Semi-supervised Learning Strategy Research Articles

Semi-MCNN: A Semisupervised Multi-CNN Ensemble Learning Method for Urban Land Cover Classification Using Submeter HRRS Images

Submeter high-resolution remote sensing image land cover classification could provide significant help for urban monitoring, management, and planning. Deep learning (DL)-based models have achieved remarkable performance in many land cover classification tasks through end-to-end supervised learning. However, the excellent performance of DL-based models relies heavily on a large number of well-annotated samples, which is impossible in practical land cover classification scenarios. Additionally, the training set could contain all of the different land cover types. To overcome these problems, in this article, a semisupervised multiple-CNN ensemble learning method, namely semi-MCNN, is proposed to solve the land cover classification problem. Considering the lack of labeled samples, a semisupervised learning strategy was adopted to leverage large amounts of unlabeled data. In the proposed approach, an automatic sample selection method called an ensembled teacher model dataset generation was adopted to select samples and generate a dataset from large amounts of unlabeled data automatically. To tackle the error propagation problem, an important strategy was adopted to correct the errors by pretraining on the selected unlabeled data, and fine-tuning on the labeled data. Moreover, the semisupervised idea together with the multi-CNN ensemble framework was integrated into an end-to-end architecture. This could significantly improve the generalization ability of the semisupervised model, as well as the classification accuracy. Experiments were conducted on Shenzhen's land cover data and two other public remote sensing datasets. These experiments confirmed the superior performance of the proposed semi-MCNN compared to the state-of-the-art land cover classification models.

Review Sharing via Deep Semi-Supervised Code Clone Detection

Code review as a typical type of user feedback has recently drawn increasing attentions for improving code quality. To carry out research on code review, sufficient review data is normally required. As a result, recent efforts commonly focus on analysis for projects with sufficient reviews (called “s-projects”), rather than projects with extremely few ones (called “f-projects”). Actually, through statistics on public platforms, the latter ones dominate open source software, in which novel approaches should be explored to improve their review-based code improvement. In this paper, we try to address the problem via building a review sharing channel where the informative review can be reasonably delivered from s-projects to the f-projects. To ensure the accuracy of shared reviews, we introduce a novel code clone detection model based on Convolutional Neural Network (CNN), and build suitable “s-projects, f-projects” pairs through the clone detection. Especially, to alleviate the dataset heterogeneity between the training and testing, an autoencoder-based semi-supervised learning strategy is employed. Furthermore, to improve the sharing experience, heuristic filtering tactics are applied to reduce the time cost. Meanwhile, the LDA (Latent Dirichlet Allocation)-based ranking algorithm is used for presenting diverse review themes. We have implemented the sharing channel as a prototype system RSharer+, which contains three representative modules: data preprocessing, code clone detection and review presentation. The collected datasets are first transformed into context-sensitive numerical vectors in the data proprecessing. Then in the clone detection, data vectors are trained and tested on the BigCloneBench and real code-review pairs. At last, the presentation module provides review classification and theme extraction for better sharing experience. Extensive comparative experiments on hundreds of real labelled code fragments demonstrate the precision of clone detection and the effectiveness of review sharing.

3D Segmentation of Pulmonary Nodules Based on Multi-View and Semi-Supervised

For large-scale CT images, the automatic segmentation of nodules is the foundation for diagnosis of various pulmonary diseases. Most existing methods have made great progress in pulmonary segmentation. But because of the similar structure between vessels and nodules in 2D, it lacks the ability to extract more discriminative features. The accuracy is still not satisfying. And the task remains challenging due to the lack of voxel labels and training strategies to balance foreground and background. To solve these problems, a 3D segmentation network of pulmonary nodules based on semi-supervised was proposed. Firstly, a framework of multi-view feature extraction was designed to solve the problem of high similarity between nodules and other tissues. It extracted features from three different views to improve precision. And three parallel dilated convolutions were added for multi-scale feature extraction. Hence, the spatial and semantic information of different sizes can be better obtained. Secondly, for the problem of identifying difficult samples, a hybrid loss function with an adjustment factor was proposed. It magnifies the loss of difficult samples, which will attract more attention from the network. And a new regularization term was introduced to avoid overfitting. The entire network was trained with a few labeled CT data set through an improved semi-supervised learning strategy, which was optimized with a new self-paced regularization. Experimental results show that the average sensitivity of the proposed method is 95.81%. It is superior to other methods in terms of precision and Dice index especially when the data set is not satisfied.

Deep nonsmooth nonnegative matrix factorization network with semi-supervised learning for SAR image change detection

In the paper, we propose a deep nonsmooth nonnegative matrix factorization (nsNMF) network with semi-supervised learning for synthetic aperture radar (SAR) image change detection. In most of the existing deep-NMF-based models, the nonnegative matrix is linearly decomposed layer by layer, which may fail to characterize the nonlinearities in complex data. As such, a nonlinear deep nsNMF model is first built for learning hierarchical, nonlinear, and localized data representations. Meanwhile, in view of its good generalization performance and low computational complexity, extreme learning machine (ELM) is integrated into the nonlinear deep nsNMF model to construct a deep nsNMF network for satisfactory classification. More importantly, since it is difficult to acquire more labeled samples in practice, semi-supervised learning strategy is proposed to make use of partially labeled data for training. The learning process of the proposed network consists of pretraining stage and fine-tuning stage, in which the former pretrains all decomposed matrices layer by layer and the latter aims to reduce the total reconstruction error by using the mini-batch gradient descent algorithm. The experimental results on four pairs of SAR images demonstrate the effectiveness of the proposed method.

Driver Danger-Level Monitoring System Using Multi-Sourced Big Driving Data

Danger-level analysis is widely used to prevent potential driving risks based on driving performance. Such analysis is essential for monitoring driver performance. Moreover, danger-level analysis is vital for automotive safety systems and driving assistance applications. However, danger-level analysis that simultaneously considers driver-, vehicle-, and road-related information from driving data has rarely been conducted. Such analysis is very challenging due to the issues associated with the high volume and high variety in multisourced driving data. In this paper, we propose a novel danger-level analysis framework for dealing with high variety and high volume problems of multisourced driving data. Built upon a feature extraction method in the proposed framework, we first profile multisourced driving features for overcoming the variety problem. Next, danger-level analysis is formulated as a multiobjective pursuit problem in a linear model. The problem is then solved using a semisupervised learning strategy to overcome the volume issue. Therefore, the danger level can be accurately estimated from multisourced driving data by using the proposed framework. The experimental results indicate that the proposed framework outperforms existing machine learning techniques for multisourced driving data.

Semi-supervised learning in multivariate calibration

Multivariate calibration has historically been associated with supervised learning where a model developed/learned from a set of labeled data is used to predict characteristics of an object by using measurements obtained from it. For example, in analytical chemistry, the characteristics are often related to composition, some form of spectroscopy is commonly used to acquire measurements, and the object is typically a sample of some material of interest. Recently in this and other contexts there has been an increased interest in semi-supervised learning where a combination of both labeled and unlabeled data are used to help produce a predictive model. The use of semi-supervised learning can be advantageous in certain situations in multivariate calibration involving both forward and inverse modeling approaches. Potential advantages of models developed via semi-supervised learning are illustrated for the two modeling approaches via a series of simulations that are derived from near-infrared reflectance spectra. The basis and conditions for benefits of a semi-supervised learning strategy for multivariate calibration are identified and discussed. In the case of both modeling approaches, the primary advantage of semi-supervised learning was found to be a reduction in conditional prediction bias. This advantage is most likely to be realized when the quantity of labeled data is small, the quantity of unlabeled data is large, and when the values of the characteristics to be predicted are distant from the centroid of associated values of the labeled data.

ALG: Adaptive low-rank graph regularization for scalable semi-supervised and unsupervised learning

Graph-based semi-supervised learning (SSL) is one of the most popular topics in the past decades. Most graph-based SSL methods utilize two stage-approach to infer the class labels of the unlabeled data, where the first stage is to construct a graph in an unsupervised way for characterizing the geometry of data manifold while the second stage is to perform SSL for classification. However, the graph construction and SSL stages are usually separate. They do not share any common information to enhance the performance of classification. In this paper, we aim to solve the above problem by proposing a unified framework for SSL. In detail, we first adopt an adaptive low-rank model for graph construction, where the coefficient matrix is calculated through an efficient procedure so that the corresponding constructed graph can capture the global structure of data manifold. Meriting from such a graph, we then propose a unified framework for SSL, where we have involved the graph construction, the SSL strategy into the whole objective function. As a result, the class labels learned by SSL can provide more discriminative information for graph construction, while the updated graph can further enhance the classification accuracies of SSL. Simulation indicates that the proposed method can achieve better classification performance compared with other state-of-the-art graph-based methods.

Direct Cup-to-Disc Ratio Estimation for Glaucoma Screening via Semi-Supervised Learning.

Glaucoma is a chronic eye disease that leads to irreversible vision loss. The Cup-to-Disc Ratio (CDR) serves as the most important indicator for glaucoma screening and plays a significant role in clinical screening and early diagnosis of glaucoma. In general, obtaining CDR is subjected to measuring on manually or automatically segmented optic disc and cup. Despite great efforts have been devoted, obtaining CDR values automatically with high accuracy and robustness is still a great challenge due to the heavy overlap between optic cup and neuroretinal rim regions. In this paper, a direct CDR estimation method is proposed based on the well-designed semi-supervised learning scheme, in which CDR estimation is formulated as a general regression problem while optic disc/cup segmentation is cancelled. The method directly regresses CDR value based on the feature representation of optic nerve head via deep learning technique while bypassing intermediate segmentation. The scheme is a two-stage cascaded approach comprised of two phases: unsupervised feature representation of fundus image with a convolutional neural networks (MFPPNet) and CDR value regression by random forest regressor. The proposed scheme is validated on the challenging glaucoma dataset Direct-CSU and public ORIGA, and the experimental results demonstrate that our method can achieve a lower average CDR error of 0.0563 and a higher correlation of around 0.726 with measurement before manual segmentation of optic disc/cup by human experts. Our estimated CDR values are also tested for glaucoma screening, which achieves the areas under curve of 0.905 on dataset of 421 fundus images. The experiments show that the proposed method is capable of state-of-the-art CDR estimation and satisfactory glaucoma screening with calculated CDR value.

Information Fusion and Semi-Supervised Deep Learning Scheme for Diagnosing Gear Faults in Induction Machine Systems

There has been an increasing interest in the design of intelligent diagnostic systems for industrial applications. The key requirement in the design of practical diagnostic systems is the ability for decision making in high-dimensional feature spaces, where the prior knowledge about the system states in terms of labels is very limited. Moreover, the problem of diagnosing simultaneous defects is rarely addressed on real industrial applications. This paper aims to develop a semi-supervised deep-learning scheme for diagnosing multiple defects including simultaneous ones in a gearbox directly connected to an induction machine shaft. This scheme consists of two main modules: information fusion and decision making. The former integrates captured multiple sensory streams into a very high dimensional feature space. The latter uses a semi-supervised deep learning procedure to minimize the human interaction during the training and maximize the diagnostic efficiency. This scheme facilitates learning and diagnosing defects under harshest conditions 1) where only a few number of labeled samples are collected together with a large number of unlabeled samples, and 2) in a very high-dimensional feature space. Several state-of-the-art semi-supervised and supervised learners have also been included in the scheme, enabling a comparative experiment for diagnosing simultaneous defects.

Nonlinear, flexible, semisupervised learning scheme for face beauty scoring

Automatic facial beauty scoring in images is an emerging research topic in face-based biometrics. All existing methods adopt fully supervised schemes. We introduce the use of semisupervised learning schemes for solving the problem of face beauty scoring. The paper has two main contributions. First, instead of using fully supervised techniques, we show that graph-based score propagation methods can enrich model learning without the need of additional labeled face images. Second, we propose a nonlinear flexible manifold embedding for solving the score propagation. This model can be used for transductive and inductive settings. The proposed semisupervised schemes were tested on three recent public datasets for face beauty analysis: SCUT-FBP, M2B, and SCUT-FBP5500. These experiments, as well as many comparisons with supervised schemes, show that the nonlinear semisupervised scheme compares favorably with many supervised schemes. They also show that its performances in terms of error prediction and Pearson correlation are better than those reported for the used datasets.

Probabilistic Representation and Inverse Design of Metamaterials Based on a Deep Generative Model with Semi-Supervised Learning Strategy.

The research of metamaterials has achieved enormous success in the manipulation of light in a prescribed manner using delicately designed subwavelength structures, so-called meta-atoms. Even though modern numerical methods allow for the accurate calculation of the optical response of complex structures, the inverse design of metamaterials, which aims to retrieve the optimal structure according to given requirements, is still a challenging task owing to the nonintuitive and nonunique relationship between physical structures and optical responses. To better unveil this implicit relationship and thus facilitate metamaterial designs, it is proposed to represent metamaterials and model the inverse design problem in a probabilistically generative manner, enabling to elegantly investigate the complex structure-performance relationship in an interpretable way, and solve the one-to-many mapping issue that is intractable in a deterministic model. Moreover, to alleviate the burden of numerical calculations when collecting data, a semisupervised learning strategy is developed that allows the model to utilize unlabeled data in addition to labeled data in an end-to-end training. On a data-driven basis, the proposed deep generative model can serve as a comprehensive and efficient tool that accelerates the design, characterization, and even new discovery in the research domain of metamaterials, and photonics in general.

A Novel Semisupervised Deep Learning Method for Human Activity Recognition

Human activity recognition (HAR) based on inertial sensors has been investigated for many industrial informatics applications, such as healthcare and ubiquitous computing. Existing methods mainly rely on supervised learning schemes, which require large labeled training data. However, labeled data are sometimes difficult to acquire, while unlabeled data are readily available. Thus, we intend to make use of both labeled and unlabeled data with semisupervised learning for accurate HAR. In this paper, we propose a semisupervised deep learning approach, using temporal ensembling of deep long short-term memory, to recognize human activities with smartphone inertial sensors. With the deep neural network processing, features are extracted for local dependencies in the recurrent framework. Besides, with an ensemble approach based on both labeled and unlabeled data, we can combine together the supervised and unsupervised losses, so as to make good use of unlabeled data that the supervised learning method cannot leverage. Experimental results indicate the effectiveness of our proposed semisupervised learning scheme, when compared to several state-of-the-art semisupervised learning approaches.

Vehicle re-identification in still images: Application of semi-supervised learning and re-ranking

Vehicle re-identification (re-ID), namely, finding exactly the same vehicle from a large number of vehicle images, remains a great challenge in computer vision. Most existing vehicle re-ID approaches follow a fully-supervised learning methodology, in which sufficient labeled training data is required. However, this limits their scalability to realistic applications, due to the high cost of data labeling. In this paper, we adopted a Generative Adversarial Network (GAN) to generate unlabeled samples and enlarge the training set. A semi-supervised learning scheme with the Convolutional Neural Networks (CNN) was proposed accordingly, which assigns a uniform label distribution to the unlabeled images to regularize the supervised model and improve the performance of the vehicle re-ID system. Besides, an improved re-ranking method based on the Jaccard distance and k-reciprocal nearest neighbors is proposed to optimize the initial rank list. Extensive experiments over the benchmark datasets VeRi-776, VehicleID and VehicleReID have demonstrated that the proposed method outperforms the state-of-the-art approaches for vehicle re-ID.

Malware Detection via Extended Label Propagation Through Graph Inference

In this paper, we model the malware detection problem as a graph inference problem, and develop a novel belief propagation approach within a semi-supervised learning scheme that fully makes use of files' and hosts' connections to detect malware. Specifically, with network download data, we build a large graph that depicts files' co-occurrence and files-hosts relationship. Different from the classical methods that heuristically define edge weights only in the file co-occurrence graph, we develop a new method to integrate homophilic host-file relationship on top of file co-occurrences. Then, by using the linear neighborhood model, we first perform propagations in the subgraph of files to achieve their stabilization, then extend the propagation to the complete file-host graph. To facilitate this propagation procedure, we develop a set of algorithmic tools that extract information for the linear neighborhood model from the link structure of download events. Also, we theoretically show that, under some mild conditions, our propagation method could reveal the actual labels of unlabeled nodes in the complete graph. Finally, we perform a set of experiments that demonstrate the effectiveness of our new method in a variety of contexts on a real-world dataset.

Variational Auto-Encoder Based Semi-supervised Learning Scheme for Learner Classification in Intelligent Tutoring System

Variational Auto-Encoder Based Semi-supervised Learning Scheme for Learner Classification in Intelligent Tutoring System

Deep learning architecture for the recursive patterns recognition model

In this work, we propose a deep learning approach for the recursive pattern recognition model, called AR2P (for its acronym in Spanish: “Algoritmo Recursivo de Reconocimiento de Patrones”), by extending its supervised learning capability towards a semi-supervised learning scheme. The deep learning architecture is composed of three phases: the first one, called discovery phase, discovers the atomic descriptors. The second one, called aggregation phase, creates a feature hierarchy (merge of descriptors) from atomic descriptors. Finally, the classification phase carries out the classification of the inputs based on the feature hierarchy. The last phase uses a supervised learning approach, while the first two follow an unsupervised learning approach. In this paper is tested the performance of the proposed model, using a dataset from the UCI Machine Learning Repository.

Semi-global Parameterization of Online Handwriting Features for Characterizing Early-Stage Alzheimer and Mild Cognitive Impairment

BackgroundBecause of the rich set of spatiotemporal features it allows to extract, online handwriting is being increasingly investigated for characterizing neurodegenerative diseases like Parkinson and Alzheimer. The state of the art on the latter is dominated by methods that extract global (average) kinematic parameters, and then apply basic classification techniques or standard statistical tests to assess the statistical significance of each parameter in discriminating a pathological population from a healthy control one. MethodsWe propose a new approach for characterizing Early-Stage Alzheimer disease (ES-AD), and Mild Cognitive Impairment (MCI) w.r.t Healthy Controls (HC) that, instead of considering average kinematic HW parameters, which discards the dynamics related to each subject, is based on a semi-global parameterization scheme encoding the distribution of each kinematic parameter over a fixed number of bins. Such a distribution characterizes the gross dynamics associated with each parameter. A semi-supervised learning is proposed, in which a Normalized Mutual Information (NMI) selection scheme guides a hierarchical clustering algorithm to choose the best tradeoff between the number of clusters and the discriminative power of each w.r.t to the three cognitive profiles. ResultsFor both global and semi-global parameters, the semi-supervised learning scheme uncovers clusters with two trends, one cluster that consists essentially of HC and MCI, and one cluster essentially composed of MCI and ES-AD. The clusters obtained with semi-global parameters are more informative than those with global parameters as reflected by a better NMI value. ConclusionA semi-global parametrization of handwriting spatiotemporal parameters allows for a better discrimination between the HC, MCI and ES-AD profiles, than a global one does. Unlike the latter, the former encodes the distribution of the dynamics of each parameter, which offers a larger parameter space in which discrimination is easier.

A semi-supervised deep learning method based on stacked sparse auto-encoder for cancer prediction using RNA-seq data

Background and objective: Cancer has become a complex health problem due to its high mortality. Over the past few decades, with the rapid development of the high-throughput sequencing technology and the application of various machine learning methods, remarkable progress in cancer research has been made based on gene expression data. At the same time, a growing amount of high-dimensional data has been generated, such as RNA-seq data, which calls for superior machine learning methods able to deal with mass data effectively in order to make accurate treatment decision.Methods: In this paper, we present a semi-supervised deep learning strategy, the stacked sparse auto-encoder (SSAE) based classification, for cancer prediction using RNA-seq data. The proposed SSAE based method employs the greedy layer-wise pre-training and a sparsity penalty term to help capture and extract important information from the high-dimensional data and then classify the samples.Results: We tested the proposed SSAE model on three public RNA-seq data sets of three types of cancers and compared the prediction performance with several commonly-used classification methods. The results indicate that our approach outperforms the other methods for all the three cancer data sets in various metrics.Conclusions: The proposed SSAE based semi-supervised deep learning model shows its promising ability to process high-dimensional gene expression data and is proved to be effective and accurate for cancer prediction.

A positive and unlabeled learning framework based on extreme learning machine for drug-drug interactions discovery

Drug-drug interactions (DDIs) lead to Adverse Drug Reactions (ADRs) in most cases, which increase medical costs tremendously, and may cause medical negligence or even fatal accidents. Many researchers introduce binary classification methods into DDIs discovery task to avoid expensive and inefficient experimental trials. However, we find that drug-drug pairs without DDI labels cannot be simply viewed as negative samples in practice, since unobserved DDIs may be discovered from unlabeled pairs. Therefore, different from traditional positive-negative classification, in this paper, we treat DDIs discovery as a positive-unlabeled learning (PU learning) problem, in which each drug-drug pair is either a positive sample (labeled as DDI) or an unlabeled one (labeled as non-DDI). We propose a PU learning framework based on Extreme Learning Machine (ELM) named PU-ELM, which consists of two components, namely reliable negative extraction and classifier learning. To improve the quality of reliable negative set, we propose a negative extraction method named OC-ELM-RN (One-Class ELM for Reliable Negative extraction). As to the learning module, we first apply the semi-supervised learning strategy in PU-ELM (denoted as PU-ELM-SS) to achieve extremely fast learning speed, and then we propose an entropy based weighted learning method named PU-ELM-EW (PU-ELM with Entropy Weighted learning) to improve the PU learning performance. Extensive experiments are conducted on a synthetic dataset with varied settings. We also apply PU-ELM to a real-world drug dataset. The results indicate that PU-ELM achieves better DDIs discovery ability compared with state-of-the-art methods.

Approximate linear dependence criteria with active learning for smart soft sensor design

As a semi-supervised machine learning strategy, active learning has recently been introduced into the soft sensing area for the performance enhancement and the save of human efforts. Active learning is capable to automatically select the most informative unlabeled samples for labeling, thus the costs related to human annotation can be reduced. Instead of randomly labeling data samples, in this paper, we employ kernel approximate linear dependence (ALD) to evaluate each unlabeled data samples, and the data samples with large evaluation values are labeled for model updating. Comparative study results show that the ALD based active learning strategy driven soft sensor obtains better prediction performance than the random selection strategy driven soft sensor.