Deep neural networks are prone to memorizing incorrect labels during training, which degrades their generalizability. Although recent methods have combined sample selection with semi-supervised learning (SSL) to exploit the memorization effect - where networks learn from clean data before noisy data - they cannot correct selection errors once a sample is misclassified. We propose asymmetric co-teaching with different architectures and machine unlearning (ACD-U), a robust learning framework that integrates different model architectures with machine unlearning. ACD-U addresses the aforementioned limitation through two core mechanisms. First, it pairs a Contrastive Language-Image Pretraining (CLIP)-pretrained vision Transformer (ViT) with a convolutional neural network (CNN), leveraging their complementary learning behaviors: the ViT provides stable predictions, whereas the CNN remains adaptive throughout training. This asymmetry, with the ViT trained only on clean samples and the CNN trained through SSL, effectively mitigates confirmation bias. Second, selective unlearning enables post hoc error correction by identifying incorrectly memorized samples through loss trajectory analysis and CLIP consistency checks, and removing their influence via Kullback-Leibler divergence-based forgetting. This approach shifts the learning paradigm from passive error avoidance to active error correction. ACD-U achieves a 35% relative improvement over DivideMix on CIFAR-100 under 90% symmetric noise, the best results across all noise rates on Red Mini-ImageNet, and top accuracy on both Clothing1M and WebVision. Ablation analyses confirm that the pretrained ViT reduces critical sample misclassifications to approximately one-sixth of those of DivideMix, and that the unlearning and asymmetric co-teaching contribute complementarily across different noise regimes. These results suggest that ACD-U can improve the reliability of deep learning models trained on large-scale web-collected datasets, where label noise is unavoidable. The code is publicly available at https://github.com/meruemon/ACD-U.

Deep neural networks excel in image classification with clean data, but their performance deteriorates significantly when trained with noisy labels. In contrast to conventional approaches, which address this challenge through sample selection or loss correction, this study proposes a fundamentally different paradigm in which machine unlearning is employed to selectively forget noisy samples. Existing unlearning methods such as scalable remembering and unlearning unbound cause detrimental feature space dispersion, which creates ambiguous decision boundaries. Thus, this paper introduces a feature space-preserving unlearning framework that maintains discriminative structures while removing the influence of noisy labels. The proposed approach computes class centroids from reliable samples and implements a dual optimization strategy in which noisy samples are pushed away from all centroids while clean samples are simultaneously pulled toward their appropriate class centers. Comprehensive experiments on the CIFAR-10 and CIFAR-100 datasets with symmetric and asymmetric noise (ranging from 10%–90%) show that our method consistently outperforms standard unlearning techniques, achieving accuracy improvements of up to 27.61 percentage points on the more complex CIFAR-100 dataset under high-noise conditions. Furthermore, we confirm that our method remains effective even when applied to models trained with state-of-the-art noisy label learning methods. Visualization analysis demonstrates that our approach preserves compact class clusters with clear boundaries, despite aggressive unlearning. The reported implementation is available at https://github.com/meruemon/FSPMU.

This paper describes a deep-learning-based methodology for assessing arteriosclerosis through the automated detection of carotid artery calcification in dental panoramic radiographs. The proposed framework introduces a detector architecture that sequentially integrates a segmentation module and a classification network, achieving both computational efficiency and high diagnostic accuracy. A TransFuse-based segmentation model is combined with a ResNet classification backbone, enabling a substantial reduction in the number of parameters compared with conventional approaches while maintaining high detection precision.

As an alternative to two-dimensional codes, there is growing interest in techniques that embed information in images and detect that information from their printed counterparts. In this paper, we propose the detection of information from data-embedded images on curved surfaces without a priori knowledge. In our proposed method, auxiliary lines are added in advance around the image in which data are embedded and information is detected by correcting the image using these lines during capture. This method mitigates limitations related to image-capturing conditions in conventional approaches.

Fake news on social media poses serious societal risks. Therefore, early detection during the initial stages of diffusion is crucial for minimizing its impact. However, existing propagation-based methods inherently rely on sufficient diffusion, which limits their effectiveness in early scenarios. To address this challenge, we propose a novel early detection framework that integrates generalized distillation and multi-task learning. Our approach transfers knowledge from a teacher model with complete diffusion information to a student model restricted to early-stage data, while jointly optimizing fake news classification and diffusion pattern prediction as auxiliary tasks. We adopt an evolving dynamic graph convolutional network with time-dependent weights to effectively model the temporal evolution of propagation structures. In comprehensive experiments on the FakeNewsNet (PolitiFact and GossipCop) and PHEME datasets, our method achieved up to 16.4% improvement in the F1 score within one hour of diffusion onset. In ablation studies, we further validated the contribution of each component, thereby highlighting the significant impact of distillation on early-stage performance. In this study, we introduce generalized distillation in the domain of propagation-based fake news detection to offer a novel solution to the early-stage data limitation problem.

Although social networking services (SNS) have enabled the free exchange of opinions and feelings, the posting of malicious content has increasingly become a problem. To solve this problem, malicious behavior detection methods based on posting behavior are being developed. Existing methods focus on semantic analysis of posts using natural language processing, and one existing approach uses graph neural networks to consider context from various elements, such as users, posts, hashtags, and entities. However, this approach does not adequately capture the complex patterns and interactions of SNS networks. In particular, it is insufficient to fully capture the complexity of heterogeneity between nodes and edges in an SNS network. In this paper, we propose a method for extended heterogeneous graph construction and an architecture for heterogeneous graph embedding learning. The proposed method focuses on and exploits the diverse heterogeneity of social networks, optimally integrates heterogeneous information from SNS posts, and analyzes the relationships in the data to improve the performance of malicious behavior detection. The effectiveness of the proposed method is demonstrated by evaluation on a newly collected large dataset.

Fake news has become a significant societal problem, and the need for automatic fake news detection techniques is growing. In recent years, graph-based methods focusing on the structure of news propagation have been proposed and significantly improved detection accuracy. Although some methods consider the temporal evolution of the propagation structure using dynamic graphs, they typically use a two-step approach, where structural features are first extracted independently of the temporal information and are then combined with temporal features in a separate step. In this study, we propose a novel fake news detection method based on a dynamic graph convolutional network that directly incorporates time series information during structural feature extraction. By introducing time series-aware structural feature extraction, our method more effectively captures the temporal evolution of the news propagation structure, improving fake news detection performance. We evaluated the effectiveness of the proposed method through experiments on two real-world datasets, FakeNewsNet and FibVID.

Calcification regions, which may be observed on dental panoramic radiographs, are a sign of vascular disease. Therefore, automatic detection methods based on semantic segmentation (SS) have been proposed. However, because of the small amount of data in the available dataset, the segmentation accuracy was insufficient. This paper proposes a method that uses adversarial features (AFs) for this problem. We extend AFs, which are an adversarial training method for discriminative problems, to SS. The proposed method can improve performance, even with a small amount of data.

Recent methods for learning with noisy labels often mitigate the effects of noisy labels by sample selection and label correction. However, high feature similarity between classes can reduce the effectiveness of these methods. In this paper, we propose a learning method that uses contrastive learning to explicitly disentangle features of highly similar classes in the feature space. Specifically, we first compute the similarity between classes to identify similar classes. Next, we introduce a new loss function that separates the features of similar class samples in the feature space. This solves the problem of the mixing of similar classes, which affected previous methods. Our proposed method can easily be integrated into the loss functions of various existing methods. Experiments on CIFAR-10, CIFAR-100, WebVision, and Clothing1M show our method achieves high accuracy on datasets with various noise patterns, outperforming existing methods significantly at high noise rates.

We can find areas of calcification in Carotid arteries on dental panoramic radiographs, which may be a sign of vascular disease. It is hoped that detection of these areas can prevent the sudden onset of vascular disorders by prompting patients to consult a physician. For this purpose, we propose an automatic detection method based on a model combining semantic segmentation and convolutional neural networks. This method uses Nash-MTL, a multi-task learning method, to solve the tradeoff between segmentation and identification tasks with different properties. In this method, the parameters for adjusting the gradient vector are obtained through an optimization method. However, in the case of two tasks, as in this problem, the exact solution can be obtained, computational complexity can be reduced without compromising accuracy. Furthermore, the computational complexity can be further reduced by simplifying the solution. The effectiveness of the proposed method is demonstrated by applying it to actual data.

This paper addresses the performance degradation of deep neural networks caused by learning with noisy labels. Recent research on this topic has exploited the memorization effect: networks fit data with clean labels during the early stages of learning and eventually memorize data with noisy labels. This property allows for the separation of clean and noisy samples from a loss distribution. In recent years, semi-supervised learning, which divides training data into a set of labeled clean samples and a set of unlabeled noisy samples, has achieved impressive results. However, this strategy has two significant problems: (1) the accuracy of dividing the data into clean and noisy samples depends strongly on the network’s performance, and (2) if the divided data are biased towards the unlabeled samples, there are few labeled samples, causing the network to overfit to the labels and leading to a poor generalization performance. To solve these problems, we propose the curriculum regularization and adaptive semi-supervised learning (CRAS) method. Its key ideas are (1) to train the network with robust regularization techniques as a warm-up before dividing the data, and (2) to control the strength of the regularization using loss weights that adaptively respond to data bias, which varies with each split at each training epoch. We evaluated the performance of CRAS on benchmark image classification datasets, CIFAR-10 and CIFAR-100, and real-world datasets, mini-WebVision and Clothing1M. The findings demonstrate that CRAS excels in handling noisy labels, resulting in a superior generalization and robustness to a range of noise rates, compared with the existing method.

The widespread dissemination of fake news on social media has substantial economic and social implications. Although traditional propagation-based methods employing graph neural networks show promise for fake news detection, they disregard the influence of confirmation bias in the spread of fake news between users with similar viewpoints. This paper presents a detection approach that accounts for opinion similarity between users by scrutinizing their stances toward news articles and users’ post interactions. Using a graph transformer network, our method simultaneously extracts global structural information and interactions of similar stances. Furthermore, it addresses the challenges of stance analysis targeting microblogs while minimizing the effect of poorly represented stance features. We evaluated our approach using custom-crawled Twitter data and the benchmark FibVID dataset. It demonstrated a marked improvement in detection performance compared with conventional methods, including state-of-the-art methods.

Fake news has become a significant social problem, and the need for automatic fake news detection techniques is growing. In recent years, graph-based methods that focus on the structure of news propagation have been proposed and have significantly improved detection accuracy. Although some methods consider the temporal evolution of the propagation structure using dynamic graphs, they typically use a two-step approach, where structural features are first extracted independently of temporal information and then combined with temporal features in a separate step. In this paper, we propose a novel fake news detection method based on a dynamic graph convolutional network that directly incorporates time-series information during structural feature extraction. By introducing time-series-aware structural feature extraction, our method more effectively captures the temporal evolution of the news propagation structure, thereby improving the fake news detection performance. We evaluated the effectiveness of the proposed method through experiments on FakeNewsNet.

User recommendation systems on social media platforms play an important role in building new friendships and sharing ideas with people who share similar interests. However, highly accurate user recommendation systems can lead to the formation of echo chambers and filter bubbles by presenting only similar users. This reduces the diversity of social relationships and can contribute to the spread of malicious rumors and conspiracy theories. In this paper, we propose a recommendation method that diversifies the user representations by applying the selection of neighborhood information and re-weighting the losses for shared relationships of users, aiming to improve diversity without significantly sacrificing accuracy. Utilizing a dataset from the popular social media platform Twitter (currently X), we demonstrate the effectiveness of this method.

Carotid arteries on dental panoramic radiographs may show areas of calcification, which is a sign of vascular disease. The sudden onset of vascular disease can be prevented by a dentist detecting these areas at a dental clinic and prompting the patient to visit a physician. A method based on semantic segmentation was proposed as an automatic detection method for this purpose. However, because of the small number of available data, the detection accuracy for calcification regions was insufficient. In this study, we propose a method that applies a type of adversarial training called latent space virtual adversarial training (LVAT). We also describe a method to improve the pre-training of a variational autoencoder, which is essential in LVAT, using healthy subjects’ data. We expect this method to improve detection accuracy, even with a limited number of data. We demonstrated the effectiveness of the proposed method by applying it to actual data.

Deep neural networks (DNNs) have proven highly effective in various computational tasks, but their success depends largely on access to large datasets with accurate labels. Obtaining such data may be challenging and costly in real-world scenarios. Common alternatives, such as the use of search engines and crowdsourcing, often result in datasets with inaccurately labeled, or “noisy,” data. This noise may significantly reduce the ability of DNNs to generalize and maintain reliability. Traditional methods for learning with noisy labels mitigate this drawback by training DNNs selectively on reliable data, but they often underutilize available data. Although data augmentation techniques are useful, they do not directly solve the noisy label problem and are limited in such contexts. This paper proposes a confidence-guided Mixup named ConfidentMix, which is a data augmentation strategy based on label confidence. Our method dynamically adjusts the intensity of data augmentation according to label confidence, to protect DNNs from the detrimental effects of noisy labels and maximize the learning potential from the most reliable portions of the dataset. ConfidentMix represents a unique blend of label confidence assessment and customized data augmentation, and improves model resilience and generalizability. Our results on standard benchmarks with synthetic noise, such as CIFAR-10 and CIFAR-100, demonstrate the superiority of ConfidentMix in high-noise environments. Furthermore, extensive experiments on Clothing1M and mini-WebVision have confirmed that ConfidentMix surpasses state-of-the-art methods in handling real-world noise.

Graph neural networks (GNNs) have significantly advanced recommender systems (RecSys) by enhancing their accuracy in complex collaborative filtering scenarios. However, this progress often comes at the cost of overlooking the diversity of recommendations, a factor in user satisfaction. Addressing this gap, this paper introduces the disentangled representation graph neural network (DRGNN). DRGNN integrates diversification into the candidate generation stage using two specialized modules. The first employs disentangled representation learning to separate item preferences from category preferences, thereby mitigating category bias in recommendations. The second module, focusing on positive sample selection, further reduces category bias. This approach not only maintains the high-order connectivity strengths of GNNs but also substantially improves the diversity of recommendations. Our extensive validation of DRGNN on three comprehensive web service datasets, Taobao, Amazon Beauty and MSD, shows that it not only matches the state-of-the-art methods in accuracy but also excels in achieving a balanced trade-off between accuracy and diversity in recommendations.

In the context of the increasing spread of misinformation via social network services, in this study, we addressed the critical challenge of detecting and explaining the spread of fake news. Early detection methods focused on content analysis, whereas recent approaches have exploited the distinctive propagation patterns of fake news to analyze network graphs of news sharing. However, these accurate methods lack accountability and provide little insight into the reasoning behind their classifications. We aimed to fill this gap by elucidating the structural differences in the spread of fake and real news, with a focus on opinion consensus within these structures. We present a novel method that improves the interpretability of graph-based propagation detectors by visualizing article topics and propagation structures using BERTopic for topic classification and analyzing the effect of topic agreement on propagation patterns. By applying this method to a real-world dataset and conducting a comprehensive case study, we not only demonstrated the effectiveness of the method in identifying characteristic propagation paths but also propose new metrics for evaluating the interpretability of the detection methods. Our results provide valuable insights into the structural behavior and patterns of news propagation, contributing to the development of more transparent and explainable fake news detection systems.

Recent character detectors have been modeled using deep neural networks and have achieved high performance in various tasks, such as text detection in natural scenes and character detection in historical documents. However, existing methods cannot achieve high detection accuracy for wooden slips because of their multi-scale character sizes and aspect ratios, high character density, and close character-to-character distance. In this study, we propose a new U-Net-based character detection and localization framework that learns character regions and boundaries between characters. The proposed method enhances the learning performance of character regions by simultaneously learning the vertical and horizontal boundaries between characters. Furthermore, by adding simple and low-cost post-processing using the learned regions of character boundaries, it is possible to more accurately detect the location of a group of characters in a close neighborhood. In this study, we construct a wooden slip dataset. Experiments demonstrated that the proposed method outperformed existing character detection methods, including state-of-the-art character detection methods for historical documents.