Baseline Model Research Articles

Deep learning for determining the difficulty of endodontic treatment: a pilot study.

To develop and validate a deep learning model for automated assessment of endodontic case difficulty from periapical radiographs. A dataset of 1,386 periapical radiographs was compiled from two clinical sites. Two dentists and two endodontists annotated the radiographs for difficulty using the "simple assessment" criteria from the American Association of Endodontists' case difficulty assessment form in the Endocase application. A classification task labeled cases as "easy" or "hard", while regression predicted overall difficulty scores. Convolutional neural networks (i.e. VGG16, ResNet18, ResNet50, ResNext50, and Inception v2) were used, with a baseline model trained via transfer learning from ImageNet weights. Other models was pre-trained using self-supervised contrastive learning (i.e. BYOL, SimCLR, MoCo, and DINO) on 20,295 unlabeled dental radiographs to learn representation without manual labels. Both models were evaluated using 10-fold cross-validation, with performance compared to seven human examiners (three general dentists and four endodontists) on a hold-out test set. The baseline VGG16 model attained 87.62% accuracy in classifying difficulty. Self-supervised pretraining did not improve performance. Regression predicted scores with ± 3.21 score error. All models outperformed human raters, with poor inter-examiner reliability. This pilot study demonstrated the feasibility of automated endodontic difficulty assessment via deep learning models.

Accuracy of Frailty Instruments in Predicting Outcomes following Perioperative Cardiac Arrest.

Frailty is associated with increased 30-day mortality and non-home discharge following perioperative cardiac arrest. We estimated the predictive accuracy of frailty when added to baseline risk prediction models. In this retrospective cohort study using 2015-2020 NSQIP data for 3,048 patients aged 50+ undergoing non-cardiac surgery and resuscitation on post-operative day 0 (i.e., intraoperatively or postoperatively on the day of surgery), baseline models including age, sex, ASA physical status, preoperative sepsis or septic shock, and emergent surgery were compared to models that added frailty indices, either RAI or mFI-5, to predict 30-day mortality and non-home discharge. Predictive accuracy was characterized by area under the receiver operating characteristic curve (AUC-ROC), integrated calibration index (ICI), and continuous net reclassification index (NRI). 1,786 patients (58.6%) died in the study cohort within 30 days, and 38.6% of eligible patients experienced non-home discharge. The baseline model showed good discrimination (AUC 0.77 for 30-day mortality and 0.74 for non-home discharge). AUC-ROC and ICI did not significantly change after adding frailty for 30-day mortality or non-home discharge. Adding RAI significantly improved NRI for 30-day mortality and non-home discharge; however, the magnitude was small and difficult to interpret, given other results including false positive and negative rates showing no difference in predictive accuracy. Incorporating frailty did not significantly improve predictive accuracy of models for 30-day mortality and non-home discharge following perioperative resuscitation. Thus, demonstrated associations between frailty and outcomes of perioperative resuscitation may not translate into improved predictive accuracy. When engaging patients in shared decision-making regarding do-not-resuscitate orders perioperatively, providers should acknowledge uncertainty in anticipating resuscitation outcomes.

TRAFMEL: Multimodal Entity Linking Based on Transformer Reranking and Multimodal Co-Attention Fusion

Multimodal entity linking aims to link mentions to target entities in the multimodal knowledge graph. The current multimodal entity linking mainly focuses on the global fusion of text and image, seldom fully exploring the correlation between modalities. In order to improve the fusion effect of multimodal feature, we propose a multimodal entity linking model based on a Multimodal Co-Attention Fusion strategy. This strategy is designed to enable text and image to guide each other for extracting features, thus making full exploration of the correlation between modalities to improve the fine-grained feature fusion effect. Furthermore, we also design a candidate entity generation strategy based on Transformer, which combines multiple candidate entity sets and adjusts the candidate entity ranking to obtain high-quality candidate entity sets. We perform experiments on domain datasets and public datasets, and the experimental results demonstrate that our model has a good performance in candidate entity generation and multimodal feature fusion, outperforming the state-of-the-art baseline models.

Beyond Event-Centric Narratives: Advancing Arabic Story Generation with Large Language Models and Beam Search

In the domain of automated story generation, the intricacies of the Arabic language pose distinct challenges. This study introduces a novel methodology that moves away from conventional event-driven narrative frameworks, emphasizing the restructuring of narrative constructs through sophisticated language models. Utilizing mBERT, our approach begins by extracting key story entities. Subsequently, XLM-RoBERTa and a BERT-based linguistic evaluation model are employed to direct beam search algorithms in the replacement of these entities. Further refinement is achieved through Low-Rank Adaptation (LoRA), which fine-tunes the extensive 3 billion-parameter BLOOMZ model specifically for generating Arabic narratives. Our methodology underwent thorough testing and validation, involving individual assessments of each submodel. The ROCStories dataset provided the training ground for our story entity extractor and new entity generator, and was also used in the fine-tuning of the BLOOMZ model. Additionally, the Arabic ComVE dataset was employed to train our commonsense evaluation model. Our extensive analyses yield crucial insights into the efficacy of our approach. The story entity extractor demonstrated robust performance with an F-score of 96.62%. Our commonsense evaluator reported an accuracy of 84.3%, surpassing the previous best by 3.1%. The innovative beam search strategy effectively produced entities that were linguistically and semantically superior to those generated using baseline models. Further subjective evaluations affirm our methodology’s capability to generate high-quality Arabic stories characterized by linguistic fluency and logical coherence.

Tradescantia response to air and soil pollution, stamen hair cells dataset and ANN color classification

Tradescantia plant is a complex system that is sensible to environmental factors such as water supply, pH, temperature, light, radiation, impurities, and nutrient availability. It can be used as a biomonitor for environmental changes; however, the bioassays are time-consuming and have a strong human interference factor that might change the result depending on who is performing the analysis. We have developed computer vision models to study color variations from Tradescantia clone 4430 plant stamen hair cells, which can be stressed due to air pollution and soil contamination. The study introduces a novel dataset, Trad-204, comprising single-cell images from Tradescantia clone 4430, captured during the Tradescantia stamen-hair mutation bioassay (Trad-SHM). The dataset contain images from two experiments, one focusing on air pollution by particulate matter and another based on soil contaminated by diesel oil. Both experiments were carried out in Curitiba, Brazil, between 2020 and 2023. The images represent single cells with different shapes, sizes, and colors, reflecting the plant's responses to environmental stressors. An automatic classification task was developed to distinguishing between blue and pink cells, and the study explores both a baseline model and three artificial neural network (ANN) architectures, namely, TinyVGG, VGG-16, and ResNet34. Tradescantia revealed sensibility to both air particulate matter concentration and diesel oil in soil. The results indicate that Residual Network architecture outperforms the other models in terms of accuracy on both training and testing sets. The dataset and findings contribute to the understanding of plant cell responses to environmental stress and provide valuable resources for further research in automated image analysis of plant cells. Discussion highlights the impact of turgor pressure on cell shape and the potential implications for plant physiology. The comparison between ANN architectures aligns with previous research, emphasizing the superior performance of ResNet models in image classification tasks. Artificial intelligence identification of pink cells improves the counting accuracy, thus avoiding human errors due to different color perceptions, fatigue, or inattention, in addition to facilitating and speeding up the analysis process. Overall, the study offers insights into plant cell dynamics and provides a foundation for future investigations like cells morphology change. This research corroborates that biomonitoring should be considered as an important tool for political actions, being a relevant issue in risk assessment and the development of new public policies relating to the environment.

Scoring Multi-hop Question Decomposition Using Masked Language Models

Question answering (QA) is a sub-field of Natural Language Processing (NLP) that focuses on developing systems capable of answering natural language queries. Within this domain, multi-hop question answering represents an advanced QA task that requires gathering and reasoning over multiple pieces of information from diverse sources or passages. To handle the complexity of multi-hop questions, question decomposition has been proven to be a valuable approach. This technique involves breaking down complex questions into simpler sub-questions, reducing the complexity of the problem. However, it’s worth noting that existing question decomposition methods often rely on training data, which may not always be readily available for low-resource languages or specialized domains. To address this issue, we propose a novel approach that utilizes pre-trained masked language models to score decomposition candidates in a zero-shot manner. The method involves generating decomposition candidates, scoring them using a pseudo-log likelihood estimation, and ranking them based on their scores. To evaluate the efficacy of the decomposition process, we conducted experiments on two datasets annotated on decomposition in two different languages, Arabic and English. Subsequently, we integrated our approach into a complete QA system and conducted a reading comprehension performance evaluation on the HotpotQA dataset. The obtained results emphasize that while the system exhibited a small drop in performance, it still maintained a significant advance compared to the baseline model. The proposed approach highlights the efficiency of the language model scoring technique in complex reasoning tasks such as multi-hop question decomposition.

ERSegDiff: a diffusion-based model for edge reshaping in medical image segmentation

Medical image segmentation is a crucial field of computer vision. Obtaining correct pathological areas can help clinicians analyze patient conditions more precisely. We have observed that both CNN-based and attention-based neural networks often produce rough segmentation results around the edges of the regions of interest. This significantly impacts the accuracy of obtaining the pathological areas. Without altering the original data and model architecture, further refining the initial segmentation outcomes can effectively address this issue and lead to more satisfactory results. Recently, diffusion models have demonstrated outstanding results in image generation, showcasing their powerful ability to model distributions. We believe that this ability can greatly enhance the accuracy of the reshaping results. This research proposes ERSegDiff, a neural network based on the diffusion model for reshaping segmentation borders. The diffusion model is trained to fit the distribution of the target edge area and is then used to modify the segmentation edge to produce more accurate segmentation results. By incorporating prior knowledge into the diffusion model, we can help it more accurately simulate the edge probability distribution of the samples. Moreover, we introduce the edge concern module, which leverages attention mechanisms to produce feature weights and further refine the segmentation outcomes. To validate our approach, we employed the COVID-19 and ISIC-2018 datasets for lung segmentation and skin cancer segmentation tasks, respectively. Compared with the baseline model, ERSegDiff improved the dice score by 3%–4% and 2%–4%, respectively, and achieved state-of-the-art scores compared to several mainstream neural networks, such as swinUNETR.

Speech Recognition for Air Traffic Control Utilizing a Multi-Head State-Space Model and Transfer Learning

Speech Recognition for Air Traffic Control Utilizing a Multi-Head State-Space Model and Transfer Learning

New Decomposition Models for Hourly Direct Normal Irradiance Estimations for Southern Africa

This research develops and validates new decomposition models for hourly direct Normal Irradiance (DNI) estimations for Southern African data. Localised models were developed using data collected from the Southern African Universities Radiometric Network (SAURAN). Clustered areas within Southern Africa were identified, and the developed cluster decomposition models highlighted the potential advantages of grouping data based on shared geographical and climatic attributes. This clustering approach could enhance decomposition model performance, particularly when local data are limited or when data are available from multiple nearby stations. Further, a regional Southern African decomposition model, which encompasses a wide spectrum of climatic regions and geographic locations, exhibited notable improvements over the baseline models despite occasional overestimation or underestimation. The results demonstrated improved DNI estimation accuracy compared to the baseline models across all testing and validation datasets. These outcomes suggest that utilising a localised model can significantly enhance DNI estimations for Southern Africa and potentially for developing similar models in diverse geographic regions worldwide. The overall metrics affirm the substantial advancement achieved with the regional model as an accurate decomposition model representing Southern Africa. Two stations were used as a validation study, as an application example where no localised model was available, and the cluster and regional models both outperformed the comparative decomposition models. This study focused on validating the model for hourly DNI in Southern Africa within a range of Kt-intervals from 0.175 to 0.875, and the range could be expanded and validated for future studies. Implementing accurate decomposition models in developing countries can accelerate the adoption of renewable energy sources, diminishing reliance on coal and fossil fuels.

An intelligent non-intrusive load monitoring model based on power encoding and convolutional state modules

Abstract Non-Intrusive Load Monitoring (NILM) identifies device power consumption or on/off states solely based on total power data, which is highly valuable for consumers to understand their appliance usage behavior and take necessary measures to reduce energy consumption, especially for the benefit of energy consumers' living production. However, a challenge faced by NILM is the tendency to focus excessively on power disaggregation while neglecting the disaggregation of on/off states, leading to lower classification accuracy, particularly owning to imbalanced states. This study proposes a model that integrates the power and on/off states to simultaneously disaggregate the power and device on/off states. The model comprises two main modules: a power encoding module for power disaggregation, and a convolutional state module for on/off state disaggregation. The power encoding module utilizes BERT-LSTM and Long Short-Term Memory (LSTM) networks for initial energy disaggregation. In contrast, the convolutional state module employs Convolutional Neural Networks (CNNs) for device state disaggregation. The output of the power-encoding module is multiplied by the probability of on/off states to obtain the final power.The proposed model is evaluated using the REDD and UK-DALE datasets. Compared to the baseline models, the results show an improvement in the device state classification average accuracy from 0.948 to 0.957, and a decrease in the average error between the real power and disaggregated power from 26.356W to 25.108W. Additionally, real-world experiments conducted using the designed platform for collecting and disaggregating power data achieve an average accuracy of 0.997. The proposed model demonstrates competitiveness in the NILM field and underscores its significance in aiding energy-consumption reduction efforts.

An improved Baldwin–Lomax algebraic wall model for high-speed canonical turbulent boundary layers using established scalings

In this work, we employ well-established relations for compressible turbulent mean flows, including the velocity transformation and algebraic temperature–velocity (TV) relation, to systematically improve the algebraic Baldwin–Lomax (BL) wall model for high-speed zero-pressure-gradient air boundary layers. Any new functions or coefficients fitted by ourselves are avoided. Twelve published direct numerical simulation (DNS) datasets are employed for a priori inspiration and a posteriori examination, with Mach numbers up to 14 under adiabatic, cold and heated walls. The baseline BL model is the widely used one with semilocal scalings. Three targeted modifications are made. First, we employ a total-stress-based transformation (Griffin et al., Proc. Natl Acad. Sci. USA, vol. 118, issue 34, 2021, e2111144118) to the inner-layer eddy viscosity for improved scaling up to the logarithmic region. Second, we utilize the van Driest transformation in the outer layer based on the compressible defect velocity scaling. Third, considering the difficulty in modelling the rapidly varying and singular turbulent Prandtl number near the temperature peak in cold-wall cases, we design a two-layer strategy and use the TV relation to formulate the inner-layer temperature. Numerical results prove that the modifications take effect as designed. The prediction accuracy for mean streamwise velocity is notably improved for diabatic cases, especially in the logarithmic region. Moreover, a significant improvement in mean temperature is realized for both adiabatic and diabatic cases. The mean relative errors of temperature to DNS for all cases are down to 0.4 % in the logarithmic wall-normal coordinate and 3.4 % in the outer coordinate, around one-third of those in the baseline model.

A hybrid method for intercity transport mode identification based on mobility features and sequential relations mined from cellular signaling data

AbstractThe proliferation of mobile phones has generated vast quantities of cellular signaling data (CSD), covering extensive spatial areas and populations. These data, containing spatiotemporal information, can be employed to identify and analyze intercity transport modes, providing valuable insights for understanding travel distribution and behavior. However, CSD are primarily intended for communication purposes and are not directly suitable for transportation research due to issues such as low spatial precision, sparse sampling granularity, and lacking traffic semantic features. This article proposes a Hybrid model for identifying individual intercity transport modes based on CSD. Several multidimensional mobility features are proposed that extract interpretable motion characteristics from CSD. A preliminary transport mode probability judgment is made based on the mobility features. Then, the complete transport mode is confirmed considering the temporal continuity correlation of the entire trace. Experiments confirm the Hybrid model's superior precision in identifying transport modes over baseline models, with an average F1 score of 0.92, maintaining high accuracy across various trajectory lengths. This model would support further studying individual intercity travel behavior patterns, aiding transportation planning and operational management decisions using CSD.

Deep convolutional neural network based synthetic minority over sampling technique: a forfending model for fraudulent credit card transactions in financial institution

Fraudulent credit card transactions are committed by unauthorized individuals and organizations employing methods such as phishing and social engineering fraud tactics. Researchers propose several Machine Learning (ML) techniques to deter the challenges of credit card fraud. However, the ML approaches are endorsed with some challenges, which makes the detection of credit card fraud extremely difficult. This study proposes a Deep Convolutional Neural Network (DCNN) with Synthetic Minority Oversampling Techniques (SMOTE) as an ideal solution. Kaggle datasets with 284,807 records and 31 features were exploited. Implementation was performed on the Google Colab cloud-based platform, embedding a Jupyter notebook setting with Graphical Processing Units (GPUs). Two experiments were conducted; the first was probed to determine suitable models among baseline models: Logistic Regression (LR), Random Forest (RF), Isolation Forest, and a single Deep Learning (DL) model of Multiple Layer Perceptron (MLP). The baseline models yielded an overfitting accuracy score, with recall, specificity, precision, and F1-score all presenting 1.00% respectively. This outcome is not sufficient in establishing findings on imbalanced data distribution as it's biased. This led to the construction of a new ML model incorporating Light Gradient Boosting Machine (LGBM), with Artificial Neural Network (ANN) and the proposed DCNN+SMOTE for the second experimental phase alongside baseline models. Experimental results via simulation show the proposed DCNN+SMOTE yielded awesome superclass performance across the board, displaying 1.00% results respectively. Its Error Rate (ER) and Null Error Rate (NER) are 0.00% distinctly. Meanwhile, the False Positive Rate (FPR) yields a 0.001% result, lesser and better than the baseline models.

A Robust DDoS Intrusion Detection System Using Convolutional Neural Network

In today’s digital age, the proliferation of network-connected devices has triggered a surge in cyberattacks. Distributed Denial-of-Service (DDoS) attacks pose a particularly formidable challenge to network security by disrupting access to vital services. While numerous researchers have proposed DDoS detection methods utilizing machine learning and deep learning techniques, developing a robust and reliable DDoS intrusion detection system remains challenging. This challenge is exacerbated by issues such as highly imbalanced data, multi-classification, and computational complexity. This paper proposes an innovative feature selection approach to create a robust intrusion detection system capable of detecting and classifying recent common DDoS attack types. We evaluate the performance of our model on the CICDDoS2019 benchmark dataset. Our experimental results demonstrate that our proposed model outperforms existing methods, achieving a detection accuracy of 96.82%, a recall of 96.82%, a precision of 96.76%, and an F1 score of 96.50%. Additionally, our model exhibits faster prediction times, with the ability to predict an attack in just 0.189 ms. Notably, our approach, combined with preprocessing and feature selection techniques, outperforms previous works and baseline models in DDoS attack classification.

Do traffic flow states follow Markov properties? A high-order spatiotemporal traffic state reconstruction approach for traffic prediction and imputation

Assessing traffic states accurately is challenging due to the complex, high-dimensional, and nonlinear nature of traffic systems. This study introduces the innovative High-Order Spatiotemporal Traffic State Reconstruction (HOSTSR) algorithm, designed to track and predict traffic flow dynamics effectively. It combines phase space reconstruction with time delays and high-order neighborhood concepts from graph theory to improve traffic state assessments' accuracy. The algorithm's effectiveness is validated using chi-square tests and the Chapman-Kolmogorov equation to confirm the Markovian properties of traffic flows. A lean autoencoder, informed by prior Markov knowledge of traffic states, is developed for mapping traffic states to real traffic data, proving highly effective for traffic data imputation due to the Markov model's memoryless property. Experimental results from the PeMSD04 and PeMSD08 datasets show that HOSTSR outperforms traditional state reconstruction methods based on delayed coordinate embedding in predicting future traffic flow state based on four key metrics. The autoencoder framework, guided by prior Markov knowledge, shows significant advantages in addressing traffic data gaps in different cases over six baseline models. Gradient sensitivity analysis further evaluates the impact of prior knowledge on improving the autoencoder's interpretability for interpolation efforts.

Soybean seedling detection and counting from UAV images based on an improved YOLOv8 Network

Abstract. The utilization of unmanned aerial vehicle (UAV) for soybean seedling detection is an effective way to estimate soybean yield, which plays a crucial role in agricultural planning and decision-making. However, the soybean seedlings objects in the UAV image are small, in clusters, and occluded each other, which makes it very challenging to achieve accurate object detection and counting. To address these issues, we optimize the YOLOv8 model and propose a GAS-YOLOv8 network, aiming to enhance the detection accuracy for the task of soybean seedling detection based on UAV images. Firstly, a global attention mechanism (GAM) is incorporated into the neck module of YOLOv8, which reallocates weights and prioritizes global information to more effectively extract soybean seedling features. Secondly, the CIOU loss function is replaced with the SIOU loss, which includes an angle loss term to guide the regression of bounding boxes. Experimental results show that, on the soybean seedling dataset, the proposed GAS-YOLOv8 model achieves a 1.3% improvement in mAP@0.5 and a 6% enhancement in detection performance in dense seedling areas, when compared to the baseline model YOLOv8s.When compared to other object detection models (YOLOv5, Faster R-CNN, etc.), the GAS-YOLOv8 model similarly achieved the best detection performance. These results demonstrate the effectiveness of the GAS-YOLOv8 in detecting dense soybean seedlings, providing more accurate theoretical support for subsequent yield estimation.

A Single-Buyer Model of Imbalance Cost Pass-Through Pricing Forecasting in the Malaysian Electricity Supply Industry

The imbalance cost pass-through (ICPT) is a flexible component of the incentive-based regulation (IBR) that empowers power producers to adjust tariffs in response to variable fuel prices, thereby enhancing the economic resilience of electricity generation. In Malaysia, the Energy Commission has conducted biannual reviews of fuel and other generation costs. Any cost savings or increases identified during these reviews will be passed on to customers in the form of rebates or surcharges. Meanwhile, if an increment in the ICPT price signal can be provided to electricity providers and consumers, early preparation for operation budgeting can be realised, and energy management program development can be properly prepared. Due to this reason, this study proposes ICPT price forecasting for the electricity market in Peninsular Malaysia that will benefit the stakeholders. The study aims to construct an ICPT-related baseline model for the peninsular generation data by employing three forecasting methods. The forecasting performance is analysed using the mean absolute percentage error (MAPE). In light of our findings, the ARIMA method is one of the most accurate forecasting methods for fuel prices compared to the moving average (MA) and LSSVM methods. The observed price differences between the ARIMA and LSSVM models for ICPT are minimal. The ICPT price for July–December 2022 and January–June 2023 is MYR 0.21/kWh for the ARIMA and MYR 0.18/kWh for LSSVM, which are close to the actual TNB’s ICPT tariff. As for forecasting, the ICPT price is expected to drop in the next announcement. The findings of this study may have a positive impact on the sustainability of the energy sector in Malaysia.

Development and validation of mortality prediction models based on the social determinants of health

BackgroundThere is no standardised approach to screening adults for social risk factors. The goal of this study was to develop mortality risk prediction models based on the social determinants of...

Multilingual Neural Machine Translation for Indic to Indic Languages

The method of translation from one language to another without human intervention is known as Machine Translation (MT). Multilingual neural machine translation (MNMT) is a technique for MT that builds a single model for multiple languages. It is preferred over other approaches, since it decreases training time and improves translation in low-resource contexts, i.e., for languages that have insufficient corpus. However, good-quality MT models are yet to be built for many scenarios such as for Indic-to-Indic Languages (IL-IL). Hence, this article is an attempt to address and develop the baseline models for low-resource languages i.e., IL-IL (for 11 Indic Languages (ILs)) in a multilingual environment. The models are built on the Samanantar corpus and analyzed on the Flores-200 corpus. All the models are evaluated using standard evaluation metrics i.e., Bilingual Evaluation Understudy (BLEU) score (with the range of 0 to 100). This article examines the effect of the grouping of related languages, namely, East Indo-Aryan (EI), Dravidian (DR), and West Indo-Aryan (WI) on the MNMT model. From the experiments, the results reveal that related language grouping is beneficial for the WI group only while it is detrimental for the EI group and it shows an inconclusive effect on the DR group. The role of pivot-based MNMT models in enhancing translation quality is also investigated in this article. Owing to the presence of large good-quality corpora from English (EN) to ILs, MNMT IL-IL models using EN as a pivot are built and examined. To achieve this, English-Indic Language (EN-IL) models are developed with and without the usage of related languages. Results show that the use of related language grouping is advantageous specifically for EN to ILs. Thus, related language groups are used for the development of pivot MNMT models. It is also observed that the usage of pivot models greatly improves MNMT baselines. Furthermore, the effect of transliteration on ILs is also analyzed in this article. To explore transliteration, the best MNMT models from the previous approaches (in most of cases pivot model using related groups) are determined and built on corpus transliterated from the corresponding scripts to a modified Indian language Transliteration script (ITRANS). The outcome of the experiments indicates that transliteration helps the models built for lexically rich languages, with the best increment of BLEU scores observed in Malayalam (ML) and Tamil (TA), i.e., 6.74 and 4.72, respectively. The BLEU score using transliteration models ranges from 7.03 to 24.29. The best model obtained is the Punjabi (PA)-Hindi (HI) language pair trained on PA-WI transliterated corpus.

Deep knowledge distillation: A self-mutual learning framework for traffic prediction

Traffic flow prediction in spatio-temporal networks is a crucial aspect of Intelligent Transportation Systems (ITS). Existing traffic flow forecasting methods, particularly those utilizing graph neural networks, encounter limitations. When processing large-scale graph data, the depth of these models can restrict their ability to effectively capture complex relationships and patterns. Additionally, these methods often focus mainly on local neighborhood information, which can limit their capability to recognize and analyze global relationships and patterns within the graph data. Therefore, we proposed a deep knowledge distillation model, tailored to effectively capture spatio-temporal patterns in traffic flow prediction. This model incorporates a bidirectional random walk process on a directed graph, enabling it to effectively capture both spatial and temporal dependencies. Utilizing a blend of mutual learning and self-distillation, our approach enhances the detection of spatio-temporal relationships within traffic data and improves the feature perception ability at both local and global levels. We tested our model on two real-world datasets, achieving notable improvements in prediction accuracy, especially for predictions within a one-hour timeframe. In comparison to the baseline model, our proposed model achieved accuracy improvements of 0.19 and 0.18 on the respective datasets. These results highlight the success of using mutual learning and self-distillation to transfer knowledge effectively within and between models and to improve the model’s capability in identifying and extracting features.