It is crucial to predict future mechanical behaviors for the prevention of structural disasters.Especially for underground construction,the structural mechanical behaviors are affected by multiple internal and externa...It is crucial to predict future mechanical behaviors for the prevention of structural disasters.Especially for underground construction,the structural mechanical behaviors are affected by multiple internal and external factors due to the complex conditions.Given that the existing models fail to take into account all the factors and accurate prediction of the multiple time series simultaneously is difficult using these models,this study proposed an improved prediction model through the autoencoder fused long-and short-term time-series network driven by the mass number of monitoring data.Then,the proposed model was formalized on multiple time series of strain monitoring data.Also,the discussion analysis with a classical baseline and an ablation experiment was conducted to verify the effectiveness of the prediction model.As the results indicate,the proposed model shows obvious superiority in predicting the future mechanical behaviors of structures.As a case study,the presented model was applied to the Nanjing Dinghuaimen tunnel to predict the stain variation on a different time scale in the future.展开更多
Electronic nose and thermal images are effective ways to diagnose the presence of gases in real-time realtime.Multimodal fusion of these modalities can result in the development of highly accurate diagnostic systems.T...Electronic nose and thermal images are effective ways to diagnose the presence of gases in real-time realtime.Multimodal fusion of these modalities can result in the development of highly accurate diagnostic systems.The low-cost thermal imaging software produces low-resolution thermal images in grayscale format,hence necessitating methods for improving the resolution and colorizing the images.The objective of this paper is to develop and train a super-resolution generative adversarial network for improving the resolution of the thermal images,followed by a sparse autoencoder for colorization of thermal images and amultimodal convolutional neural network for gas detection using electronic nose and thermal images.The dataset used comprises 6400 thermal images and electronic nose measurements for four classes.A multimodal Convolutional Neural Network(CNN)comprising an EfficientNetB2 pre-trainedmodel was developed using both early and late feature fusion.The Super Resolution Generative Adversarial Network(SRGAN)model was developed and trained on low and high-resolution thermal images.Asparse autoencoder was trained on the grayscale and colorized thermal images.The SRGAN was trained on lowand high-resolution thermal images,achieving a Structural Similarity Index(SSIM)of 90.28,a Peak Signal-to-Noise Ratio(PSNR)of 68.74,and a Mean Absolute Error(MAE)of 0.066.The autoencoder model produced an MAE of 0.035,a Mean Squared Error(MSE)of 0.006,and a Root Mean Squared Error(RMSE)of 0.0705.The multimodal CNN,trained on these images and electronic nose measurements using both early and late fusion techniques,achieved accuracies of 97.89% and 98.55%,respectively.Hence,the proposed framework can be of great aid for the integration with low-cost software to generate high quality thermal camera images and highly accurate detection of gases in real-time.展开更多
Virtual Power Plants(VPPs)are integral to modern energy systems,providing stability and reliability in the face of the inherent complexities and fluctuations of solar power data.Traditional anomaly detection methodolo...Virtual Power Plants(VPPs)are integral to modern energy systems,providing stability and reliability in the face of the inherent complexities and fluctuations of solar power data.Traditional anomaly detection methodologies often need to adequately handle these fluctuations from solar radiation and ambient temperature variations.We introduce the Memory-Enhanced Autoencoder with Adversarial Training(MemAAE)model to overcome these limitations,designed explicitly for robust anomaly detection in VPP environments.The MemAAE model integrates three principal components:an LSTM-based autoencoder that effectively captures temporal dynamics to distinguish between normal and anomalous behaviors,an adversarial training module that enhances system resilience across diverse operational scenarios,and a prediction module that aids the autoencoder during the reconstruction process,thereby facilitating precise anomaly identification.Furthermore,MemAAE features a memory mechanism that stores critical pattern information,mitigating overfitting,alongside a dynamic threshold adjustment mechanism that adapts detection thresholds in response to evolving operational conditions.Our empirical evaluation of the MemAAE model using real-world solar power data shows that the model outperforms other comparative models on both datasets.On the Sopan-Finder dataset,MemAAE has an accuracy of 99.17%and an F1-score of 95.79%,while on the Sunalab Faro PV 2017 dataset,it has an accuracy of 97.67%and an F1-score of 93.27%.Significant performance advantages have been achieved on both datasets.These results show that MemAAE model is an effective method for real-time anomaly detection in virtual power plants(VPPs),which can enhance robustness and adaptability to inherent variables in solar power generation.展开更多
We analyze the suitability of existing pre-trained transformer-based language models(PLMs)for abstractive text summarization on German technical healthcare texts.The study focuses on the multilingual capabilities of t...We analyze the suitability of existing pre-trained transformer-based language models(PLMs)for abstractive text summarization on German technical healthcare texts.The study focuses on the multilingual capabilities of these models and their ability to perform the task of abstractive text summarization in the healthcare field.The research hypothesis was that large language models could perform high-quality abstractive text summarization on German technical healthcare texts,even if the model is not specifically trained in that language.Through experiments,the research questions explore the performance of transformer language models in dealing with complex syntax constructs,the difference in performance between models trained in English and German,and the impact of translating the source text to English before conducting the summarization.We conducted an evaluation of four PLMs(GPT-3,a translation-based approach also utilizing GPT-3,a German language Model,and a domain-specific bio-medical model approach).The evaluation considered the informativeness using 3 types of metrics based on Recall-Oriented Understudy for Gisting Evaluation(ROUGE)and the quality of results which is manually evaluated considering 5 aspects.The results show that text summarization models could be used in the German healthcare domain and that domain-independent language models achieved the best results.The study proves that text summarization models can simplify the search for pre-existing German knowledge in various domains.展开更多
Pulse pile-up is a problem in nuclear spectroscopy and nuclear reaction studies that occurs when two pulses overlap and distort each other,degrading the quality of energy and timing information.Different methods have ...Pulse pile-up is a problem in nuclear spectroscopy and nuclear reaction studies that occurs when two pulses overlap and distort each other,degrading the quality of energy and timing information.Different methods have been used for pile-up rejection,both digital and analogue,but some pile-up events may contain pulses of interest and need to be reconstructed.The paper proposes a new method for reconstructing pile-up events acquired with a neutron detector array(NEDA)using an one-dimensional convolutional autoencoder(1D-CAE).The datasets for training and testing the 1D-CAE are created from data acquired from the NEDA.The new pile-up signal reconstruction method is evaluated from the point of view of how similar the reconstructed signals are to the original ones.Furthermore,it is analysed considering the result of the neutron-gamma discrimination based on charge comparison,comparing the result obtained from original and reconstructed signals.展开更多
Significant advancements have beenwitnessed in visual tracking applications leveragingViT in recent years,mainly due to the formidablemodeling capabilities of Vision Transformer(ViT).However,the strong performance of ...Significant advancements have beenwitnessed in visual tracking applications leveragingViT in recent years,mainly due to the formidablemodeling capabilities of Vision Transformer(ViT).However,the strong performance of such trackers heavily relies on ViT models pretrained for long periods,limitingmore flexible model designs for tracking tasks.To address this issue,we propose an efficient unsupervised ViT pretraining method for the tracking task based on masked autoencoders,called TrackMAE.During pretraining,we employ two shared-parameter ViTs,serving as the appearance encoder and motion encoder,respectively.The appearance encoder encodes randomly masked image data,while the motion encoder encodes randomly masked pairs of video frames.Subsequently,an appearance decoder and a motion decoder separately reconstruct the original image data and video frame data at the pixel level.In this way,ViT learns to understand both the appearance of images and the motion between video frames simultaneously.Experimental results demonstrate that ViT-Base and ViT-Large models,pretrained with TrackMAE and combined with a simple tracking head,achieve state-of-the-art(SOTA)performance without additional design.Moreover,compared to the currently popular MAE pretraining methods,TrackMAE consumes only 1/5 of the training time,which will facilitate the customization of diverse models for tracking.For instance,we additionally customize a lightweight ViT-XS,which achieves SOTA efficient tracking performance.展开更多
First,we propose a cross-domain authentication architecture based on trust evaluation mechanism,including registration,certificate issuance,and cross-domain authentication processes.A direct trust evaluation mechanism...First,we propose a cross-domain authentication architecture based on trust evaluation mechanism,including registration,certificate issuance,and cross-domain authentication processes.A direct trust evaluation mechanism based on the time decay factor is proposed,taking into account the influence of historical interaction records.We weight the time attenuation factor to each historical interaction record for updating and got the new historical record data.We refer to the beta distribution to enhance the flexibility and adaptability of the direct trust assessment model to better capture time trends in the historical record.Then we propose an autoencoder-based trust clustering algorithm.We perform feature extraction based on autoencoders.Kullback leibler(KL)divergence is used to calculate the reconstruction error.When constructing a convolutional autoencoder,we introduce convolutional neural networks to improve training efficiency and introduce sparse constraints into the hidden layer of the autoencoder.The sparse penalty term in the loss function measures the difference through the KL divergence.Trust clustering is performed based on the density based spatial clustering of applications with noise(DBSCAN)clustering algorithm.During the clustering process,edge nodes have a variety of trustworthy attribute characteristics.We assign different attribute weights according to the relative importance of each attribute in the clustering process,and a larger weight means that the attribute occupies a greater weight in the calculation of distance.Finally,we introduced adaptive weights to calculate comprehensive trust evaluation.Simulation experiments prove that our trust evaluation mechanism has excellent reliability and accuracy.展开更多
With the rapid development of the Internet,network security and data privacy are increasingly valued.Although classical Network Intrusion Detection System(NIDS)based on Deep Learning(DL)models can provide good detecti...With the rapid development of the Internet,network security and data privacy are increasingly valued.Although classical Network Intrusion Detection System(NIDS)based on Deep Learning(DL)models can provide good detection accuracy,but collecting samples for centralized training brings the huge risk of data privacy leakage.Furthermore,the training of supervised deep learning models requires a large number of labeled samples,which is usually cumbersome.The“black-box”problem also makes the DL models of NIDS untrustworthy.In this paper,we propose a trusted Federated Learning(FL)Traffic IDS method called FL-TIDS to address the above-mentioned problems.In FL-TIDS,we design an unsupervised intrusion detection model based on autoencoders that alleviates the reliance on marked samples.At the same time,we use FL for model training to protect data privacy.In addition,we design an improved SHAP interpretable method based on chi-square test to perform interpretable analysis of the trained model.We conducted several experiments to evaluate the proposed FL-TIDS.We first determine experimentally the structure and the number of neurons of the unsupervised AE model.Secondly,we evaluated the proposed method using the UNSW-NB15 and CICIDS2017 datasets.The exper-imental results show that the unsupervised AE model has better performance than the other 7 intrusion detection models in terms of precision,recall and f1-score.Then,federated learning is used to train the intrusion detection model.The experimental results indicate that the model is more accurate than the local learning model.Finally,we use an improved SHAP explainability method based on Chi-square test to analyze the explainability.The analysis results show that the identification characteristics of the model are consistent with the attack characteristics,and the model is reliable.展开更多
For data mining tasks on large-scale data,feature selection is a pivotal stage that plays an important role in removing redundant or irrelevant features while improving classifier performance.Traditional wrapper featu...For data mining tasks on large-scale data,feature selection is a pivotal stage that plays an important role in removing redundant or irrelevant features while improving classifier performance.Traditional wrapper feature selection methodologies typically require extensive model training and evaluation,which cannot deliver desired outcomes within a reasonable computing time.In this paper,an innovative wrapper approach termed Contribution Tracking Feature Selection(CTFS)is proposed for feature selection of large-scale data,which can locate informative features without population-level evolution.In other words,fewer evaluations are needed for CTFS compared to other evolutionary methods.We initially introduce a refined sparse autoencoder to assess the prominence of each feature in the subsequent wrapper method.Subsequently,we utilize an enhanced wrapper feature selection technique that merges Mutual Information(MI)with individual feature contributions.Finally,a fine-tuning contribution tracking mechanism discerns informative features within the optimal feature subset,operating via a dominance accumulation mechanism.Experimental results for multiple classification performance metrics demonstrate that the proposed method effectively yields smaller feature subsets without degrading classification performance in an acceptable runtime compared to state-of-the-art algorithms across most large-scale benchmark datasets.展开更多
This paper addresses the challenge of identifying abnormal states in Lithium-ion Battery(LiB)time series data.As the energy sector increasingly focuses on integrating distributed energy resources,Virtual Power Plants(...This paper addresses the challenge of identifying abnormal states in Lithium-ion Battery(LiB)time series data.As the energy sector increasingly focuses on integrating distributed energy resources,Virtual Power Plants(VPP)have become a vital new framework for energy management.LiBs are key in this context,owing to their high-efficiency energy storage capabilities essential for VPP operations.However,LiBs are prone to various abnormal states like overcharging,over-discharging,and internal short circuits,which impede power transmission efficiency.Traditional methods for detecting such abnormalities in LiB are too broad and lack precision for the dynamic and irregular nature of LiB data.In response,we introduce an innovative method:a Long Short-Term Memory(LSTM)autoencoder based on Dynamic Frequency Memory and Correlation Attention(DFMCA-LSTM-AE).This unsupervised,end-to-end approach is specifically designed for dynamically monitoring abnormal states in LiB data.The method starts with a Dynamic Frequency Fourier Transform module,which dynamically captures the frequency characteristics of time series data across three scales,incorporating a memory mechanism to reduce overgeneralization of abnormal frequencies.This is followed by integrating LSTM into both the encoder and decoder,enabling the model to effectively encode and decode the temporal relationships in the time series.Empirical tests on a real-world LiB dataset demonstrate that DFMCA-LSTM-AE outperforms existing models,achieving an average Area Under the Curve(AUC)of 90.73%and an F1 score of 83.83%.These results mark significant improvements over existing models,ranging from 2.4%–45.3%for AUC and 1.6%–28.9%for F1 score,showcasing the model’s enhanced accuracy and reliability in detecting abnormal states in LiB data.展开更多
Wayside monitoring is a promising cost-effective alternative to predict damage in the rolling stock. The main goal of this work is to present an unsupervised methodology to identify out-of-roundness(OOR) damage wheels...Wayside monitoring is a promising cost-effective alternative to predict damage in the rolling stock. The main goal of this work is to present an unsupervised methodology to identify out-of-roundness(OOR) damage wheels, such as wheel flats and polygonal wheels. This automatic damage identification algorithm is based on the vertical acceleration evaluated on the rails using a virtual wayside monitoring system and involves the application of a two-step procedure. The first step aims to define a confidence boundary by using(healthy) measurements evaluated on the rail constituting a baseline. The second step of the procedure involves classifying damage of predefined scenarios with different levels of severities. The proposed procedure is based on a machine learning methodology and includes the following stages:(1) data collection,(2) damage-sensitive feature extraction from the acquired responses using a neural network model, i.e., the sparse autoencoder(SAE),(3) data fusion based on the Mahalanobis distance, and(4) unsupervised feature classification by implementing outlier and cluster analysis. This procedure considers baseline responses at different speeds and rail irregularities to train the SAE model. Then, the trained SAE is capable to reconstruct test responses(not trained) allowing to compute the accumulative difference between original and reconstructed signals. The results prove the efficiency of the proposed approach in identifying the two most common types of OOR in railway wheels.展开更多
Network security problems bring many imperceptible threats to the integrity of data and the reliability of device services,so proposing a network intrusion detection model with high reliability is of great research si...Network security problems bring many imperceptible threats to the integrity of data and the reliability of device services,so proposing a network intrusion detection model with high reliability is of great research significance for network security.Due to the strong generalization of invalid features during training process,it is more difficult for single autoencoder intrusion detection model to obtain effective results.A network intrusion detection model based on the Ensemble of Denoising Adversarial Autoencoder(EDAAE)was proposed,which had higher accuracy and reliability compared to the traditional anomaly detection model.Using the adversarial learning idea of Adversarial Autoencoder(AAE),the discriminator module was added to the original model,and the encoder part was used as the generator.The distribution of the hidden space of the data generated by the encoder matched with the distribution of the original data.The generalization of the model to the invalid features was also reduced to improve the detection accuracy.At the same time,the denoising autoencoder and integrated operation was introduced to prevent overfitting in the adversarial learning process.Experiments on the CICIDS2018 traffic dataset showed that the proposed intrusion detection model achieves an Accuracy of 95.23%,which out performs traditional self-encoders and other existing intrusion detection models methods in terms of overall performance.展开更多
In the context of rapid digitization in industrial environments,how effective are advanced unsupervised learning models,particularly hybrid autoencoder models,at detecting anomalies in industrial control system(ICS)da...In the context of rapid digitization in industrial environments,how effective are advanced unsupervised learning models,particularly hybrid autoencoder models,at detecting anomalies in industrial control system(ICS)datasets?This study is crucial because it addresses the challenge of identifying rare and complex anomalous patterns in the vast amounts of time series data generated by Internet of Things(IoT)devices,which can significantly improve the reliability and safety of these systems.In this paper,we propose a hybrid autoencoder model,called ConvBiLSTMAE,which combines convolutional neural network(CNN)and bidirectional long short-term memory(BiLSTM)to more effectively train complex temporal data patterns in anomaly detection.On the hardware-in-the-loopbased extended industrial control system dataset,the ConvBiLSTM-AE model demonstrated remarkable anomaly detection performance,achieving F1 scores of 0.78 and 0.41 for the first and second datasets,respectively.The results suggest that hybrid autoencoder models are not only viable,but potentially superior alternatives for unsupervised anomaly detection in complex industrial systems,offering a promising approach to improving their reliability and safety.展开更多
Although speech emotion recognition is challenging,it has broad application prospects in human-computer interaction.Building a system that can accurately and stably recognize emotions from human languages can provide ...Although speech emotion recognition is challenging,it has broad application prospects in human-computer interaction.Building a system that can accurately and stably recognize emotions from human languages can provide a better user experience.However,the current unimodal emotion feature representations are not distinctive enough to accomplish the recognition,and they do not effectively simulate the inter-modality dynamics in speech emotion recognition tasks.This paper proposes a multimodal method that utilizes both audio and semantic content for speech emotion recognition.The proposed method consists of three parts:two high-level feature extractors for text and audio modalities,and an autoencoder-based feature fusion.For audio modality,we propose a structure called Temporal Global Feature Extractor(TGFE)to extract the high-level features of the timefrequency domain relationship from the original speech signal.Considering that text lacks frequency information,we use only a Bidirectional Long Short-Term Memory network(BLSTM)and attention mechanism to simulate an intra-modal dynamic.Once these steps have been accomplished,the high-level text and audio features are sent to the autoencoder in parallel to learn their shared representation for final emotion classification.We conducted extensive experiments on three public benchmark datasets to evaluate our method.The results on Interactive Emotional Motion Capture(IEMOCAP)and Multimodal EmotionLines Dataset(MELD)outperform the existing method.Additionally,the results of CMU Multi-modal Opinion-level Sentiment Intensity(CMU-MOSI)are competitive.Furthermore,experimental results show that compared to unimodal information and autoencoderbased feature level fusion,the joint multimodal information(audio and text)improves the overall performance and can achieve greater accuracy than simple feature concatenation.展开更多
Fault diagnosis of traction systems is important for the safety operation of high-speed trains.Long-term operation of the trains will degrade the performance of systems,which decreases the fault detection accuracy.To ...Fault diagnosis of traction systems is important for the safety operation of high-speed trains.Long-term operation of the trains will degrade the performance of systems,which decreases the fault detection accuracy.To solve this problem,this paper proposes a fault detection method developed by a Generalized Autoencoder(GAE)for systems with performance degradation.The advantage of this method is that it can accurately detect faults when the traction system of high-speed trains is affected by performance degradation.Regardless of the probability distribution,it can handle any data,and the GAE has extremely high sensitivity in anomaly detection.Finally,the effectiveness of this method is verified through the Traction Drive Control System(TDCS)platform.At different performance degradation levels,our method’s experimental results are superior to traditional methods.展开更多
To enhance the accuracy and efficiency of bridge damage identification,a novel data-driven damage identification method was proposed.First,convolutional autoencoder(CAE)was used to extract key features from the accele...To enhance the accuracy and efficiency of bridge damage identification,a novel data-driven damage identification method was proposed.First,convolutional autoencoder(CAE)was used to extract key features from the acceleration signal of the bridge structure through data reconstruction.The extreme gradient boosting tree(XGBoost)was then used to perform analysis on the feature data to achieve damage detection with high accuracy and high performance.The proposed method was applied in a numerical simulation study on a three-span continuous girder and further validated experimentally on a scaled model of a cable-stayed bridge.The numerical simulation results show that the identification errors remain within 2.9%for six single-damage cases and within 3.1%for four double-damage cases.The experimental validation results demonstrate that when the tension in a single cable of the cable-stayed bridge decreases by 20%,the method accurately identifies damage at different cable locations using only sensors installed on the main girder,achieving identification accuracies above 95.8%in all cases.The proposed method shows high identification accuracy and generalization ability across various damage scenarios.展开更多
Sentence classification is the process of categorizing a sentence based on the context of the sentence.Sentence categorization requires more semantic highlights than other tasks,such as dependence parsing,which requir...Sentence classification is the process of categorizing a sentence based on the context of the sentence.Sentence categorization requires more semantic highlights than other tasks,such as dependence parsing,which requires more syntactic elements.Most existing strategies focus on the general semantics of a conversation without involving the context of the sentence,recognizing the progress and comparing impacts.An ensemble pre-trained language model was taken up here to classify the conversation sentences from the conversation corpus.The conversational sentences are classified into four categories:information,question,directive,and commission.These classification label sequences are for analyzing the conversation progress and predicting the pecking order of the conversation.Ensemble of Bidirectional Encoder for Representation of Transformer(BERT),Robustly Optimized BERT pretraining Approach(RoBERTa),Generative Pre-Trained Transformer(GPT),DistilBERT and Generalized Autoregressive Pretraining for Language Understanding(XLNet)models are trained on conversation corpus with hyperparameters.Hyperparameter tuning approach is carried out for better performance on sentence classification.This Ensemble of Pre-trained Language Models with a Hyperparameter Tuning(EPLM-HT)system is trained on an annotated conversation dataset.The proposed approach outperformed compared to the base BERT,GPT,DistilBERT and XLNet transformer models.The proposed ensemble model with the fine-tuned parameters achieved an F1_score of 0.88.展开更多
We present an approach to classify medical text at a sentence level automatically.Given the inherent complexity of medical text classification,we employ adapters based on pre-trained language models to extract informa...We present an approach to classify medical text at a sentence level automatically.Given the inherent complexity of medical text classification,we employ adapters based on pre-trained language models to extract information from medical text,facilitating more accurate classification while minimizing the number of trainable parameters.Extensive experiments conducted on various datasets demonstrate the effectiveness of our approach.展开更多
Network embedding(NE)tries to learn the potential properties of complex networks represented in a low-dimensional feature space.However,the existing deep learningbased NE methods are time-consuming as they need to tra...Network embedding(NE)tries to learn the potential properties of complex networks represented in a low-dimensional feature space.However,the existing deep learningbased NE methods are time-consuming as they need to train a dense architecture for deep neural networks with extensive unknown weight parameters.A sparse deep autoencoder(called SPDNE)for dynamic NE is proposed,aiming to learn the network structures while preserving the node evolution with a low computational complexity.SPDNE tries to use an optimal sparse architecture to replace the fully connected architecture in the deep autoencoder while maintaining the performance of these models in the dynamic NE.Then,an adaptive simulated algorithm to find the optimal sparse architecture for the deep autoencoder is proposed.The performance of SPDNE over three dynamical NE models(i.e.sparse architecture-based deep autoencoder method,DynGEM,and ElvDNE)is evaluated on three well-known benchmark networks and five real-world networks.The experimental results demonstrate that SPDNE can reduce about 70%of weight parameters of the architecture for the deep autoencoder during the training process while preserving the performance of these dynamical NE models.The results also show that SPDNE achieves the highest accuracy on 72 out of 96 edge prediction and network reconstruction tasks compared with the state-of-the-art dynamical NE algorithms.展开更多
Many types of real-world information systems, including social media and e-commerce platforms, can be modelled by means of attribute-rich, connected networks. The goal of anomaly detection in artificial intelligence i...Many types of real-world information systems, including social media and e-commerce platforms, can be modelled by means of attribute-rich, connected networks. The goal of anomaly detection in artificial intelligence is to identify illustrations that deviate significantly from the main distribution of data or that differ from known cases. Anomalous nodes in node-attributed networks can be identified with greater precision if both graph and node attributes are taken into account. Almost all of the studies in this area focus on supervised techniques for spotting outliers. While supervised algorithms for anomaly detection work well in theory, they cannot be applied to real-world applications owing to a lack of labelled data. Considering the possible data distribution, our model employs a dual variational autoencoder (VAE), while a generative adversarial network (GAN) assures that the model is robust to adversarial training. The dual VAEs are used in another capacity: as a fake-node generator. Adversarial training is used to ensure that our latent codes have a Gaussian or uniform distribution. To provide a fair presentation of the graph, the discriminator instructs the generator to generate latent variables with distributions that are more consistent with the actual distribution of the data. Once the model has been learned, the discriminator is used for anomaly detection via reconstruction loss which has been trained to distinguish between the normal and artificial distributions of data. First, using a dual VAE, our model simultaneously captures cross-modality interactions between topological structure and node characteristics and overcomes the problem of unlabeled anomalies, allowing us to better understand the network sparsity and nonlinearity. Second, the proposed model considers the regularization of the latent codes while solving the issue of unregularized embedding techniques that can quickly lead to unsatisfactory representation. Finally, we use the discriminator reconstruction loss for anomaly detection as the discriminator is well-trained to separate the normal and generated data distributions because reconstruction-based loss does not include the adversarial component. Experiments conducted on attributed networks demonstrate the effectiveness of the proposed model and show that it greatly surpasses the previous methods. The area under the curve scores of our proposed model for the BlogCatalog, Flickr, and Enron datasets are 0.83680, 0.82020, and 0.71180, respectively, proving the effectiveness of the proposed model. The result of the proposed model on the Enron dataset is slightly worse than other models;we attribute this to the dataset’s low dimensionality as the most probable explanation.展开更多
基金National Key Research and Development Program of China,Grant/Award Number:2018YFB2101003National Natural Science Foundation of China,Grant/Award Numbers:51991395,U1806226,51778033,51822802,71901011,U1811463,51991391Science and Technology Major Project of Beijing,Grant/Award Number:Z191100002519012。
文摘It is crucial to predict future mechanical behaviors for the prevention of structural disasters.Especially for underground construction,the structural mechanical behaviors are affected by multiple internal and external factors due to the complex conditions.Given that the existing models fail to take into account all the factors and accurate prediction of the multiple time series simultaneously is difficult using these models,this study proposed an improved prediction model through the autoencoder fused long-and short-term time-series network driven by the mass number of monitoring data.Then,the proposed model was formalized on multiple time series of strain monitoring data.Also,the discussion analysis with a classical baseline and an ablation experiment was conducted to verify the effectiveness of the prediction model.As the results indicate,the proposed model shows obvious superiority in predicting the future mechanical behaviors of structures.As a case study,the presented model was applied to the Nanjing Dinghuaimen tunnel to predict the stain variation on a different time scale in the future.
基金funded by the Centre for Advanced Modelling and Geospatial Information Systems(CAMGIS),Faculty of Engineering and IT,University of Technology Sydneysupported by the Researchers Supporting Project,King Saud University,Riyadh,Saudi Arabia,under Project RSP2025 R14.
文摘Electronic nose and thermal images are effective ways to diagnose the presence of gases in real-time realtime.Multimodal fusion of these modalities can result in the development of highly accurate diagnostic systems.The low-cost thermal imaging software produces low-resolution thermal images in grayscale format,hence necessitating methods for improving the resolution and colorizing the images.The objective of this paper is to develop and train a super-resolution generative adversarial network for improving the resolution of the thermal images,followed by a sparse autoencoder for colorization of thermal images and amultimodal convolutional neural network for gas detection using electronic nose and thermal images.The dataset used comprises 6400 thermal images and electronic nose measurements for four classes.A multimodal Convolutional Neural Network(CNN)comprising an EfficientNetB2 pre-trainedmodel was developed using both early and late feature fusion.The Super Resolution Generative Adversarial Network(SRGAN)model was developed and trained on low and high-resolution thermal images.Asparse autoencoder was trained on the grayscale and colorized thermal images.The SRGAN was trained on lowand high-resolution thermal images,achieving a Structural Similarity Index(SSIM)of 90.28,a Peak Signal-to-Noise Ratio(PSNR)of 68.74,and a Mean Absolute Error(MAE)of 0.066.The autoencoder model produced an MAE of 0.035,a Mean Squared Error(MSE)of 0.006,and a Root Mean Squared Error(RMSE)of 0.0705.The multimodal CNN,trained on these images and electronic nose measurements using both early and late fusion techniques,achieved accuracies of 97.89% and 98.55%,respectively.Hence,the proposed framework can be of great aid for the integration with low-cost software to generate high quality thermal camera images and highly accurate detection of gases in real-time.
基金supported by“Regional Innovation Strategy(RIS)”through the National Research Foundation of Korea(NRF)funded by the Ministry of Education(MOE)(2021RIS-002)the Technology Development Program(RS-2023-00266141)funded by the Ministry of SMEs and Startups(MSS,Republic of Korea).
文摘Virtual Power Plants(VPPs)are integral to modern energy systems,providing stability and reliability in the face of the inherent complexities and fluctuations of solar power data.Traditional anomaly detection methodologies often need to adequately handle these fluctuations from solar radiation and ambient temperature variations.We introduce the Memory-Enhanced Autoencoder with Adversarial Training(MemAAE)model to overcome these limitations,designed explicitly for robust anomaly detection in VPP environments.The MemAAE model integrates three principal components:an LSTM-based autoencoder that effectively captures temporal dynamics to distinguish between normal and anomalous behaviors,an adversarial training module that enhances system resilience across diverse operational scenarios,and a prediction module that aids the autoencoder during the reconstruction process,thereby facilitating precise anomaly identification.Furthermore,MemAAE features a memory mechanism that stores critical pattern information,mitigating overfitting,alongside a dynamic threshold adjustment mechanism that adapts detection thresholds in response to evolving operational conditions.Our empirical evaluation of the MemAAE model using real-world solar power data shows that the model outperforms other comparative models on both datasets.On the Sopan-Finder dataset,MemAAE has an accuracy of 99.17%and an F1-score of 95.79%,while on the Sunalab Faro PV 2017 dataset,it has an accuracy of 97.67%and an F1-score of 93.27%.Significant performance advantages have been achieved on both datasets.These results show that MemAAE model is an effective method for real-time anomaly detection in virtual power plants(VPPs),which can enhance robustness and adaptability to inherent variables in solar power generation.
文摘We analyze the suitability of existing pre-trained transformer-based language models(PLMs)for abstractive text summarization on German technical healthcare texts.The study focuses on the multilingual capabilities of these models and their ability to perform the task of abstractive text summarization in the healthcare field.The research hypothesis was that large language models could perform high-quality abstractive text summarization on German technical healthcare texts,even if the model is not specifically trained in that language.Through experiments,the research questions explore the performance of transformer language models in dealing with complex syntax constructs,the difference in performance between models trained in English and German,and the impact of translating the source text to English before conducting the summarization.We conducted an evaluation of four PLMs(GPT-3,a translation-based approach also utilizing GPT-3,a German language Model,and a domain-specific bio-medical model approach).The evaluation considered the informativeness using 3 types of metrics based on Recall-Oriented Understudy for Gisting Evaluation(ROUGE)and the quality of results which is manually evaluated considering 5 aspects.The results show that text summarization models could be used in the German healthcare domain and that domain-independent language models achieved the best results.The study proves that text summarization models can simplify the search for pre-existing German knowledge in various domains.
基金partially supported by MICIU MCIN/AEI/10.13039/501100011033Spain with grant PID2020-118265GB-C42,-C44,PRTR-C17.I01+1 种基金Generalitat Valenciana,Spain with grant CIPROM/2022/54,ASFAE/2022/031,CIAPOS/2021/114the EU NextGenerationEU,ESF funds,and the National Science Centre (NCN),Poland (grant No.2020/39/D/ST2/00466)
文摘Pulse pile-up is a problem in nuclear spectroscopy and nuclear reaction studies that occurs when two pulses overlap and distort each other,degrading the quality of energy and timing information.Different methods have been used for pile-up rejection,both digital and analogue,but some pile-up events may contain pulses of interest and need to be reconstructed.The paper proposes a new method for reconstructing pile-up events acquired with a neutron detector array(NEDA)using an one-dimensional convolutional autoencoder(1D-CAE).The datasets for training and testing the 1D-CAE are created from data acquired from the NEDA.The new pile-up signal reconstruction method is evaluated from the point of view of how similar the reconstructed signals are to the original ones.Furthermore,it is analysed considering the result of the neutron-gamma discrimination based on charge comparison,comparing the result obtained from original and reconstructed signals.
基金supported in part by National Natural Science Foundation of China(No.62176041)in part by Excellent Science and Technique Talent Foundation of Dalian(No.2022RY21).
文摘Significant advancements have beenwitnessed in visual tracking applications leveragingViT in recent years,mainly due to the formidablemodeling capabilities of Vision Transformer(ViT).However,the strong performance of such trackers heavily relies on ViT models pretrained for long periods,limitingmore flexible model designs for tracking tasks.To address this issue,we propose an efficient unsupervised ViT pretraining method for the tracking task based on masked autoencoders,called TrackMAE.During pretraining,we employ two shared-parameter ViTs,serving as the appearance encoder and motion encoder,respectively.The appearance encoder encodes randomly masked image data,while the motion encoder encodes randomly masked pairs of video frames.Subsequently,an appearance decoder and a motion decoder separately reconstruct the original image data and video frame data at the pixel level.In this way,ViT learns to understand both the appearance of images and the motion between video frames simultaneously.Experimental results demonstrate that ViT-Base and ViT-Large models,pretrained with TrackMAE and combined with a simple tracking head,achieve state-of-the-art(SOTA)performance without additional design.Moreover,compared to the currently popular MAE pretraining methods,TrackMAE consumes only 1/5 of the training time,which will facilitate the customization of diverse models for tracking.For instance,we additionally customize a lightweight ViT-XS,which achieves SOTA efficient tracking performance.
基金This work is supported by the 2022 National Key Research and Development Plan“Security Protection Technology for Critical Information Infrastructure of Distribution Network”(2022YFB3105100).
文摘First,we propose a cross-domain authentication architecture based on trust evaluation mechanism,including registration,certificate issuance,and cross-domain authentication processes.A direct trust evaluation mechanism based on the time decay factor is proposed,taking into account the influence of historical interaction records.We weight the time attenuation factor to each historical interaction record for updating and got the new historical record data.We refer to the beta distribution to enhance the flexibility and adaptability of the direct trust assessment model to better capture time trends in the historical record.Then we propose an autoencoder-based trust clustering algorithm.We perform feature extraction based on autoencoders.Kullback leibler(KL)divergence is used to calculate the reconstruction error.When constructing a convolutional autoencoder,we introduce convolutional neural networks to improve training efficiency and introduce sparse constraints into the hidden layer of the autoencoder.The sparse penalty term in the loss function measures the difference through the KL divergence.Trust clustering is performed based on the density based spatial clustering of applications with noise(DBSCAN)clustering algorithm.During the clustering process,edge nodes have a variety of trustworthy attribute characteristics.We assign different attribute weights according to the relative importance of each attribute in the clustering process,and a larger weight means that the attribute occupies a greater weight in the calculation of distance.Finally,we introduced adaptive weights to calculate comprehensive trust evaluation.Simulation experiments prove that our trust evaluation mechanism has excellent reliability and accuracy.
基金supported by National Natural Science Fundation of China under Grant 61972208National Natural Science Fundation(General Program)of China under Grant 61972211+2 种基金National Key Research and Development Project of China under Grant 2020YFB1804700Future Network Innovation Research and Application Projects under Grant No.2021FNA020062021 Jiangsu Postgraduate Research Innovation Plan under Grant No.KYCX210794.
文摘With the rapid development of the Internet,network security and data privacy are increasingly valued.Although classical Network Intrusion Detection System(NIDS)based on Deep Learning(DL)models can provide good detection accuracy,but collecting samples for centralized training brings the huge risk of data privacy leakage.Furthermore,the training of supervised deep learning models requires a large number of labeled samples,which is usually cumbersome.The“black-box”problem also makes the DL models of NIDS untrustworthy.In this paper,we propose a trusted Federated Learning(FL)Traffic IDS method called FL-TIDS to address the above-mentioned problems.In FL-TIDS,we design an unsupervised intrusion detection model based on autoencoders that alleviates the reliance on marked samples.At the same time,we use FL for model training to protect data privacy.In addition,we design an improved SHAP interpretable method based on chi-square test to perform interpretable analysis of the trained model.We conducted several experiments to evaluate the proposed FL-TIDS.We first determine experimentally the structure and the number of neurons of the unsupervised AE model.Secondly,we evaluated the proposed method using the UNSW-NB15 and CICIDS2017 datasets.The exper-imental results show that the unsupervised AE model has better performance than the other 7 intrusion detection models in terms of precision,recall and f1-score.Then,federated learning is used to train the intrusion detection model.The experimental results indicate that the model is more accurate than the local learning model.Finally,we use an improved SHAP explainability method based on Chi-square test to analyze the explainability.The analysis results show that the identification characteristics of the model are consistent with the attack characteristics,and the model is reliable.
基金supported in part by the National Key Research and Development Program of China under Grant(No.2021YFB3300900)the NSFC Key Supported Project of the Major Research Plan under Grant(No.92267206)+2 种基金the National Natural Science Foundation of China under Grant(Nos.72201052,62032013,62173076)the Fundamental Research Funds for the Central Universities under Grant(No.N2204017)the Fundamental Research Funds for State Key Laboratory of Synthetical Automation for Process Industries under Grant(No.2013ZCX11).
文摘For data mining tasks on large-scale data,feature selection is a pivotal stage that plays an important role in removing redundant or irrelevant features while improving classifier performance.Traditional wrapper feature selection methodologies typically require extensive model training and evaluation,which cannot deliver desired outcomes within a reasonable computing time.In this paper,an innovative wrapper approach termed Contribution Tracking Feature Selection(CTFS)is proposed for feature selection of large-scale data,which can locate informative features without population-level evolution.In other words,fewer evaluations are needed for CTFS compared to other evolutionary methods.We initially introduce a refined sparse autoencoder to assess the prominence of each feature in the subsequent wrapper method.Subsequently,we utilize an enhanced wrapper feature selection technique that merges Mutual Information(MI)with individual feature contributions.Finally,a fine-tuning contribution tracking mechanism discerns informative features within the optimal feature subset,operating via a dominance accumulation mechanism.Experimental results for multiple classification performance metrics demonstrate that the proposed method effectively yields smaller feature subsets without degrading classification performance in an acceptable runtime compared to state-of-the-art algorithms across most large-scale benchmark datasets.
基金supported by“Regional Innovation Strategy(RIS)”through the National Research Foundation of Korea(NRF)funded by the Ministry of Education(MOE)(2021RIS-002)the Technology Development Program(RS-2023-00278623)funded by the Ministry of SMEs and Startups(MSS,Korea).
文摘This paper addresses the challenge of identifying abnormal states in Lithium-ion Battery(LiB)time series data.As the energy sector increasingly focuses on integrating distributed energy resources,Virtual Power Plants(VPP)have become a vital new framework for energy management.LiBs are key in this context,owing to their high-efficiency energy storage capabilities essential for VPP operations.However,LiBs are prone to various abnormal states like overcharging,over-discharging,and internal short circuits,which impede power transmission efficiency.Traditional methods for detecting such abnormalities in LiB are too broad and lack precision for the dynamic and irregular nature of LiB data.In response,we introduce an innovative method:a Long Short-Term Memory(LSTM)autoencoder based on Dynamic Frequency Memory and Correlation Attention(DFMCA-LSTM-AE).This unsupervised,end-to-end approach is specifically designed for dynamically monitoring abnormal states in LiB data.The method starts with a Dynamic Frequency Fourier Transform module,which dynamically captures the frequency characteristics of time series data across three scales,incorporating a memory mechanism to reduce overgeneralization of abnormal frequencies.This is followed by integrating LSTM into both the encoder and decoder,enabling the model to effectively encode and decode the temporal relationships in the time series.Empirical tests on a real-world LiB dataset demonstrate that DFMCA-LSTM-AE outperforms existing models,achieving an average Area Under the Curve(AUC)of 90.73%and an F1 score of 83.83%.These results mark significant improvements over existing models,ranging from 2.4%–45.3%for AUC and 1.6%–28.9%for F1 score,showcasing the model’s enhanced accuracy and reliability in detecting abnormal states in LiB data.
基金a result of project WAY4SafeRail—Wayside monitoring system FOR SAFE RAIL transportation, with reference NORTE-01-0247-FEDER-069595co-funded by the European Regional Development Fund (ERDF), through the North Portugal Regional Operational Programme (NORTE2020), under the PORTUGAL 2020 Partnership Agreement+3 种基金financially supported by Base Funding-UIDB/04708/2020Programmatic Funding-UIDP/04708/2020 of the CONSTRUCT—Instituto de Estruturas e Constru??esfunded by national funds through the FCT/ MCTES (PIDDAC)Grant No. 2021.04272. CEECIND from the Stimulus of Scientific Employment, Individual Support (CEECIND) - 4th Edition provided by “FCT – Funda??o para a Ciência, DOI : https:// doi. org/ 10. 54499/ 2021. 04272. CEECI ND/ CP1679/ CT0003”。
文摘Wayside monitoring is a promising cost-effective alternative to predict damage in the rolling stock. The main goal of this work is to present an unsupervised methodology to identify out-of-roundness(OOR) damage wheels, such as wheel flats and polygonal wheels. This automatic damage identification algorithm is based on the vertical acceleration evaluated on the rails using a virtual wayside monitoring system and involves the application of a two-step procedure. The first step aims to define a confidence boundary by using(healthy) measurements evaluated on the rail constituting a baseline. The second step of the procedure involves classifying damage of predefined scenarios with different levels of severities. The proposed procedure is based on a machine learning methodology and includes the following stages:(1) data collection,(2) damage-sensitive feature extraction from the acquired responses using a neural network model, i.e., the sparse autoencoder(SAE),(3) data fusion based on the Mahalanobis distance, and(4) unsupervised feature classification by implementing outlier and cluster analysis. This procedure considers baseline responses at different speeds and rail irregularities to train the SAE model. Then, the trained SAE is capable to reconstruct test responses(not trained) allowing to compute the accumulative difference between original and reconstructed signals. The results prove the efficiency of the proposed approach in identifying the two most common types of OOR in railway wheels.
文摘Network security problems bring many imperceptible threats to the integrity of data and the reliability of device services,so proposing a network intrusion detection model with high reliability is of great research significance for network security.Due to the strong generalization of invalid features during training process,it is more difficult for single autoencoder intrusion detection model to obtain effective results.A network intrusion detection model based on the Ensemble of Denoising Adversarial Autoencoder(EDAAE)was proposed,which had higher accuracy and reliability compared to the traditional anomaly detection model.Using the adversarial learning idea of Adversarial Autoencoder(AAE),the discriminator module was added to the original model,and the encoder part was used as the generator.The distribution of the hidden space of the data generated by the encoder matched with the distribution of the original data.The generalization of the model to the invalid features was also reduced to improve the detection accuracy.At the same time,the denoising autoencoder and integrated operation was introduced to prevent overfitting in the adversarial learning process.Experiments on the CICIDS2018 traffic dataset showed that the proposed intrusion detection model achieves an Accuracy of 95.23%,which out performs traditional self-encoders and other existing intrusion detection models methods in terms of overall performance.
基金supported by the Culture,Sports,and Tourism R&D Program through the Korea Creative Content Agency grant funded by the Ministry of Culture,Sports,and Tourism in 2024(Project Name:Development of Distribution and Management Platform Technology and Human Resource Development for Blockchain-Based SW Copyright Protection,Project Number:RS-2023-00228867,Contribution Rate:100%)and also supported by the Soonchunhyang University Research Fund.
文摘In the context of rapid digitization in industrial environments,how effective are advanced unsupervised learning models,particularly hybrid autoencoder models,at detecting anomalies in industrial control system(ICS)datasets?This study is crucial because it addresses the challenge of identifying rare and complex anomalous patterns in the vast amounts of time series data generated by Internet of Things(IoT)devices,which can significantly improve the reliability and safety of these systems.In this paper,we propose a hybrid autoencoder model,called ConvBiLSTMAE,which combines convolutional neural network(CNN)and bidirectional long short-term memory(BiLSTM)to more effectively train complex temporal data patterns in anomaly detection.On the hardware-in-the-loopbased extended industrial control system dataset,the ConvBiLSTM-AE model demonstrated remarkable anomaly detection performance,achieving F1 scores of 0.78 and 0.41 for the first and second datasets,respectively.The results suggest that hybrid autoencoder models are not only viable,but potentially superior alternatives for unsupervised anomaly detection in complex industrial systems,offering a promising approach to improving their reliability and safety.
基金funded in part by the MOE(Ministry of Education in China)Project of Humanities and Social Sciences(No.19YJC880068)the Hubei Provincial Natural Science Foundation of China(No.2019CFB347)+5 种基金the China Postdoctoral Science Foundation(No.2018M632889,No.2022T150250)the National Natural Science Foundation of China(No.61977027)the Hubei Province Technological Innovation Major Project Q2(No.2019AAA044)the Science&Technology Major Project of Hubei Province Next-Generation AI Technologies(No.2021BEA159)the Research Funds of CCNU from the Colleges'Basic Research and Operation of MOE(No.30106220491)in party by the Key Program of National Natural Science Foundation of China(No.61937001)。
文摘Although speech emotion recognition is challenging,it has broad application prospects in human-computer interaction.Building a system that can accurately and stably recognize emotions from human languages can provide a better user experience.However,the current unimodal emotion feature representations are not distinctive enough to accomplish the recognition,and they do not effectively simulate the inter-modality dynamics in speech emotion recognition tasks.This paper proposes a multimodal method that utilizes both audio and semantic content for speech emotion recognition.The proposed method consists of three parts:two high-level feature extractors for text and audio modalities,and an autoencoder-based feature fusion.For audio modality,we propose a structure called Temporal Global Feature Extractor(TGFE)to extract the high-level features of the timefrequency domain relationship from the original speech signal.Considering that text lacks frequency information,we use only a Bidirectional Long Short-Term Memory network(BLSTM)and attention mechanism to simulate an intra-modal dynamic.Once these steps have been accomplished,the high-level text and audio features are sent to the autoencoder in parallel to learn their shared representation for final emotion classification.We conducted extensive experiments on three public benchmark datasets to evaluate our method.The results on Interactive Emotional Motion Capture(IEMOCAP)and Multimodal EmotionLines Dataset(MELD)outperform the existing method.Additionally,the results of CMU Multi-modal Opinion-level Sentiment Intensity(CMU-MOSI)are competitive.Furthermore,experimental results show that compared to unimodal information and autoencoderbased feature level fusion,the joint multimodal information(audio and text)improves the overall performance and can achieve greater accuracy than simple feature concatenation.
基金supported by the National Natural Science Foundation of China(Grant Nos.U20A20186 and 62372063).
文摘Fault diagnosis of traction systems is important for the safety operation of high-speed trains.Long-term operation of the trains will degrade the performance of systems,which decreases the fault detection accuracy.To solve this problem,this paper proposes a fault detection method developed by a Generalized Autoencoder(GAE)for systems with performance degradation.The advantage of this method is that it can accurately detect faults when the traction system of high-speed trains is affected by performance degradation.Regardless of the probability distribution,it can handle any data,and the GAE has extremely high sensitivity in anomaly detection.Finally,the effectiveness of this method is verified through the Traction Drive Control System(TDCS)platform.At different performance degradation levels,our method’s experimental results are superior to traditional methods.
基金The National Natural Science Foundation of China(No.52361165658,52378318,52078459).
文摘To enhance the accuracy and efficiency of bridge damage identification,a novel data-driven damage identification method was proposed.First,convolutional autoencoder(CAE)was used to extract key features from the acceleration signal of the bridge structure through data reconstruction.The extreme gradient boosting tree(XGBoost)was then used to perform analysis on the feature data to achieve damage detection with high accuracy and high performance.The proposed method was applied in a numerical simulation study on a three-span continuous girder and further validated experimentally on a scaled model of a cable-stayed bridge.The numerical simulation results show that the identification errors remain within 2.9%for six single-damage cases and within 3.1%for four double-damage cases.The experimental validation results demonstrate that when the tension in a single cable of the cable-stayed bridge decreases by 20%,the method accurately identifies damage at different cable locations using only sensors installed on the main girder,achieving identification accuracies above 95.8%in all cases.The proposed method shows high identification accuracy and generalization ability across various damage scenarios.
文摘Sentence classification is the process of categorizing a sentence based on the context of the sentence.Sentence categorization requires more semantic highlights than other tasks,such as dependence parsing,which requires more syntactic elements.Most existing strategies focus on the general semantics of a conversation without involving the context of the sentence,recognizing the progress and comparing impacts.An ensemble pre-trained language model was taken up here to classify the conversation sentences from the conversation corpus.The conversational sentences are classified into four categories:information,question,directive,and commission.These classification label sequences are for analyzing the conversation progress and predicting the pecking order of the conversation.Ensemble of Bidirectional Encoder for Representation of Transformer(BERT),Robustly Optimized BERT pretraining Approach(RoBERTa),Generative Pre-Trained Transformer(GPT),DistilBERT and Generalized Autoregressive Pretraining for Language Understanding(XLNet)models are trained on conversation corpus with hyperparameters.Hyperparameter tuning approach is carried out for better performance on sentence classification.This Ensemble of Pre-trained Language Models with a Hyperparameter Tuning(EPLM-HT)system is trained on an annotated conversation dataset.The proposed approach outperformed compared to the base BERT,GPT,DistilBERT and XLNet transformer models.The proposed ensemble model with the fine-tuned parameters achieved an F1_score of 0.88.
文摘We present an approach to classify medical text at a sentence level automatically.Given the inherent complexity of medical text classification,we employ adapters based on pre-trained language models to extract information from medical text,facilitating more accurate classification while minimizing the number of trainable parameters.Extensive experiments conducted on various datasets demonstrate the effectiveness of our approach.
基金National Natural Science Foundation of China,Grant/Award Numbers:62173236,61876110,61806130,61976142,82304204.
文摘Network embedding(NE)tries to learn the potential properties of complex networks represented in a low-dimensional feature space.However,the existing deep learningbased NE methods are time-consuming as they need to train a dense architecture for deep neural networks with extensive unknown weight parameters.A sparse deep autoencoder(called SPDNE)for dynamic NE is proposed,aiming to learn the network structures while preserving the node evolution with a low computational complexity.SPDNE tries to use an optimal sparse architecture to replace the fully connected architecture in the deep autoencoder while maintaining the performance of these models in the dynamic NE.Then,an adaptive simulated algorithm to find the optimal sparse architecture for the deep autoencoder is proposed.The performance of SPDNE over three dynamical NE models(i.e.sparse architecture-based deep autoencoder method,DynGEM,and ElvDNE)is evaluated on three well-known benchmark networks and five real-world networks.The experimental results demonstrate that SPDNE can reduce about 70%of weight parameters of the architecture for the deep autoencoder during the training process while preserving the performance of these dynamical NE models.The results also show that SPDNE achieves the highest accuracy on 72 out of 96 edge prediction and network reconstruction tasks compared with the state-of-the-art dynamical NE algorithms.
文摘Many types of real-world information systems, including social media and e-commerce platforms, can be modelled by means of attribute-rich, connected networks. The goal of anomaly detection in artificial intelligence is to identify illustrations that deviate significantly from the main distribution of data or that differ from known cases. Anomalous nodes in node-attributed networks can be identified with greater precision if both graph and node attributes are taken into account. Almost all of the studies in this area focus on supervised techniques for spotting outliers. While supervised algorithms for anomaly detection work well in theory, they cannot be applied to real-world applications owing to a lack of labelled data. Considering the possible data distribution, our model employs a dual variational autoencoder (VAE), while a generative adversarial network (GAN) assures that the model is robust to adversarial training. The dual VAEs are used in another capacity: as a fake-node generator. Adversarial training is used to ensure that our latent codes have a Gaussian or uniform distribution. To provide a fair presentation of the graph, the discriminator instructs the generator to generate latent variables with distributions that are more consistent with the actual distribution of the data. Once the model has been learned, the discriminator is used for anomaly detection via reconstruction loss which has been trained to distinguish between the normal and artificial distributions of data. First, using a dual VAE, our model simultaneously captures cross-modality interactions between topological structure and node characteristics and overcomes the problem of unlabeled anomalies, allowing us to better understand the network sparsity and nonlinearity. Second, the proposed model considers the regularization of the latent codes while solving the issue of unregularized embedding techniques that can quickly lead to unsatisfactory representation. Finally, we use the discriminator reconstruction loss for anomaly detection as the discriminator is well-trained to separate the normal and generated data distributions because reconstruction-based loss does not include the adversarial component. Experiments conducted on attributed networks demonstrate the effectiveness of the proposed model and show that it greatly surpasses the previous methods. The area under the curve scores of our proposed model for the BlogCatalog, Flickr, and Enron datasets are 0.83680, 0.82020, and 0.71180, respectively, proving the effectiveness of the proposed model. The result of the proposed model on the Enron dataset is slightly worse than other models;we attribute this to the dataset’s low dimensionality as the most probable explanation.
