其具有相同的分布.而在实际应用中,这一假设通常并不成立,特别当训练集和测试集来自不同的数据集时,即跨域人脸表情识别问题.为了解决这一问题,本...     

Non-negative and sparse spectral clustering

Spectral clustering aims to partition a data set into several groups by using the Laplacian of the graph such that data points in the same group are similar while data points in different groups are dissimilar to each other. Spectral clustering is very simple to implement and has many advantages over the traditional clustering algorithms such as k-means. Non-negative matrix factorization (NMF) factorizes a non-negative data matrix into a product of two non-negative (lower rank) matrices so as to achieve dimension reduction and part-based data representation. In this work, we proved that the spectral clustering under some conditions is equivalent to NMF. Unlike the previous work, we formulate the spectral clustering as a factorization of data matrix (or scaled data matrix) rather than the symmetrical factorization of the symmetrical pairwise similarity matrix as the previous study did. Under the NMF framework, where regularization can be easily incorporated into the spectral clustering, we propose several non-negative and sparse spectral clustering algorithms. Empirical studies on real world data show much better clustering accuracy of the proposed algorithms than some state-of-the-art methods such as ratio cut and normalized cut spectral clustering and non-negative Laplacian embedding.     

Semi-supervised transfer discriminant analysis based on cross-domain mean constraint

In this paper, a novel semi-supervised feature extraction algorithm, i.e., semi-supervised transfer discriminant analysis (STDA) with knowledge transfer capability is proposed, based on the traditional algorithm that cannot get adapted in the change of the learning environment. By using both the pseudo label information from target domain samples and the actual label information from source domain samples in the label iterative refinement process, not only the between-class scatter is maximized while that within-class scatter is minimized, but also the original space structure is maintained via Laplacian matrix, and the distribution difference is reduced by using maximum mean discrepancy as well. Moreover, semi-supervised transfer discriminant analysis based on cross-domain mean constraint (STDA-CMC) is proposed. In this algorithm, the cross-domain mean constraint term is incorporated into STDA, such that knowledge transfer between domains is facilitated by making source and target samples after being projected are located more closely in the low-dimensional feature subspace. The proposed algorithm is proved efficient and feasible from experiments on several datasets.     

基于多对多生成对抗网络的非对称跨域迁移行人再识别

无监督跨域迁移学习是行人再识别中一个非常重要的任务. 给定一个有标注的源域和一个没有标注的目标域, 无监督跨域迁移的关键点在于尽可能地把源域的知识迁移到目标域. 然而, 目前的跨域迁移方法忽略了域内各视角分布的差异性, 导致迁移效果不好. 针对这个缺陷, 本文提出了一个基于多视角的非对称跨域迁移学习的新问题. 为了实现这种非对称跨域迁移, 提出了一种基于多对多生成对抗网络(Many-to-many generative adversarial network, M2M-GAN)的迁移方法. 该方法嵌入了指定的源域视角标记和目标域视角标记作为引导信息, 并增加了视角分类器用于鉴别不同的视角分布, 从而使模型能自动针对不同的源域视角和目标域视角组合采取不同的迁移方式. 在行人再识别基准数据集Market1501、DukeMTMC-reID和MSMT17上, 实验验证了本文的方法能有效提升迁移效果, 达到更高的无监督跨域行人再识别准确率.     

自适应图正则化的低秩非负矩阵分解算法

图正则化(nonnegative matrix factorization,NMF)算法(graph regularization nonnegative matrix factorization,GNMF)仍存在一些不足之处:GNMF算法并没有考虑数据的低秩结构;在GNMF算法中,其拉普拉斯图是使用K近邻(K nea...     

Robust non-negative matrix factorization via joint sparse and graph regularization for transfer learning

In real-world applications, we often have to deal with some high-dimensional, sparse, noisy, and non-independent identically distributed data. In this paper, we aim to handle this kind of complex data in a transfer learning framework, and propose a robust non-negative matrix factorization via joint sparse and graph regularization model for transfer learning. First, we employ robust non-negative matrix factorization via sparse regularization model (RSNMF) to handle source domain data and then learn a meaningful matrix, which contains much common information between source domain and target domain data. Second, we treat this learned matrix as a bridge and transfer it to target domain. Target domain data are reconstructed by our robust non-negative matrix factorization via joint sparse and graph regularization model (RSGNMF). Third, we employ feature selection technique on new sparse represented target data. Fourth, we provide novel efficient iterative algorithms for RSNMF model and RSGNMF model and also give rigorous convergence and correctness analysis separately. Finally, experimental results on both text and image data sets demonstrate that our REGTL model outperforms existing start-of-art methods.     

深度融合辅助信息的跨域推荐算法

协同过滤已成功用于为用户提供个性化的产品和服务,然而它面临数据稀疏和冷启动的问题。一种解决方案是结合辅助信息,另一种是从相关领域学习知识。综合考虑了这两个方面,提出一种深度融合辅助信息的跨域推荐算法CICDR,它集成了集体矩阵分解和深度迁移学习。该算法通过Semi-SDAE和矩阵分解（MF）在源域和目标域中进行建模,学习评分信息和辅助信息中的有效特征向量,并利用用户的隐式反馈信息来做出更准确的推荐。通过这种方式,在两个领域中学习到的用户和项目潜在因素为推荐保留了更多的语义信息。通过非完备正交非负矩阵三分解（IONMTF）产生桥接两个相关领域的公共潜在因素,以缓解目标域中的冷启动和数据稀疏问题。在三个真实数据集上与四种经典算法进行对比,验证了提出算法的有效性,进一步提高了推荐精度和用户满意度。     

基于集成一致性的多源跨领域情感分类模型

现有的跨领域情感分类方法大多只利用了单个源领域到目标域的迁移特征,没有充分考虑目标域实例与不同源域之间的联系。针对此问题,本文提出一种无监督的多源跨领域情感分类模型。首先利用单个源域到目标域的迁移特征训练基分类器,并对不同的基分类器加权;然后将不同基分类器对目标域实例预测的集成一致性作为目标函数,优化该目标函数,得到不同基分类器的权重;最后利用加权后的基分类器得到目标域的情感分类结果。该模型在Amazon数据集上进行了多源域情感迁移实验,取得了较好的实验结果,相对其他基线模型,在4组实验中平均提升了0.75%。     

一种基于局部流形结构的无监督特征学习方法(英文)

Unsupervised feature selection is fundamental in statistical pattern recognition, and has drawn persistent attention in the past several decades. Recently, much work has shown that feature selection can be formulated as nonlinear dimensionality reduction with discrete constraints. This line of research emphasizes utilizing the manifold learning techniques, where feature selection and learning can be studied based on the manifold assumption in data distribution. Many existing feature selection methods such as Laplacian score, SPEC(spectrum decomposition of graph Laplacian), TR(trace ratio) criterion, MSFS(multi-cluster feature selection) and EVSC(eigenvalue sensitive criterion) apply the basic properties of graph Laplacian, and select the optimal feature subsets which best preserve the manifold structure defined on the graph Laplacian. In this paper, we propose a new feature selection perspective from locally linear embedding(LLE), which is another popular manifold learning method. The main difficulty of using LLE for feature selection is that its optimization involves quadratic programming and eigenvalue decomposition, both of which are continuous procedures and different from discrete feature selection. We prove that the LLE objective can be decomposed with respect to data dimensionalities in the subset selection problem, which also facilitates constructing better coordinates from data using the principal component analysis(PCA) technique. Based on these results, we propose a novel unsupervised feature selection algorithm,called locally linear selection(LLS), to select a feature subset representing the underlying data manifold. The local relationship among samples is computed from the LLE formulation, which is then used to estimate the contribution of each individual feature to the underlying manifold structure. These contributions, represented as LLS scores, are ranked and selected as the candidate solution to feature selection. We further develop a locally linear rotation-selection(LLRS) algorithm which extends LLS to identify the optimal coordinate subset from a new space. Experimental results on real-world datasets show that our method can be more effective than Laplacian eigenmap based feature selection methods.     

图嵌入正则化投影非负矩阵分解人脸图像特征提取

目的 针对投影非负矩阵分解（PNMF）不能揭示数据空间的流形几何结构和判别信息的缺点,提出一种图嵌入正则化投影非负矩阵分解（GEPNMF）人脸图像特征提取方法。 方法 首先构建了描述数据空间的流形几何结构和类间分离度的两个近邻图,然后采用它们的拉普拉斯矩阵设计了一个图嵌入正则项,并将该图嵌入正则项与PNMF的目标函数融合以建立GEPNMF的目标函数。由于引入了图嵌入正则项,GEPNMF求得的子空间能在保持数据空间的流形几何结构的同时,类间间距最大。此外,在GEPNMF目标函数中引入了一个正交正则项,以确保GEPNMF子空间基向量具有数据局部表示能力。最后,对求解GEPNMF目标函数的累乘更新规则（MUR）进行了详细推导,并从理论上证明了其收敛性。结果 在ORL、Yale和CMU PIE人脸图像数据库上分别进行了人脸识别实验,识别率分别达到了94.00%、64.33%和98.58%。结论 实验结果表明,GEPNMF提取的人脸图像特征用于人脸识别时,具有较高的识别率。     

注意力迁移的联合平衡领域自适应

目的 现有的图像识别方法应用于从同一分布中提取的训练数据和测试数据时具有良好性能,但这些方法在实际场景中并不适用,从而导致识别精度降低。使用领域自适应方法是解决此类问题的有效途径,领域自适应方法旨在解决来自两个领域相关但分布不同的数据问题。方法 通过对数据分布的分析,提出一种基于注意力迁移的联合平衡自适应方法,将源域有标签数据中提取的图像特征迁移至无标签的目标域。首先,使用注意力迁移机制将有标签源域数据的空间类别信息迁移至无标签的目标域。通过定义卷积神经网络的注意力,使用关注信息来提高图像识别精度。其次,基于目标数据集引入网络参数的先验分布,并且赋予网络自动调整每个领域对齐层特征对齐的能力。最后,通过跨域偏差来描述特定领域的特征对齐层的输入分布,定量地表示每层学习到的领域适应性程度。结果 该方法在数据集Office-31上平均识别准确率为77.6%,在数据集Office-Caltech上平均识别准确率为90.7%,不仅大幅领先于传统手工特征方法,而且取得了与目前最优的方法相当的识别性能。结论 注意力迁移的联合平衡领域自适应方法不仅可以获得较高的识别精度,而且能够自动学习领域间特征的对齐程度,同时也验证了进行域间特征迁移可以提高网络优化效果这一结论。     

动态迁移实体块信息的跨领域中文实体识别模型

由于中文文本之间没有分隔符,难以识别中文命名实体的边界.此外,在垂直领域中难以获取充足的标记完整的语料,例如医疗领域和金融领域等垂直领域.为解决上述不足,提出一种动态迁移实体块信息的跨领域中文实体识别模型(TES-NER),将跨领域共享的实体块信息(entity span)通过基于门机制(gate mechanism)的动态融合层,从语料充足的通用领域(源领域)动态迁移到垂直领域(目标领域)上的中文命名实体模型,其中,实体块信息用于表示中文命名实体的范围.TES-NER模型首先通过双向长短期记忆神经网络(BiLSTM)和全连接网络(FCN)构建跨领域共享实体块识别模块,用于识别跨领域共享的实体块信息以确定中文命名实体的边界;然后,通过独立的基于字的双向长短期记忆神经网络和条件随机场(BiLSTM-CRF)构建中文命名实体识别模块,用于识别领域指定的中文命名实体;最后构建动态融合层,将实体块识别模块抽取得到的跨领域共享实体块信息通过门机制动态决定迁移到领域指定的命名实体识别模型上的量.设置通用领域(源领域)数据集为标记语料充足的新闻领域数据集(MSRA),垂直领域(目标领域)数据集为混合领域(OntoNotes 5.0)、金融领域(Resume)和医学领域(CCKS 2017)这3个数据集,其中,混合领域数据集(OntoNotes 5.0)是融合了6个不同垂直领域的数据集.实验结果表明,提出的模型在OntoNotes 5.0、Resume和CCKS 2017这3个垂直领域数据集上的F1值相比于双向长短期记忆和条件随机场模型(BiLSTM-CRF)分别高出2.18%、1.68%和0.99%.     

A study of graph spectra for comparing graphs and trees

The spectrum of a graph has been widely used in graph theory to characterise the properties of a graph and extract information from its structure. It has also been employed as a graph representation for pattern matching since it is invariant to the labelling of the graph. There are, however, a number of potential drawbacks in using the spectrum as a representation of a graph. Firstly, more than one graph may share the same spectrum. It is well known, for example, that very few trees can be uniquely specified by their spectrum. Secondly, the spectrum may change dramatically with a small change structure.There are a wide variety of graph matrix representations from which the spectrum can be extracted. Among these are the adjacency matrix, combinatorial Laplacian, normalised Laplacian and unsigned Laplacian. Spectra can also be derived from the heat kernel matrix and path length distribution matrix. The choice of matrix representation clearly has a large effect on the suitability of spectrum in a number of pattern recognition tasks.In this paper we investigate the performance of the spectra as a graph representation in a variety of situations. Firstly, we investigate the cospectrality of the various matrix representations over large graph and tree sets, extending the work of previous authors. We then show that the Euclidean distance between spectra tracks the edit distance between graphs over a wide range of edit costs, and we analyse the accuracy of this relationship. We then use the spectra to both cluster and classify the graphs and demonstrate the effect of the graph matrix formulation on error rates. These results are produced using both synthetic graphs and trees and graphs derived from shape and image data.     

Effective human action recognition by combining manifold regularization and pairwise constraints

The ever-growing popularity of mobile networks and electronics has prompted intensive research on multimedia data (e.g. text, image, video, audio, etc.) management. This leads to the researches of semi-supervised learning that can incorporate a small number of labeled and a large number of unlabeled data by exploiting the local structure of data distribution. Manifold regularization and pairwise constraints are representative semi-supervised learning methods. In this paper, we introduce a novel local structure preserving approach by considering both manifold regularization and pairwise constraints. Specifically, we construct a new graph Laplacian that takes advantage of pairwise constraints compared with the traditional Laplacian. The proposed graph Laplacian can better preserve the local geometry of data distribution and achieve the effective recognition. Upon this, we build the graph regularized classifiers including support vector machines and kernel least squares as special cases for action recognition. Experimental results on a multimodal human action database (CAS-YNU-MHAD) show that our proposed algorithms outperform the general algorithms.