Pub Date : 2024-06-18DOI: 10.1109/TAI.2024.3415551
Jonathan Cui;David A. Araujo;Suman Saha;Md Faisal Kabir
Despite simpler architectural designs compared with vision transformers (ViTs) and convolutional neural networks, vision multilayer perceptrons (MLPs) have demonstrated strong performance and high data efficiency for image classification and semantic segmentation. Following pioneering works such as MLP-Mixers and gMLPs, later research proposed a plethora of vision MLP architectures that achieve token-mixing with specifically engineered convolution- or attentionlike mechanisms. However, existing methods such as $text{S}^{text{2}}$ -MLPs and PoolFormers typically model spatial information in equal-sized spatial regions and do not consider cross-scale spatial interactions, thus delivering subpar performance compared with transformer models that employ global token mixing. Further, these MLP token-mixers, along with most ViTs, only model one- or two-axis correlations among space and channels, avoiding simultaneous three-axis spatial–channel mixing due to its computational demands. We, therefore, propose CS-Mixer, a hierarchical vision MLP that learns dynamic low-rank transformations for tokens aggregated across scales, both locally and globally. Such aggregation allows for token-mixing that explicitly models spatial–channel interactions, made computationally possible by a multihead design that projects to low-dimensional subspaces. The proposed methodology achieves competitive results on popular image recognition benchmarks without incurring substantially more computing. Our largest model, CS-Mixer-L, reaches 83.2% top-1 accuracy on ImageNet-1k with 13.7 GFLOPs and 94 M parameters.
{"title":"CS-Mixer: A Cross-Scale Vision Multilayer Perceptron With Spatial–Channel Mixing","authors":"Jonathan Cui;David A. Araujo;Suman Saha;Md Faisal Kabir","doi":"10.1109/TAI.2024.3415551","DOIUrl":"https://doi.org/10.1109/TAI.2024.3415551","url":null,"abstract":"Despite simpler architectural designs compared with vision transformers (ViTs) and convolutional neural networks, vision multilayer perceptrons (MLPs) have demonstrated strong performance and high data efficiency for image classification and semantic segmentation. Following pioneering works such as MLP-Mixers and gMLPs, later research proposed a plethora of vision MLP architectures that achieve token-mixing with specifically engineered convolution- or attentionlike mechanisms. However, existing methods such as \u0000<inline-formula><tex-math>$text{S}^{text{2}}$</tex-math></inline-formula>\u0000-MLPs and PoolFormers typically model spatial information in equal-sized spatial regions and do not consider cross-scale spatial interactions, thus delivering subpar performance compared with transformer models that employ global token mixing. Further, these MLP token-mixers, along with most ViTs, only model one- or two-axis correlations among space and channels, avoiding simultaneous three-axis spatial–channel mixing due to its computational demands. We, therefore, propose CS-Mixer, a hierarchical vision MLP that learns dynamic low-rank transformations for tokens aggregated across scales, both locally and globally. Such aggregation allows for token-mixing that explicitly models spatial–channel interactions, made computationally possible by a multihead design that projects to low-dimensional subspaces. The proposed methodology achieves competitive results on popular image recognition benchmarks without incurring substantially more computing. Our largest model, CS-Mixer-L, reaches 83.2% top-1 accuracy on ImageNet-1k with 13.7 GFLOPs and 94 M parameters.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-06-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142443002","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Educational data mining (EDM) offers an effective solution to predict students’ course grades in the next term. Conventional grade prediction methods can be viewed as regressing an expectation of the probability distribution of the student's grade, typically called single-value grade prediction. The reliable prediction outcomes of these methods depend on the complete input information related to students. However, next-term grade prediction often encounters the challenge of incomplete input information due to the inaccessibility of future data and the privacy of data. In this scenario, single-value grade prediction struggles to assess students’ academic status, as it may not be represented and assessed by relying on a singular expectation value. This limitation increases the risk of misjudgment, and may lead to errors in educational decision-making. Considering the challenge of collecting complete input information, we shift from traditional single-value predictions to forecasting the explicit probability distribution of the course grade. The probability distribution of the grade can assess the students’ academic status by providing probabilities corresponding to all possible grade values rather than relying solely on an expectation value, which offers the foundation to support the educators’ decision-making. In this article, the course grade distribution prediction (CGDP) model is proposed, aiming to estimate an explicit conditional probability distribution of course grades in the next term. This model can identify at-risk students, offering comprehensive decision-making information for educators and students. To ensure precise distribution predictions, a calibration method is also employed to improve the alignment between predicted and actual probabilities. Experimental results verify the effectiveness of the proposed model in early grade warning for undergraduates, based on real university data.
{"title":"A Novel Grades Prediction Method for Undergraduate Students by Learning Explicit Conditional Distribution","authors":"Na Zhang;Ming Liu;Lin Wang;Shuangrong Liu;Runyuan Sun;Bo Yang;Shenghui Zhu;Chengdong Li;Cheng Yang;Yuhu Cheng","doi":"10.1109/TAI.2024.3416077","DOIUrl":"https://doi.org/10.1109/TAI.2024.3416077","url":null,"abstract":"Educational data mining (EDM) offers an effective solution to predict students’ course grades in the next term. Conventional grade prediction methods can be viewed as regressing an expectation of the probability distribution of the student's grade, typically called single-value grade prediction. The reliable prediction outcomes of these methods depend on the complete input information related to students. However, next-term grade prediction often encounters the challenge of incomplete input information due to the inaccessibility of future data and the privacy of data. In this scenario, single-value grade prediction struggles to assess students’ academic status, as it may not be represented and assessed by relying on a singular expectation value. This limitation increases the risk of misjudgment, and may lead to errors in educational decision-making. Considering the challenge of collecting complete input information, we shift from traditional single-value predictions to forecasting the explicit probability distribution of the course grade. The probability distribution of the grade can assess the students’ academic status by providing probabilities corresponding to all possible grade values rather than relying solely on an expectation value, which offers the foundation to support the educators’ decision-making. In this article, the course grade distribution prediction (CGDP) model is proposed, aiming to estimate an explicit conditional probability distribution of course grades in the next term. This model can identify at-risk students, offering comprehensive decision-making information for educators and students. To ensure precise distribution predictions, a calibration method is also employed to improve the alignment between predicted and actual probabilities. Experimental results verify the effectiveness of the proposed model in early grade warning for undergraduates, based on real university data.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-06-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142169721","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-06-18DOI: 10.1109/TAI.2024.3415549
Huai-Ning Wu;Xiao-Yan Jiang;Mi Wang
Since human beings are of limited reasoning ability as well as the machines do not usually know human intentions, how to learn human cognitive levels in shared control to enhance the machines’ intelligence is a challenging issue. In this study, this issue is addressed in the context of human–machine shared control for a class of human-in-the-loop (HiTL) systems based on a differential game with bounded rationality and incomplete information. Initially, we formulate the human–machine shared control problem as a two-player nonzero-sum linear quadratic dynamic game (LQDG), where the weighting matrix of the cost function representing the human intention is unknown for the machine. To model the human bounded rationality, the level- $boldsymbol{k}$ (LK) approach is employed to set up the LK control policies of two players performing the corresponding steps of strategic thinking. To infer the human intention, an online adaptive inverse optimal control (IOC) algorithm is then developed by using the system state data, so that the control policies of different cognitive levels can be computed. In addition, a reinforcement learning method is proposed for the machine to identify the distribution of the human cognitive levels while providing a proactive collaborative control to assist the human in a probabilistic switching way. Finally, simulation results on a cooperative shared control driver assistance system (DAS) illustrate the efficacy of the proposed approach.
{"title":"Human Cognitive Learning in Shared Control via Differential Game With Bounded Rationality and Incomplete Information","authors":"Huai-Ning Wu;Xiao-Yan Jiang;Mi Wang","doi":"10.1109/TAI.2024.3415549","DOIUrl":"https://doi.org/10.1109/TAI.2024.3415549","url":null,"abstract":"Since human beings are of limited reasoning ability as well as the machines do not usually know human intentions, how to learn human cognitive levels in shared control to enhance the machines’ intelligence is a challenging issue. In this study, this issue is addressed in the context of human–machine shared control for a class of human-in-the-loop (HiTL) systems based on a differential game with bounded rationality and incomplete information. Initially, we formulate the human–machine shared control problem as a two-player nonzero-sum linear quadratic dynamic game (LQDG), where the weighting matrix of the cost function representing the human intention is unknown for the machine. To model the human bounded rationality, the level-\u0000<inline-formula><tex-math>$boldsymbol{k}$</tex-math></inline-formula>\u0000 (LK) approach is employed to set up the LK control policies of two players performing the corresponding steps of strategic thinking. To infer the human intention, an online adaptive inverse optimal control (IOC) algorithm is then developed by using the system state data, so that the control policies of different cognitive levels can be computed. In addition, a reinforcement learning method is proposed for the machine to identify the distribution of the human cognitive levels while providing a proactive collaborative control to assist the human in a probabilistic switching way. Finally, simulation results on a cooperative shared control driver assistance system (DAS) illustrate the efficacy of the proposed approach.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-06-18","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142443001","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Ridge regression (RR)-based methods aim to obtain a low-dimensional subspace for feature extraction. However, the subspace's dimensionality does not exceed the number of data categories, hence compromising its capability of feature representation. Moreover, these methods with $L_{2}$ -norm metric and regularization cannot extract highly robust features from data with corruption. To address these problems, in this article, we propose generalized jointly sparse linear discriminant regression (GOAL), a novel regression method based on joint $L_{2,1}$ -norm and capped- $L_{2}$ -norm, which can integrate sparsity, locality, and discriminability into one model to learn a full-rank robust feature extractor. The sparsely selected discriminative features are robust enough to characterize the decision boundary between classes. Locality is related to manifold structure and Laplacian smoothing, which can enhance the robustness of the model. By using the multinorm metric and regularization regression framework, the proposed method obtains the projection with joint sparsity and guarantees that the rank of the projection matrix will not be limited by the number of classes. An iterative algorithm is proposed to compute the optimal solution. Complexity analysis and proofs of convergence are also given in the article. Experiments on well-known datasets demonstrate our model's superiority and generalization ability.
{"title":"GOAL: Generalized Jointly Sparse Linear Discriminant Regression for Feature Extraction","authors":"Haoquan Lu;Zhihui Lai;Junhong Zhang;Zhuozhen Yu;Jiajun Wen","doi":"10.1109/TAI.2024.3412862","DOIUrl":"https://doi.org/10.1109/TAI.2024.3412862","url":null,"abstract":"Ridge regression (RR)-based methods aim to obtain a low-dimensional subspace for feature extraction. However, the subspace's dimensionality does not exceed the number of data categories, hence compromising its capability of feature representation. Moreover, these methods with \u0000<inline-formula><tex-math>$L_{2}$</tex-math></inline-formula>\u0000-norm metric and regularization cannot extract highly robust features from data with corruption. To address these problems, in this article, we propose generalized jointly sparse linear discriminant regression (GOAL), a novel regression method based on joint \u0000<inline-formula><tex-math>$L_{2,1}$</tex-math></inline-formula>\u0000-norm and capped-\u0000<inline-formula><tex-math>$L_{2}$</tex-math></inline-formula>\u0000-norm, which can integrate sparsity, locality, and discriminability into one model to learn a full-rank robust feature extractor. The sparsely selected discriminative features are robust enough to characterize the decision boundary between classes. Locality is related to manifold structure and Laplacian smoothing, which can enhance the robustness of the model. By using the multinorm metric and regularization regression framework, the proposed method obtains the projection with joint sparsity and guarantees that the rank of the projection matrix will not be limited by the number of classes. An iterative algorithm is proposed to compute the optimal solution. Complexity analysis and proofs of convergence are also given in the article. Experiments on well-known datasets demonstrate our model's superiority and generalization ability.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-06-11","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142443106","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-06-07DOI: 10.1109/TAI.2024.3410934
Shusen Ma;Yun-Bo Zhao;Yu Kang;Peng Bai
Many real-world scenarios require accurate predictions of time series, especially in the case of long sequence time-series forecasting (LSTF), such as predicting traffic flow and electricity consumption. However, existing time-series prediction models encounter certain limitations. First, they struggle with mapping the multidimensional information present in each time step to high dimensions, resulting in information coupling and increased prediction difficulty. Second, these models fail to effectively decompose the intertwined temporal patterns within the time series, which hinders their ability to learn more predictable features. To overcome these challenges, we propose a novel end-to-end LSTF model with parallelized convolution and decomposed sparse-Transformer (PCDformer). PCDformer achieves the decoupling of input sequences by parallelizing the convolutional layers, enabling the simultaneous processing of different variables within the input sequence. To decompose distinct temporal patterns, PCDformer incorporates a temporal decomposition module within the encoder–decoder structure, effectively separating the input sequence into predictable seasonal and trend components. Additionally, to capture the correlation between variables and mitigate the impact of irrelevant information, PCDformer utilizes a sparse self-attention mechanism. Extensive experimentation conducted on five diverse datasets demonstrates the superior performance of PCDformer in LSTF tasks compared to existing approaches, particularly outperforming encoder–decoder-based models.
{"title":"Multivariate Time-Series Modeling and Forecasting With Parallelized Convolution and Decomposed Sparse-Transformer","authors":"Shusen Ma;Yun-Bo Zhao;Yu Kang;Peng Bai","doi":"10.1109/TAI.2024.3410934","DOIUrl":"https://doi.org/10.1109/TAI.2024.3410934","url":null,"abstract":"Many real-world scenarios require accurate predictions of time series, especially in the case of long sequence time-series forecasting (LSTF), such as predicting traffic flow and electricity consumption. However, existing time-series prediction models encounter certain limitations. First, they struggle with mapping the multidimensional information present in each time step to high dimensions, resulting in information coupling and increased prediction difficulty. Second, these models fail to effectively decompose the intertwined temporal patterns within the time series, which hinders their ability to learn more predictable features. To overcome these challenges, we propose a novel end-to-end LSTF model with parallelized convolution and decomposed sparse-Transformer (PCDformer). PCDformer achieves the decoupling of input sequences by parallelizing the convolutional layers, enabling the simultaneous processing of different variables within the input sequence. To decompose distinct temporal patterns, PCDformer incorporates a temporal decomposition module within the encoder–decoder structure, effectively separating the input sequence into predictable seasonal and trend components. Additionally, to capture the correlation between variables and mitigate the impact of irrelevant information, PCDformer utilizes a sparse self-attention mechanism. Extensive experimentation conducted on five diverse datasets demonstrates the superior performance of PCDformer in LSTF tasks compared to existing approaches, particularly outperforming encoder–decoder-based models.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-06-07","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142442963","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-06-04DOI: 10.1109/TAI.2024.3409516
Suman De;Pabitra Mitra
Manufacturing plants are highly dependent on machines and involve a significant number of equipment to produce a finished product. Industry 4.0 helps structure the processes involved in such setups and enables the functionalities of how the equipment and machines interact with each other. With the advancement of visualizing these types of equipment as digital twins, multiple opportunities have developed for automating processes and optimizing various aspects of the assembly, especially for original equipment manufacturers (OEMs). One problem that concerns a network of manufacturers is the availability of equipment and spare parts data which are sometimes confidential but are required by a new member in the network for several analytical applications. This article looks at this problem statement to turn this into an opportunity by introducing a novel concept of AASGAN that combines the knowledge representation of a digital twin data in the asset administration shell (AAS) and a synthetic data generation technique of generative adversarial network (GAN) to generate fake data that is identical to real data. This article also explains how this concept helps perform analytical operations using industry grade solutions for the automotive industry available for managing digital twins and other scenarios for industrial automation.
{"title":"A Novel Technique of Synthetic Data Generation for Asset Administration Shells in Industry 4.0 Scenarios","authors":"Suman De;Pabitra Mitra","doi":"10.1109/TAI.2024.3409516","DOIUrl":"https://doi.org/10.1109/TAI.2024.3409516","url":null,"abstract":"Manufacturing plants are highly dependent on machines and involve a significant number of equipment to produce a finished product. Industry 4.0 helps structure the processes involved in such setups and enables the functionalities of how the equipment and machines interact with each other. With the advancement of visualizing these types of equipment as digital twins, multiple opportunities have developed for automating processes and optimizing various aspects of the assembly, especially for original equipment manufacturers (OEMs). One problem that concerns a network of manufacturers is the availability of equipment and spare parts data which are sometimes confidential but are required by a new member in the network for several analytical applications. This article looks at this problem statement to turn this into an opportunity by introducing a novel concept of AASGAN that combines the knowledge representation of a digital twin data in the asset administration shell (AAS) and a synthetic data generation technique of generative adversarial network (GAN) to generate fake data that is identical to real data. This article also explains how this concept helps perform analytical operations using industry grade solutions for the automotive industry available for managing digital twins and other scenarios for industrial automation.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-06-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142443114","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-06-04DOI: 10.1109/TAI.2024.3408717
Kyle Worrall;Zongyu Yin;Tom Collins
There have been many attempts to model the ability of human musicians to take a score and perform or render it expressively, by adding tempo, timing, loudness, and articulation changes to nonexpressive music data. While expressive rendering models exist in academic research, most of these are not open source or accessible, meaning they are difficult to evaluate empirically and have not been widely adopted in professional music software. Systematic comparative evaluation of such algorithms stopped after the last performance rendering contest (RENCON) in 2013, making it difficult to compare newer models to existing work in a fair and valid way. In this article, we introduce the first transformer-based model for expressive rendering, cue-free express + pedal (CFE + P), which predicts expressive attributes such as notewise dynamics and micro-timing adjustments, and beatwise tempo and sustain pedal use based only on the start and end times and pitches of notes (e.g., inexpressive musical instrument digital interface (MIDI) input). We perform two comparative evaluations on our model against a nonmachine learning baseline taken from professional music software and two open-source algorithms—a feedforward neural network (FFNN) and hierarchical recurrent neural network (HRNN). The results of two listening studies indicate that our model renders passages that outperform what can be done in professional music software such as Logic Pro and Ableton Live. 1
All data and preexisting hypotheses can be accessed via the Open Science Foundation: https://osf.io/6uwjk/.
{"title":"Comparative Evaluation in the Wild: Systems for the Expressive Rendering of Music","authors":"Kyle Worrall;Zongyu Yin;Tom Collins","doi":"10.1109/TAI.2024.3408717","DOIUrl":"https://doi.org/10.1109/TAI.2024.3408717","url":null,"abstract":"There have been many attempts to model the ability of human musicians to take a score and perform or render it expressively, by adding tempo, timing, loudness, and articulation changes to nonexpressive music data. While expressive rendering models exist in academic research, most of these are not open source or accessible, meaning they are difficult to evaluate empirically and have not been widely adopted in professional music software. Systematic comparative evaluation of such algorithms stopped after the last performance rendering contest (RENCON) in 2013, making it difficult to compare newer models to existing work in a fair and valid way. In this article, we introduce the first transformer-based model for expressive rendering, cue-free express + pedal (CFE + P), which predicts expressive attributes such as notewise dynamics and micro-timing adjustments, and beatwise tempo and sustain pedal use based only on the start and end times and pitches of notes (e.g., inexpressive musical instrument digital interface (MIDI) input). We perform two comparative evaluations on our model against a nonmachine learning baseline taken from professional music software and two open-source algorithms—a feedforward neural network (FFNN) and hierarchical recurrent neural network (HRNN). The results of two listening studies indicate that our model renders passages that outperform what can be done in professional music software such as Logic Pro and Ableton Live.\u0000<xref><sup>1</sup></xref>\u0000<fn><label><sup>1</sup></label><p>All data and preexisting hypotheses can be accessed via the Open Science Foundation: <uri>https://osf.io/6uwjk/</uri>.</p></fn>","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-06-04","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142443129","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pancreatic cancer is a highly fatal cancer type. Patients are typically in an advanced stage at their first diagnosis, mainly due to the absence of distinctive early stage symptoms and lack of effective early diagnostic methods. In this work, we propose an automated method for pancreatic cancer diagnosis using noncontrast computed tomography (CT), taking advantage of its widespread availability in clinic. Currently, a primary challenge limiting the clinical value of intelligent systems is low generalization, i.e., the difficulty of achieving stable performance across datasets from different medical sources. To address this challenge, a novel causality-informed graph intervention model is developed based on a multi-instance-learning framework integrated with graph neural network (GNN) for the extraction of local discriminative features. Within this model, we develop a graph causal intervention scheme with three levels of intervention for graph nodes, structures, and representations. This scheme systematically suppresses noncausal factors and thus lead to generalizable predictions. Specifically, first, a target node perturbation strategy is designed to capture target-region features. Second, a causal-structure separation module is developed to automatically identify the causal graph structures for obtaining stable representations of whole target regions. Third, a graph-level feature consistency mechanism is proposed to extract invariant features. Comprehensive experiments on large-scale datasets validated the promising early diagnosis performance of our proposed model. The model generalizability was confirmed on three independent datasets, where the classification accuracy reached 86.3%, 80.4%, and 82.2%, respectively. Overall, we provide a valuable potential tool for pancreatic cancer screening and early diagnosis.
{"title":"A Causality-Informed Graph Intervention Model for Pancreatic Cancer Early Diagnosis","authors":"Xinyue Li;Rui Guo;Hongzhang Zhu;Tao Chen;Xiaohua Qian","doi":"10.1109/TAI.2024.3395586","DOIUrl":"https://doi.org/10.1109/TAI.2024.3395586","url":null,"abstract":"Pancreatic cancer is a highly fatal cancer type. Patients are typically in an advanced stage at their first diagnosis, mainly due to the absence of distinctive early stage symptoms and lack of effective early diagnostic methods. In this work, we propose an automated method for pancreatic cancer diagnosis using noncontrast computed tomography (CT), taking advantage of its widespread availability in clinic. Currently, a primary challenge limiting the clinical value of intelligent systems is low generalization, i.e., the difficulty of achieving stable performance across datasets from different medical sources. To address this challenge, a novel causality-informed graph intervention model is developed based on a multi-instance-learning framework integrated with graph neural network (GNN) for the extraction of local discriminative features. Within this model, we develop a graph causal intervention scheme with three levels of intervention for graph nodes, structures, and representations. This scheme systematically suppresses noncausal factors and thus lead to generalizable predictions. Specifically, first, a target node perturbation strategy is designed to capture target-region features. Second, a causal-structure separation module is developed to automatically identify the causal graph structures for obtaining stable representations of whole target regions. Third, a graph-level feature consistency mechanism is proposed to extract invariant features. Comprehensive experiments on large-scale datasets validated the promising early diagnosis performance of our proposed model. The model generalizability was confirmed on three independent datasets, where the classification accuracy reached 86.3%, 80.4%, and 82.2%, respectively. Overall, we provide a valuable potential tool for pancreatic cancer screening and early diagnosis.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-04-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142169722","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-04-30DOI: 10.1109/TAI.2024.3395574
Yufei Liu;Jia Wu;Jie Cao
Social behavior prediction on social media is attracting significant attention from researchers. Social e-commerce focuses on engagement marketing, which emphasizes social behavior because it effectively increases brand recognition. Currently, existing works on social behavior prediction suffer from two main problems: 1) they assume that social influence probabilities can be learned independently of each other, and their calculations do not include any influence probability estimations based on friends’ behavior; and 2) negative sampling of subgraphs is usually ignored in social behavior prediction work. To the best of our knowledge, introducing graph contrastive learning (GCL) to social behavior prediction is novel and interesting. In this article, we propose a framework, social behavior prediction via graph contrastive learning with attention named SBP-GCA, to promote social behavior prediction. First, two methods are designed to extract subgraphs from the original graph, and their structural features are learned by GCL. Then, it models how a user's behavior is influenced by neighbors and learns influence features via graph attention networks (GATs). Furthermore, it combines structural features, influence features, and intrinsic features to predict social behavior. Extensive and systematic experiments on three datasets validate the superiority of the proposed SBP-GCA.
社交媒体上的社交行为预测正引起研究人员的极大关注。社交电子商务侧重于参与式营销，强调社交行为，因为它能有效提高品牌认知度。目前，有关社交行为预测的现有研究存在两个主要问题：1）假设社交影响概率可以独立学习，其计算不包括任何基于好友行为的影响概率估计；2）社交行为预测工作通常忽略子图的负采样。据我们所知，将图对比学习（GCL）引入社交行为预测是一项新颖而有趣的工作。在本文中，我们提出了一个通过图对比学习（graph contrastive learning with attention）进行社会行为预测的框架，命名为 SBP-GCA，以促进社会行为预测。首先，我们设计了两种方法从原始图中提取子图，并通过 GCL 学习子图的结构特征。然后，它对用户行为如何受邻居影响进行建模，并通过图注意力网络（GAT）学习影响特征。此外，它还结合了结构特征、影响特征和内在特征来预测社交行为。在三个数据集上进行的广泛而系统的实验验证了所提出的 SBP-GCA 的优越性。
{"title":"SBP-GCA: Social Behavior Prediction via Graph Contrastive Learning With Attention","authors":"Yufei Liu;Jia Wu;Jie Cao","doi":"10.1109/TAI.2024.3395574","DOIUrl":"https://doi.org/10.1109/TAI.2024.3395574","url":null,"abstract":"Social behavior prediction on social media is attracting significant attention from researchers. Social e-commerce focuses on engagement marketing, which emphasizes social behavior because it effectively increases brand recognition. Currently, existing works on social behavior prediction suffer from two main problems: 1) they assume that social influence probabilities can be learned independently of each other, and their calculations do not include any influence probability estimations based on friends’ behavior; and 2) negative sampling of subgraphs is usually ignored in social behavior prediction work. To the best of our knowledge, introducing graph contrastive learning (GCL) to social behavior prediction is novel and interesting. In this article, we propose a framework, social behavior prediction via graph contrastive learning with attention named SBP-GCA, to promote social behavior prediction. First, two methods are designed to extract subgraphs from the original graph, and their structural features are learned by GCL. Then, it models how a user's behavior is influenced by neighbors and learns influence features via graph attention networks (GATs). Furthermore, it combines structural features, influence features, and intrinsic features to predict social behavior. Extensive and systematic experiments on three datasets validate the superiority of the proposed SBP-GCA.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-04-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142169601","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
Pub Date : 2024-04-30DOI: 10.1109/TAI.2024.3394792
Israq Ahmed Anik;A. H. M. Kamal;Muhammad Ashad Kabir;Shahadat Uddin;Mohammad Ali Moni
Major depressive disorder (MDD), commonly called depression, is a prevalent psychiatric condition diagnosed via questionnaire-based mental status assessments. However, this method often yields inconsistent and inaccurate results. Furthermore, there is currently a lack of a comprehensive diagnostic framework for MDD that assesses various brainwaves (alpha, theta, gamma, etc.) of electroencephalogram (EEG) signals as potential biomarkers, aiming to identify the most effective one for achieving accurate and robust diagnostic outcomes. To address this issue, we propose an innovative approach employing a deep convolutional neural network (DCNN) for MDD diagnosis utilizing the brainwaves present in EEG signals. Our proposed model, an extended 11-layer 1-D convolutional neural network (Ex-1DCNN), is designed to automatically learn from input EEG signals, foregoing the need for manual feature selection. By harnessing intrinsic brainwave patterns, our model demonstrates adaptability in classifying EEG signals into depressive and healthy categories. We have conducted an extensive analysis to identify optimal brainwave features and epoch duration for accurate MDD diagnosis. Leveraging EEG data from 34 MDD patients and 30 healthy subjects, we have identified that the Gamma brainwave at a 15-s epoch duration is the most effective configuration, achieving an accuracy of 99.60%, sensitivity of 100%, specificity of 99.21%, and an F1 score of 99.60%. This study highlights the potential of deep-learning techniques in streamlining the diagnostic process for MDD and offering a reliable aid to clinicians in MDD diagnosis.
{"title":"A Robust Deep-Learning Model to Detect Major Depressive Disorder Utilizing EEG Signals","authors":"Israq Ahmed Anik;A. H. M. Kamal;Muhammad Ashad Kabir;Shahadat Uddin;Mohammad Ali Moni","doi":"10.1109/TAI.2024.3394792","DOIUrl":"https://doi.org/10.1109/TAI.2024.3394792","url":null,"abstract":"Major depressive disorder (MDD), commonly called depression, is a prevalent psychiatric condition diagnosed via questionnaire-based mental status assessments. However, this method often yields inconsistent and inaccurate results. Furthermore, there is currently a lack of a comprehensive diagnostic framework for MDD that assesses various brainwaves (alpha, theta, gamma, etc.) of electroencephalogram (EEG) signals as potential biomarkers, aiming to identify the most effective one for achieving accurate and robust diagnostic outcomes. To address this issue, we propose an innovative approach employing a deep convolutional neural network (DCNN) for MDD diagnosis utilizing the brainwaves present in EEG signals. Our proposed model, an extended 11-layer 1-D convolutional neural network (Ex-1DCNN), is designed to automatically learn from input EEG signals, foregoing the need for manual feature selection. By harnessing intrinsic brainwave patterns, our model demonstrates adaptability in classifying EEG signals into depressive and healthy categories. We have conducted an extensive analysis to identify optimal brainwave features and epoch duration for accurate MDD diagnosis. Leveraging EEG data from 34 MDD patients and 30 healthy subjects, we have identified that the Gamma brainwave at a 15-s epoch duration is the most effective configuration, achieving an accuracy of 99.60%, sensitivity of 100%, specificity of 99.21%, and an F1 score of 99.60%. This study highlights the potential of deep-learning techniques in streamlining the diagnostic process for MDD and offering a reliable aid to clinicians in MDD diagnosis.","PeriodicalId":73305,"journal":{"name":"IEEE transactions on artificial intelligence","volume":null,"pages":null},"PeriodicalIF":0.0,"publicationDate":"2024-04-30","publicationTypes":"Journal Article","fieldsOfStudy":null,"isOpenAccess":false,"openAccessPdf":"","citationCount":null,"resultStr":null,"platform":"Semanticscholar","paperid":"142443088","PeriodicalName":null,"FirstCategoryId":null,"ListUrlMain":null,"RegionNum":0,"RegionCategory":"","ArticlePicture":[],"TitleCN":null,"AbstractTextCN":null,"PMCID":"","EPubDate":null,"PubModel":null,"JCR":null,"JCRName":null,"Score":null,"Total":0}
京公网安备 11010802042870号
京ICP备2023020795号-1
扫码关注我们
求助内容：
应助结果提醒方式：