Semi-supervised Model Research Articles

Neural machine translation for low resource Indian language: Hindi-Kangri

Neural Machine Translation (NMT) for low resource languages is a challenging task due to unavailability of large parallel corpus. The efficacy of Transformer based NMT models largely depends on scale of the parallel corpus and the configuration of hyperparameters implemented during model training. This study aims to delve into and elucidate the impact of hyperparameters on the performance of NMT models for low resource languages. To accomplish this, a series of experiments are conducted using an open-source Hindi-Kangri corpus to train both supervised and semi-supervised NMT models. Throughout the experimentation process, a significant number of discrepancies were identified within the data-set, necessitating manual correction. The best translation performance evaluated with respect to the metrics such as BLEU (0–1), SacreBLEU (0–100), Chrf (0–100), Chrf+ (0–100), Chrf++ (0–100) and TER (%) is (0.15, 14.98, 41.43, 41.49, 38.77, 68.20) for Hindi to Kangri direction, and (0.283, 28.17, 49.71, 50.64, 48.63, 51.25) for Kangri to Hindi direction.

Reconstruction of elasticity modulus distribution base on semi-supervised neural network

Accurate reconstruction of tissue elasticity modulus distribution has always been an important challenge in ultrasound elastography. Considering that existing deep learning-based supervised reconstruction methods only use simulated displacement data with random noise in training, which cannot fully provide the complexity and diversity brought by in-vivo ultrasound data, this study introduces the use of displacement data obtained by tracking in-vivo ultrasound radio frequency signals (i.e., real displacement data) during training, employing a semi-supervised approach to enhance the prediction accuracy of the model. Experimental results indicate that in phantom experiments, the semi-supervised model augmented with real displacement data provides more accurate predictions, with mean absolute errors and mean relative errors both around 3%, while the corresponding data for the fully supervised model are around 5%. When processing real displacement data, the area of prediction error of semi-supervised model was less than that of fully supervised model. The findings of this study confirm the effectiveness and practicality of the proposed approach, providing new insights for the application of deep learning methods in the reconstruction of elastic distribution from in-vivo ultrasound data.

Does the Gender of Doctors Change a Patient’s Perception?

ABSTRACT In the past, researchers have used various data sources like social media, admission applications, or letters of recommendation to identify gender-based differences in linguistic data. One such avenue in healthcare is the online physician reviews website. Such websites, for example, RateMyDoctors.com, ZocDoc, or even Yelp.com have become a go-to place for patients when choosing their physicians. In the current research, we used two different natural language processing (NLP) approaches: semi-supervised and unsupervised topic modeling to analyze the text of the reviews to identify gender-based linguistic differences from patients’ perspectives. We found that female physicians receive more reviews on their personable skills and warmth, aligning with the Stereotype Content Model. We also found other popular topics discussing bedside manners and overall patient experiences, where the reviews suggested that patients were happier with their experience with female physicians and perceived them to have more positive traits than their male counterparts. Although our study did not reflect significant linguistic differences; it highlights the importance for patients and doctors to be more aware of potential gender stereotypes and perceptions.

Semi-supervised learning model synergistically utilizing labeled and unlabeled data for failure detection in optical networks

In optical networks, reliable failure detection is essential for maintaining quality of service. The methodology has evolved from traditional performance threshold-driven approaches to contemporary data-driven AI algorithms, predominantly employing supervised and unsupervised learning. However, with the advent of second-level telemetry, optical transport networks have amassed a wealth of unlabeled performance data, while labeled data remains limited due to the intensive effort required for annotation. In this scenario, to address the challenges of scarce labeled data in supervised learning and the accuracy issues in unsupervised methods, we propose an OpenFE-VIME semi-supervised model. This model synergizes the robustness of supervised approaches with the flexibility of unsupervised approaches. It not only leverages the abundant reservoir of unlabeled data but also addresses the challenges posed by the limited availability of labeled data, enabling reliable and efficient failure detection. Upon evaluation using performance data from OTN node devices in the operator’s optical backbone network, the OpenFE-VIME model demonstrates remarkable performance, achieving an F1-score of 0.947 and accuracy of 0.946, while significantly reducing false negative and false positive rates to 0.073 and 0.035, respectively. Moreover, our research explores the model’s capabilities in utilizing both labeled and unlabeled data and investigates the threshold for training convergence across various data ratios. Additionally, the model’s internal mechanisms and decision-making processes are interpreted using t-SNE visualization, offering enhanced insights into its operational efficacy.

EEG power spectra parameterization and adaptive channel selection towards semi-supervised seizure prediction

Background:The seizure prediction algorithms have demonstrated their potential in mitigating epilepsy risks by detecting the pre-ictal state using ongoing electroencephalogram (EEG) signals. However, most of them require high-density EEG, which is burdensome to the patients for daily monitoring. Moreover, prevailing seizure models require extensive training with significant labeled data which is very time-consuming and demanding for the epileptologists. Method:To address these challenges, here we propose an adaptive channel selection strategy and a semi-supervised deep learning model respectively to reduce the number of EEG channels and to limit the amount of labeled data required for accurate seizure prediction. Our channel selection module is centered on features from EEG power spectra parameterization that precisely characterize the epileptic activities to identify the seizure-associated channels for each patient. The semi-supervised model integrates generative adversarial networks and bidirectional long short-term memory networks to enhance seizure prediction. Results:Our approach is evaluated on the CHB-MIT and Siena epilepsy datasets. With utilizing only 4 channels, the method demonstrates outstanding performance with an AUC of 93.15% on the CHB-MIT dataset and an AUC of 88.98% on the Siena dataset. Experimental results also demonstrate that our selection approach reduces the model parameters and training time. Conclusions:Adaptive channel selection coupled with semi-supervised learning can offer the possible bases for a light weight and computationally efficient seizure prediction system, making the daily monitoring practical to improve patients’ quality of life.

Statistical Inference For Noisy Matrix Completion Incorporating Auxiliary Information

This article investigates statistical inference for noisy matrix completion in a semi-supervised model when auxiliary covariates are available. The model consists of two parts. One part is a low-rank matrix induced by unobserved latent factors; the other part models the effects of the observed covariates through a coefficient matrix which is composed of high-dimensional column vectors. We model the observational pattern of the responses through a logistic regression of the covariates, and allow its probability to go to zero as the sample size increases. We apply an iterative least squares (LS) estimation approach in our considered context. The iterative LS methods in general enjoy a low computational cost, but deriving the statistical properties of the resulting estimators is a challenging task. We show that our method only needs a few iterations, and the resulting entry-wise estimators of the low-rank matrix and the coefficient matrix are guaranteed to have asymptotic normal distributions. As a result, individual inference can be conducted for each entry of the unknown matrices. We also propose a simultaneous testing procedure with multiplier bootstrap for the high-dimensional coefficient matrix. This simultaneous inferential tool can help us further investigate the effects of covariates for the prediction of missing entries. Supplementary materials for this article are available online, including a standardized description of the materials available for reproducing the work.

3D medical image segmentation based on semi-supervised learning using deep co-training

In recent years, artificial intelligence has been applied to 3D COVID-19 medical image diagnosis, which reduces detection costs and missed diagnosis rates with higher predictive accuracy, and diagnostic efficiency. However, the limited size and low quality of clinical 3D medical image samples have hindered the segmentation performance of 3D models. Therefore, we propose a 3D medical image segmentation model based on semi-supervised learning using co-training. Multi-view and multi-modal images are generated using spatial flipping and windowing techniques to enhance the spatial diversity of 3D image samples. A pseudo label generation module based on confidence-weights is employed to generate reliable pseudo labels for non-annotated data, thereby increasing the sample size and reducing overfitting. The proposed approach utilizes a three-stage training process: firstly, training a single network based on annotated data; secondly, incorporating non-annotated data to train a dual-modal network and generate pseudo labels; finally, jointly training six models in three dimensions using both annotated and pseudo labels generated from multi-view and multi-modal images, aiming to enhance segmentation accuracy and generalization performance. Additionally, a consistency regularization loss is applied to reduce noises and accelerate convergence of the training. Moreover, a heatmap visualization method is employed to focus on the attention of features at each stage of training, providing effective reference for clinical diagnosis. Experiments were conducted on an open dataset of 3D COVID-19 CT samples and a non-annotated dataset from TCIA, including 771 NIFTI-format CT images from 661 COVID-19 patients. The results of 5-fold cross-validation show that the proposed model achieves a segmentation accuracy of Dice=73.30 %, ASD=10.633, Sensitivity=63.00 %, and Specificity=99.60 %. Compared to various typical semi-supervised learning 3D segmentation models, it demonstrates better segmentation accuracy and generalization performance.

Reject inference in credit scoring based on cost-sensitive learning and joint distribution adaptation method

As traditional credit evaluation methods generally only use accepted sample modeling, the rejected data is omitted, which means the model's prediction of new customers is biased. However, reject inference can be used to solve this credit evaluation sample selection bias. This paper proposes a new reject inference method based on joint distribution adaptation (JDA) and cost-sensitive semi-supervised support vector machines (CS4VM). First, this method uses both accepted (labeled) samples and rejected (unlabeled) samples modeling, which overcomes the deviations in traditional credit evaluation methods. Second, as the accepted sample and the rejected sample distributions are different, this method reduces the distribution differences between the accepted and rejected sample sets, which ensures that the sample data conforms to the basic assumptions in the semi-supervised model, and improves the performance of the classification model. Third, this method reduces the overall cost in the actual credit business by considering both the traditional misclassification costs when mining the default samples and the different decision weights for the accepted and rejected samples. Finally, an empirical study verifies the excellent predictive performance of the proposed method and effectively reduces the total credit costs.

Landslide susceptibility assessment based on frequency ratio and semi-supervised heterogeneous ensemble learning model.

Epistemic uncertainty in data-driven landslide susceptibility assessment often tends to be increased by the limited accuracy of an individual model, as well as uncertainties associated with the selection of non-landslide samples. To address these issues, this paper centers on the landslide disaster in Ji'an City, China, and proposes a heterogeneous ensemble learning method incorporating frequency ratio (FR) and semi-supervised sample expansion. Based on the superimposed results of 12 environmental factor frequency ratios (FFR), non-landslide samples were selected and input into light gradient boosting machine (LightGBM), random forest (RF), and convolutional neural network (CNN) models for prediction along with historical landslide samples. The predicted probability values are integrated by four heterogeneous ensemble strategies to expand samples from high-confidence results. The model's performance is evaluated using the area under the receiver operating characteristic curve (AUC), partition frequency ratio (PFR), and other verification methods. The results demonstrate that the negative sample based on FFR sampling is more accurate than the random sampling method, and the FR-SSELR model based on frequency ratio sampling and semi-supervised ensemble strategy exhibits the highest performance (AUC = 0.971, ACC = 0.941). A more reasonable landslide susceptibility map was drawn based on this model, with the lowest percentage of landslides in the low and very low susceptibility zones (sum of PFR = 0.194), as well as the highest percentage of landslides in the high and very high susceptibility zones (sum of PFR = 6.800). Furthermore, the FR-SSELR model improved economic benefits by 3.82-14.2%, offering valuable guidance for decision-making regarding landslide management and the sustainability of Ji'an City.

A novel semi-supervised model for pre-impact fall detection with limited fall data

Due to the high prevalence and severe consequences of falls among older people, pre-impact fall detection algorithms which aim to detect falls before body-ground impacts are of critical importance. The success of existing algorithms based on wearable sensors relies heavily on numerous labeled data from both falls and activities of daily living (ADLs). However, fall data is much more difficult to acquire compared with ADLs, and moreover, labeling falling period apart from ADLs is laborious and time-consuming. To this end, we proposed a novel semi-supervised model for pre-impact fall detection, called Semi-PFD, which maximizes the utilization of easily obtainable ADL data and reduces the dependency on labeled fall data. This model was evaluated on two large-scale public fall datasets (KFall and SisFall) and compared with its supervised baseline under different ratios of fall data. Cross-validation results showed that Semi-PFD outperformed the supervised baseline across all conditions on both datasets. More importantly, Semi-PFD achieved considerable improvement of 2–4% F1-score when the ratios of fall data were low (≤23.1%). Remarkably, with only 76.9% of fall data, Semi-PFD achieved comparable and even higher F1-scores than the state-of-the-art benchmark models which utilized 100% of fall data. We further observed that the incorporation of unsupervised training in Semi-PFD mitigated the typical overconfidence problem associated with supervised training without affecting lead time. These findings underscore Semi-PFD's practical potential, reducing the burden of fall data collection and labeling while maintaining superior performance.

Multilingual Hate Speech Detection: A Semi-Supervised Generative Adversarial Approach.

Social media platforms have surpassed cultural and linguistic boundaries, thus enabling online communication worldwide. However, the expanded use of various languages has intensified the challenge of online detection of hate speech content. Despite the release of multiple Natural Language Processing (NLP) solutions implementing cutting-edge machine learning techniques, the scarcity of data, especially labeled data, remains a considerable obstacle, which further requires the use of semisupervised approaches along with Generative Artificial Intelligence (Generative AI) techniques. This paper introduces an innovative approach, a multilingual semisupervised model combining Generative Adversarial Networks (GANs) and Pretrained Language Models (PLMs), more precisely mBERT and XLM-RoBERTa. Our approach proves its effectiveness in the detection of hate speech and offensive language in Indo-European languages (in English, German, and Hindi) when employing only 20% annotated data from the HASOC2019 dataset, thereby presenting significantly high performances in each of multilingual, zero-shot crosslingual, and monolingual training scenarios. Our study provides a robust mBERT-based semisupervised GAN model (SS-GAN-mBERT) that outperformed the XLM-RoBERTa-based model (SS-GAN-XLM) and reached an average F1 score boost of 9.23% and an accuracy increase of 5.75% over the baseline semisupervised mBERT model.

Randomized feature selection based semi-supervised latent Dirichlet allocation for microbiome analysis

Health and disease are fundamentally influenced by microbial communities and their genes (the microbiome). An in-depth analysis of microbiome structure that enables the classification of individuals based on their health can be crucial in enhancing diagnostics and treatment strategies to improve the overall well-being of an individual. In this paper, we present a novel semi-supervised methodology known as Randomized Feature Selection based Latent Dirichlet Allocation (RFSLDA) to study the impact of the gut microbiome on a subject’s health status. Since the data in our study consists of fuzzy health labels, which are self-reported, traditional supervised learning approaches may not be suitable. As a first step, based on the similarity between documents in text analysis and gut-microbiome data, we employ Latent Dirichlet Allocation (LDA), a topic modeling approach which uses microbiome counts as features to group subjects into relatively homogeneous clusters, without invoking any knowledge of observed health status (labels) of subjects. We then leverage information from the observed health status of subjects to associate these clusters with the most similar health status making it a semi-supervised approach. Finally, a feature selection technique is incorporated into the model to improve the overall classification performance. The proposed method provides a semi-supervised topic modelling approach that can help handle the high dimensionality of the microbiome data in association studies. Our experiments reveal that our semi-supervised classification algorithm is effective and efficient in terms of high classification accuracy compared to popular supervised learning approaches like SVM and multinomial logistic model. The RFSLDA framework is attractive because it (i) enhances clustering accuracy by identifying key bacteria types as indicators of health status, (ii) identifies key bacteria types within each group based on estimates of the proportion of bacteria types within the groups, and (iii) computes a measure of within-group similarity to identify highly similar subjects in terms of their health status.

A semi-supervised mixture model of visual language multitask for vehicle recognition

Effective vehicle detection is extremely vital in cities. However, existing vehicle recognition models exhibit limited functionality and have not been effectively integrated with language models to aid in vehicle monitoring within cities. To address these challenges, this paper proposes a semi-supervised mixture model with detection, segmentation, and question-answering capabilities, referred to as the Multi-task Visual Language Model (MTVLM), for vehicle recognition. Further, a vehicle dataset containing 8425 vehicle images from 13 different categories is introduced, which is collected across various urban scenarios and weather conditions. MTVLM consists of a detector, a language model, and a segmentation model. The detector YOLODS is based on an improved You Only Look Once algorithm (YOLOv4), which has a lower number of parameters and higher detection accuracy. The detection results of YOLODS are then used as prompts for the pre-trained language model, a distilled version of Bidirectional Encoder Representations from Transformers (DistilBERT), to answer questions related to the detected vehicles. Furthermore, the coordinates of bounding boxes obtained by YOLODS are input into the Segment Anything Model (SAM) as prompts to segment vehicles from images. Experimental results show that MTVLM achieves a mean average precision (mAP) of 97.12 % on the vehicle dataset, surpassing other state-of-the-art (SOTA) detectors. The high-precision detector also ensures highly accurate text prompts for input to DistilBERT and SAM. By employing text prompts, the potential and performance of the pre-trained models DistilBERT and SAM can be fully explored and utilized to better complete question-answering and segmentation tasks.

Semi-supervised multiview fuzzy broad learning

Semi-supervised learning models often rely on restricted assumptions, and can easily suffer from covariate shift or noise. Few studies have investigated the use of fuzzy rule-based methods in the semi-supervised discipline. To improve model accuracy against covariate shift and to introduce fuzzy methods for interpretability, we first build a semi-supervised fuzzy broad learning model named SSFBLS, which employs a Mean-Teacher framework. Then, a trusted multiview semi-supervised classification method, termed TMSSC, is proposed by integrating the SSFBLS with a multiview fusion network to enhance the robustness of the model. Under the Mean-Teacher framework, SSFBLS involves the Takagi-Sugeno-Kang fuzzy model which can effectively deal with imprecision and uncertainty, and broad learning system which has strong learning ability and high computational efficiency. TMSSC utilizes a trusted mechanism to blend multiple views, so as to enhance its learning ability in semi-supervised scenarios. Experiments on the benchmark datasets demonstrate that the proposed methods have better anti-noise ability, competitive classification accuracy, as well as fast running speed.

Semi-supervised latent diffusion model for Biliary Atresia class-imbalanced image recognition

Biliary Atresia (BA) is a congenital condition affecting the bile ducts in newborn babies worldwide, typically diagnosed by the color of infant stools in clinical practice. Due to the limited availability of pediatricians and the lack of effective diagnostic tools for parents, computer vision is devised as a non-invasive and convenient means for early detection of BA. However, the limited recognition accuracy of BA and the challenge of imbalanced classes in real-world scenarios restrict the generalizability of existing methods, posing a challenge to their decision-making performance. To address those challenges, we introduce a Latent Diffusion Biliary Atresia method (LDBA) to leverage imbalanced data to enhance the recognition of BA. Specifically, we propose a novel pre-trained Biliary Atresia network (PBNet) to predict labels and effectively filter out high-threshold data. Then, a new data augmentation method is proposed to focus attention on the color and anomaly points of images. Subsequently, a new label fusion method and dynamic thresholds strategy are devised to optimize the utilization of generated data by the Latent Diffusion Model (LDM). The entire process is trained in a semi-supervised manner. The experimental results indicate that our proposed LDBA outperforms competing methods with a superior performance of 78.43%. Our method can improve the accuracy and reliability of BA clinical diagnosis.

Structural entropy minimization combining graph representation for money laundering identification

Money laundering identification (MLI) is a challenging task for financial AI research and application due to its massive transaction volume, label sparseness, and label bias. Most of the existing MLI methods focus on individual-level abnormal behavior while neglecting the community factor that money laundering is a collaborative group crime. Furthermore, the massive volume of transactions and the issue of label shifting also impede the application of supervised or semi-supervised models. To this end, this paper proposes an efficient community-oriented algorithm, namely SEGE, to identify money laundering based on structural entropy minimization (SEM) with graph embedding in an unsupervised approach. Experiments on both a private real-world money laundering network and a public synthetic dataset show that our SEGE algorithm derives prominent performance and outperforms the parameterized learning-based graph representation methods. Moreover, we find that there are pervasive sub-communities in the real-world money laundering network. Based on our local algorithm, we propose a real combat strategy against the money laundering group, in which when we have several scattered suspicious accounts in the transaction network, we are able to retrieve the whole money laundering group by the union of sub-communities with both high precision and high recall rates.

Multi-rate modulation encoding via unsupervised learning for audio event detection

Technologies in healthcare, smart homes, security, ecology, and entertainment all deploy audio event detection (AED) in order to detect sound events in an audio recording. Effective AED techniques rely heavily on supervised or semi-supervised models to capture the wide range of dynamics spanned by sound events in order to achieve temporally precise boundaries and accurate event classification. These methods require extensive collections of labeled or weakly labeled in-domain data, which is costly and labor-intensive. Importantly, these approaches do not fully leverage the inherent variability and range of dynamics across sound events, aspects that can be effectively identified through unsupervised methods. The present work proposes an approach based on multi-rate autoencoders that are pretrained in an unsupervised way to leverage unlabeled audio data and ultimately learn the rich temporal dynamics inherent in natural sound events. This approach utilizes parallel autoencoders that achieve decompositions of the modulation spectrum along different bands. In addition, we introduce a rate-selective temporal contrastive loss to align the training objective with event detection metrics. Optimizing the configuration of multi-rate encoders and the temporal contrastive loss leads to notable improvements in domestic sound event detection in the context of the DCASE challenge.

Murmur identification and outcome prediction in phonocardiograms using deep features based on Stockwell transform

Traditionally, heart murmurs are diagnosed through cardiac auscultation, which requires specialized training and experience. The purpose of this study is to predict patients' clinical outcomes (normal or abnormal) and identify the presence or absence of heart murmurs using phonocardiograms (PCGs) obtained at different auscultation points. A semi-supervised model tailored to PCG classification is introduced in this study, with the goal of improving performance using time–frequency deep features. The study begins by investigating the behavior of PCGs in the time–frequency domain, utilizing the Stockwell transform to convert the PCG signal into two-dimensional time–frequency maps (TFMs). A deep network named AlexNet is then used to derive deep feature sets from these TFMs. In feature reduction, redundancy is eliminated and the number of deep features is reduced to streamline the feature set. The effectiveness of the extracted features is evaluated using three different classifiers using the CinC/Physionet challenge 2022 dataset. For Task I, which focuses on heart murmur detection, the proposed approach achieved an average accuracy of 93%, sensitivity of 91%, and F1-score of 91%. According to Task II of the CinC/Physionet challenge 2022, the approach showed a clinical outcome cost of 5290, exceeding the benchmark set by leading methods in the challenge.

Contemporary attitudes and beliefs on coronary artery calcium from social media using artificial intelligence

Coronary artery calcium (CAC) is a powerful tool to refine atherosclerotic cardiovascular disease (ASCVD) risk assessment. Despite its growing interest, contemporary public attitudes around CAC are not well-described in literature and have important implications for shared decision-making around cardiovascular prevention. We used an artificial intelligence (AI) pipeline consisting of a semi-supervised natural language processing model and unsupervised machine learning techniques to analyze 5,606 CAC-related discussions on Reddit. A total of 91 discussion topics were identified and were classified into 14 overarching thematic groups. These included the strong impact of CAC on therapeutic decision-making, ongoing non-evidence-based use of CAC testing, and the patient perceived downsides of CAC testing (e.g., radiation risk). Sentiment analysis also revealed that most discussions had a neutral (49.5%) or negative (48.4%) sentiment. The results of this study demonstrate the potential of an AI-based approach to analyze large, publicly available social media data to generate insights into public perceptions about CAC, which may help guide strategies to improve shared decision-making around ASCVD management and public health interventions.

Semi-supervised semantic labeling of remote sensing images with improved image-level selection retraining

In recent years, image semantic segmentation technology has developed rapidly, but image annotation usually requires a significant amount of human and financial resources, especially for remote sensing image annotation, which can be expensive and sometimes even unaffordable. To address this issue, this paper integrates the idea of curriculum learning into the self-training method and screens reliable pseudo-labels through computing image-level confidence, significantly reducing the confirmation error problem. Furthermore, the semi-supervised model in this paper combines implicit semantic enhancement with strong data augmentation, which can reduce the coupling between the teacher model and the student model’s prediction distribution and enhance the model’s robustness. Finally, the proposed semi-supervised method is experimentally verified using the ISPRS competition dataset and compared with existing state-of-the-art (SOTA) methods. Experimental results show that the proposed semi-supervised segmentation method achieves higher segmentation accuracy compared to self-training methods. Moreover, despite not using iterative training to simplify the training process, the proposed method still yields satisfactory segmentation results.