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.     

Online clustering via finite mixtures of Dirichlet and minimum message length

This paper presents an online algorithm for mixture model-based clustering. Mixture modeling is the problem of identifying and modeling components in a given set of data. The online algorithm is based on unsupervised learning of finite Dirichlet mixtures and a stochastic approach for estimates updating. For the selection of the number of clusters, we use the minimum message length (MML) approach. The proposed method is validated by synthetic data and by an application concerning the dynamic summarization of image databases.     

Model-based approach for high-dimensional non-Gaussian visual data clustering and feature weighting

Clustering is the task of classifying patterns or observations into clusters or groups. Generally, clustering in high-dimensional feature spaces has a lot of complications such as: the unidentified or unknown data shape which is typically non-Gaussian and follows different distributions; the unknown number of clusters in the case of unsupervised learning; and the existence of noisy, redundant, or uninformative features which normally compromise modeling capabilities and speed. Therefore, high-dimensional data clustering has been a subject of extensive research in data mining, pattern recognition, image processing, computer vision, and other areas for several decades. However, most of existing researches tackle one or two problems at a time which is unrealistic because all problems are connected and should be tackled simultaneously. Thus, in this paper, we propose two novel inference frameworks for unsupervised non-Gaussian feature selection, in the context of finite asymmetric generalized Gaussian (AGG) mixture-based clustering. The choice of the AGG distribution is mainly due to its ability not only to approximate a large class of statistical distributions (e.g. impulsive, Laplacian, Gaussian and uniform distributions) but also to include the asymmetry. In addition, the two frameworks simultaneously perform model parameters estimation as well as model complexity (i.e., both model and feature selection) determination in the same step. This was done by incorporating a minimum message length (MML) penalty in the model learning step and by fading out the redundant densities in the mixture using the rival penalized EM (RPEM) algorithm, for first and second frameworks, respectively. Furthermore, for both algorithms, we tackle the problem of noisy and uninformative features by determining a set of relevant features for each data cluster. The efficiencies of the proposed algorithms are validated by applying them to real challenging problems namely action and facial expression recognition.     

Self-representation and PCA embedding for unsupervised feature selection

Feature selection is an important preprocessing step for dealing with high dimensional data. In this paper, we propose a novel unsupervised feature selection method by embedding a subspace learning regularization (i.e., principal component analysis (PCA)) into the sparse feature selection framework. Specifically, we select informative features via the sparse learning framework and consider preserving the principal components (i.e., the maximal variance) of the data at the same time, such that improving the interpretable ability of the feature selection model. Furthermore, we propose an effective optimization algorithm to solve the proposed objective function which can achieve stable optimal result with fast convergence. By comparing with five state-of-the-art unsupervised feature selection methods on six benchmark and real-world datasets, our proposed method achieved the best result in terms of classification performance.     

BYY Harmony Learning on Finite Mixture: Adaptive Gradient Implementation and A Floating RPCL Mechanism

In tackling the learning problem on a set of finite samples, Bayesian Ying-Yang (BYY) harmony learning has developed a new learning mechanism that makes model selection implemented either automatically during parameter learning or in help of evaluating a new class of model selection criteria. In this paper, parameter learning with automated model selection has been studied for finite mixture model via an adaptive gradient learning algorithm for BYY harmony learning on a specific bidirectional architecture (BI-architecture). Via theoretical analysis, it has shown that the adaptive gradient learning implements a mechanism of floating rival penalized competitive learning (RPCL) among the components in the mixture. Also, the simulation results are demonstrated well for the adaptive gradient algorithm on the sample data sets from Gaussian mixtures with certain degree of overlap. Moreover, the adaptive gradient algorithm is applied to classification of the Iris data and unsupervised color image segmentation.     

Simultaneous Localized Feature Selection and Model Detection for Gaussian Mixtures

In this paper, we propose a novel approach of simultaneous localized feature selection and model detection for unsupervised learning. In our approach, local feature saliency, together with other parameters of Gaussian mixtures, are estimated by Bayesian variational learning. Experiments performed on both synthetic and real-world data sets demonstrate that our approach is superior over both global feature selection and subspace clustering methods.     

Scale selection for supervised image segmentation

Finding the right scales for feature extraction is crucial for supervised image segmentation based on pixel classification. There are many scale selection methods in the literature; among them the one proposed by Lindeberg is widely used for image structures such as blobs, edges and ridges. Those schemes are usually unsupervised, as they do not take into account the actual segmentation problem at hand. In this paper, we consider the problem of selecting scales, which aims at an optimal discrimination between user-defined classes in the segmentation. We show the deficiency of the classical unsupervised scale selection paradigms and present a supervised alternative. In particular, the so-called max rule is proposed, which selects a scale for each pixel to have the largest confidence in the classification across the scales. In interpreting the classifier as a complex image filter, we can relate our approach back to Lindeberg's original proposal. In the experiments, the max rule is applied to artificial and real-world image segmentation tasks, which is shown to choose the right scales for different problems and lead to better segmentation results.     

An unsupervised approach to learn the kernel functions: from global influence to local similarity

Recently there has been a steep growth in the development of kernel-based learning algorithms. The intrinsic problem in such algorithms is the selection of the optimal kernel for the learning task of interest. In this paper, we propose an unsupervised approach to learn a linear combination of kernel functions, such that the resulting kernel best serves the objectives of the learning task. This is achieved through measuring the influence of each point on the structure of the dataset. This measure is calculated by constructing a weighted graph on which a random walk is performed. The measure of influence in the feature space is probabilistically related to the input space that yields an optimization problem to be solved. The optimization problem is formulated in two different convex settings, namely linear and semidefinite programming, dependent on the type of kernel combination considered. The contributions of this paper are twofold: first, a novel unsupervised approach to learn the kernel function, and second, a method to infer the local similarity represented by the kernel function by measuring the global influence of each point toward the structure of the dataset. The proposed approach focuses on the kernel selection which is independent of the kernel-based learning algorithm. The empirical evaluation of the proposed approach with various datasets shows the effectiveness of the algorithm in practice.     

字典重建和空间分布关系约束下的特征选择与图像拼接

目的 针对大型图像检索领域中,复杂图像中SIFT特征描述子的冗余和高维问题,提出了一种基于字典重建和空间分布关系约束的特征选择的方法,来消除冗余特征并保留最具表现力的、保留原始空间结构性的SIFT特征描述子。方法 首先,实验发现了特征选择和字典学习方法在稀疏表示方面的内在联系,将特征选择问题转化为字典重构任务;其次,在SIFT特征选择问题中,为了保证特征空间中特征的鲁棒性,设计了新型的字典学习模型,并采用模拟退火算法进行迭代求解;最后,在字典学习的过程中,加入熵理论来约束特征的空间分布,使学习到的特征描述子能最大限度保持原始SIFT特征空间的空间拓扑关系。结果 在公开数据集Holiday大型场景图片检索数据库上,通过与国际公认的特征选择方法进行实验对比,本文提出的特征选择方法在节省内存空间和提高时间效率（30%~ 50%）的同时,还能保证所筛选的特征描述子的检索准确率比同类特征提高8%~ 14.1%;在国际通用的大型场景图片拼接数据库IPM上,验证本文方法在图像拼接应用中特征提取和特征匹配上的有效性,实验表明本文方法能节省（50% ~70%）图像拼接时间。结论 与已有的方法比较,本文的特征选择方法既不依赖训练数据集,也不丢失重要的空间结构和纹理信息,在大型图像检索、图像拼接领域和3D检索领域中,能够精简特征,提高特征匹配效率和准确率。     

维吾尔文无监督自动切分及无监督特征选择

维吾尔文常用切分方法会产生大量的语义抽象甚至多义的词特征,因此学习算法难以发现高维数据中隐藏的结构.提出一种无监督切分方法dme-TS和一种无监督特征选择方法UMRMR-UFS.dme-TS从大规模生语料中自动获取单词Bi-gram及上下文语境信息,并将相邻单词间的t-测试差、互信息及双词上下文邻接对熵的线性融合作为一个组合统计量(dme)来评价单词间的结合能力,从而将文本切分成语义具体的独立语言单位的特征集合.UMRMR-UFS用一种综合考虑最大相关度和最小冗余的无监督特征选择标准(UMRMR)来评价每一个特征的重要性,并将最重要的特征依次移入到特征子集中.实验结果表明dme-TS能有效控制原始特征集的规模,提高特征项本身的质量,用UMRMR-UFS的输出来表征文本时,学习算法也表现出其最高的性能.     

基于遗传算法的无监督分形属性规约技术

属性规约是应对“维数灾难”的有效技术,分形属性规约FDR（Fractal Dimensionality Reduction）是近年来出现的一种无监督属性选择技术,令人遗憾的是其需要多遍扫描数据集,因而难于应对高维数据集情况;基于遗传算法的属性规约技术对于高维数据而言优越于传统属性选择技术,但其无法应用于无监督学习领域。为此,结合遗传算法内在随机并行寻优机制及分形属性选择的无监督特点,设计并实现了基于遗传算法的无监督分形属性子集选择算法GABUFSS（Genetic Algorithm Based Unsupervised Feature Subset Selection）。基于合成与实际数据集的实验对比分析了GABUFSS算法与FDR算法的性能,结果表明GABUFSS相对优于FDR算法,并具有发现等价结果属性子集的特点。     

Subspace learning for unsupervised feature selection via matrix factorization

Dimensionality reduction is an important and challenging task in machine learning and data mining. Feature selection and feature extraction are two commonly used techniques for decreasing dimensionality of the data and increasing efficiency of learning algorithms. Specifically, feature selection realized in the absence of class labels, namely unsupervised feature selection, is challenging and interesting. In this paper, we propose a new unsupervised feature selection criterion developed from the viewpoint of subspace learning, which is treated as a matrix factorization problem. The advantages of this work are four-fold. First, dwelling on the technique of matrix factorization, a unified framework is established for feature selection, feature extraction and clustering. Second, an iterative update algorithm is provided via matrix factorization, which is an efficient technique to deal with high-dimensional data. Third, an effective method for feature selection with numeric data is put forward, instead of drawing support from the discretization process. Fourth, this new criterion provides a sound foundation for embedding kernel tricks into feature selection. With this regard, an algorithm based on kernel methods is also proposed. The algorithms are compared with four state-of-the-art feature selection methods using six publicly available datasets. Experimental results demonstrate that in terms of clustering results, the proposed two algorithms come with better performance than the others for almost all datasets we experimented with here.     

A countably infinite mixture model for clustering and feature selection

Mixture modeling is one of the most useful tools in machine learning and data mining applications. An important challenge when applying finite mixture models is the selection of the number of clusters which best describes the data. Recent developments have shown that this problem can be handled by the application of non-parametric Bayesian techniques to mixture modeling. Another important crucial preprocessing step to mixture learning is the selection of the most relevant features. The main approach in this paper, to tackle these problems, consists on storing the knowledge in a generalized Dirichlet mixture model by applying non-parametric Bayesian estimation and inference techniques. Specifically, we extend finite generalized Dirichlet mixture models to the infinite case in which the number of components and relevant features do not need to be known a priori. This extension provides a natural representation of uncertainty regarding the challenging problem of model selection. We propose a Markov Chain Monte Carlo algorithm to learn the resulted infinite mixture. Through applications involving text and image categorization, we show that infinite mixture models offer a more powerful and robust performance than classic finite mixtures for both clustering and feature selection.     

Unsupervised learning in neural computation

In this article, we consider unsupervised learning from the point of view of applying neural computation on signal and data analysis problems. The article is an introductory survey, concentrating on the main principles and categories of unsupervised learning. In neural computation, there are two classical categories for unsupervised learning methods and models: first, extensions of principal component analysis and factor analysis, and second, learning vector coding or clustering methods that are based on competitive learning. These are covered in this article. The more recent trend in unsupervised learning is to consider this problem in the framework of probabilistic generative models. If it is possible to build and estimate a model that explains the data in terms of some latent variables, key insights may be obtained into the true nature and structure of the data. This approach is also briefly reviewed.     

FEAST: Feature evaluation and selection technique for deployment in unsupervised nonparametric environments

The problem of feature selection in a totally unsupervised, distribution free environment being conceptually ill-defined, the problem has been studied in an artifically evolved pseudosupervised environment. The evolution of such an environment is achieved by formulating a unified approach to the twin problems of feature selection and unsupervised learning. The solution of the latter problem leads to the pseudosupervised environment in which the features are evaluated by employing a multistate-choice automaton model as the feature selector. The methodology developed here is intended to be deployed in conjunction with any one of the numerous recursive schemes of clustering in which the crudely formed initial clusters are refined in a recursive fashion by successively determining the centroids of the different clusters and reallocating the samples to the clusters defined by these centroids. This allocation is carried out on the basis of distance measures (Euclidean or modifications thereof) and is in parallel progress with the feature-evaluation process. The clusters, as formulated at each stage of the recursive process, provide the pseudosupervised environment for the feature selector. The track record of the automaton in terms of probabilities of penalized action provides a measure of the efficiency of the different feature subsets in the unsupervised environment.An earlier version of this paper was presented at the Fifth Annual Symposium on Automatic Imagery Pattern Recognition, University of Maryland, College Park, Maryland, April 1975.     

基于深度学习的单目深度估计技术综述

场景的深度估计问题是计算机视觉领域中的经典问题之一,也是3维重建和图像合成等应用中的一个重要环节。基于深度学习的单目深度估计技术高速发展,各种网络结构相继提出。本文对基于深度学习的单目深度估计技术最新进展进行了综述,回顾了基于监督学习和基于无监督学习方法的发展历程。重点关注单目深度估计的优化思路及其在深度学习网络结构中的表现,将监督学习方法分为多尺度特征融合的方法、结合条件随机场(conditional random field,CRF)的方法、基于序数关系的方法、结合多元图像信息的方法和其他方法等5类;将无监督学习方法分为基于立体视觉的方法、基于运动恢复结构(structure from motion,SfM)的方法、结合对抗性网络的方法、基于序数关系的方法和结合不确定性的方法等5类。此外,还介绍了单目深度估计任务中常用的数据集和评价指标,并对目前基于深度学习的单目深度估计技术在精确度、泛化性、应用场景和无监督网络中不确定性研究等方面的现状和面临的挑战进行了讨论,为相关领域的研究人员提供一个比较全面的参考。     

Semi-supervised feature selection via hierarchical regression for web image classification

Feature selection is an important step for large-scale image data analysis, which has been proved to be difficult due to large size in both dimensions and samples. Feature selection firstly eliminates redundant and irrelevant features and then chooses a subset of features that performs as efficient as the complete set. Generally, supervised feature selection yields better performance than unsupervised feature selection because of the utilization of labeled information. However, labeled data samples are always expensive to obtain, which constraints the performance of supervised feature selection, especially for the large web image datasets. In this paper, we propose a semi-supervised feature selection algorithm that is based on a hierarchical regression model. Our contribution can be highlighted as: (1) Our algorithm utilizes a statistical approach to exploit both labeled and unlabeled data, which preserves the manifold structure of each feature type. (2) The predicted label matrix of the training data and the feature selection matrix are learned simultaneously, making the two aspects mutually benefited. Extensive experiments are performed on three large-scale image datasets. Experimental results demonstrate the better performance of our algorithm, compared with the state-of-the-art algorithms.     

Discrete data clustering using finite mixture models

Finite mixture models have been applied for different computer vision, image processing and pattern recognition tasks. The majority of the work done concerning finite mixture models has focused on mixtures for continuous data. However, many applications involve and generate discrete data for which discrete mixtures are better suited. In this paper, we investigate the problem of discrete data modeling using finite mixture models. We propose a novel, well motivated mixture that we call the multinomial generalized Dirichlet mixture. The novel model is compared with other discrete mixtures. We designed experiments involving spatial color image databases modeling and summarization, and text classification to show the robustness, flexibility and merits of our approach.     

基于l2,0范数稀疏性和模糊相似性的图优化无监督组特征选择方法

基于图的无监督特征选择方法大多选择投影矩阵的l_2,1范数稀疏正则化代替非凸的l_2,0范数约束,然而l_2,1范数正则化方法根据得分高低逐个选择特征,未考虑特征的相关性.因此,文中提出基于l_2,0范数稀疏性和模糊相似性的图优化无监督组特征选择方法,同时进行图学习和特征选择.在图学习中,学习具有精确连通分量的相似性矩阵.在特征选择过程中,约束投影矩阵的非零行个数,实现组特征选择.为了解决非凸的l_2,0范数约束,引入元素为0或1的特征选择向量,将l_2,0范数约束问题转化为0-1整数规划问题,并将离散的0-1整数约束转化为2个连续约束进行求解.最后,引入模糊相似性因子,拓展文中方法,学习更精确的图结构.在真实数据集上的实验表明文中方法的有效性.     

Efficient greedy feature selection for unsupervised learning

Reducing the dimensionality of the data has been a challenging task in data mining and machine learning applications. In these applications, the existence of irrelevant and redundant features negatively affects the efficiency and effectiveness of different learning algorithms. Feature selection is one of the dimension reduction techniques, which has been used to allow a better understanding of data and improve the performance of other learning tasks. Although the selection of relevant features has been extensively studied in supervised learning, feature selection in the absence of class labels is still a challenging task. This paper proposes a novel method for unsupervised feature selection, which efficiently selects features in a greedy manner. The paper first defines an effective criterion for unsupervised feature selection that measures the reconstruction error of the data matrix based on the selected subset of features. The paper then presents a novel algorithm for greedily minimizing the reconstruction error based on the features selected so far. The greedy algorithm is based on an efficient recursive formula for calculating the reconstruction error. Experiments on real data sets demonstrate the effectiveness of the proposed algorithm in comparison with the state-of-the-art methods for unsupervised feature selection.