Graph-based semi-supervised learning

Recent years have witnessed a surge of interest in graph-based semi-supervised learning. However, two of the major problems in graph-based semi-supervised learning are: (1) how to set the hyperparameter in the Gaussian similarity; and (2) how to make the algorithm scalable. In this article, we introduce a general framework for graphbased learning. First, we propose a method called linear neighborhood propagation, which can automatically construct the optimal graph. Then we introduce a novel multilevel scheme to make our algorithm scalable for large data sets. The applications of our algorithm to various real-world problems are also demonstrated.     

Visual instance mining from the graph perspective

In this paper, we address the problem of visual instance mining, which is to automatically discover frequently appearing visual instances from a large collection of images. We propose a scalable mining method by leveraging the graph structure with images as vertices. Different from most existing approaches that focus on either instance-level similarities or image-level context properties, our method captures both information. In the proposed framework, the instance-level information is integrated during the construction of a sparse instance graph based on the similarity between augmented local features, while the image-level context is explored with a greedy breadth-first search algorithm to discover clusters of visual instances from the graph. This framework can tackle the challenges brought by small visual instances, diverse intra-class variations, as well as noise in large-scale image databases. To further improve the robustness, we integrate two techniques into the basic framework. First, to better cope with the increasing noise of large databases, weak geometric consistency is adopted to efficiently combine the geometric information of local matches into the construction of the instance graph. Second, we propose the layout embedding algorithm, which leverages the algorithm originally designed for graph visualization to fully explore the image database structure. The proposed method was evaluated on four annotated data sets with different characteristics, and experimental results showed the superiority over state-of-the-art algorithms on all data sets. We also applied our framework on a one-million Flickr data set and proved its scalability.     

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.     

Template Assembly for Detailed Urban Reconstruction

We propose a new framework to reconstruct building details by automatically assembling 3D templates on coarse textured building models. In a preprocessing step, we generate an initial coarse model to approximate a point cloud computed using Structure from Motion and Multi View Stereo, and we model a set of 3D templates of facade details. Next, we optimize the initial coarse model to enforce consistency between geometry and appearance (texture images). Then, building details are reconstructed by assembling templates on the textured faces of the coarse model. The 3D templates are automatically chosen and located by our optimization‐based template assembly algorithm that balances image matching and structural regularity. In the results, we demonstrate how our framework can enrich the details of coarse models using various data sets.     

DISCOVERING ROBUST EMBEDDINGS IN (DIS)SIMILARITY SPACE FOR HIGH‐DIMENSIONAL LINGUISTIC FEATURES

Recent research has shown the effectiveness of rich feature representation for tasks in natural language processing (NLP). However, exceedingly large number of features do not always improve classification performance. They may contain redundant information, lead to noisy feature presentations, and also render the learning algorithms intractable. In this paper, we propose a supervised embedding framework that modifies the relative positions between instances to increase the compatibility between the input features and the output labels and meanwhile preserves the local distribution of the original data in the embedded space. The proposed framework attempts to support flexible balance between the preservation of intrinsic geometry and the enhancement of class separability for both interclass and intraclass instances. It takes into account characteristics of linguistic features by using an inner product‐based optimization template. (Dis)similarity features, also known as empirical kernel mapping, is employed to enable computationally tractable processing of extremely high‐dimensional input, and also to handle nonlinearities in embedding generation when necessary. Evaluated on two NLP tasks with six data sets, the proposed framework provides better classification performance than the support vector machine without using any dimensionality reduction technique. It also generates embeddings with better class discriminability as compared to many existing embedding algorithms.     

Prediction-based reversible data hiding

For some applications such as satellite and medical images, reversible data hiding is the best solution to provide copyright protection or authentication. Being reversible, the decoder can extract the hidden data and recover the original image without distortion. In this paper, a reversible data hiding scheme based on prediction error expansion is proposed. The predictive value is computed by using various predictors. The secret data is embedded in the cover image by exploiting the expansion of the difference between a pixel and its predictive value. Experimental results show that our method is capable of providing a great embedding capacity without making noticeable distortion. In addition, the proposed scheme is also applicable to various predictors.     

基于图嵌入的用户加权Slope One算法

针对传统Slope One推荐算法在稀疏数据集上预测准确率较低的问题,提出一种基于图嵌入的加权Slope One算法。本文算法首先以融合时间信息的用户相似度为边权建立用户关联图,对该图进行图嵌入得到用户特征向量,然后基于Canopy聚类对用户进行类内加权Slope One推荐。另外,为优化算法性能,本文算法基于Spark计算框架实现。实验结果表明,对比传统的加权Slope One,本文算法在稀疏数据集和显式、隐式评分数据集上的推荐效果和评分预测准确率都更优。     

CageR: Cage‐Based Reverse Engineering of Animated 3D Shapes

We present CageR: A novel framework for converting animated 3D shape sequences into compact and stable cage‐based representations. Given a raw animated sequence with one‐to‐one point correspondences together with an initial cage embedding, our algorithm automatically generates smoothly varying cage embeddings which faithfully reconstruct the enclosed object deformation. Our technique is fast, automatic, oblivious to the cage coordinate system, provides controllable error and exploits a GPU implementation. At the core of our method, we introduce a new algebraic algorithm based on maximum volume sub‐matrices (maxvol) to speed up and stabilize the deformation inversion. We also present a new spectral regularization algorithm that can apply arbitrary regularization terms on selected subparts of the inversion spectrum. This step allows to enforce a highly localized cage regularization, guaranteeing its smooth variation along the sequence. We demonstrate the speed, accuracy and robustness of our framework on various synthetic and acquired data sets. The benefits of our approach are illustrated in applications such as animation compression and post‐editing.     

A data hiding scheme using the varieties of pixel-value differencing in multimedia images

In this paper, a capacity promoting technique is proposed for embedding data in an image using pixel-value differencing (PVD). The PVD scheme embeds data by changing the difference value between two adjacent pixels so that more data is embedded into two pixels located in the edge area, than in the smooth area. In order to increase the embedding capacity, a new approach is proposed in this paper by searching edge area more flexibly. Instead of processing a pair of pixels at a time as proposed by Wu and Tsai, two pairs of pixels in a block are processed at the same time. In addition, we proposed a pixel-value shifting scheme to further increase the chances for embedding data. Our scheme exploits the edge areas more efficiently, thus leading to an increase in embedding capacity as shown by experimental results compared to Wu and Tsai's method. Also, the embedding result of our scheme passes the Fridrich et al.’s detection. Besides, according to the distribution of difference values, more practical range partitions are suggested for improving capacity.     

Security and Robustness Enhancement for Image Data Hiding

In many applications, data hiding can be viewed as a tradeoff between capacity, robustness (against attacks), and embedding induced distortion. In this paper, we consider a fourth parameter: the security of the hidden information. Specifically, we propose a hash-based randomized embedding algorithm (HRE) that increases the security of the hidden data. We then optimize this algorithm against JPEG attacks. We derive a mathematical expression for the security of our algorithm, using which we show that the security of our algorithm can be increased independent of capacity, robustness, and embedding induced distortion. The maximum security depends only on the length of the key sequence, which is limited only by the size of the host image. Using a joint security and capacity measure, we show that the proposed scheme performs better than current secure quantization based data hiding schemes. We also derive the optimal value of distortion compensation factor of the HRE algorithm against JPEG compression attack. Experimental results show that the operating points achieved by the proposed scheme are 7 dB better than current blind data hiding schemes against the JPEG attack     

Distance learning for similarity estimation

In this paper, we present a general guideline to find a better distance measure for similarity estimation based on statistical analysis of distribution models and distance functions. A new set of distance measures are derived from the harmonic distance, the geometric distance, and their generalized variants according to the Maximum Likelihood theory. These measures can provide a more accurate feature model than the classical Euclidean and Manhattan distances. We also find that the feature elements are often from heterogeneous sources that may have different influence on similarity estimation. Therefore, the assumption of single isotropic distribution model is often inappropriate. To alleviate this problem, we use a boosted distance measure framework that finds multiple distance measures which fit the distribution of selected feature elements best for accurate similarity estimation. The new distance measures for similarity estimation are tested on two applications: stereo matching and motion tracking in video sequences. The performance of boosted distance measure is further evaluated on several benchmark data sets from the UCI repository and two image retrieval applications. In all the experiments, robust results are obtained based on the proposed methods.     

RiMOM: A Dynamic Multistrategy Ontology Alignment Framework

Ontology alignment identifies semantically matching entities in different ontologies. Various ontology alignment strategies have been proposed; however, few systems have explored how to automatically combine multiple strategies to improve the matching effectiveness. This paper presents a dynamic multistrategy ontology alignment framework, named RiMOM. The key insight in this framework is that similarity characteristics between ontologies may vary widely. We propose a systematic approach to quantitatively estimate the similarity characteristics for each alignment task and propose a strategy selection method to automatically combine the matching strategies based on two estimated factors. In the approach, we consider both textual and structural characteristics of ontologies. With RiMOM, we participated in the 2006 and 2007 campaigns of the Ontology Alignment Evaluation Initiative (OAEI). Our system is among the top three performers in benchmark data sets.     

Kernelized locality-sensitive hashing

Fast retrieval methods are critical for many large-scale and data-driven vision applications. Recent work has explored ways to embed high-dimensional features or complex distance functions into a low-dimensional Hamming space where items can be efficiently searched. However, existing methods do not apply for high-dimensional kernelized data when the underlying feature embedding for the kernel is unknown. We show how to generalize locality-sensitive hashing to accommodate arbitrary kernel functions, making it possible to preserve the algorithm's sublinear time similarity search guarantees for a wide class of useful similarity functions. Since a number of successful image-based kernels have unknown or incomputable embeddings, this is especially valuable for image retrieval tasks. We validate our technique on several data sets, and show that it enables accurate and fast performance for several vision problems, including example-based object classification, local feature matching, and content-based retrieval.     

MR-SimLab: Scalable subgraph selection with label similarity for big data

With the increasing size and complexity of available databases, existing machine learning and data mining algorithms are facing a scalability challenge. In many applications, the number of features describing the data could be extremely high. This hinders or even could make any further exploration infeasible. In fact, many of these features are redundant or simply irrelevant. Hence, feature selection plays a key role in helping to overcome the problem of information overload especially in big data applications. Since many complex datasets could be modeled by graphs of interconnected labeled elements, in this work, we are particularly interested in feature selection for subgraph patterns. In this paper, we propose MR-SimLab, a MapReduce-based approach for subgraph selection from large input subgraph sets. In many applications, it is easy to compute pairwise similarities between labels of the graph nodes. Our approach leverages such rich information to measure an approximate subgraph matching by aggregating the elementary label similarities between the matched nodes. Based on the aggregated similarity scores, our approach selects a small subset of informative representative subgraphs. We provide a distributed implementation of our algorithm on top of the MapReduce framework that optimizes the computational efficiency of our approach for big data applications. We experimentally evaluate MR-SimLab on real datasets. The obtained results show that our approach is scalable and that the selected subgraphs are informative.     

Graph grammars with neighbourhood-controlled embedding

A central feature that distinguishes graph grammars (we consider grammars generating sets of node-labelled undirected graphs only) from string grammars is that in the former one has to provide a mechanism by which a daughter graph (the right-hand side of a production) can be embedded in the rest of the mother graph, while in the latter this embedding is provided automatically by the structure that all strings possess (left-to-right orientation). In this paper we consider a possible classification of embedding mechanisms for (node-rewriting) graph grammars. This classification originates from the basic ideas of [9]. On the one hand it allows one to fit a number of existing notions of a graph grammar into a common framework and on the other hand it points out new “natural” possibilities for defining the embedding mechanism in a graph grammar. The relationship between the graph-language generating power of graph grammars using various embedding mechanisms is established.     

Efficient many-to-many feature matching under the l1 norm

Matching configurations of image features, represented as attributed graphs, to configurations of model features is an important component in many object recognition algorithms. Noisy segmentation of images and imprecise feature detection may lead to graphs that represent visually similar configurations that do not admit an injective matching. In previous work, we presented a framework which computed an explicit many-to-many vertex correspondence between attributed graphs of features configurations. The framework utilized a low distortion embedding function to map the nodes of the graphs into point sets in a vector space. The Earth Movers Distance (EMD) algorithm was then used to match the resulting points, with the computed flows specifying the many-to-many vertex correspondences between the input graphs. In this paper, we will present a distortion-free embedding, which represents input graphs as metric trees and then embeds them isometrically in the geometric space under the l₁ norm. This not only improves the representational power of graphs in the geometric space, it also reduces the complexity of the previous work using recent developments in computing EMD under l₁. Empirical evaluation of the algorithm on a set of recognition trials, including a comparison with previous approaches, demonstrates the effectiveness and robustness of the proposed framework.     

Semi-automatic computer aided lesion detection in dental X-rays using variational level set

A semi-automatic lesion detection framework is proposed to detect areas of lesions from periapical dental X-rays using level set method. In this framework, first, a new proposed competitive coupled level set method is used to segment the image into three pathologically meaningful regions using two coupled level set functions. Tailored for the dental clinical setting, a two-stage clinical segmentation acceleration scheme is used. The method uses a trained support vector machine (SVM) classifier to provide an initial contour for two coupled level sets. Then, based on the segmentation results, an analysis scheme is applied. Firstly, the scheme builds an uncertainty map from which those areas with radiolucent will be automatically emphasized by a proposed color emphasis scheme. Those radiolucent in the teeth or jaw usually suggested possible lesions. Secondly, the scheme employs a method based on the average intensity profile to isolate the teeth and locate two types of lesions: periapical lesion (PL) and bifurcation lesion (BL). Experimental results show that our proposed segmentation method is able to segment the image into pathological meaningful regions for further analysis; our proposed framework is able to automatically provide direct visual cues for the lesion detection; and when given the orientation of the teeth, it is able to automatically locate the PL and BL with a seriousness level marked for further dental diagnosis. When used in the clinical setting, the framework enables dentist to improve interpretation and to focus their attention on critical areas.     

A parallel scheme for accelerating parameter sweep applications on a GPU

This paper proposes a parallel scheme for accelerating parameter sweep applications on a graphics processing unit. By using hundreds of cores on the graphics processing unit, we found that our scheme simultaneously processes multiple parameters rather than a single parameter. The simultaneous sweeps exploit the similarity of computing behaviors shared by different parameters, thus allowing memory accesses to be coalesced into a single access if similar irregularities appear among the parameters’ computational tasks. In addition, our scheme reduces the amount of off‐chip memory access by unifying the data that are commonly referenced by multiple parameters and by placing the unified data in the fast on‐chip memory. In several experiments, we applied our scheme to practical applications and found that our scheme can perform up to 8.5 times faster than a naive scheme that processes a single parameter at a time. We also include a discussion on application characteristics that are required for our scheme to outperform the naive scheme. Copyright © 2013 John Wiley & Sons, Ltd.     

Histogram-based embedding for learning on statistical manifolds

A novel binning and learning framework is presented for analyzing and applying large data sets that have no explicit knowledge of distribution parameterizations, and can only be assumed generated by the underlying probability density functions (PDFs) lying on a nonparametric statistical manifold. For models’ discretization, the uniform sampling-based data space partition is used to bin flat-distributed data sets, while the quantile-based binning is adopted for complex distributed data sets to reduce the number of under-smoothed bins in histograms on average. The compactified histogram embedding is designed so that the Fisher–Riemannian structured multinomial manifold is compatible to the intrinsic geometry of nonparametric statistical manifold, providing a computationally efficient model space for information distance calculation between binned distributions. In particular, without considering histogramming in optimal bin number, we utilize multiple random partitions on data space to embed the associated data sets onto a product multinomial manifold to integrate the complementary bin information with an information metric designed by factor geodesic distances, further alleviating the effect of over-smoothing problem. Using the equipped metric on the embedded submanifold, we improve classical manifold learning and dimension estimation algorithms in metric-adaptive versions to facilitate lower-dimensional Euclidean embedding. The effectiveness of our method is verified by visualization of data sets drawn from known manifolds, visualization and recognition on a subset of ALOI object database, and Gabor feature-based face recognition on the FERET database.     

Effective Tag Recommendation System Based on Topic Ontology Using Wikipedia and WordNet

In this paper, we proposed a novel approach based on topic ontology for tag recommendation. The proposed approach intelligently generates tag suggestions to blogs. In this approach, we construct topic ontology through enriching the set of categories in existing small ontology called as Open Directory Project. To construct topic ontology, a set of topics and their associated semantic relationships is identified automatically from the corpus‐based external knowledge resources such as Wikipedia and WordNet. The construction relies on two folds such as concept acquisition and semantic relation extraction. In the first fold, a topic‐mapping algorithm is developed to acquire the concepts from the semantic of Wikipedia. A semantic similarity‐clustering algorithm is used to compute the semantic similarity measure to group the set of similar concepts. The second is the semantic relation extraction algorithm, which derives associated semantic relations between the set of extracted topics from the lexical patterns between synsets in WordNet. A suitable software prototype is created to implement the topic ontology construction process. A Jena API framework is used to organize the set of extracted semantic concepts and their corresponding relationship in the form of knowledgeable representation of Web ontology language. Thus, Protégé tool provides the platform to visualize the automatically constructed topic ontology successfully. Using the constructed topic ontology, we can generate and suggest the most suitable tags for the new resource to users. The applicability of topic ontology with a spreading activation algorithm supports efficient recommendation in practice that can recommend the most popular tags for a specific resource. The spreading activation algorithm can assign the interest scores to the existing extracted blog content and tags. The weight of the tags is computed based on the activation score determined from the similarity between the topics in constructed topic ontology and content of the existing blogs. High‐quality tags that has the highest activation score is recommended to the users. Finally, we conducted experimental evaluation of our tag recommendation approach using a large set of real‐world data sets. Our experimental results explore and compare the capabilities of our proposed topic ontology with the spreading activation tag recommendation approach with respect to the existing AutoTag mechanism. And also discuss about the improvement in precision and recall of recommended tags on the data sets of Delicious and BibSonomy. The experiment shows that tag recommendation using topic ontology results in the folksonomy enrichment. Thus, we report the results of an experiment mean to improve the performance of the tag recommendation approach and its quality.