Labeled Data Set Research Articles

Data-Centric AI for Healthcare Fraud Detection.

Automated methods for detecting fraudulent healthcare providers have the potential to save billions of dollars in healthcare costs and improve the overall quality of patient care. This study presents a data-centric approach to improve healthcare fraud classification performance and reliability using Medicare claims data. Publicly available data from the Centers for Medicare & Medicaid Services (CMS) are used to construct nine large-scale labeled data sets for supervised learning. First, we leverage CMS data to curate the 2013-2019 Part B, Part D, and Durable Medical Equipment, Prosthetics, Orthotics, and Supplies (DMEPOS) Medicare fraud classification data sets. We provide a review of each data set and data preparation techniques to create Medicare data sets for supervised learning and we propose an improved data labeling process. Next, we enrich the original Medicare fraud data sets with up to 58 new provider summary features. Finally, we address a common model evaluation pitfall and propose an adjusted cross-validation technique that mitigates target leakage to provide reliable evaluation results. Each data set is evaluated on the Medicare fraud classification task using extreme gradient boosting and random forest learners, multiple complementary performance metrics, and 95% confidence intervals. Results show that the new enriched data sets consistently outperform the original Medicare data sets that are currently used in related works. Our results encourage the data-centric machine learning workflow and provide a strong foundation for data understanding and preparation techniques for machine learning applications in healthcare fraud.

A novel density ratio‐based batch active learning fault diagnosis method integrated with adaptive Laplacian graph trimming

AbstractIn actual industrial processes, although a large number of original data are easy to obtain, only a few samples are effectively labelled, which is insufficient to construct a supervised fault diagnostic model. Facing the industrial demand of fault diagnosis, in this paper, a novel density ratio (DR)‐based batch active learning (BAL) fault diagnosis method integrated with adaptive Laplacian graph trimming (ALGT) method is proposed. First, under the active learning framework, a new index DR‐based on local reachability density (LRD) is proposed to search the low density and high uncertainty samples, in which the local outliers factor (LOF) is used to search the samples in low density region and the ratio of LRD and intra‐class LRD is calculated to search the samples with high uncertainty. Second, the samples are selected and manually labelled in batches according to the proposed index DR, and the labelled data set and the unlabelled data set are updated and reconstructed. Third, based on the reconstructed labelled dataset and remaining unlabelled dataset, a semi‐supervised classifier ALGT is constructed for fault diagnosis. In ALGT, the Laplacian weighted graph is initialized and iteratively optimized by ALGT. Finally, the proposed DR‐based BAL‐ALGT (DRBAL‐ALGT) fault diagnosis method is verified by the Tennessee Eastman process (TEP) and applied to grid‐connected photovoltaic systems (GPVS). The experimental results show that the proposed DRBAL‐ALGT method can achieve higher accuracy for fault diagnosis.

Unsupervised Clustering and Supervised Machine Learning for Lightning Classification: Application to Identifying EIPs for Ground‐Based TGF Detection

AbstractWe developed a framework merging unsupervised and supervised machine learning to classify lightning radio signals, and applied it to the possible detection of terrestrial gamma‐ray flashes (TGFs). Recent studies have established a tight connection between energetic in‐cloud pulses (EIPs, >150 kA) and a subset of TGFs, enabling continuous and large‐scale ground‐based TGF detection. However, even with a high peak current threshold, it is time‐consuming to manually search for EIPs in a background of many non‐EIP events, and it becomes even more difficult when a lower peak current threshold is used. Machine learning classifiers are an effective tool. Beginning with unsupervised learning, spectral clustering is performed on the low‐dimensional features extracted by an autoencoder from raw radio waveforms, showing that +EIPs naturally constitute a distinct class of waveform and 6%–7% of the total population. The clustering results are used to form a labeled data set (∼10,000 events) to further train supervised convolutional neural network (CNN) that targets for +EIPs. Our CNN models identify on average 95.2% of true +EIPs with accuracy up to 98.7%, representing a powerful tool for +EIP classification. The pretrained CNN classifier is further applied to identify lower peak current EIPs (LEIPs, >50 kA) from a larger data set (∼30,000 events). Among 10 LEIPs coincident with Fermi TGF observations, 2 previously reported TGFs and 2 unreported but suspected TGFs are found, while the majority are not associated with detectable TGFs. In addition, unsupervised clustering is found to reflect characteristics of the ionosphere reflection height and its effect on radio wave propagation.

Cotton Stubble Detection Based on Improved YOLOv3

The stubble after cotton harvesting was used as the detection object to achieve the visual navigation operation for residual film recovery after autumn. An improved (You Only Look Once v3) YOLOv3-based target detection algorithm was proposed to detect cotton stubble. First, field images of residual film recycling were collected. Considering the inconsistency between stubble size and shape, a segmented labeling data set of stubble is proposed. Secondly, the Darknet-53 backbone of the original YOLOv3 network is improved to accommodate tiny targets. Next, the prediction anchor box of the improved detection backbone is clustered using K-means++, and the size of the prediction anchor box suitable for improved YOLOv3 is determined. Finally, for the false detection points after detection, a mean value denoising method is used to remove the false detection points. Feature points are extracted from the denoised stubble, and the candidate points are fitted by the least square method to obtain the navigation line. The optimal model with a mean average precision (mAP) of 0.925 is selected for testing at the test stage. The test results show that the algorithm in this article can detect the stubble of residual film recovery images at different locations, different time periods, and different camera depression angles without misdetection. The detection speed of a single image is 98.6 ms. Based on an improvement over YOLOv3, the improved model has a significantly higher detection rate in different scenarios than YOLOv3. This can provide practical technical support for the visual navigation of residual film recovery.

Semisupervised semantic segmentation for seismic interpretation

Seismic interpretation plays an essential role in locating subsurface horizons and understanding geologic formations. Traditionally, constructing facies models for the entire reservoir often requires domain experts’ manual interpretations of many seismic images, which is time-consuming. Although labeled horizons are limited, we have many unlabeled seismic images. Therefore, we develop a semisupervised segmentation framework that uses unlabeled seismic data through a reconstruction loss to learn a robust encoder and improve horizon label predictions. To further mitigate the limited data problem, we incorporate data augmentation using the time-shifting method to mimic similar deformations in a reservoir. Finally, we investigate the prediction uncertainty using deep ensembles. Results indicate that unlabeled seismic data help us learn a better latent representation and achieve a higher prediction accuracy and a lower prediction uncertainty than solely using labeled data sets. We also compare semisupervised and supervised semantic segmentation further and understand when semisupervised learning performs better in terms of the number of labels, facies, and locations of predicted sections. We develop an active learning framework to label the most valuable unlabeled sections given the uncertainty estimation. We believe our work helps geophysicists reduce the amount of labeling efforts and achieve a higher facies classification accuracy with the same amount of labeling work.

Modern Artificial Neural Networks: Is Evolution Cleverer?

Machine learning tools, particularly artificial neural networks (ANN), have become ubiquitous in many scientific disciplines, and machine learning-based techniques flourish not only because of the expanding computational power and the increasing availability of labeled data sets but also because of the increasingly powerful training algorithms and refined topologies of ANN. Some refined topologies were initially motivated by neuronal network architectures found in the brain, such as convolutional ANN. Later topologies of neuronal networks departed from the biological substrate and began to be developed independently as the biological processing units are not well understood or are not transferable to in silico architectures. In the field of neuroscience, the advent of multichannel recordings has enabled recording the activity of many neurons simultaneously and characterizing complex network activity in biological neural networks (BNN). The unique opportunity to compare large neuronal network topologies, processing, and learning strategies with those that have been developed in state-of-the-art ANN has become a reality. The aim of this review is to introduce certain basic concepts of modern ANN, corresponding training algorithms, and biological counterparts. The selection of these modern ANN is prone to be biased (e.g., spiking neural networks are excluded) but may be sufficient for a concise overview.

A Near-Real-Time Flood Detection Method Based on Deep Learning and SAR Images

Owning to the nature of flood events, near-real-time flood detection and mapping is essential for disaster prevention, relief, and mitigation. In recent years, the rapid advancement of deep learning has brought endless possibilities to the field of flood detection. However, deep learning relies heavily on training samples and the availability of high-quality flood datasets is rather limited. The present study collected 16 flood events in the Yangtze River Basin and divided them into three categories for different purpose: training, testing, and application. An efficient methodology of dataset-generation for training, testing, and application was proposed. Eight flood events were used to generate strong label datasets with 5296 tiles as flood training samples along with two testing datasets. The performances of several classic convolutional neural network models were evaluated with those obtained datasets, and the results suggested that the efficiencies and accuracies of convolutional neural network models were obviously higher than that of the threshold method. The effects of VH polarization, VV polarization, and the involvement of auxiliary DEM on flood detection were investigated, which indicated that VH polarization was more conducive to flood detection, while the involvement of DEM has a limited effect on flood detection in the Yangtze River Basin. Convolutional neural network trained by strong datasets were used in near-real-time flood detection and mapping for the remaining eight flood events, and weak label datasets were generated to expand the flood training samples to evaluate the possible effects on deep learning models in terms of flood detection and mapping. The experiments obtained conclusions consistent with those previously made on experiments with strong datasets.

Sentiment analysis of clinical narratives: A scoping review

A clinical sentiment is a judgment, thought or attitude promoted by an observation with respect to the health of an individual. Sentiment analysis has drawn attention in the healthcare domain for secondary use of data from clinical narratives, with a variety of applications including predicting the likelihood of emerging mental illnesses or clinical outcomes. The current state of research has not yet been summarized. This study presents results from a scoping review aiming at providing an overview of sentiment analysis of clinical narratives in order to summarize existing research and identify open research gaps. The scoping review was carried out in line with the PRISMA-ScR (Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews) guideline. Studies were identified by searching 4 electronic databases (e.g., PubMed, IEEE Xplore) in addition to conducting backward and forward reference list checking of the included studies. We extracted information on use cases, methods and tools applied, used datasets and performance of the sentiment analysis approach. Of 1,200 citations retrieved, 29 unique studies were included in the review covering a period of 8 years. Most studies apply general domain tools (e.g. TextBlob) and sentiment lexicons (e.g. SentiWordNet) for realizing use cases such as prediction of clinical outcomes; others proposed new domain-specific sentiment analysis approaches based on machine learning. Accuracy values between 71.5–88.2% are reported. Data used for evaluation and test are often retrieved from MIMIC databases or i2b2 challenges. Latest developments related to artificial neural networks are not yet fully considered in this domain. We conclude that future research should focus on developing a gold standard sentiment lexicon, adapted to the specific characteristics of clinical narratives. Efforts have to be made to either augment existing or create new high-quality labeled data sets of clinical narratives. Last, the suitability of state-of-the-art machine learning methods for natural language processing and in particular transformer-based models should be investigated for their application for sentiment analysis of clinical narratives.

An Empirical Survey of Data Augmentation for Limited Data Learning in NLP

AbstractNLP has achieved great progress in the past decade through the use of neural models and large labeled datasets. The dependence on abundant data prevents NLP models from being applied to low-resource settings or novel tasks where significant time, money, or expertise is required to label massive amounts of textual data. Recently, data augmentation methods have been explored as a means of improving data efficiency in NLP. To date, there has been no systematic empirical overview of data augmentation for NLP in the limited labeled data setting, making it difficult to understand which methods work in which settings. In this paper, we provide an empirical survey of recent progress on data augmentation for NLP in the limited labeled data setting, summarizing the landscape of methods (including token-level augmentations, sentence-level augmentations, adversarial augmentations, and hidden-space augmentations) and carrying out experiments on 11 datasets covering topics/news classification, inference tasks, paraphrasing tasks, and single-sentence tasks. Based on the results, we draw several conclusions to help practitioners choose appropriate augmentations in different settings and discuss the current challenges and future directions for limited data learning in NLP.

Real-Time Wildfire Detection Algorithm Based on VIIRS Fire Product and Himawari-8 Data

Wildfires have a significant impact on the atmosphere, terrestrial ecosystems, and society. Real-time monitoring of wildfire locations is crucial in fighting wildfires and reducing human casualties and property damage. Geostationary satellites offer the advantage of high temporal resolution and are gradually being used for real-time fire detection. In this study, we constructed a fire label dataset using the stable VNP14IMG fire product and used the random forest (RF) model for fire detection based on Himawari-8 multiband data. The band calculation features related brightness temperature, spatial features, and auxiliary data as input used in this framework for model training. We also used a recursive feature elimination method to evaluate the impact of these features on model accuracy and to exclude redundant features. The daytime and nighttime RF models (RF-D/RF-N) are separately constructed to analyze their applicability. Finally, we extensively evaluated the model performance by comparing them with the Japan Aerospace Exploration Agency (JAXA) wildfire product. The RF models exhibited higher accuracy, with recall and precision rates of 95.62% and 59%, respectively, and the recall rate for small fires was 19.44% higher than that of the JAXA wildfire product. Adding band calculation features and spatial features, as well as feature selection, effectively reduced the overfitting and improved the model’s generalization ability. The RF-D model had higher fire detection accuracy than the RF-N model. Omission errors and commission errors were mainly concentrated in the adjacent pixels of the fire clusters. In conclusion, our VIIRS fire product and Himawari-8 data-based fire detection model can monitor the fire location in real time and has excellent detection capability for small fires, making it highly significant for fire detection.

Identifying Active Galactic Nuclei at z ∼ 3 from the HETDEX Survey Using Machine Learning

We used data from the Hobby–Eberly Telescope Dark Energy Experiment (HETDEX) to study the incidence of AGN in continuum-selected galaxies at z ∼ 3. From optical and infrared imaging in the 24 deg2 Spitzer HETDEX Exploratory Large Area survey, we constructed a sample of photometric-redshift selected z ∼ 3 galaxies. We extracted HETDEX spectra at the position of 716 of these sources and used machine-learning methods to identify those which exhibited AGN-like features. The dimensionality of the spectra was reduced using an autoencoder, and the latent space was visualized through t-distributed stochastic neighbor embedding. Gaussian mixture models were employed to cluster the encoded data and a labeled data set was used to label each cluster as either AGN, stars, high-redshift galaxies, or low-redshift galaxies. Our photometric redshift (photoz) sample was labeled with an estimated 92% overall accuracy, an AGN accuracy of 83%, and an AGN contamination of 5%. The number of identified AGN was used to measure an AGN fraction for different magnitude bins. The ultraviolet (UV) absolute magnitude where the AGN fraction reaches 50% is M UV = −23.8. When combined with results in the literature, our measurements of AGN fraction imply that the bright end of the galaxy luminosity function exhibits a power law rather than exponential decline, with a relatively shallow faint-end slope for the z ∼ 3 AGN luminosity function.

Pride: Prioritizing Documentation Effort Based on a PageRank-Like Algorithm and Simple Filtering Rules

Code documentation can be helpful in many software quality assurance tasks. However, due to resource constraints (e.g., time, human resources, and budget), programmers often cannot document their work completely and timely. In the literature, two approaches (one is supervised and the other is unsupervised) have been proposed to prioritize documentation effort to ensure the most important classes to be documented first. However, both of them contain several limitations. The supervised approach overly relies on a difficult-to-obtain labeled data set and has high computation cost. The unsupervised one depends on a graph representation of the software structure, which is inaccurate since it neglects many important couplings between classes. In this paper, we propose an improved approach, named Pride, to prioritize documentation effort. First, Pride uses a weighted directed class coupling network to precisely describe classes and their couplings. Second, we propose a PageRank-like algorithm to quantify the importance of classes in the whole class coupling network. Third, we use a set of software metrics to quantify source code complexity and further propose a simple but easy-to-operate filtering rule. Fourth, we sort all the classes according to their importance in descending order and use the filtering rule to filter out unimportant classes. Finally, a threshold <inline-formula><tex-math notation="LaTeX">$k$</tex-math></inline-formula> is utilized, and the top- <inline-formula><tex-math notation="LaTeX">$k$</tex-math></inline-formula> % ranked classes are the identified important classes to be documented first. Empirical results on a set of nine software systems show that, according to the average ranking of the Friedman test, Pride is superior to the existing approaches in the whole data set.

Synthetic images generation for semantic understanding in facility management

PurposeThis study aims to introduce a new methodology for generating synthetic images for facility management purposes. The method starts by leveraging the existing 3D open-source BIM models and using them inside a graphic engine to produce a photorealistic representation of indoor spaces enriched with facility-related objects. The virtual environment creates several images by changing lighting conditions, camera poses or material. Moreover, the created images are labeled and ready to be trained in the model.Design/methodology/approachThis paper focuses on the challenges characterizing object detection models to enrich digital twins with facility management-related information. The automatic detection of small objects, such as sockets, power plugs, etc., requires big, labeled data sets that are costly and time-consuming to create. This study proposes a solution based on existing 3D BIM models to produce quick and automatically labeled synthetic images.FindingsThe paper presents a conceptual model for creating synthetic images to increase the performance in training object detection models for facility management. The results show that virtually generated images, rather than an alternative to real images, are a powerful tool for integrating existing data sets. In other words, while a base of real images is still needed, introducing synthetic images helps augment the model’s performance and robustness in covering different types of objects.Originality/valueThis study introduced the first pipeline for creating synthetic images for facility management. Moreover, this paper validates this pipeline by proposing a case study where the performance of object detection models trained on real data or a combination of real and synthetic images are compared.

Development of a COVID-19-Related Anti-Asian Tweet Data Set: Quantitative Study.

Since the advent of the COVID-19 pandemic, individuals of Asian descent (colloquial usage prevalent in North America, where "Asian" is used to refer to people from East Asia, particularly China) have been the subject of stigma and hate speech in both offline and online communities. One of the major venues for encountering such unfair attacks is social networks, such as Twitter. As the research community seeks to understand, analyze, and implement detection techniques, high-quality data sets are becoming immensely important. In this study, we introduce a manually labeled data set of tweets containing anti-Asian stigmatizing content. We sampled over 668 million tweets posted on Twitter from January to July 2020 and used an iterative data construction approach that included 3 different stages of algorithm-driven data selection. Finally, we found volunteers who manually annotated the tweets by hand to arrive at a high-quality data set of tweets and a second, more sampled data set with higher-quality labels from multiple annotators. We presented this final high-quality Twitter data set on stigma toward Chinese people during the COVID-19 pandemic. The data set and instructions for labeling can be viewed in the Github repository. Furthermore, we implemented some state-of-the-art models to detect stigmatizing tweets to set initial benchmarks for our data set. Our primary contributions are labeled data sets. Data Set v3.0 contained 11,263 tweets with primary labels (unknown/irrelevant, not-stigmatizing, stigmatizing-low, stigmatizing-medium, stigmatizing-high) and tweet subtopics (eg, wet market and eating habits, COVID-19 cases, bioweapon). Data Set v3.1 contained 4998 (44.4%) tweets randomly sampled from Data Set v3.0, where a second annotator labeled them only on the primary labels and then a third annotator resolved conflicts between the first and second annotators. To demonstrate the usefulness of our data set, preliminary experiments on the data set showed that the Bidirectional Encoder Representations from Transformers (BERT) model achieved the highest accuracy of 79% when detecting stigma on unseen data with traditional models, such as a support vector machine (SVM) performing at 73% accuracy. Our data set can be used as a benchmark for further qualitative and quantitative research and analysis around the issue. It first reaffirms the existence and significance of widespread discrimination and stigma toward the Asian population worldwide. Moreover, our data set and subsequent arguments should assist other researchers from various domains, including psychologists, public policy authorities, and sociologists, to analyze the complex economic, political, historical, and cultural underlying roots of anti-Asian stigmatization and hateful behaviors. A manually annotated data set is of paramount importance for developing algorithms that can be used to detect stigma or problematic text, particularly on social media. We believe this contribution will help predict and subsequently design interventions that will significantly help reduce stigma, hate, and discrimination against marginalized populations during future crises like COVID-19.

Detecting Human Trafficking: Automated Classification of Online Customer Reviews of Massage Businesses

Problem definition: Approximately 11,000 alleged illicit massage businesses (IMBs) exist across the United States hidden in plain sight among legitimate businesses. These illicit businesses frequently exploit workers, many of whom are victims of human trafficking, forced or coerced to provide commercial sex. Academic/practical relevance: Although IMB review boards like Rubmaps.ch can provide first-hand information to identify IMBs, these sites are likely to be closed by law enforcement. Open websites like Yelp.com provide more accessible and detailed information about a larger set of massage businesses. Reviews from these sites can be screened for risk factors of trafficking. Methodology: We develop a natural language processing approach to detect online customer reviews that indicate a massage business is likely engaged in human trafficking. We label data sets of Yelp reviews using knowledge of known IMBs. We develop a lexicon of key words/phrases related to human trafficking and commercial sex acts. We then build two classification models based on this lexicon. We also train two classification models using embeddings from the bidirectional encoder representations from transformers (BERT) model and the Doc2Vec model. Results: We evaluate the performance of these classification models and various ensemble models. The lexicon-based models achieve high precision, whereas the embedding-based models have relatively high recall. The ensemble models provide a compromise and achieve the best performance on the out-of-sample test. Our results verify the usefulness of ensemble methods for building robust models to detect risk factors of human trafficking in reviews on open websites like Yelp. Managerial implications: The proposed models can save countless hours in IMB investigations by automatically sorting through large quantities of data to flag potential illicit activity, eliminating the need for manual screening of these reviews by law enforcement and other stakeholders. Funding: This work was supported by the National Science Foundation [Grant 1936331]. Supplemental Material: The online appendix is available at https://doi.org/10.1287/msom.2023.1196 .

Comparative Analysis of Different CNN Models for Building Segmentation from Satellite and UAV Images

Building segmentation has numerous application areas such as urban planning and disaster management. In this study, 12 CNN models (U-Net, FPN, and LinkNet using EfficientNet-B5 backbone, U-Net, SegNet, FCN, and six Residual U-Net models) were generated and used for building segmentation. Inria Aerial Image Labeling Data Set was used to train models, and three data sets (Inria Aerial Image Labeling Data Set, Massachusetts Buildings Data Set, and Syedra Archaeological Site Data Set) were used to evaluate trained models. On the Inria test set, Residual-2 U-Net has the highest F1 and Intersection over Union (IoU) scores with 0.824 and 0.722, respectively. On the Syedra test set, LinkNet-EfficientNet-B5 has F1 and IoU scores of 0.336 and 0.246. On the Massachusetts test set, Residual-4 U-Net has F1 and IoU scores of 0.394 and 0.259. It has been observed that, for all sets, at least two of the top three models used residual connections. Therefore, for this study, residual connections are more successful than conventional convolutional layers.

Using self-supervised feature learning to improve the use of pulse oximeter signals to predict paediatric hospitalization.

Background: The success of many machine learning applications depends on knowledge about the relationship between the input data and the task of interest (output), hindering the application of machine learning to novel tasks. End-to-end deep learning, which does not require intermediate feature engineering, has been recommended to overcome this challenge but end-to-end deep learning models require large labelled training data sets often unavailable in many medical applications. In this study, we trained self-supervised learning (SSL) models for automatic feature extraction from raw photoplethysmography (PPG) obtained using a pulse oximeter, with the aim of predicting paediatric hospitalization. Methods: We compared logistic regression models fitted using features extracted using SSL with models trained using both clinical and SSL features. In addition, we compared end-to-end deep learning models initialized randomly or using weights from the SSL models. We also compared the performance of SSL models trained on labelled data alone (n=1,031) with SSL trained using both labelled and unlabelled signals (n=7,578). Results: Logistic regression models were more predictive of hospitalization when trained on features extracted using labelled PPG signals only compared to SSL models trained on both labelled and unlabelled signals (AUC 0.83 vs 0.80). However, features extracted using SSL model trained on both labelled and unlabelled PPG signals were more predictive of hospitalization when concatenated with clinical features (AUC 0.89 vs 0.87). The end-to-end deep learning model had an AUC of 0.80 when initialized using the SSL model trained on all PPG signals, 0.77 when initialized using SSL trained on labelled data only, and 0.73 when initialized randomly. Conclusions: This study shows that SSL can extract features from PPG signals that are predictive of hospitalization or initialize end-to-end deep learning models. Furthermore, SSL can leverage larger unlabelled data sets to improve performance of models fitted using small labelled data sets.

Physics-driven self-supervised learning system for seismic velocity inversion

Seismic velocity inversion plays a vital role in various applied seismology processes. A series of deep learning methods have been developed that rely purely on manually provided labels for supervision; however, their performances depend heavily on using large training data sets with corresponding velocity models. Because no physical laws are used in the training phase, it is usually challenging to generalize trained neural networks to a new data domain. To mitigate these issues, we have embedded a seismic forward modeling step at the end of a network to remap the inversion result back to seismic data and thus train the neural network through self-supervised loss, i.e., the misfit between the network input and output. As a result, we eliminate the need for many labeled velocity models, and physical laws are introduced when back-propagating gradients through the seismic forward modeling step. We verify the effectiveness of our approach through comprehensive experiments on synthetic data sets, where self-supervised learning outperforms the fully supervised approach, which accesses much more labeled data. The superior performance is even more significant when compared with a new data domain that has velocity models with faults and more geologic layers. Finally, in case of unknown and more complex data types, we develop a network-constrained full-waveform inversion (FWI) method. This method refines the initial prediction of the network by iteratively optimizing network parameters other than the velocity model, as found with the conventional FWI method, and demonstrates clear advantages in terms of interface and velocity accuracy. With these measures (self-supervised learning and network-constrained FWI), our physics-driven self-supervised learning system successfully mitigates issues such as the dependence on large labeled data sets, the absence of physical laws, and the difficulty in adapting to new data domains.

Disambiguation-based partial label feature selection via feature dependency and label consistency

Partial label learning refers to the issue that each training sample corresponds to a candidate label set containing only one valid label. Feature selection can be viewed as an effective pre-processing technology to improve the generalization performance of learning models, while the partial label feature selection task is challenging due to ambiguous labeling information. To this end, this paper utilizes granular ball computing and neighborhood rough sets to put forward a disambiguation-based partial label feature selection algorithm via feature dependency and label consistency. Firstly, the proposed algorithm performs an adaptive neighbor aggregation based on granular balls to disambiguate candidate labels. Adaptive neighbors are flexible to make more efficient learning performance. Then, considering the labeling confidence induced by disambiguation, the significance of each feature is evaluated by fusing the feature dependency from the neighborhood granularity and the label consistency among the nearest neighbors. Neighborhood rough sets directly handle continuous features that can reduce the adverse impact of data discretization on feature selection. The label consistency among adjacent samples is important for measuring feature discrimination. The fusion of feature dependency and label consistency can facilitate performance improvement. Finally, experiments conducted on eight controlled UCI and five real-world partial label datasets demonstrate that the proposed algorithm can improve the generalization performance of partial label learning and achieve superior performance compared to the partial label feature selection methods.

An Anomaly Detection-Based Domain Adaptation Framework for Cross-Domain Building Extraction from Remote Sensing Images

Deep learning-based building extraction methods have achieved a high accuracy in closed remote sensing datasets. In fact, the distribution bias between the source and target domains can lead to a dramatic decrease in their building extraction effect in the target domain. However, the mainstream domain adaptation methods that specifically address this domain bias problem require the reselection of many unlabeled samples and retraining in other target domains. This is time-consuming and laborious and even impossible at small regions. To address this problem, a novel domain adaptation framework for cross-domain building extraction is proposed from a perspective of anomaly detection. First, the initial extraction results of images in the target domain are obtained by a source domain-based pre-trained model, and then these results are classified into building mixed and non-building layers according to the predicted probability. Second, anomalous objects in the building layer are detected using the isolation forest method. Subsequently, the remaining objects in the building layer and the objects in the non-building layer are used as positive and negative samples, respectively, to reclassify the mixed layer using the random forest classifier. The newly extracted objects are fused with the remaining objects in the building layer as the final result. Four different experiments are performed on different semantic segmentation models and target domains. Some experimental results indicate that our framework can improve cross-domain building extraction compared to the pre-trained model, with an 8.7% improvement in the F1 metric when migrating from the Inria Aerial Image Labeling dataset to the Wuhan University dataset. Furthermore, experimental results show that our framework can be applied to multiple target domains without retraining and can achieve similar results to domain adaptation models based on adversarial learning.