Relaxational metric adaptation and its application to semi-supervised clustering and content-based image retrieval

The performance of many supervised and unsupervised learning algorithms is very sensitive to the choice of an appropriate distance metric. Previous work in metric learning and adaptation has mostly been focused on classification tasks by making use of class label information. In standard clustering tasks, however, class label information is not available. In order to adapt the metric to improve the clustering results, some background knowledge or side information is needed. One useful type of side information is in the form of pairwise similarity or dissimilarity information. Recently, some novel methods (e.g., the parametric method proposed by Xing et al.) for learning global metrics based on pairwise side information have been shown to demonstrate promising results. In this paper, we propose a nonparametric method, called relaxational metric adaptation (RMA), for the same metric adaptation problem. While RMA is local in the sense that it allows locally adaptive metrics, it is also global because even patterns not in the vicinity can have long-range effects on the metric adaptation process. Experimental results for semi-supervised clustering based on both simulated and real-world data sets show that RMA outperforms Xing et al.'s method under most situations. Besides applying RMA to semi-supervised learning, we have also used it to improve the performance of content-based image retrieval systems through metric adaptation. Experimental results based on two real-world image databases show that RMA significantly outperforms other methods in improving the image retrieval performance.     

Non-linear metric learning using pairwise similarity and dissimilarity constraints and the geometrical structure of data

The problem of clustering with side information has received much recent attention and metric learning has been considered as a powerful approach to this problem. Until now, various metric learning methods have been proposed for semi-supervised clustering. Although some of the existing methods can use both positive (must-link) and negative (cannot-link) constraints, they are usually limited to learning a linear transformation (i.e., finding a global Mahalanobis metric). In this paper, we propose a framework for learning linear and non-linear transformations efficiently. We use both positive and negative constraints and also the intrinsic topological structure of data. We formulate our metric learning method as an appropriate optimization problem and find the global optimum of this problem. The proposed non-linear method can be considered as an efficient kernel learning method that yields an explicit non-linear transformation and thus shows out-of-sample generalization ability. Experimental results on synthetic and real-world data sets show the effectiveness of our metric learning method for semi-supervised clustering tasks.     

Distance metric learning guided adaptive subspace semi-supervised clustering

Most existing semi-supervised clustering algorithms are not designed for handling high-dimensional data. On the other hand, semi-supervised dimensionality reduction methods may not necessarily improve the clustering performance, due to the fact that the inherent relationship between subspace selection and clustering is ignored. In order to mitigate the above problems, we present a semi-supervised clustering algorithm using adaptive distance metric learning (SCADM) which performs semi-supervised clustering and distance metric learning simultaneously. SCADM applies the clustering results to learn a distance metric and then projects the data onto a low-dimensional space where the separability of the data is maximized. Experimental results on real-world data sets show that the proposed method can effectively deal with high-dimensional data and provides an appealing clustering performance.     

Semi-supervised image clustering with multi-modal information

How to organize and retrieve images is now a great challenge in various domains. Image clustering is a key tool in some practical applications including image retrieval and understanding. Traditional image clustering algorithms consider a single set of features and use ad hoc distance functions, such as Euclidean distance, to measure the similarity between samples. However, multi-modal features can be extracted from images. The dimension of multi-modal data is very high. In addition, we usually have several, but not many labeled images, which lead to semi-supervised learning. In this paper, we propose a framework of image clustering based on semi-supervised distance learning and multi-modal information. First we fuse multiple features and utilize a small amount of labeled images for semi-supervised metric learning. Then we compute similarity with the Gaussian similarity function and the learned metric. Finally, we construct a semi-supervised Laplace matrix for spectral clustering and propose an effective clustering method. Extensive experiments on some image data sets show the competent performance of the proposed algorithm.     

Semi-supervised clustering with metric learning: An adaptive kernel method

Most existing representative works in semi-supervised clustering do not sufficiently solve the violation problem of pairwise constraints. On the other hand, traditional kernel methods for semi-supervised clustering not only face the problem of manually tuning the kernel parameters due to the fact that no sufficient supervision is provided, but also lack a measure that achieves better effectiveness of clustering. In this paper, we propose an adaptive Semi-supervised Clustering Kernel Method based on Metric learning (SCKMM) to mitigate the above problems. Specifically, we first construct an objective function from pairwise constraints to automatically estimate the parameter of the Gaussian kernel. Then, we use pairwise constraint-based K-means approach to solve the violation issue of constraints and to cluster the data. Furthermore, we introduce metric learning into nonlinear semi-supervised clustering to improve separability of the data for clustering. Finally, we perform clustering and metric learning simultaneously. Experimental results on a number of real-world data sets validate the effectiveness of the proposed method.     

基于特征选择和半监督学习的遥感图像检索

随着卫星遥感技术的不断发展,基于内容的遥感图像检索技术越来越受到关注。目前该方向的研究主要集中在对遥感图像中不同特征的提取和融合方面,这些方法普遍忽略了这样一个事实：对于不同类型的检索目标,特征应该是不同的。另外,小样本问题也是遥感图像检索中一个较为突出的问题。基于以上两方面考虑,本文提出一种基于特征选择和半监督学习的遥感图像检索新方法,该方法主要包括4个方面：1）利用最小描述长度准则自动确定聚类数目;2）结合聚类方法和适当的聚类有效性指标选择最能表示检索目标的特征,在计算聚类有效性指数时,针对遥感图像检索特点对原有的Davies-Bouldin指数进行了改进;3）动态确定最优颜色特征和最优纹理特征之间的权重;4）根据最优颜色特征和最优纹理特征的权重自动确定半监督学习方法,并进行遥感图像的检索。实验结果表明,与相关反馈方法的检索效果相比,该算法在土壤侵蚀区域检索以及其他一般地表覆盖目标检索中均获得了相近的检索效果,但不需要用户多次反馈。     

Manifold-ranking based retrieval using k-regular nearest neighbor graph

Manifold-ranking is a powerful method in semi-supervised learning, and its performance heavily depends on the quality of the constructed graph. In this paper, we propose a novel graph structure named k-regular nearest neighbor (k-RNN) graph as well as its constructing algorithm, and apply the new graph structure in the framework of manifold-ranking based retrieval. We show that the manifold-ranking algorithm based on our proposed graph structure performs better than that of the existing graph structures such as k-nearest neighbor (k-NN) graph and connected graph in image retrieval, 2D data clustering as well as 3D model retrieval. In addition, the automatic sample reweighting and graph updating algorithms are presented for the relevance feedback of our algorithm. Experiments demonstrate that the proposed algorithm outperforms the state-of-the-art algorithms.     

Probability based document clustering and image clustering using content-based image retrieval

Clustering of related or similar objects has long been regarded as a potentially useful contribution of helping users to navigate an information space such as a document collection. Many clustering algorithms and techniques have been developed and implemented but as the sizes of document collections have grown these techniques have not been scaled to large collections because of their computational overhead. To solve this problem, the proposed system concentrates on an interactive text clustering methodology, probability based topic oriented and semi-supervised document clustering. Recently, as web and various documents contain both text and large number of images, the proposed system concentrates on content-based image retrieval (CBIR) for image clustering to give additional effect to the document clustering approach. It suggests two kinds of indexing keys, major colour sets (MCS) and distribution block signature (DBS) to prune away the irrelevant images to given query image. Major colour sets are related with colour information while distribution block signatures are related with spatial information. After successively applying these filters to a large database, only small amount of high potential candidates that are somewhat similar to that of query image are identified. Then, the system uses quad modelling method (QM) to set the initial weight of two-dimensional cells in query image according to each major colour and retrieve more similar images through similarity association function associated with the weights. The proposed system evaluates the system efficiency by implementing and testing the clustering results with Dbscan and K-means clustering algorithms. Experiment shows that the proposed document clustering algorithm performs with an average efficiency of 94.4% for various document categories.     

Locally adaptive subspace and similarity metric learning for visual data clustering and retrieval

Subspace and similarity metric learning are important issues for image and video analysis in the scenarios of both computer vision and multimedia fields. Many real-world applications, such as image clustering/labeling and video indexing/retrieval, involve feature space dimensionality reduction as well as feature matching metric learning. However, the loss of information from dimensionality reduction may degrade the accuracy of similarity matching. In practice, such basic conflicting requirements for both feature representation efficiency and similarity matching accuracy need to be appropriately addressed. In the style of “Thinking Globally and Fitting Locally”, we develop Locally Embedded Analysis (LEA) based solutions for visual data clustering and retrieval. LEA reveals the essential low-dimensional manifold structure of the data by preserving the local nearest neighbor affinity, and allowing a linear subspace embedding through solving a graph embedded eigenvalue decomposition problem. A visual data clustering algorithm, called Locally Embedded Clustering (LEC), and a local similarity metric learning algorithm for robust video retrieval, called Locally Adaptive Retrieval (LAR), are both designed upon the LEA approach, with variations in local affinity graph modeling. For large size database applications, instead of learning a global metric, we localize the metric learning space with kd-tree partition to localities identified by the indexing process. Simulation results demonstrate the effective performance of proposed solutions in both accuracy and speed aspects.     

Partially supervised speaker clustering

Content-based multimedia indexing, retrieval, and processing as well as multimedia databases demand the structuring of the media content (image, audio, video, text, etc.), one significant goal being to associate the identity of the content to the individual segments of the signals. In this paper, we specifically address the problem of speaker clustering, the task of assigning every speech utterance in an audio stream to its speaker. We offer a complete treatment to the idea of partially supervised speaker clustering, which refers to the use of our prior knowledge of speakers in general to assist the unsupervised speaker clustering process. By means of an independent training data set, we encode the prior knowledge at the various stages of the speaker clustering pipeline via 1) learning a speaker-discriminative acoustic feature transformation, 2) learning a universal speaker prior model, and 3) learning a discriminative speaker subspace, or equivalently, a speaker-discriminative distance metric. We study the directional scattering property of the Gaussian mixture model (GMM) mean supervector representation of utterances in the high-dimensional space, and advocate exploiting this property by using the cosine distance metric instead of the euclidean distance metric for speaker clustering in the GMM mean supervector space. We propose to perform discriminant analysis based on the cosine distance metric, which leads to a novel distance metric learning algorithm—linear spherical discriminant analysis (LSDA). We show that the proposed LSDA formulation can be systematically solved within the elegant graph embedding general dimensionality reduction framework. Our speaker clustering experiments on the GALE database clearly indicate that 1) our speaker clustering methods based on the GMM mean supervector representation and vector-based distance metrics outperform traditional speaker clustering methods based on the “bag of acoustic features” representation and statistical model-based distance metrics, 2) our advocated use of the cosine distance metric yields consistent increases in the speaker clustering performance as compared to the commonly used euclidean distance metric, 3) our partially supervised speaker clustering concept and strategies significantly improve the speaker clustering performance over the baselines, and 4) our proposed LSDA algorithm further leads to state-of-the-art speaker clustering performance.     

一种改进的局部线性嵌套方法

局部线性嵌套LLE(locally linear embedding)是一种经典的流形学习方法.对于从单个流形上采样得到的数据集,它能够有效地学习其内在低维结构,然而当数据集是从多个流形上采样得到时,U正的效果并不理想.提出了一种基于距离度量学习的改进方法:Metric LLE,它利用部分数据点的相似信息来学习距离度量.实验结果表明Metric LLE在应用中有很好的性能:分类能力比LLE好;在可视化方面,效果比Supervised LLE好.     

基于混合样本自动数据增强技术的半监督学习方法

基于一致性的半监督学习方法通常使用简单的数据增强方法来实现对原始输入和扰动输入的一致性预测.在有标签数据的比例较低的情况下,该方法的效果难以得到保证.将监督学习中一些先进的数据增强方法扩展到半监督学习环境中,是解决该问题的思路之一.基于一致性的半监督学习方法MixMatch,提出了基于混合样本自动数据增强技术的半监督学...     

基于卷积与稀疏编码的半监督学习方法

卷积神经网络（CNN）在半监督学习中取得了良好的成绩,其在训练阶段既利用有标记样本,也利用无标记样本帮助规范化学习模型。为进一步加强半监督模型的特征学习能力,提高其在图像分类时的性能表现,本文提出一种联合深度半监督卷积神经网络和字典学习的端到端半监督学习方法,称为Semi-supervised Learning based on Sparse Coding and Convolution（SSSConv）;该算法框架旨在学习到鉴别性更强的图像特征表示。SSSConv首先利用CNN提取特征,并对所提取特征进行正交投影变换,下一步通过学习其稀疏编码的低维嵌入以得到图像的特征表示,最后据此进行分类。整个模型框架可进行端到端的半监督学习训练,CNN提取特征部分和稀疏编码字典学习部分具有统一的损失函数,目标一致。本文利用共轭梯度下降算法、链式法则和反向传播等算法对目标函数的参数进行优化,将稀疏编码的相关参数约束于流形上,CNN参数既可定义在欧氏空间,也可以进一步定义在正交空间中。基于半监督分类任务的实验结果验证了所提出SSSConv框架的有效性,与现有方法相比具有较强的竞争力。     

Pairwise ranking component analysis

Uncovering the latent structure of the data is an active research topic in data mining. However, in the distance metric learning framework, previous studies have mainly focused on the classification performance. In this work, we consider the distance metric learning problem in the ranking setting, where predicting the order between the data vectors is more important than predicting the class labels. We focus on two problems: improving the ranking prediction accuracy and identifying the latent structure of the data. The core of our model consists of ranking the data using a Mahalanobis distance function. The additional use of non-negativity constraints and an entropy-based cost function allows us to simultaneously minimize the ranking error while identifying useful meta-features. To demonstrate its usefulness for information retrieval applications, we compare the performance of our method with four other methods on four UCI data sets, three text data sets, and four image data sets. Our approach shows good ranking accuracies, especially when few training data are available. We also use our model to extract and interpret the latent structure of the data sets. In addition, our approach is simple to implement and computationally efficient and can be used for data embedding and visualization.     

Semi-supervised classification based on random subspace dimensionality reduction

Graph structure is vital to graph based semi-supervised learning. However, the problem of constructing a graph that reflects the underlying data distribution has been seldom investigated in semi-supervised learning, especially for high dimensional data. In this paper, we focus on graph construction for semi-supervised learning and propose a novel method called Semi-Supervised Classification based on Random Subspace Dimensionality Reduction, SSC-RSDR in short. Different from traditional methods that perform graph-based dimensionality reduction and classification in the original space, SSC-RSDR performs these tasks in subspaces. More specifically, SSC-RSDR generates several random subspaces of the original space and applies graph-based semi-supervised dimensionality reduction in these random subspaces. It then constructs graphs in these processed random subspaces and trains semi-supervised classifiers on the graphs. Finally, it combines the resulting base classifiers into an ensemble classifier. Experimental results on face recognition tasks demonstrate that SSC-RSDR not only has superior recognition performance with respect to competitive methods, but also is robust against a wide range of values of input parameters.     

Neighborhood preserving regression for image retrieval

In content-based image retrieval (CBIR), relevance feedback has been proven to be a powerful tool for bridging the gap between low level visual features and high level semantic concepts. Traditionally, relevance feedback driven CBIR is often considered as a supervised learning problem where the user provided feedbacks are used to learn a distance metric or classification function. However, CBIR is intrinsically a semi-supervised learning problem in which the testing samples (images in the database) are present during the learning process. Moreover, when there are no sufficient feedbacks, these methods may suffer from the overfitting problem. In this paper, we propose a novel neighborhood preserving regression algorithm which makes efficient use of both labeled and unlabeled images. By using the unlabeled images, the geometrical structure of the image space can be incorporated into the learning system through a regularizer. Specifically, from all the functions which minimize the empirical loss on the labeled images, we select the one which best preserves the local neighborhood structure of the image space. In this way, our method can obtain a regression function which respects both semantic and geometrical structures of the image database. We present experimental evidence suggesting that our algorithm is able to use unlabeled data effectively for image retrieval.     

Noise-robust semi-supervised learning via fast sparse coding

This paper presents a novel noise-robust graph-based semi-supervised learning algorithm to deal with the challenging problem of semi-supervised learning with noisy initial labels. Inspired by the successful use of sparse coding for noise reduction, we choose to give new L₁-norm formulation of Laplacian regularization for graph-based semi-supervised learning. Since our L₁-norm Laplacian regularization is explicitly defined over the eigenvectors of the normalized Laplacian matrix, we formulate graph-based semi-supervised learning as an L₁-norm linear reconstruction problem which can be efficiently solved by sparse coding. Furthermore, by working with only a small subset of eigenvectors, we develop a fast sparse coding algorithm for our L₁-norm semi-supervised learning. Finally, we evaluate the proposed algorithm in noise-robust image classification. The experimental results on several benchmark datasets demonstrate the promising performance of the proposed algorithm.     

半监督鲁棒联机聚类算法

将监督信息引入到聚类算法中去,在先前提出的鲁棒联机聚类算法(ROC)的基础上,通过引入以样本类标号形式给出的监督信息,提出了一种半监督的鲁棒联机聚类算法(Semi-ROC).在算法的聚类精度和鲁棒性能上,算法Semi-ROC比ROC和AddC有着更好的性能,在人工数据集和UCI标准数据集上的实验结果表明,Semi-ROC能有效地利用少量的监督信息来提高算法的聚类性能,得到较优的结果.另外,在添加噪声的情况下,算法Semi-ROC比原始的联机聚类算法AddC和ROC都更加鲁棒.     

Semi-supervised hyperspectral image classification algorithm based on graph embedding and discriminative spatial information

Image classification is one of the important techniques in computer vision. Due to the limited access of labeled samples in hyperspectral images, semi-supervised learning (SSL) methods have been widely applied in hyperspectral image classification. Graph based semi-supervised learning provides an effective solution to model data in classification problems, of which graph construction is the critical step. In this paper we employ the graphs constructed with a typical manifold learning method-locally linear embedding (LLE), based on which semi-supervised classification is then conducted. To exploit the valuable spatial information contained in hyperspectral images, discriminative spatial information (DSI) is then extracted. The proposed classification method is evaluated using three real hyperspectral data sets, revealing state-of-art performance when compared with different classification methods.     

Active hashing and its application to image and text retrieval

In recent years, hashing-based methods for large-scale similarity search have sparked considerable research interests in the data mining and machine learning communities. While unsupervised hashing-based methods have achieved promising successes for metric similarity, they cannot handle semantic similarity which is usually given in the form of labeled point pairs. To overcome this limitation, some attempts have recently been made on semi-supervised hashing which aims at learning hash functions from both metric and semantic similarity simultaneously. Existing semi-supervised hashing methods can be regarded as passive hashing since they assume that the labeled pairs are provided in advance. In this paper, we propose a novel framework, called active hashing, which can actively select the most informative labeled pairs for hash function learning. Specifically, it identifies the most informative points to label and constructs labeled pairs accordingly. Under this framework, we use data uncertainty as a measure of informativeness and develop a batch mode algorithm to speed up active selection. We empirically compare our method with a state-of-the-art passive hashing method on two benchmark data sets, showing that the proposed method can reduce labeling cost as well as overcome the limitations of passive hashing.