Model Training Research Articles

Efficient Distributed Sparse Relative Similarity Learning

Learning a good similarity measure for large-scale high-dimensional data is a crucial task in machine learning applications, yet it poses a significant challenge. Distributed minibatch stochastic gradient descent (SGD) serves as an efficient optimization method in large-scale distributed training, allowing linear speedup in proportion to the number of workers. However, communication efficiency in distributed SGD requires a sufficiently large minibatch size, presenting two distinct challenges. Firstly, a large minibatch size leads to high memory usage and computational complexity during parallel training of high-dimensional models. Second, a larger batch size of data reduces the convergence rate. To overcome these challenges, we propose an efficient distributed sparse relative similarity learning framework EDSRSL . This framework integrates two strategies: local minibatch SGD and sparse relative similarity learning. By effectively reducing the number of updates through synchronous delay while maintaining a large batch size, we address the issue of high computational cost. Additionally, we incorporate sparse model learning into the training process, significantly reducing computational cost. This paper also provides theoretical proof that the convergence rate does not decrease significantly with increasing batch size. Various experiments on six high-dimensional real-world datasets demonstrate the efficacy and efficiency of the proposed algorithms, with a communication cost reduction of up to \(90.89\%\) and a maximum wall time speedup of \(5.66\times\) compared to the baseline methods.

Transfer Learning-Based Health Monitoring of Robotic Rotate Vector Reducer Under Variable Working Conditions

Due to their precision, compact size, and high torque transfer, Rotate vector (RV) reducers are becoming more popular in industrial robots. However, repetitive operations and varying speed conditions mean that these components are prone to mechanical failure. Therefore, it is important to develop effective health monitoring (HM) strategies. Traditional approaches for HM, including those using vibration and acoustic emission sensors, encounter such challenges as noise interference, data inconsistency, and high computational costs. Deep learning-based techniques, which use current electrical data embedded within industrial robots, address these issues, offering a more efficient solution. This research provides transfer learning (TL) models for the HM of RV reducers, which eliminate the need to train models from scratch. Fine-tuning pre-trained architectures on operational data for the three different reducers of health conditions, which are healthy, faulty, and faulty aged, improves fault classification across different motion profiles and variable speed conditions. Four TL models, EfficientNet, MobileNet, GoogleNet, and ResNET50v2, are considered. The classification accuracy and generalization capabilities of the suggested models were assessed across diverse circumstances, including low speed, high speed, and speed fluctuations. Compared to the other models, the proposed EfficientNet model showed the most promising results, achieving a testing accuracy and an F1-score of 98.33% each, which makes it best suited for the HM of robotic reducers.

Efficient Pest Detection Through Advanced Machine Learning Technique

Crop Protection is the key element to achieve food security. Many studies have been conducted over the decades to avoid crop losses during pre- harvest and post-harvest stages. Crop losses due to pest attack and plant disease spread, reduces the agriculture production and possess a direct impact on the economy of a country. Deployment of Artificial Intelligence (AI) based pest control strategies to detect pest species is under research. In this manuscript, deep learning based EfficientNetB7 architecture and transfer learning methodology is used to develop a real-time resource-efficient pest detection system. EfficientNetB7's innovative compound scaling technique has managed to balance efficiency in terms of computations and accuracy to effectively classify pests through images with minimal resources. The proposed system uses appropriate fine-tuning of training parameters and regularization mechanisms such as optimizers, data augmentation, to effectively develop a pest detection system. The trained model is ported on to STM32 microcontroller using CubeAI in STM32CubeIDE.The model is assessed by using a publicly available dataset, and a classification accuracy of 93.5% is achieved. This system not only automates pest detection but also provides a match percentage for the identified pests, thus supporting precision agriculture. Future work includes increasing the dataset and exploring edge AI techniques for decentralized decision-making. This manuscript showcases the transformative potential of EfficientNetB7 in precision agriculture, offering a scalable, cost-effective, and sustainable solution for the detection of pests with relevant broader impacts on agricultural automation practices.

Estimation of barrier layer thickness from satellite-derived sea surface parameters in the north Indian Ocean using machine learning

ABSTRACT Oceanic barrier layer thickness (BLT) has been estimated from sea surface parameters through four machine learning (ML) models. The BLT used for training and testing of ML models has been computed from 19 years (2005–2023) of Argo data of vertical temperature and salinity profile in the north Indian ocean. It is observed that thick barrier layer (BL) exists in three regions within the north Indian ocean: the Bay of Bengal (BoB), Eastern Arabian Sea (EAS) and the East Equatorial Indian Ocean (EEIO) off the Sumatra. The observed BLT ranges from 0 to 80 m in the north Indian ocean with a seasonal cycle. Four ML models, i.e. multiple linear regression, artificial neural network (ANN), Random Forest and XGBoost have been trained for 2010–2018 to learn the relationship between sea surface parameters (input features) and BLT (target/output variable) in three thick BL regions within the north Indian Ocean. Thereafter, the trained models have been tested for the estimation of BLT in the same regions for the time period of 2019–2023. The ML models are well able to estimate the BLT from sea surface parameters (SST, SSS) with RMSE less than 5 m in the BoB and EAS, which is less than 8 percentile of the model testing datasets. The ‘Random Forest Regressor’ turns out to perform better than the other three ML models for estimation of BLT. Further, the Random Forest estimated BLT has been compared with the BLT computed using ORAS5 reanalysis data of vertical temperature and salinity profile. This study explores the capability of the ML models for capturing the signature of a subsurface variable on the surface owing to the limitation of costly subsurface measurement.

Metrologically Interpretable Soft-Sensing Technique for Non-Invasive Liquid Flow Estimation from Vibration Data

This paper proposes a metrologically interpretable soft sensing method for estimating the liquid flow rates in hydraulic systems from non-invasive vibration frequency power band data. Despite considerable interest in non-invasive flow estimation, state-of-the-art methods provide little to no metrological capabilities. In this work, a dedicated test rig was developed to automatically acquire vibration and flow rate data from a centrifugal pump, in a flow rate range between 0.05 × 10−5m3/s and 9.11 × 10−5m3/s. The vibration data were processed into power bands, which were subsequently used to optimize and train a multilayer perceptron neural network for flow soft sensing. The trained model was compared with models with different vibration processing methods from literature. The power band processing model resulted in a root mean squared error 75.4% smaller than the second-best model in cross-validation, and 51.5% smaller with test data. The uncertainty of the proposed regression model was estimated using a combination of ensemble learning and Monte Carlo simulations, and combined with the reference flow sensor uncertainty to obtain the total combined uncertainty of the soft sensor, found to be between 3.9 × 10−6m3/s and 6.1 × 10−6m3/s throughout the measured flow range. The reference flow sensor accuracy was found to be the largest individual contribution for the final uncertainty, closely followed by the regression model uncertainty.

Prediction of outpatient rehabilitation patient preferences and optimization of graded diagnosis and treatment based on XGBoost machine learning algorithm

BackgroundThe Department of Rehabilitation Medicine is key to improving patients’ quality of life. Driven by chronic diseases and an aging population, there is a need to enhance the efficiency and resource allocation of outpatient facilities. This study aims to analyze the treatment preferences of outpatient rehabilitation patients by using data and a grading tool to establish predictive models. The goal is to improve patient visit efficiency and optimize resource allocation through these predictive models.MethodsData were collected from 38 Chinese institutions, including 4,244 patients visiting outpatient rehabilitation clinics. Data processing was conducted using Python software. The pandas library was used for data cleaning and preprocessing, involving 68 categorical and 12 continuous variables. The steps included handling missing values, data normalization, and encoding conversion. The data were divided into 80% training and 20% test sets using the Scikit-learn library to ensure model independence and prevent overfitting. Performance comparisons among XGBoost, random forest, and logistic regression were conducted using metrics, including accuracy and receiver operating characteristic (ROC) curves. The imbalanced learning library’s SMOTE technique was used to address the sample imbalance during model training. The model was optimized using a confusion matrix and feature importance analysis, and partial dependence plots (PDP) were used to analyze the key influencing factors.ResultsXGBoost achieved the highest overall accuracy of 80.21% with high precision and recall in Category 1. random forest showed a similar overall accuracy. Logistic Regression had a significantly lower accuracy, indicating difficulties with nonlinear data. The key influencing factors identified include distance to medical institutions, arrival time, length of hospital stay, and specific diseases, such as cardiovascular, pulmonary, oncological, and orthopedic conditions. The tiered diagnosis and treatment tool effectively helped doctors assess patients’ conditions and recommend suitable medical institutions based on rehabilitation grading.ConclusionThis study confirmed that ensemble learning methods, particularly XGBoost, outperform single models in classification tasks involving complex datasets. Addressing class imbalance and enhancing feature engineering can further improve model performance. Understanding patient preferences and the factors influencing medical institution selection can guide healthcare policies to optimize resource allocation, improve service quality, and enhance patient satisfaction. Tiered diagnosis and treatment tools play a crucial role in helping doctors evaluate patient conditions and make informed recommendations for appropriate medical care.

Cultural Sensitivity in AI Language Learning: Using NLP to Enhance Language Understanding Across Cultural Contexts

The aim of this paper is to demonstrate the integration of cultural awareness into AI language learning systems using Natural Language Processing (NLP) models. Recognising that the traditional methods of language learning tend not to capture cultural variations, the studies focus on how culturally relevant NLP models (GPT-3, BERT, RNNs) promote language acquisition and cross-cultural integration. Data preprocessing, cultural data augmentation and model training was used to encode cultural variables such as greetings, politeness, and contextually relevant expressions into training datasets. Experimental results suggest that culturally sensitive models outperform the generic model on tasks that involve cultural awareness, such as politeness detection, cultural phrase recognition, and tone sensitivity. GPT-3 for example improved the accuracy of politeness detection tasks from 72.4% to 85.7%. These results highlight the need to bring cultural awareness into AI to develop engaging, context-sensitive language learning. It concludes that culturally based NLP models enhance learner capacity for real-world communication and foster global cultural awareness. Future work should focus on expanding culturally diverse datasets and refining AI models to address subtle cultural complexities

Clinical and imaging spectrum of non-congenital dominant ACTN2 myopathy.

Alpha-actinin-2, a protein with high expression in cardiac and skeletal muscle, is located in the Z-disc and plays a key role in sarcomere stability. Mutations in ACTN2 have been associated with both hypertrophic and dilated cardiomyopathy and, more recently, with skeletal myopathy. Genetic, clinical, and muscle imaging data were collected from 37 patients with an autosomal dominant ACTN2 myopathy belonging to 11 families from Spain and Belgium. Haplotypes were studied to confirm a common ancestry for the two most common recurrent variants identified in this study. A trained machine learning model was used to compare muscle MRI scans in ACTN2 myopathy with other neuromuscular diseases to identify a specific pattern of muscle involvement. The clinical phenotype ranged from asymptomatic to limb-girdle weakness and facial involvement and was depending on genotype. Cardiac and respiratory involvement were not common. Belgian families carrying the p.Ile134Asn variant and Basque-Spanish families carrying the p.Cys487Arg variant each showed unique haplotypes supporting respective common ancestry. Available muscle biopsies showed non-specific changes. In muscle imaging, the most affected muscles were the glutei minor, glutei medius, hamstrings, tibialis anterior, and soleus. A machine learning model showed that the most differentiating features in dominant ACTN2 myopathy were the involvement of the tibialis anterior and gluteus medius muscles and preservation of the quadratus femoris, gastrocnemius lateralis, and tensor fasciae latae muscles. We provide new insights into genetic, clinical, and muscle imaging characteristics of non-congenital dominant ACTN2 myopathy, broadening the phenotypic spectrum of muscle disorders caused by ACTN2 variants.

Adaptive Month Matching: A Phenological Alignment Method for Transfer Learning in Cropland Segmentation

The increasing demand for food and rapid population growth have made advanced crop monitoring essential for sustainable agriculture. Deep learning models leveraging multispectral satellite imagery, like Sentinel-2, provide valuable solutions. However, transferring these models to diverse regions is challenging due to phenological differences in crop growth stages between training and target areas. This study proposes the Adaptive Month Matching (AMM) method to align the phenological stages of crops between training and target areas for enhanced transfer learning in cropland segmentation. In the AMM method, an optimal Sentinel-2 monthly time series is identified in the training area based on deep learning model performance for major crops common to both areas. A month-matching process then selects the optimal Sentinel-2 time series for the target area by aligning the phenological stages between the training and target areas. In this study, the training area covered part of the Mississippi River Delta, while the target areas included diverse regions across the US and Canada. The evaluation focused on major crops, including corn, soybeans, rice, and double-cropped winter wheat/soybeans. The trained deep learning model was transferred to the target areas, and accuracy metrics were compared across different time series chosen by various phenological alignment methods. The AMM method consistently demonstrated strong performance, particularly in transferring to rice-growing regions, achieving an overall accuracy of 98%. It often matched or exceeded other phenological matching techniques in corn segmentation, with an average overall accuracy across all target areas exceeding 79% for cropland segmentation.

A Cloud Detection System for UAV Sense and Avoid: Analysis of a Monocular Approach in Simulation and Flight Tests

In order to contribute to the operation of unmanned aerial vehicles (UAVs) according to visual flight rules (VFR), this article proposes a monocular approach for cloud detection using an electro-optical sensor. Cloud avoidance is motivated by several factors, including improving visibility for collision prevention and reducing the risks of icing and turbulence. The described workflow is based on parallelized detection, tracking and triangulation of features with prior segmentation of clouds in the image. As output, the system generates a cloud occupancy grid of the aircraft’s vicinity, which can be used for cloud avoidance calculations afterwards. The proposed methodology was tested in simulation and flight experiments. With the aim of developing cloud segmentation methods, datasets were created, one of which was made publicly available and features 5488 labeled, augmented cloud images from a real flight experiment. The trained segmentation models based on the YOLOv8 framework are able to separate clouds from the background even under challenging environmental conditions. For a performance analysis of the subsequent cloud position estimation stage, calculated and actual cloud positions are compared and feature evaluation metrics are applied. The investigations demonstrate the functionality of the approach, even if challenges become apparent under real flight conditions.

Evaluation and Future Prospects of Data‐Driven Intelligence‐Based Framework for Predicting Cyclic Behavior of Reconstituted Sand

ABSTRACTMost of the robust artificial intelligence (AI)‐based constitutive models are developed with synthetic datasets generated from traditional constitutive models. Therefore, they fundamentally rely on the traditional constitutive models rather than laboratory test results. Also, their potential use within geotechnical engineering communities is limited due to the unavailability of datasets along with the model code files. In this study, the data‐driven constitutive models are developed using only laboratory test databases and deep learning (DL) techniques. The laboratory database was prepared by conducting cyclic direct simple shear (CDSS) tests on reconstituted sand, that is, PDX sand. The stacked long short‐term memory (LSTM) network and its variants are considered for developing the predictive models of the shear strain (γ [%]) and excess pore pressure ratio (ru) time histories. The suitable input parameters (IPs) are selected based on the physics behind the generation of ru and γ (%) of the liquefiable sands. The predicted responses of γ (%) and ru agree well in most cases and are used to predict the dynamic soil properties of the PDX sand. The same modeling framework is extended for other sand and compared with existing AI‐based constitutive models to verify its practical applicability. In summary, it is observed that though the trained models predicted the time histories of ru and γ reasonably well; however, they struggled to predict the hysteresis loops at higher cycles. Therefore, more research is needed to verify and enhance the predictability of existing AI‐based models in the future before using them in practice for simulating cyclic response.

Real-time detection of powder bed defects in laser powder bed fusion using deep learning on 3D point clouds

ABSTRACT Powder bed defects are critical factors affecting the print quality and stability in Laser Powder Bed Fusion (LPBF). However, traditional 2D image-based powder bed defect monitoring methods are limited by sensitivity to lighting conditions and insufficient data capture. This study proposes a real-time defect monitoring system based on 3D point cloud data and deep learning approach. The system uses binocular vision to capture point cloud data in real time, enabling high-precision defect segmentation with advanced deep learning models. However, direct deep learning on point clouds can result in the loss of small defect features during downsampling. To address this, an indirect point cloud deep learning method based on 2D projection is introduced, which improves segmentation accuracy for small defects while reducing inference time. By deploying the trained model, this study establishes a closed-loop control system for powder bed defect detection and conducts real-world printing tests, demonstrating effective defect remediation capabilities. Although larger-scale industrial testing is still required, this research illustrates the significant potential of 3D point cloud-based deep learning in enhancing defect detection and quality control in additive manufacturing.

Advanced Data Engineering and Machine Learning Approaches for Accurate Yield Prediction in Semiconductor Manufacturing Processes.

In semiconductor manufacturing, yield prediction is a critical area where advancements in data engineering and machine learning can significantly impact production efficiency and cost-effectiveness. This paper explores the integration of data engineering pipelines with machine learning models to improve yield prediction. Using a scalable and efficient framework, we demonstrate how leveraging Databricks for data processing and model training enhances the prediction accuracy and provides actionable insights for process optimization. Keywords: Data Engineering, Machine Learning, Semiconductor Manufacturing, Yield Prediction, Databricks

Combining Differential Privacy with FedAvgM to Optimize Data Access Control in the Information Platform

In response to the privacy leakage risk and low-efficiency issues of traditional data access control, especially in distributed environments, data transmission and centralized processing may expose sensitive information. To improve data access control in information platforms, ensure data security, and accelerate global model training, this article introduced a joint mechanism of differential privacy and FedAvgM (Federated Averaging with Momentum) optimization to enhance privacy protection and improve training efficiency. Firstly, the Laplacian noise mechanism was adopted to prevent user privacy information from being leaked during global model training by adding noise to the dataset, ensuring the privacy of each data access. Then, combined with the FedAvgM optimization algorithm, distributed nodes calculated local model parameters separately and merged these parameters through weighted averaging to reduce the training time of the global model and improve efficiency. Finally, layered encryption technology was adopted to add multiple layers of encryption during data transmission to ensure the security of the transmission link. At the same time, a dynamic permission allocation mechanism was introduced to limit the frequency of access to sensitive data. The experimental results demonstrate that when the privacy budget at ε=20, the privacy protection method of Laplace mechanism still maintains 73% accuracy in data transmission, with a privacy leakage risk of only 0.27. Under the same training epochs, the FedAvgM optimization algorithm achieved a correct data transmission accuracy of 97%. In comparison, the convergence speed increased by 0.29%/min in 1-20 training epochs, showing a faster convergence speed. Under four different noise attack methods, the anti-noise ability of layered encryption is the lowest at 15.4 dB, 18.5 dB, 17.1 dB, and 17.3 dB, respectively. This method effectively improves the problems of privacy leakage risk and low efficiency in traditional data access control.

Self-supervised learning fabric defect segmentation using anomaly generation

PurposeThe problem of fabric defect detection is a particularly challenging task, as the fabric defects occupy only a small portion of the image pixels and it is difficult to collect sufficient samples for training deep learning-based models. The purpose of this work is to present a novel self-supervised learning method to address this problem.Design/methodology/approachIn order to solve the problem of lack of defect samples, based on the fabric-specific degree of texture regularity, we propose an anomaly generation method to create synthetic fabric defects by destroying the normal fabric texture. To improve the detection of defects of different sizes, a global–local parallel detection mechanism is proposed. A self-supervised model including an anomaly generation module, a reconstruction subnetwork and a discriminative subnetwork is established to achieve model training without prior anomaly information.FindingsThe proposed method features self-supervised training, does not require no labelled anomaly data and detects anomalies by distinguishing their distance from normal samples at both global and local levels. When tested on four fabric datasets, our approach outperforms state-of-the-art unsupervised and self-supervised methods and achieves significantly higher localization accuracy.Originality/valueA high-fidelity fabric defects synthesis method is presented to alleviate the problem of collecting numerous fabric defects, providing a reference for other surface anomaly detection problems. The proposed global–local parallel detection mechanism can serve as a reference for other methods dedicated to detecting particular anomalies. The proposed self-supervised network model can effectively locate fabric anomalies without prior labelling information, which could be used as a framework for other model designs.

Characterization of adrenal glands on computed tomography with a 3D V-Net-based model

ObjectivesTo evaluate the performance of a 3D V-Net-based segmentation model of adrenal lesions in characterizing adrenal glands as normal or abnormal.MethodsA total of 1086 CT image series with focal adrenal lesions were retrospectively collected, annotated, and used for the training of the adrenal lesion segmentation model. The dice similarity coefficient (DSC) of the test set was used to evaluate the segmentation performance. The other cohort, consisting of 959 patients with pathologically confirmed adrenal lesions (external validation dataset 1), was included for validation of the classification performance of this model. Then, another consecutive cohort of patients with a history of malignancy (N = 479) was used for validation in the screening population (external validation dataset 2). Parameters of sensitivity, accuracy, etc., were used, and the performance of the model was compared to the radiology report in these validation scenes.ResultsThe DSC of the test set of the segmentation model was 0.900 (0.810–0.965) (median (interquartile range)). The model showed sensitivities and accuracies of 99.7%, 98.3% and 87.2%, 62.2% in external validation datasets 1 and 2, respectively. It showed no significant difference comparing to radiology reports in external validation datasets 1 and lesion-containing groups of external validation datasets 2 (p = 1.000 and p > 0.05, respectively).ConclusionThe 3D V-Net-based segmentation model of adrenal lesions can be used for the binary classification of adrenal glands.Critical relevance statementA 3D V-Net-based segmentation model of adrenal lesions can be used for the detection of abnormalities of adrenal glands, with a high accuracy in the pre-surgical scene as well as a high sensitivity in the screening scene.Key PointsAdrenal lesions may be prone to inter-observer variability in routine diagnostic workflow.The study developed a 3D V-Net-based segmentation model of adrenal lesions with DSC 0.900 in the test set.The model showed high sensitivity and accuracy of abnormalities detection in different scenes.Graphical

Deep generative modeling of annotated bacterial biofilm images

Biofilms are critical for understanding environmental processes, developing biotechnology applications, and progressing in medical treatments of various infections. Nowadays, a key limiting factor for biofilm analysis is the difficulty in obtaining large datasets with fully annotated images. This study introduces a versatile approach for creating synthetic datasets of annotated biofilm images with employing deep generative modeling techniques, including VAEs, GANs, diffusion models, and CycleGAN. Synthetic datasets can significantly improve the training of computer vision models for automated biofilm analysis, as demonstrated with the application of Mask R-CNN detection model. The approach represents a key advance in the field of biofilm research, offering a scalable solution for generating high-quality training data and working with different strains of microorganisms at different stages of formation. Terabyte-scale datasets can be easily generated on personal computers. A web application is provided for the on-demand generation of biofilm images.

Hierarchical skin lesion image classification with prototypical decision tree

Traditional disease classification models often disregard the clinical significance of misclassifications and lack interpretability. To overcome these challenges, we propose a hierarchical prototypical decision tree (HPDT) for skin lesion classification. HPDT combines prototypical networks and decision trees, leveraging a class hierarchy to guide interpretable predictions from general to specific categories. By incorporating a hierarchy-based distance matrix, the model prioritizes less severe misclassifications while maintaining diagnostic accuracy. Evaluated on a dataset of 235,268 dermoscopic images across 65 conditions, HPDT outperforms flat classifiers and existing hierarchical methods in accuracy, error severity reduction, and interpretability. It also generalizes effectively to unseen classes. These results highlight the value of integrating clinical hierarchies into model design and training to improve diagnostic reliability and decision transparency, demonstrating HPDT’s potential for clinical decision support.

An intrusion detection model based on Convolutional Kolmogorov-Arnold Networks

The application of artificial neural networks (ANNs) can be found in numerous fields, including image and speech recognition, natural language processing, and autonomous vehicles. As well, intrusion detection, the subject of this paper, relies heavily on it. Different intrusion detection models have been constructed using ANNs. While ANNs are relatively mature to construct intrusion detection models, some challenges remain. Among the most notorious of these are the bloated models caused by the large number of parameters, and the non-interpretability of the models. Our paper presents Convolutional Kolmogorov-Arnold Networks (CKANs), which are designed to overcome these difficulties and provide an interpretable and accurate intrusion detection model. Kolmogorov-Arnold Networks (KANs) are developed from the Kolmogorov-Arnold representation theorem. Meanwhile, CKAN incorporates a convolutional computational mechanism based on KAN. The model proposed in this paper is constructed by incorporating attention mechanisms into CKAN’s computational logic. The datasets CICIoT2023 and CICIoMT2024 were used for model training and validation. From the results of evaluating the performance indicators of the experiments, the intrusion detection model constructed based on CKANs has an attractive application prospect. As compared with other methods, the model can predict a much higher level of accuracy with significantly fewer parameters. However, it is not superior in terms of memory usage, execution speed and energy consumption.

A Data-Driven Deep Learning Framework for Prediction of Traffic Crashes at Road Intersections

Traffic crash prediction (TCP) is a fundamental problem for intelligent transportation systems in smart cities. Improving the accuracy of traffic crash prediction is important for road safety and effective traffic management. Owing to recent advances in artificial neural networks, several new deep-learning models have been proposed for TCP. However, these works mainly focus on accidents in regions, which are typically pre-determined using a grid map. We argue that TCP for roads, especially for crashes at or near road intersections which account for more than 50% of the fatal or injury crashes based on the Federal Highway Administration, has a significant practical and research value and thus deserves more research. In this paper, we formulate TCP at Road Intersections as a classification problem and propose a three-phase data-driven deep learning model, called Road Intersection Traffic Crash Prediction (RoadInTCP), to predict traffic crashes at intersections by exploiting publicly available heterogeneous big data. In Phase I we extract discriminative latent features called topological-relational features (tr-features), of intersections using a neural network model by exploiting topological information of the road network and various relationships amongst nearby intersections. In Phase II, in addition to tr-features which capture some inherent properties of the road network, we also explore additional thematic information in terms of environmental, traffic, weather, risk, and calendar features associated with intersections. In order to incorporate the potential correlation in nearby intersections, we utilize a Graph Convolution Network (GCN) to aggregate features from neighboring intersections based on a message-passing paradigm for TCP. While Phase II serves well as a TCP model, we further explore the signals embedded in the sequential feature changes over time for TCP in Phase III, by exploring RNN or 1DCNN which have known success on sequential data. Additionally, to address the serious issues of imbalanced classes in TCP and large-scale heterogeneous big data, we propose an effective data sampling approach in data preparation to facilitate model training. We evaluate the proposed RoadInTCP model via extensive experiments on a real-world New York City traffic dataset. The experimental results show that the proposed RoadInTCP robustly outperforms existing methods.