Search

The key idea of active learning is that it can perform better with less data or costs if a machine learner is allowed to choose the data actively. However, the relation between labeling cost and model performance is seldom studied in the literature. In this paper, we thoroughly study this problem and give a criterion called as cost-performance to balance this relation. Based on the criterion, a cost-driven active SSC algorithm is proposed, which can stop the active process automatically. Empirical results show that our method outperforms active SVM and co-EMT. [ABSTRACT FROM AUTHOR]

Given a frequency band, we propose a semi-supervised SVM algorithm to select a set of channels for motor imagery-based BCIs in this paper. Both training data with labels and test data without labels are used for training a classifier in this approach. Hence, it is suitable for the case of small training data set. To test our algorithm, it is applied to a BCI competition data set. Data analysis results demonstrate the effectiveness of our algorithm. [ABSTRACT FROM AUTHOR]

The k-nearest neighbor (KNN) classifier is a simple and effective method for image classification. However, its performance significantly depends on how the distance between samples is calculated. Therefore, learning an appropriate distance metric is the most important issue for the KNN-based classifiers. The distance metric can be learned from either labeled or unlabeled data. Labeled images are expensive to generate, while unlabeled images are abundant, and the label information is crucial for the performance of the learned metric. In this work, we present a semi-supervised method for learning the distance metric. We propose a semi-supervised extension to the Neighborhood Component Analysis (NCA) method, which is a supervised method especially tailored for KNN classifiers. Then, we use the learned distance metric to classify images using the KNN method. Experiment shows that our proposed method outperforms both the traditional supervised and unsupervised methods. [ABSTRACT FROM AUTHOR]

In a real stream environment, labeled data may be very scarce and labeling all data is very difficult and expensive. Our goal is to derive a model to predict future instances΄ label as accurately as possible. Active learning selectively labels instances and can tackle the challenges raised by highly dynamic nature of data stream, but it ignores the effect of unlabeled instances utilization that can help to strength supervised learning. In this paper, we propose a framework that combines active and semi-supervised learning to get advantage of both methods, to boost the performance of learning algorithm. This framework solves the active learning problem in addition to the challenges of evolving data streams. Experimental results on real data sets prove the effectiveness of our proposed framework. [ABSTRACT FROM AUTHOR]

An algorithm for intrusion detection based on improved evolutionary semi- supervised fuzzy clustering is proposed which is suited for situation that gaining labeled data is more difficulty than unlabeled data in intrusion detection systems. The algorithm requires a small number of labeled data only and a large number of unlabeled data and class labels information provided by labeled data is used to guide the evolution process of each fuzzy partition on unlabeled data, which plays the role of chromosome. This algorithm can deal with fuzzy label, uneasily plunges locally optima and is suited to implement on parallel architecture. Experiments show that the algorithm can improve classification accuracy and has high detection efficiency. [ABSTRACT FROM AUTHOR]

Color information carried by the matting components in spectral matting is used in our method to solve the matting problem. By comparing the color similarity of each matting component, the matting components were separated into two groups: the foreground and the background. There are two advantages of doing so. Firstly, there is no need to use the trimap which is a necessary tool in traditional matting solving methods to accomplish the process of semi-supervised learning. Instead, the color similarity information could be used as constrain or heuristic cue in finishing the process of unsupervised learning to lower the blindness of automatically extracting the foreground object. Secondly, combining the color similarity information with a little user provided foreground and background cue could further reduce the number of matting components whose group have not been known yet, and thus a semi-supervised learning process which could only be done by using a lot user provided cue in traditional matting solving methods could be accomplished. [ABSTRACT FROM AUTHOR]

Recommendation, in the filed of machine learning, is known as a technique of identifying user preferences to new items with ratings from recommender systems. Recently, one novel recommendation model using Green΄s function treats recommendation as the process of label propagation. Although this model outperforms many standard recommendation methods, it suffers from information loss during graph construction because of data sparsity. In this paper, aiming at solving this problem and improving prediction accuracy, we propose an enhanced semi-supervised Green΄s function recommendation model. The main contributions are two-fold: 1) To reduce information loss, we propose a novel graph construction method with global and local consistent similarity; 2) We enhance the recommendation algorithm with the multi-class semi-supervised learning framework. Finally, experimental results on real world data demonstrate the effectiveness of our model. [ABSTRACT FROM AUTHOR]

Label propagation (LP) is used in the framework of semi-supervised learning. In this paper, we propose a novel method of logistic label propagation (LLP). The proposed method employs logistic functions for accurately estimating the label values as the posterior probabilities. In LLP, the label of newly input sample is efficiently estimated by using the optimized coefficients in the logistic function, without such recomputation of all label values as in original LP. In the experiments on classification, the proposed method produced more reliable label values at the high degree of confidence than LP and ordinary logistic regression. In addition, even for a small portion of the labeled samples, the error rates by LLP were lower than those by the logistic regression. [ABSTRACT FROM AUTHOR]

Zadeh proposes that there are three basic concepts underlying human cognition: granulation, organization and causation and that a granule is a clump of points (objects) drawn together by indistinguishability, similarity, proximity or functionality. Tolerance relation can describe the concept of Granular systems. In this paper, a novel definition of Granular System(GS), which is described by metric function under the framework of tolerance relation, is presented, concepts are created upon GS, and we introduce semi-supervised learning into the Granular computing for concepts creating. For this purpose, a novel back propagation approach is developed for concepts learning. The experiment shows that the new BP is better than traditional EM algorithm when samples do not come from a random source, which has the density we want to estimate. [ABSTRACT FROM AUTHOR]

Semi-supervised learning has become an active area of recent research in machine learning. To date,many approaches to semi-supervised learning are presented. In this paper,Consistency method and its some variants are deeply studied. The proof about the important condition for convergence of consistency method is given in detail. Moreover,we further study the validity of some variants of consistency method. Finally we conduct the experimental study on the parameters involved in consistency method to face recognition. Meanwhile, the performance of Consistency method and its some variants are compared with that of support vector machine supervised learning methods. [ABSTRACT FROM AUTHOR]

To deal with the problem of sensitivity to noise in semi-supervised learning for biometrics, this paper proposes a robust Gaussian-Laplacian Regularized (GLR) framework based on maximum correntropy criterion (MCC), called GLR-MCC, along with its convergence analysis. The half quadratic (HQ) optimization technique is used to simplify the correntropy optimization problem to a standard semi-supervised problem in each iteration. Experimental results show that the proposed GRL-MCC can effectively improve the semi-supervised learning performance and is robust to mislabeling noise and occlusion as compared with GLR. [ABSTRACT FROM AUTHOR]

Graph-based semi-supervised learning strategy plays an important role in the semi-supervised learning area. This paper presents a novel label propagation algorithm based on nonnegative sparse representation (NSR) for bioinformatics and biometrics. Firstly, we construct a sparse probability graph (SPG) whose nonnegative weight coefficients are derived by nonnegative sparse representation algorithm. The weights of SPG naturally reveal the clustering relationship of labeled and unlabeled samples; meanwhile automatically select appropriate adjacency structure as compared to traditional semi-supervised learning algorithm. Then the labels of unlabeled samples are propagated until algorithm converges. Extensive experimental results on biometrics, UCI machine learning and TDT2 text datasets demonstrate that label propagation algorithm based on NSR outperforms the standard label propagation algorithm. [ABSTRACT FROM AUTHOR]

In order to alleviate user fatigue and improve the performance of interactive genetic algorithms(IGAs) in searching, we introduce a co-training semi-supervised learning(CSSL)algorithm into interval fitness IGAs with large and variational population size. The CSSL is adopted to model the user΄s preference so as to estimate abundant of unevaluated individuals΄ fitness. First, the method to select the labeled and unlabeled samples for CSSL is proposed according to the clustering results of the large size population. Combined with the approximation precision of two co-training learners, an efficient strategy for selecting high reliable unlabeled samples to label is given. Then, we adopt the CSSL mechanism to train two RBF neural networks for establishing the surrogate model with high precision and generalization. In the evolution, the surrogate model estimates individuals΄ fitness and it is managed to guarantee the approximation precision based on its estimation error. The proposed algorithm is applied to a fashion evolutionary design system, and the experimental results show its efficiency. [ABSTRACT FROM AUTHOR]

Multi-view semi-supervised learning is a hot research topic recently. In this paper, we consider the regularization problem in multi-view semi-supervised learning. A regularization method adaptive to the given data is proposed, which can use unlabeled data to adjust the degree of regularization automatically. This new regularization method comprises two levels of regularization simultaneously. Experimental evidence on real word dataset shows its effectivity. [ABSTRACT FROM AUTHOR]

We address the problem of within-network classification in sparsely labeled networks. Recent work has demonstrated success with statistical relational learning (SRL) and semi-supervised learning (SSL) on such problems. However, both approaches rely on the availability of labeled nodes to infer the values of missing labels. When few labels are available, the performance of these approaches can degrade. In addition, many such approaches are sensitive to the specific set of nodes labeled. So, although average performance may be acceptable, the performance on a specific task may not. We explore a complimentary approach to within-network classification, based on the use of label-independent (LI) features – i.e., features calculated without using the values of class labels. While previous work has made some use of LI features, the effects of these features on classification performance have not been extensively studied. Here, we present an empirical study in order to better understand these effects. Through experiments on several real-world data sets, we show that the use of LI features produces classifiers that are less sensitive to specific label assignments and can lead to performance improvements of over 40% for both SRL- and SSL-based classifiers. We also examine the relative utility of individual LI features; and show that, in many cases, it is a combination of a few diverse network-based structural characteristics that is most informative. [ABSTRACT FROM AUTHOR]

In the paper, we present an improved approach based on Semi-supervised Discriminant Analysis (SDA), called semi-supervised local discriminant embedding (SLDE), for reducing the dimensionality of the feature space. We take the manifold structure into account and try to learn a subspace in which the Euclidean distances can better reflect class structure of the images. The weight matrix and the scatter matrices in SDA are improved to make efficient use of both labeled and unlabeled images. After being embedded into a low-dimensional subspace, the similar images maintain their intrinsic neighbor relations, whereas the dissimilarity neighboring images no longer stick to one another. Experiments have been carried out to validate our approach. [ABSTRACT FROM AUTHOR]

This paper presents a novel semi-supervised approach that determines a linear predictor using Support Vector Machines (SVMs) and incorporates information on rejected loans, assuming that the labeled data (accepted applicants) and unlabeled data (rejected applicants) are not drawn from the same distribution. We use a self-training algorithm in order to predict how likely a rejected applicant would have repaid had the applicant received credit. A modification to the self-training algorithm based on Platt΄s probabilistic output for SVMs is introduced. Experiments with two toy data sets; one well-known benchmark Credit Scoring data set, and one project performed for a Chilean financial institution demonstrate that our approach accomplishes the best classification performance compared to well-known reject inference alternatives and another state-of-the-art semi-supervised method for SVMs (Transductive SVM). [ABSTRACT FROM AUTHOR]

Intrusion Detection Systems (IDSs) which have been deployed in computer networks to detect a wide variety of attacks are suffering how to manage of a large number of triggered alerts. Thus, reducing false alarms efficiently has become the most important issue in IDS. In this paper, we introduce the semi-supervised learning mechanism to build an alert filter, which will reduce up to 85% false alarms and still keep a high detection rate. In our semi-supervised learning approach, we only need a very small amount of label information. This will save a huge security officer΄s effort and make the alert filter be more practical for the real systems. Numerical comparison with conventional supervised learning approach with the same small portion labeled data, our method has significantly superior detection rate as well as in the false alarm reduction rate. [ABSTRACT FROM AUTHOR]

Text classification is essential for narrowing down the number of documents relevant to a particular topic for further pursual, especially when searching through large biomedical databases. Protein-protein interactions are an example of such a topic with databases being devoted specifically to them. This paper proposed a semi-supervised learning algorithm via local learning with class priors (LL-CP) for biomedical text classification where unlabeled data points are classified in a vector space based on their proximity to labeled nodes. The algorithm has been evaluated on a corpus of biomedical documents to identify abstracts containing information about protein-protein interactions with promising results. Experimental results show that LL-CP outperforms the traditional semi-supervised learning algorithms such as SVM and it also performs better than local learning without incorporating class priors. [ABSTRACT FROM AUTHOR]

In this paper, we propose the topic-constrained hierarchical clustering, which organizes document datasets into hierarchical trees consistant with a given set of topics. The proposed algorithm is based on a constrained agglomerative clustering framework and a semi-supervised criterion function that emphasizes the relationship between documents and topics and the relationship among documents themselves simultaneously. The experimental evaluation show that our algorithm outperformed the traditional agglomerative algorithm by 7.8% to 11.4%. [ABSTRACT FROM AUTHOR]

Semi-supervised learning has been the focus of machine learning and data mining research in the past few years. Various algorithms and techniques have been proposed, from generative models to graph-based algorithms. In this work, we focus on the Cluster-and-Label approaches for semi-supervised classification. Existing cluster-and-label algorithms are based on some underlying models and/or assumptions. When the data fits the model well, the classification accuracy will be high. Otherwise, the accuracy will be low. In this paper, we propose a refinement approach to address the model misfit problem in semi-supervised classification. We show that we do not need to change the cluster-and-label technique itself to make it more flexible. Instead, we propose to use successive refinement clustering of the dataset to correct the model misfit. A series of experiments on UCI benchmarking data sets have shown that the proposed approach outperforms existing cluster-and-label algorithms, as well as traditional semi-supervised classification techniques including Selftraining and Tri-training. [ABSTRACT FROM AUTHOR]

Many tasks of information extraction or natural language processing have a property that the data naturally consist of several views–disjoint subsets of features. Specifically, a semantic relationship can be represented with some entity pairs or contexts surrounding the entity pairs. For example, the Person-Birthplace relation can be recognized from the entity pair view, such as (Albert Einstein, Ulm), (Pablo Picasso, Malaga) and so on. On the other side, this relation can be identified with some contexts, such as ˵A was born in B″, ˵B, the birth place of A″ and so on. To leverage the unlabeled data in the training stage, semi-supervised learning has been applied to relation extraction task. In this paper, we propose a multi-view semi-supervised learning algorithm, Co-Label Propagation, to combine the `information΄ from both the entity pair view and the context view. In propagation process, the label scores of classes are spread not only in the entity pair view and the context view, but also between the two views. The proposed algorithm is evaluated using semantic relation classification tasks. The experiment results validate its effectiveness. [ABSTRACT FROM AUTHOR]

A novel version of spectral mapping for partially labeled sample classification is proposed in this paper. This new method adds the label information into the mapping process, and adopts the geodesic distance rather than Euclidean distance as the measure of the difference between two data points. The experimental results show that the proposed method yields significant benefits for partially labeled classification with respect to the previous methods. [ABSTRACT FROM AUTHOR]

We mainly generalize consistency method in semi-supervised learning by expanding kernel matrix,denoted by GCM(Generalized Consistency Method), and study its convergence. Aimed at GCM,we give the detailed proof for condition of convergence. Moreover,we further study the validity of some variants of GCM. Finally we conduct the experimental study on the parameters involved in GCM to face recognition. Meanwhile, the performance of GCM and its some variants are compared with that of support vector machine methods. [ABSTRACT FROM AUTHOR]

Fingerprint segmentation is an important pre-processing step in automatic fingerprint identification system. Traditional fingerprint segmentation methods either highly depend on empirical thresholds sophistically chosen by experts or a learned model trained by elements generated from manually segmented fingerprints. It is manpower and time consuming. They always try their best to tune their fingerprint segmentation methods to be universal to all unseen fingerprints. However, one fingerprint may have a significantly distinct distribution from another in feature space because fingerprint acquisition is affected by several factors, such as pressure, the types of sensors, finger tip condition (dry, wet etc.). As a result, the delicate threshold and the well trained model may not be suitable to the new input fingerprints from a new finger or a new person. And it makes worse when automatic fingerprint identification systems meet sensor interoperability. To solve the problem, we propose a personalized fingerprint segmentation method: Automatic Labeling based Linear Neighborhood Propagation (ALLNP), which learns a segmentation model special for each input fingerprint image based on the input image only. The proposed method is tested with typical fingerprint images from four heterogeneous data bases of FVC2000. Experimental results show its effectiveness and encouraging strength when fingerprint segmentation meets sensor interoperability. [ABSTRACT FROM AUTHOR]

Usually, semi-supervised learning requires a number of prior knowledge to supervise the learning process, such as, seeds in Seeded-Kmeans, pair-wise constraints in COP-Kmeans, and labeled data for training an initial useful classifier in S3VM. Such prior knowledge is generally provided by the domain expert, so it is very expensive. In this paper, we propose a new automatical document labeling strategy to derive much more prior knowledge based on the very limited labeled data and the whole data set. Experimental results on 20-Newsgroup text data have shown that the new strategy is helpful for semi-supervised document categorization and improves the learning performance. [ABSTRACT FROM AUTHOR]

In Independent Factor Analysis (IFA), latent components (or sources) are recovered from only their linear observed mixtures. Both the mixing process and the source densities (that are assumed to be generated according to mixtures of Gaussians) are learned from observed data. This paper investigates the possibility of estimating the IFA model in its noiseless setting when two kinds of prior information are incorporated: constraints on the mixing process and partial knowledge on the cluster membership of some examples. Semi-supervised or partially supervised learning frameworks can thus be handled. These two proposals have been initially motivated by a real-world application that concerns fault diagnosis of a railway device. Results from this application are provided to demonstrate the ability of our approach to enhance estimation accuracy and remove indeterminacy commonly encountered in unsupervised IFA such as source permutations. [ABSTRACT FROM AUTHOR]

In this paper we present a novel approach to multi–view object recognition based on kernel methods with constraints. Differently from many previous approaches, we describe a system that is able to exploit a set of views of an input object to recognize it. Views are acquired by cameras located around the object and each view is modeled by a specific classifier. The relationships among different views are formulated as constraints that are exploited by a sort of collaborative learning process. The proposed approach applies the constraints on unlabeled data in a semi–supervised framework. The results collected on the COIL benchmark show that constraint based learning can improve the quality of the recognition system and of each single classifier, both on the original and noisy data, and it can increase the invariance with respect to object orientation. [ABSTRACT FROM AUTHOR]

Within-network classification, where the goal is to classify the nodes of a partly labeled network, is a semi-supervised learning problem that has applications in several important domains like image processing, the classification of documents, and the detection of malicious activities. While most methods for this problem infer the missing labels collectively based on the hypothesis that linked or nearby nodes are likely to have the same labels, there are many types of networks for which this assumption fails, e.g., molecular graphs, trading networks, etc. In this paper, we present a collective classification method, based on relaxation labeling, that classifies entities of a network using their local structure. This method uses a marginalized similarity kernel that compares the local structure of two nodes with random walks in the network. Through experimentation on different datasets, we show our method to be more accurate than several state-of-the-art approaches for this problem. [ABSTRACT FROM AUTHOR]

Classification task involves inducing a predictive model using a set of labeled samples. The accuracy of the model usually increases as more labeled samples are available. When one has only few samples, the obtained model tends to offer poor results. Even when labeled samples are difficult to get, a lot of unlabeled samples are generally available on which unsupervised learning can be done. In this chapter, a way to combine supervised and unsupervised learning in order to use both labeled and unlabeled samples is explored. The efficiency of the method is evaluated on various UCI datasets and on the classification of a very high resolution remote sensing image when the number of labeled samples is very low. [ABSTRACT FROM AUTHOR]

Intravascular Ultrasound (IVUS) data validation is usually performed by comparing post-mortem (in-vitro) IVUS data and corresponding histological analysis of the tissue, obtaining a reliable ground truth. The main drawback of this method is the few number of available study cases due to the complex procedure of histological analysis. In this work we propose a novel semi-supervised approach to enhance the in-vitro training set by including examples from in-vivo coronary plaques data set. For this purpose, a Sequential Floating Forward Selection method is applied on in-vivo data and plaque characterization performances are evaluated by Leave-One-Patient-Out cross-validation technique. Supervised data inclusion improves global classification accuracy from 89.39% to 91.82%. [ABSTRACT FROM AUTHOR]

Support Vector Machines (SVMs) have been dominant learning techniques for more than ten years, and mostly applied to supervised learning problems. These years two-class unsupervised and semi-supervised classification algorithms based on Bounded C-SVMs, Bounded ν-SVMs, Lagrangian SVMs (LSVMs) and robust version to Bounded C − SVMs respectively, which are relaxed to Semi-definite Programming (SDP), get good classification results. But the parameter C in Bounded C-SVMs has no specific in quantification. Therefore we proposed robust version to unsupervised and semi-supervised classification algorithms based on Bounded ν− Support Vector Machines (Bν−SVMs). Numerical results confirm the robustness of proposed methods and show that our new algorithms based on robust version to Bν−SVM often obtain more accurate results than other algorithms. [ABSTRACT FROM AUTHOR]

In the k-nearest neighbor (KNN) classifier, nearest neighbors involve only labeled data. That makes it inappropriate for the data set that includes very few labeled data. In this paper, we aim to solve the classification problem by applying transduction to the KNN algorithm. We consider two groups of nearest neighbors for each data point – one from labeled data, and the other from unlabeled data. A kernel function is used to assign weights to neighbors. We derive the recurrence relation of neighboring data points, and then present two solutions to the classification problem. One solution is to solve it by matrix computation for small or medium-size data sets. The other is an iterative algorithm for large data sets, and in the iterative process an energy function is minimized. Experiments show that our solutions achieve high performance and our iterative algorithm converges quickly. [ABSTRACT FROM AUTHOR]

Semi-Supervised Learning is one of the most popular and emerging issues in Machine Learning. Since it is very costly to label large amounts of data, it is useful to use data sets without labels. For doing that, normally we uses Semi-Supervised Learning to improve the performance or efficiency of the classification algorithms. This paper intends to use the techniques of Semi-Supervised Learning to boost the performance of the Robust Alternating AdaBoost algorithm. We introduce the algorithm RADA+ and compare it with RADA, reporting the performance results using synthetic and real data sets, the latter obtained from a benchmark site. [ABSTRACT FROM AUTHOR]

In many real-world applications, there only exist very few labeled samples, while a large number of unlabeled samples are available. Therefore, it is difficult for some traditional semi-supervised algorithms to generate the useful classifiers to evaluate the labeling confidence of unlabeled samples. In this paper, a new semi-supervised classification based on clustering ensembles named SSCCE is proposed. It takes advantages of clustering ensembles to generate multiple partitions for a given dataset, and then uses the clustering consistency index to determine the labeling confidence of unlabeled samples. The algorithm can overcome some defects about the traditional semi-supervised classification algorithms, and enhance the performance of the hypothesis trained on very few labeled samples by exploiting a large number of unlabeled samples. Experiments carried out on ten public data sets from UCI machine learning repository show that this method is effective and feasible. [ABSTRACT FROM AUTHOR]

With the technologies of blind steganalysis becoming increasingly popular, a growing number of researchers concern in this domain. Supervised learning for classification is widely used, but this method is often time consuming and effort costing to obtain the labeled data. In this paper, an improved semi-supervised learning method: path-based transductive support vector machines (TSVM) algorithm with Mahalanobis distance is proposed for blind steganalysis classification, by using modified connectivity kernel matrix to improve the classification accuracy. Experimental results show that our proposed algorithm achieves the highest accuracy among all examined semi-supervised TSVM methods, especially for a small labeled data set. [ABSTRACT FROM AUTHOR]

Genetic interaction, in which two mutations have a combined effect not exhibited by either mutation alone, is a powerful and widespread tool for establishing functional linkages between genes. However, little is known about how genes genetic interact to produce phenotypes and the comprehensive identification of genetic interaction in genome-scale by experiment is a laborious and time-consuming work. In this paper, we present a computational method of system biology to analyze synthetic genetic interactions. We firstly constructed a high-quality functional gene network by integrating protein interaction, protein complex and microarray gene expression data together. Then we extracted the network properties such as network centrality degree, clustering coefficient, etc., which reflect the local connectivity and global position of a gene and are supposed to correlate with its functional properties. Finally we find relationships between synthetic genetic interactions and function network properties using the graph-based semi-supervised learning which incorporates labeled and unlabeled data together. Experimental results showed that Semi-supervised method outperformed standard supervised learning algorithms and reached 97.1% accuracy at a maximum. Especially, the semi-supervised method largely outperformed when the number of training samples is very small. [ABSTRACT FROM AUTHOR]

Dimension reduction is very important for applications in data mining and machine learning. Dependence maximization based supervised feature extraction (SDMFE) is an effective dimension reduction method proposed recently. A shortcoming of SDMFE is that it can only use labeled data, and does not work well when labeled data are limited. However, in many applications, it is a common case. In this paper, we propose a novel feature extraction method, called Semi-Supervised Dependence Maximization Feature Extraction (SSDMFE), which can utilize simultaneously both labeled and unlabeled data to perform feature extraction. The labeled data are used to maximize the dependence and the unlabeled data are used as regulations with respect to the intrinsic geometric structure of the data. Experiments on several datasets are presented and the results demonstrate that SSDMFE achieves much higher classification accuracy than SDMFE when the amount of labeled data are limited. [ABSTRACT FROM AUTHOR]

We often face the situation where very limited labeled data are available to learn an effective classifier in target domain while there exist large amounts of labeled data with similar feature or distribution in certain relevant domains. Transfer learning aims at improving the performance of a learner in target domain given labeled data in one or more source domains. In this paper, we present an algorithm to learn effective classifier without or a few labeled data on target domain, given some labeled data with same features and similar distribution in source domain. Our algorithm uses the data edit technique to approach distribution from the source domain to the target domain by removing ˵mismatched″ examples in source domain and adding ˵matched″ examples in target domain. Experimental results on email classification problem have confirmed the effectiveness of the proposed algorithm. [ABSTRACT FROM AUTHOR]

Active learning and semi-supervised learning are two approaches to alleviate the burden of labeling large amounts of data. In active learning, user is asked to label the most informative examples in the domain. In semi-supervised learning, labeled data is used together with unlabeled data to boost the performance of learning algorithms. We focus here to combine them together. We first introduce a new active learning strategy, then we propose an algorithm to take the advantage of both active learning and semi-supervised learning. We discuss several advantages of our method. Experimental results show that it is efficient and robust to noise. [ABSTRACT FROM AUTHOR]

A semi-supervised learning algorithm is proposed based on label propagation through submanifold. The algorithm assumes that samples lying in a local neighborhood share the same labels and the global labels changing among submanifolds is sufficiently smooth. The algorithm firstly introduces a k-nearest neighbor graph to describe local neighborhood among the data set. And then, a cost function and a constraint equation are proposed, which stand for the global smoothness of the class labels΄ changing and the labeled samples΄ information respectively. The final semi-supervised learning task is converted to a typical quadratic program, whose optimal solution can minimize the cost function and satisfy the supervised constraint. Experimental results of the algorithm on toy data, digit recognition, and text classification demonstrate the feasibility and efficiency of the proposed algorithm. [ABSTRACT FROM AUTHOR]

In this paper, a new co-training style semi-supervised algorithm is proposed, which employs Bagging based multimodal perturbation to label the unlabeled data. In detail, through perturbing the training data, input attributes and learning parameters together, the algorithm generates accurate but diversity k-nearest neighbor classifiers. These classifiers are refined using unlabeled examples which are labeled if the other classifiers agree on the labeling. Experimental results show that the semi-supervised algorithm could effectively improve the classification generalization by utilizing the unlabeled data. [ABSTRACT FROM AUTHOR]

Learning-based ranking is a promising approach to a variety of search tasks, which is aimed at automatically creating the ranking model based on training samples and machine learning techniques. However, the problem of lacking training samples labeled with relevancy degree or ranking orders is frequently encountered. To address this problem, we propose a novel graph-based learning to rank (GLRank) for video search by leveraging the vast amount of unlabeled samples. A relation graph is constructed by using sample (i.e., video shot) pairs rather than individual samples as vertices. Each vertex in this graph represents the ˵relevancy relation″ between two samples in a pair (i.e., which sample is more relevant to the given query). Such relevancy relation is discovered through a set of pre-trained concept detectors and then propagated among the pairs. When all the pairs, constructed with the samples to be searched, receive the propagated relevancy relation, a round robin criterion is proposed to obtain the final ranking list. We have conducted comprehensive experiments on automatic video search task over TRECVID 2005-2007 benchmarks and shown significant and consistent improvements over the other state-of-the-art ranking approaches. [ABSTRACT FROM AUTHOR]