Learning Datasets Research Articles

A large-scale multivariate soccer athlete health, performance, and position monitoring dataset

Data analysis for athletic performance optimization and injury prevention is of tremendous interest to sports teams and the scientific community. However, sports data are often sparse and hard to obtain due to legal restrictions, unwillingness to share, and lack of personnel resources to be assigned to the tedious process of data curation. These constraints make it difficult to develop automated systems for analysis, which require large datasets for learning. We therefore present SoccerMon, the largest soccer athlete dataset available today containing both subjective and objective metrics, collected from two different elite women’s soccer teams over two years. Our dataset contains 33,849 subjective reports and 10,075 objective reports, the latter including over six billion GPS position measurements. SoccerMon can not only play a valuable role in developing better analysis and prediction systems for soccer, but also inspire similar data collection activities in other domains which can benefit from subjective athlete reports, GPS position information, and/or time-series data in general.

EEGDiR: Electroencephalogram denoising network for temporal information storage and global modeling through Retentive Network

Electroencephalogram (EEG) signals are pivotal in clinical medicine, brain research, and neurological disorder studies. However, their susceptibility to contamination from physiological and environmental noise challenges the precision of brain activity analysis. Advances in deep learning have yielded superior EEG signal denoising techniques that eclipse traditional approaches. In this research, we deploy the Retentive Network architecture – initially crafted for large language models (LLMs) – for EEG denoising, exploiting its robust feature extraction and comprehensive modeling prowess. Furthermore, its inherent temporal structure alignment makes the Retentive Network particularly well-suited for the time-series nature of EEG signals, offering an additional rationale for its adoption. To conform the Retentive Network to the unidimensional characteristic of EEG signals, we introduce a signal embedding tactic that reshapes these signals into a two-dimensional embedding space conducive to network processing. This avant-garde method not only carves a novel trajectory in EEG denoising but also enhances our comprehension of brain functionality and the accuracy in diagnosing neurological ailments. Moreover, in response to the labor-intensive creation of deep learning datasets, we furnish a standardized, preprocessed dataset poised to streamline deep learning advancements in this domain.

RFLSE: Joint radiomics feature‐enhanced level‐set segmentation for low‐contrast SPECT/CT tumour images

AbstractDoctors typically use non‐contrast‐enhanced computed tomography (NCECT) in the treatment of kidney cancer to map kidney and tumour structural information to functional imaging single‐photon emission computed tomography, which is then used to assess patient kidney function and predict postoperative recovery. However, the assessment of kidney function and formulation of surgical plans is constrained by the low contrast of tumours in NCECT, which hinders the acquisition of accurate tumour boundaries. Therefore, this study designed a radiomics feature‐enhanced level‐set evolution (RFLSE) to precisely segment small‐sample low‐contrast kidney tumours. Integration of high‐dimensional radiomics features into the level‐set energy function enhances the edge detection capability of low‐contrast kidney tumours. The use of sensitive radiomics features to control the regional term parameters achieves adaptive adjustment of the curve evolution amplitude, improving the level‐set segmentation process. The experimental data used low‐contrast, limited‐sample tumours provided by hospitals, as well as the public datasets BUSI18 and KiTS19. Comparative results with advanced energy functionals and deep learning models demonstrate the precision and robustness of RFLSE segmentation. Additionally, the application value of RFLSE in assisting doctors with accurately marking tumours and generating high‐quality pseudo‐labels for deep learning datasets is demonstrated.

In-situ verification of a deep-learning-based larval identification system for the Pacific oyster Magallana gigas

Natural seedlings are collected for use at Pacific oyster (Magallana gigas) aquaculture sites in Japan. For seedling collection to be successful, aquaculture farmers identify target larvae through microscopic observation, using morphological features such as the shell height/shell length ratio, lateral asymmetry of the shell, and umbo development. To reduce the time and labor associated with larval identification, we developed a deep-learning-based Pacific oyster larvae identification system, which consisted of a tablet connecting a new photographic device to capture images for identification. The final accuracies of the trained identification system for the learning datasets for measuring accuracy were 82.5% and 80.5% for precision (P) and recall (R), respectively, for all sizes of larvae (shell height ≥100 µm). We verified our identification system at an oyster aquaculture site. The obtained accuracy in the verification was low for some samples (63.6%–100.0% and 41.9%–100.0% for P and R, respectively, for pre-attachment stage larvae), indicating a decrease in identification accuracy for the in-situ samples. The decrease likely occurred owing to the natural variability of larvae. We modified the learning dataset with high variation in oyster larval images and retrained the identification system. The accuracy of the retrained identification system was increased to 83.9%–100.0% for P and R and thus satisfied the requirements for practical application for pre-attachment stage larvae. For small-sized larvae, P was improved but R was relatively low, indicating that although the detection level of our identification model for oyster larvae is low, the accuracy of the detected oyster larvae is high.

Machine learning for the prediction of problems in steel tube bending process

Tube bending is a widely used process in marine, automotive, construction and other industries. Different methods are available for this process during which various problems and defects are encountered. This paper aims to develop the most relevant model for bending the tubes using a computer numerical controlled (CNC) bending machine on the first attempt without any defects. To achieve this goal, the parameters affecting the tube bending process on the bending machine are defined. Based on these parameters, 150 bending experiments are conducted in an industrial plant to collect data and their results are used as dataset for machine learning (ML) algorithms. Seven different state-of-the-art ML algorithms –logistic regression, decision tree, k-nearest neighbor, random forest, Naïve Bayes, support vector machine and eXtreme gradient boosting (XGBoost)– are implemented using Python's Scikit-Learn library. Their performance is compared using confusion matrix and classification metrics such as accuracy, precision, recall, F1-score, receiver operating characteristic (ROC) curve and area under the ROC curve (AUC) value. Considering all the performance metrics, logistic regression performed best in terms of prediction on our dataset. XGBoost and support vector machine were quiet successful based on F1 score and AUC, respectively. Overall, logistic regression, Naïve Bayes, Support Vector Machine, and XGBoost showed performances above 90% across all metrics. Decision tree, k-nearest neighbor and random forest performed poorly compared to other algorithms for our data. The proposed ML method is expected to save costs by reducing material waste and labor time and to increase the process efficiency and the product quality.

An open flame and smoke detection dataset for deep learning in remote sensing based fire detection

ABSTRACT The early warning of fires is pivotal for preventing substantial economic losses and ecological damage. However, the enhancement of fire model generalization performance still faces challenges due to limitations in existing fire datasets, such as image quantity and heterogeneity. Taking the advantage of rapid advancements in remote sensing technology and artificial intelligence, we introduce the Flame And Smoke Detection Dataset (FASDD), a pioneering collection comprising over 120,000 heterogeneous images covering various fire scenarios, aiming to promote the evolution of fire detection models. FASDD serves as a challenging benchmark for object detection tasks and features three sub-datasets: FASDD_CV, FASDD_UAV, and FASDD_RS. These sub-datasets comprise cross-domain images from ground-based, airborne, and spaceborne sensors, respectively. Extensive experiments using advanced Swin Transformer models demonstrate satisfactory fire detection performance, with mAP scores of 84.9%, 89.7%, and 74.0% on the respective sub-datasets. Notably, the application case in wildfire localization highlights the proficiency of FASDD_RS models for monitoring and tracking forest fires. Trained on FASDD_RS, FASDD_UAV, and FASDD_CV, deep learning models can be individually deployed on edge devices including satellites, drones, and terrestrial sensors, thus empowering collaborative fire warning within the space-air-ground integrated observation network. Additionally, a standardized procedure for data preparation, data procedure, and quality assurance is established, providing off-the-shelf annotations with four common configurations to facilitate machine learning model training. FASDD stands as a pivotal benchmark, providing valuable resources for training and evaluating deep learning fire detection models, especially Large Vision Models (LVM). It holds the potential to accelerate research progress in urban firefighting or forest fire detection, providing vital security measures for early fire warning and emergency response tasks.

InsectSound1000 An insect sound dataset for deep learning based acoustic insect recognition

InsectSound1000 is a dataset comprising more than 169000 labelled sound samples of 12 insects. The insect sound level spans from very loud (Bombus terrestris) to inaudible to human ears (Aphidoletes aphidimyza). The samples were extracted from more than 1000 h of recordings made in an anechoic box with a four-channel low-noise measurement microphone array. Each sample is a four-channel wave-file of 2500 kHz length, at 16 kHz sample rate and 32 bit resolution. Acoustic insect recognition holds great potential to form the basis of a digital insect sensor. Such sensors are desperately needed to automate pest monitoring and ecological monitoring. With its significant size and high-quality recordings, InsectSound1000 can be used to train data-hungry deep learning models. Used to pretrain models, it can also be leveraged to enable the development of acoustic insect recognition systems on different hardware or for different insects. Further, the methodology employed to create the dataset is presented in detail to allow for the extension of the published dataset.

Simulation of gamma spectra from soil samples by using MCNP: A case study

A comparison between the MCNP simulated gamma spectrums based on the nuclear data from NuDat with the current version 3.0 and Nucléide-Lara against measured spectra based on the IAEA reference samples has been performed to assess the influence of nuclear data of photon decay processes and cross-section of photon interactions on the quality of the simulated spectra. As the results, the appearance of some abnormal energy peaks; namely 215 keV, 571 keV, 675 keV, 1227 keV of the NuDat-based simulated spectra and 90 keV, 94 keV, 106 keV, 416 keV of the Nucléide-Lara-based simulated spectra, which were present in neither the measured spectrum nor remaining simulated spectra, indicating issues with accuracy and completeness of these dataset. In addition, the good correlation between the combined dataset-based simulated spectra and reference samples-based measured spectra within the range of 50 keV to 2620 keV suggests that this MCNP simulation configuration can be used to generate a large simulated dataset for Machine Learning (ML) models that automatically identify and qualify radioactive isotope from gamma-ray spectra, overcoming the practical limitation of number of reference samples to sufficiently generate data for training and testing ML algorithms in the field of environmental radiation [1].

A new sub-class linear discriminant for miniature spectrometer based food analysis

The well-known and extensively studied Linear Discriminant Analysis (LDA) can have its performance lowered in scenarios where data is not homoscedastic or not Gaussian. That is, the classical assumptions when LDA models are built are not applicable, and consequently LDA projections would not be able to extract the needed features to explain the intrinsic structure of data and for classes to be separated. As with many real word data sets, data obtained using miniature spectrometers can suffer from such drawbacks which would limit the deployment of such technology needed for food analysis. The solution presented in the paper is to divide classes into subclasses and to use means of sub classes, classes, and data in the suggested between classes scatter metric. Further, samples belonging to the same subclass are used to build a measure of within subclass scatterness. Such a solution solves the shortcoming of the classical LDA. The obtained results when using the proposed solution on food data and on general machine learning datasets show that the work in this paper compares well to and is very competitive with similar sub-class LDA algorithms in the literature. An extension to a Hilbert space is also presented; and the kernel version of the presented solution can be fused with its linear counter parts to yield improved classification rates.

Enhanced Cost-sensitive Ensemble Learning for Imbalanced Class in Medical Data

The challenge of imbalanced datasets in machine learning, particularly in the medical and public health sectors, necessitates innovative solutions that enhance predictive accuracy and reduce misclassification costs. This paper introduces the Enhanced Cost Sensitive Ensemble Learning (ECSEL), a novel approach that combines cost-sensitive learning with ensemble techniques to address the imbalance inherent in many critical datasets. We evaluate ECSEL on two specific datasets: the Framingham Heart Study dataset, which is pivotal in cardiovascular disease prediction, and a dataset focused on COVID-19 infection forecasting, relevant for public health responses. The results demonstrate that ECSEL significantly outperforms traditional methods like SMOTE, D-SMOTE, and BP-SMOTE by improving accuracy, precision, recall, and ROC-AUC values, particularly reducing Type II errors in scenarios where the cost of false negatives is exceptionally high. The method’s effectiveness is showcased through its superior performance in predicting cardiovascular risks and its robustness in forecasting the spread of COVID-19, reflecting its applicability and potential in real-world settings.

MISATO: machine learning dataset of protein–ligand complexes for structure-based drug discovery

Large language models have greatly enhanced our ability to understand biology and chemistry, yet robust methods for structure-based drug discovery, quantum chemistry and structural biology are still sparse. Precise biomolecule–ligand interaction datasets are urgently needed for large language models. To address this, we present MISATO, a dataset that combines quantum mechanical properties of small molecules and associated molecular dynamics simulations of ~20,000 experimental protein–ligand complexes with extensive validation of experimental data. Starting from the existing experimental structures, semi-empirical quantum mechanics was used to systematically refine these structures. A large collection of molecular dynamics traces of protein–ligand complexes in explicit water is included, accumulating over 170 μs. We give examples of machine learning (ML) baseline models proving an improvement of accuracy by employing our data. An easy entry point for ML experts is provided to enable the next generation of drug discovery artificial intelligence models.

Robust Representation Learning for Power System Short-Term Voltage Stability Assessment Under Diverse Data Loss Conditions.

With the help of neural network-based representation learning, significant progress has been recently made in data-driven online dynamic stability assessment (DSA) of complex electric power systems. However, without sufficient attention to diverse data loss conditions in practice, the existing data-driven DSA solutions' performance could be largely degraded due to practical defective input data. To address this problem, this work develops a robust representation learning approach to enhance DSA performance against multiple input data loss conditions in practice. Specifically, focusing on the short-term voltage stability (SVS) issue, an ensemble representation learning scheme (ERLS) is carefully designed to achieve data loss-tolerant online SVS assessment: 1) based on an efficient data masking technique, various missing data conditions are handled and augmented in a unified manner for lossy learning dataset preparation; 2) the emerging spatial-temporal graph convolutional network (STGCN) is leveraged to derive multiple diversified base learners with strong capability in SVS feature learning and representation; and 3) with massive SVS scenarios deeply grouped into a number of clusters, these STGCN-enabled base learners are distinctly assembled for each cluster via multilinear regression (MLR) to realize ensemble SVS assessment. Such a divide-and-conquer ensemble strategy results in highly robust SVS assessment performance when faced with various severe data loss conditions. Numerical tests on the benchmark Nordic test system illustrate the efficacy of the proposed approach.

Crescent Moon Visibility: A New Criterion using Deep learned Artificial Neural-Network

Many authors investigated the problem of the early visibility of the new crescent moon after the conjunction and proposed many criteria addressing this issue in the literature. This article presented a proposed criterion for early crescent moon sighting based on a deep-learned pattern recognizer artificial neural network (ANN) performance. Moon sight datasets were collected from various sources and used to learn the ANN. The new criterion relied on the crescent width and the arc of vision from the edge of the crescent bright limb. The result of that criterion was a control value indicating the moon's visibility condition, which separated the datasets into four regions: invisible, telescope only, probably visible, and certainly visible. This criterion was used on the dataset for ANN learning to compare its efficiency with the actual moon visibility events.

TorchQL: A Programming Framework for Integrity Constraints in Machine Learning

Finding errors in machine learning applications requires a thorough exploration of their behavior over data. Existing approaches used by practitioners are often ad-hoc and lack the abstractions needed to scale this process. We present TorchQL, a programming framework to evaluate and improve the correctness of machine learning applications. TorchQL allows users to write queries to specify and check integrity constraints over machine learning models and datasets. It seamlessly integrates relational algebra with functional programming to allow for highly expressive queries using only eight intuitive operators. We evaluate TorchQL on diverse use-cases including finding critical temporal inconsistencies in objects detected across video frames in autonomous driving, finding data imputation errors in time-series medical records, finding data labeling errors in real-world images, and evaluating biases and constraining outputs of language models. Our experiments show that TorchQL enables up to 13x faster query executions than baselines like Pandas and MongoDB, and up to 40% shorter queries than native Python. We also conduct a user study and find that TorchQL is natural enough for developers familiar with Python to specify complex integrity constraints.

Analysis and Modelling Of English Learning Behaviour Based on Data Mining Technology

This paper presents an in-depth exploration into the analysis and modelling of English learning behaviour leveraging data mining technology. In today's digital age, understanding the intricacies of how individuals acquire a second language is paramount for optimizing educational strategies and enhancing learning outcomes. By employing advanced data mining techniques, we aim to uncover patterns, trends, and underlying dynamics within English learning datasets, shedding light on the diverse behaviours exhibited by learners. Furthermore, this research transcends analysis, as we endeavour to construct robust models capable of simulating and predicting English learning trajectories. By leveraging insights derived from extensive data exploration, we aim to develop predictive models that offer personalized recommendations and adaptive learning pathways tailored to individual learner profiles. Ultimately, this paper contributes to the ongoing discourse surrounding language acquisition by providing valuable insights into English learning behaviours. By harnessing the power of data mining technology, we aspire to empower educators, policymakers, and digital learning platforms with the tools necessary to cultivate more efficient and inclusive approaches to English language education.

Which Explanation Should be Selected: A Method Agnostic Model Class Reliance Explanation for Model and Explanation Multiplicity

Feature importance techniques offer valuable insights into machine learning (ML) models by conducting quantitative assessments of the individual contributions of variables to the model’s predictive outcomes. This quantification differs across various explanation methods and multiple almost equally accurate models (Rashomon models), creating explanation and model multiplicities. This resulted in a novel framework called method agnostic model class reliance range (MAMCR) for identifying a unified explanation across methods for multiple models. This consensus explanation provides each feature’s importance range for a class of models. Using state-of-the-art feature importance methods, experiments on popular machine learning datasets are conducted with a ε-\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$\\varepsilon -$$\\end{document} threshold value of 0.1. The dataset-specific Rashomon set with 200 models, and the prediction accuracy of concerned reference models (m∗\\documentclass[12pt]{minimal} \\usepackage{amsmath} \\usepackage{wasysym} \\usepackage{amsfonts} \\usepackage{amssymb} \\usepackage{amsbsy} \\usepackage{mathrsfs} \\usepackage{upgreek} \\setlength{\\oddsidemargin}{-69pt} \\begin{document}$$m^*$$\\end{document}) have produced encouraging results in obtaining a consensus model reliance explanation that is consistent across multiple methods. The experiment results ensure whether the prediction accuracy level of models has an impact on the importance range estimation of features. Also, the order of features suggested by MAMCR leads to better performance of models consistently in all the experimented datasets, than the state-of-the-art methods.

Machine learning-based multi-performance prediction and analysis of Earth-Air Heat Exchanger

The Earth-Air Heat Exchanger (EAHE) plays a crucial role in global renewable energy utilization. Despite various EAHE models, a machine learning-based multi-performance prediction model is lacking. This study introduces a framework based on diverse machine learning models for precise prediction and comprehensive analysis of EAHE's multiple performance indicators. Initially, the theoretical EAHE model is established and validated. Six feature variables and four target variables are input to generate a dataset for machine learning. Six machine learning models are then selected for training, optimization, and validation. Subsequently, model interpretation and analysis are conducted using SHapley additive exPlanations and Partial dependence plots. Results indicate the optimized-extreme gradient boosting model performs the best, with a determination coefficient of 0.98, root mean squared error of 22.98, and mean absolute error of 8.88, identified as the top multi-performance prediction model for EAHE. Notably, pipe diameter, air velocity, and ground type significantly influence EAHE performance. Complex correlations and mutual exclusions exist among target variables, with a strong negative correlation (−0.59) between the coefficient of performance and payback time, highlighting the challenge of simultaneous improvement. This study presents a valuable framework for predicting and analyzing EAHE performance, crucial for promoting its widespread application and enhancing renewable energy utilization.

LaANIL: ANIL with Look-Ahead Meta-Optimization and Data Parallelism

Meta-few-shot learning algorithms, such as Model-Agnostic Meta-Learning (MAML) and Almost No Inner Loop (ANIL), enable machines to learn complex tasks quickly with limited data and based on previous experience. By maintaining the inner loop head of the neural network, ANIL leads to simpler computations and reduces the complexity of MAML. Despite its benefits, ANIL suffers from issues like accuracy variance, slow initial learning, and overfitting, hardening its adaptation and generalization. This work proposes “Look-Ahead ANIL” (LaANIL), an enhancement to ANIL for better learning. LaANIL reorganizes ANIL’s internal architecture, integrating parallel computing techniques (to process multiple training examples simultaneously across computing units) and incorporating Nesterov momentum (which accelerates convergence by adjusting the learning rate based on past gradient information and extracting informative features for look-ahead gradient computation). These additional features make our model more state-of-the-art capable and better edge-compatible and thus improve few-short learning by enabling models to quickly adapt to new information and tasks. LaANIL’s effectiveness is validated on established meta-few-shot learning datasets, including FC100, CIFAR-FS, Mini-ImageNet, CUBirds-200-2011, and Tiered-ImageNet. The proposed model achieved an increased validation accuracy by 7 ± 0.7% and a variance reduction by 44 ± 4% in two-way two-shot classification as well as increased validation by 5 ± 0.4% and a variance reduction by 18 ± 2% in five-way five-shot classification on the FC100 dataset and similarly performed well on other datasets.

From MNIST to ImageNet and back: benchmarking continual curriculum learning

Continual learning (CL) is one of the most promising trends in recent machine learning research. Its goal is to go beyond classical assumptions in machine learning and develop models and learning strategies that present high robustness in dynamic environments. This goal is realized by designing strategies that simultaneously foster the incorporation of new knowledge while avoiding forgetting past knowledge. The landscape of CL research is fragmented into several learning evaluation protocols, comprising different learning tasks, datasets, and evaluation metrics. Additionally, the benchmarks adopted so far are still distant from the complexity of real-world scenarios, and are usually tailored to highlight capabilities specific to certain strategies. In such a landscape, it is hard to clearly and objectively assess models and strategies. In this work, we fill this gap for CL on image data by introducing two novel CL benchmarks that involve multiple heterogeneous tasks from six image datasets, with varying levels of complexity and quality. Our aim is to fairly evaluate current state-of-the-art CL strategies on a common ground that is closer to complex real-world scenarios. We additionally structure our benchmarks so that tasks are presented in increasing and decreasing order of complexity—according to a curriculum—in order to evaluate if current CL models are able to exploit structure across tasks. We devote particular emphasis to providing the CL community with a rigorous and reproducible evaluation protocol for measuring the ability of a model to generalize and not to forget while learning. Furthermore, we provide an extensive experimental evaluation showing that popular CL strategies, when challenged with our proposed benchmarks, yield sub-par performance, high levels of forgetting, and present a limited ability to effectively leverage curriculum task ordering. We believe that these results highlight the need for rigorous comparisons in future CL works as well as pave the way to design new CL strategies that are able to deal with more complex scenarios.

National-scale digital soil mapping performances are related to covariates and sampling density: Lessons from France

Accurate soil property and class predictions through spatial modelling necessitate a thoughtful selection of explanatory variables and sample size, as their choice greatly impacts model performance. Within the framework of Global Soil Nutrient and Nutrient Budget maps (GSNmap), the FAO Global Soil Partnership (GSP) launched a country-driven digital soil mapping (DSM) approach. The GSP asked the countries if they could implement the DSM prediction of ten soil properties, using their national point data and a set of widely available covariates (GSP_Cov). In this study, we examined the effect of including additional national-based covariates and soil observations on the performance of the prediction models using mainland France as a pilot. The learning soil dataset was based on a systematic 16-to-16 km grid. For a subset of soil properties, we also assessed using repeated k-fold cross-validation the effect of adding to this dataset many other irregularly spread measurements. The GSP_Cov included common widely available covariates that represented information about terrain, climate, and organisms. The second set of covariates consisted of the GSP_Cov, extended to extra covariates available at a national level, such as previously existing soil maps, geological maps, remote sensing products and others. Random Forest approach in combination with the Boruta selection method was employed for mapping ten soil properties: soil organic carbon (SOC), pH (water), total nitrogen (N), available phosphorus (P), available potassium (K), cation exchange capacity (CEC), bulk density (BD), and texture (clay, silt, and sand). The results revealed noteworthy enhancements in prediction performance for more than half of the properties, although, for some of them, the improvements were negligible. The most significant improvements were obtained for pH, CEC and texture, where geological variables and a previous pH map significantly contributed to the increase in accuracy. Adding numerous points (around 25,000) to the learning dataset improved the performance of soil particle-size fractions predictions. By broadening the spectrum of covariates and better covering the feature and geographical spaces considered in soil prediction models, this research underscores the importance of implementing a more diverse range of covariates at a national scale and of densifying soil information to enlarge the feature and geographical spaces of multidimensional soil/covariates combinations. This information should be taken into account in national and continental digital soil mapping endeavours.