Image clustering is a fundamental task in unsupervised visual learning. While recent self-supervised methods have explored various pretext tasks to generate supervision signals for clustering, they typically depend exclusively on raw images, resulting in insufficient supervision signals that are inherently constrained by limited visual semantics. In this paper, we propose a novel Semantic-Augmented image Clustering (SAC) method, which transcends the inherent limitations of purely visual representations through the integration of external knowledge. Specifically, SAC utilizes Vision-Language pre-trained Models (VLMs) to flexibly generate textual descriptions for each image, providing external semantic cues to supplement the visual information. By integrating both visual and textual information, SAC achieves image clustering through a multi-modal learning framework. To mitigate the negative impact of inaccurate textual information, SAC designs an uncertainty-driven adaptive weighting mechanism that explores both intra-modal and inter-modal neighborhood structures, and incorporates the adaptive weights into intra-modal and inter-modal contrastive learning, which improves the robustness against noisy image-text correspondences. Experiments on several popular datasets demonstrate the superiority of SAC compared to state-of-the-art methods.

Multimodal learning has achieved remarkable success by integrating heterogeneous information from multiple sources. However, most existing methods either treat all samples uniformly or rely on deterministic prediction correctness to guide cross-modal alignment, overlooking the varying degrees of epistemic uncertainty inherent in each modality. Such assumptions often lead to noise propagation, especially when models exhibit over-confident yet unreliable predictions. To address this limitation, we propose Adaptive Predictive Contrastive Learning (AdaPCL), an evidential uncertainty modulated framework designed to regulate cross-modal interactions based on modality-specific reliability. Specifically, we first employ evidential deep learning to explicitly quantify modality-specific epistemic uncertainty, establishing a dual-view assessment that calibrates discriminative confidence with evidential reliability. Building upon this, we introduce a continuous reliability-modulated mechanism that synthesizes these dual perspectives to assign soft, instance-specific weights. This design enables AdaPCL to formulate a unified and adaptive contrastive objective encompassing three complementary alignment strategies: (i) Symmetric Alignment for mutually reliable modality pairs, (ii) Directional Distillation where a reliable modality provides pseudo-supervision for its uncertain counterpart, and (iii) Reliability-Aware Slack Regularization that adaptively attenuates the influence of mutually unreliable samples without enforcing rigid geometric constraints. Extensive experiments on multiple benchmark datasets demonstrate that AdaPCL consistently outperforms baseline multimodal classification methods. Code is available at https://github.com/yuhongcqupt/AdaPCL.

Face age synthesis (FAS) predicts a person's future or past facial appearance. In FAS, modifying one facial attribute usually affects the generation of other attributes during face image generation. Current models directly learn entangled representations of age-related features, resulting in insufficient feature disentanglement, which consequently impairs their causal reasoning capability for FAS tasks. To this end, we propose a hierarchical causal learning model for face age synthesis (HCFace), which integrates hierarchical structures and causal relationships into the facial generative model. Specifically, we propose to leverage hierarchical causal relationships to align with facial features for feature disentanglement. Furthermore, we design a novel nonlinear mapping function that captures the true patterns of facial attribute changes with age, enhancing the disentanglement of these attributes. We conduct extensive experiments to validate the superiority of our proposed model. Compared to other advanced baseline methods, HCFace improves overall accuracy by 2.47%, with improvements of 9.75% and 9.69% in certain age-related attributes, such as skin and hair. Our source code is available at https://github.com/SE-hash/HCFace.

The reconstruction of gene regulatory networks (GRNs) is crucial for uncovering regulatory relationships between genes and understanding the mechanisms of gene expression within cells. With advancements in single-cell RNA sequencing (scRNA-seq) technology, researchers have sought to infer GRNs at the single-cell level. However, existing methods primarily construct global models encompassing entire gene networks. While these approaches aim to capture genome-wide interactions, they frequently suffer from decreased accuracy due to challenges such as network scale, noise interference, and data sparsity. This study proposes GRANet (Graph Residual Attention Network), a novel deep learning framework for inferring GRNs. GRANet leverages residual attention mechanisms to adaptively learn complex gene regulatory relationships while integrating multi-dimensional biological features for a more comprehensive inference process. We evaluated GRANet across multiple datasets, benchmarking its performance against state-of-the-art methods. The experimental results demonstrate that GRANet consistently outperforms existing methods in GRN inference tasks. In addition, in our case study on EGR1, CBFB, and ELF1, GRANet achieved high prediction accuracy, effectively identifying both known and novel regulatory interactions. These findings highlight GRANet's potential to advance research in gene regulation and disease mechanisms.

Transfer learning (TL) is a crucial technique to improve the cross-system analysis ability of data-driven transient stability assessment (TSA) approaches. However, existing transfer learning methods for TSA in new scenarios generally depend on a larger number of target samples, and the computational performance is limited by the inconsistency of temporal features in the transfer process. To resolve this issue, this paper proposes a temporal transfer learning framework (TTLF) for TSA, which can effectively improve the performance of cross-system TSA with zero target samples. The TTLF is sequentially composed of two modules. The first module is referred to as source sample characterization (SSC), which aims to better characterize the distribution information of transient time-series curves. The second module is termed temporal distribution matching (TDM), which aims to maximize the interclass dispersion and minimize the intra-class scatter in each hidden layer at each time. What’s more, the framework can also be extended to the situation with a few cross-system samples, which can significantly improve the assessment performance. The simulation results demonstrate the efficiency and scalability of the proposed framework in several utility benchmark systems with different data-driven models.

EEG emotion recognition is crucial in both human-machine interaction and healthcare. However, recognizing emotions across different subjects remains challenging due to individual variability. While existing multi-source domain adaptation methods have been utilized for cross-subject EEG emotion decoding, they often struggle with irrelevant or weakly relevant source domains, leading to negative transfer. Additionally, variations within subdomains are often neglected in these studies. We propose a joint domain adaptation method, Adaptive Source Joint Domain Adaptation (ASJDA) to address these issues. ASJDA utilizes an unsupervised adaptive source selection strategy to select a subset of source domains by evaluating the Jensen-Shannon divergence between the source and target domains, choosing those most relevant to the target. Subsequently, it implements joint domain adaptation with these chosen sources at both the domain and category subdomain levels. Our proposed method outperforms existing state-of-the-art methods, achieving cross-subject accuracies of 96.81% in SEED, 89.69% in SEED-IV, and 69.31% in DEAP. This work significantly advances the state of the art in EEG emotion recognition by effectively addressing the challenges of cross-subject variability.

Multimodal Emotion Recognition in Conversations (MERC) is an important topic in human-computer interaction. In the MERC task, conversations exhibit dynamic emotional dependency, including inter-speaker and intra-speaker emotional dependency, both are vital in understanding the content. However, current research primarily integrates these two emotional dependencies into one unified module, limiting the accuracy of MERC. In this paper, we propose a dynamic emotion-dependent network with relational subgraph interaction named DEDNet. DEDNet introduces relational subgraphs to separately model two emotional dependencies, enabling structured learning paths for utterances based on distinct emotional dependency types. Specifically, nodes indicate the utterances at different moments in the conversation, while edges define the emotional dependency and temporal relationships between nodes. To explicitly capture the differences between these two emotional dependencies, distinct subgraphs are designed for comprehensive representations. Furthermore, we propose an incremental interactive strategy, sequentially leveraging two emotional dependencies to learn the changes in dependency relationships. We find that modeling inter-speaker emotional dependency can better identify negative emotions and modeling intra-speaker emotional dependency can better recognize positive emotions. Experimental results demonstrate that our model outperforms current state-of-the-art methods on three benchmark datasets, IEMOCAP, MELD and DailyDialog.

Outliers are intrinsic to recommender systems (RSs) due to user uncertainty and bring large deviations to the total loss. Classical loss functions such as$L_{1}$and$L_{2}$do not consider this issue, thus biasing the model toward abnormal ratings. Advanced ones such as Logcosh, Huber, and Wing losses try to control model biasing; however, they do not consider the rating ranges of recommender systems. In this article, we propose a Seagull loss combining quadratic and softsign functions to handle this issue. The quadratic function aims to preserve or appropriately increase losses caused by small deviations, while softsign function aims to suppress losses caused by large deviations. It is distinguished from three types of popular losses. Compared with$L_{1}$and$L_{2}$, it controls the upper bound, therefore suppressing the influence of outliers. Compared with Logcosh and Huber, it not only adjusts the loss of small deviations but also suppresses the loss of large deviations. Compared with Wing loss, it considers the characteristics of data in RSs and therefore is more suitable to the matrix factorization model. Experiments are undertaken on six real-world recommendation datasets in comparison with six popular loss functions. Results show that Seagull loss function has superiority to all losses above on mean absolute error (MAE) and root mean square error (RMSE) and slight inferiority to$L_{2}$and Wing on hit ratio (HR), mean average precision (MAP), and normalized discounted cumulative gain (NDCG).

Precisely identifying and explaining subpopulations in heterogeneous populations is essential to understanding the disease subtype. Using community detection to identify subpopulations is a promising way. However, there remains an issue in the existing community detection: Current methods for identifying subpopulations in medical data rely solely on separate attribute values, ignoring the important association rules between attribute values. Association rules are crucial in medical diagnosis to determine disease subtypes. Thus, We propose a rule-guided community detection (RGCD) method for precisely identifying homogeneous subpopulations. Specifically, the RGCD incorporates association rules into the original network, thereby constructing an augmented network. It proves that decomposing the embedding vectors obtained from biased random walks on the augmented network is equivalent to decomposing the transition probability matrix. Based on this proof, we enhance the transition probability matrix through rule-guided biased random walks, resulting in the rule-augmented matrix. By performing matrix decomposition and clustering on this matrix, we achieve precise identification of subpopulations. To the best of our knowledge, this is the first work that introduces the incorporation of association rules into community detection. Extensive experiments on 10 real-world datasets from medical fields fully show that the RGCD is more competitive than six state-of-the-art community detection methods. The weighted F1 of RGCD increases by up to 22.62%, compared to the best existing community detection methods. Furthermore, We provide a qualitative depiction of the subpopulations obtained through RGCD and acquire medically significant insights.

In the natural language processing task of clinical named entity recognition (CNER), accurately identifying the boundaries and categories of medical entities is crucial. However, traditional methods struggle to recognize a large number of clinical terms and symbols that have never been encountered before, ultimately limiting the performance of CNER. Besides, there exist some easy-to-confuse Chinese clinical entities that are semantically similar but belong to quite different categories, such as "" (pulmonary nodules, a symptom entity) and "" (pulmonary tuberculosis, a disease entity), which can lead to entity misidentification. To address these problems, we propose a novel NER model called Cross-Interaction of Chinese characters structures and Boundary Features (CCS). The proposed model leverages Chinese character structural features and boundary information to comprehensively and accurately identify confusing entities. We further design a Cross-Attention mechanism to capture dependency relationships between different entities and radicals of characters, enhancing the model's semantic understanding of specialized terms and symbols, as well as improving its ability to recognize boundaries. Our experimental results show that our proposed model outperforms other state-of-the-art models on various public medical datasets, achieving significant improvements on the CCKS2020, CMeEE, CMI, and IMCS datasets, respectively.

The causal structure learning for streaming features (CSLSFs) faces the following challenges: 1) the precision of learned causal structures is limited due to the score-based learning method and 2) they fail to detect the latent confounders. To address the challenges, this article proposes a novel causal structure learning method with linear non-Gaussian acyclic models for streaming features (LiNGAM-SFs), which utilizes the causal identifiability of the data. It is the first time to utilize LiNGAMs for online causal structure learning. First, we utilize the classical SF algorithm to learn the causal skeleton. This article provides the property of this skeleton, proving that two adjacent variables on an edge is one of three possible structures. Second, we give two propositions and identify the causal directions in the presence of latent variables (ICDPLV) subalgorithm to distinguish among the three structures and precisely identify the causal directions. In addition, the subalgorithm can output a candidate set of latent confounders from a local perspective. Finally, the detecting latent confounder (DLC) subalgorithm detects the latent confounders in the candidate set with a global perspective. The precision of the proposed method is increased at least by 11% on average than those of the state-of-the-art method. Furthermore, the experiments verify that the LiNGAM-SF method is able to detect the latent confounders.

Replaying episodic memory hippocampus-based is a promising class incremental learning (CIL) method, and it must address the problem of catastrophic forgetting. However, most current studies have ignored background information provided by the parahippocampal gyrus. Therefore, in this paper, the topological data analysis (TDA) technology is proposed for the first time to simulate the biological function of the parahippocampal gyrus in order to obtain multi-scale background information such as geometric and topological. Based on the idea, a novel CIL method via semantic information mapping and background information calibrating (SiBiCIL) is proposed. It takes categorical prototypes as semantic information and uses the constructed mapping function to generate the prototype classifiers. The classification results are then calibrated using background information extracted from the point-clouded data. In addition to the cross-entropy loss and the knowledge distillation loss, a difference loss is defined to maintain the distinctiveness between the prototype classifiers. These losses collectively contribute to updating the model and retaining previous classes’ knowledge, as well as a way to mitigate catastrophic forgetting. Compared to other methods with the total number of exemplars no more than 500, the experimental results on CIFAR-10, CIFAR-100, ImageNet-100, and Protein_family datasets show that highest percentage increases of average classification accuracies are 49.72%, 7.15%, 1.96%, and 12.35% respectively, while the forgetting rate has concurrently decreased by 32.35%, 9.80%, 12.50%, and 5.41%. Moreover, SiBiCIL significantly reduces the episodic memory buffer budget by an average of 76.39%.

Virtual sample generation (VSG) is an important technology for dealing with small sample learning in some industries. Using evolutionary computation algorithms to solve virtual sample generation is a promising way. However, two issues remain unaddressed in the existing VSG approaches: estimating the distribution of original samples and ensuring the authenticity of virtual samples. Thus, this paper proposes a novel Virtual Sample Generation approach based on genetic algorithm (GA) combing with Information Gain and Acceptance-rejection Sampling, abbreviated as VSG3A2. First, this work develops the ARS-VAD subalgorithm, by integrating the acceptance rejection sampling method into the crossover and mutation operations of GA. The algorithm ensures that the distribution of offspring attribute values is close to the distribution of original samples at attribute level. Second, this work presents the IG-VSS subalgorithm, which is combined with the idea of minimizing absolute information gain, to find the optimal offspring sample as a virtual sample in each loop, ensuring the authenticity of virtual samples at the sample level. To the best of our knowledge, this is the first work that introduces the concept of information gain into VSG. Extensive experiments on four public datasets from various fields fully demonstrate that the VSG3A2 is more competitive than six state-of-the-art VSG approaches. The MAE, the RMSE and the MAPE metrics of prediction models, trained on virtual samples generated by the proposed VSG3A2, are reduced at least by 19.78%, 19.56%, 20.16% on average than that of the best compared VSG approach, respectively.

OBJECT: Transformer-based neural networks have been applied to the electroencephalography (EEG) decoding for motor imagery (MI). However, most networks focus on applying the self-attention mechanism to extract global temporal information, while the cross-frequency coupling features between different frequencies have been neglected. Additionally, effectively integrating different neural networks poses challenges for the advanced design of decoding algorithms. METHODS: This study proposes a novel end-to-end Multi-Scale Vision Transformer Neural Network (MSVTNet) for MI-EEG classification. MSVTNet first extracts local spatio-temporal features at different filtered scales through convolutional neural networks (CNNs). Then, these features are concatenated along the feature dimension to form local multi-scale spatio-temporal feature tokens. Finally, Transformers are utilized to capture cross-scale interaction information and global temporal correlations, providing more distinguishable feature embeddings for classification. Moreover, auxiliary branch loss is leveraged for intermediate supervision to ensure the effective integration of CNNs and Transformers. RESULTS: The performance of MSVTNet was assessed through subject-dependent (session-dependent and session-independent) and subject-independent experiments on three MI datasets, i.e., the BCI competition IV 2a, 2b and OpenBMI datasets. The experimental results demonstrate that MSVTNet achieves state-of-the-art performance in all analyses. CONCLUSION: MSVTNet shows superiority and robustness in enhancing MI decoding performance.

A dilemma faced by classification is that the data is not collected at the same frequency in some applications. We investigate the mixed frequency data in a new way and recognize them as a special style of multi-view data, in which each view data is collected at a different sampling frequency. This paper proposes a discriminative dictionary pair learning method constrained by ordinal locality for mixed frequency data classification (shorted by DPLOL-MF). This method integrates synthesis dictionary and analysis dictionary into a dictionary pair, which not only improves computational cost caused by the ℓ0or ℓ1-norm constraint, but also can deal with the sampling frequency inconsistency. The DPLOL-MF utilizes a synthesis dictionary to learn class-specified reconstruction information and employs an analysis dictionary to generate coding coefficients by analyzing samples. Particularly, the ordinal locality preserving term is leveraged to constrain the atoms of dictionaries pair to further facilitate the learned dictionary pair to be more discriminative. Besides, we design a specific classification scheme for the inconsistent sample size of mixed frequency data. This paper illustrates a novel idea to solve the classification task of mixed frequency data and the experimental results demonstrate the effectiveness of the proposed method.

Co-training is a popular semi-supervised learning method. The learners exchange pseudo-labels obtained from different views to reduce the accumulation of errors. One of the key issues is how to ensure the quality of pseudo-labels. However, the pseudo-labels obtained during the co-training process may be inaccurate. In this paper, we propose a safe co-training (SaCo) algorithm for regression with two new characteristics. First, the safe labeling technique obtains pseudo-labels that are certified by both views to ensure their reliability. It differs from popular techniques of using two views to assign pseudo-labels to each other. Second, the label dynamic adjustment strategy updates the previous pseudo-labels to keep them up-to-date. These pseudo-labels are predicted using the augmented training data. Experiments are conducted on twelve datasets commonly used for regression testing. Results show that SaCo is superior to other co-training style regression algorithms and state-of-the-art semi-supervised regression algorithms.

To predict and optimize the billet heating process in the reheating furnace for rolling mills, this article proposes a hybrid model that combines the data-driven model with traditional mechanism knowledge, abbreviated as HMDM. By examining the heat conduction mechanism, a billet temperature distribution equation is established. Then, the billet temperature distribution in each heating zone is calculated and spliced with the corresponding process parameters. The stacked-autoencoder is utilized to extract the features of process parameters, and the long short-term memory model is employed to predict the temperature. Finally, using the previous predictions, the parameters of the subsequent heating stage are optimized and adjusted during the heating process. The experimental results on the real steel plant verify the effectiveness of HMDM. For example, the temperature prediction error has been reduced to less than 4$^{\circ }$C, and the number of billets with abnormal tapping temperature has been decreased by 42.9%.

A dilemma faced by classification is that the data is not collected at the same frequency in some applications. We investigate the mixed frequency data in a new way and recognize them as a special style of multi-view data, in which each view data is collected at a different sampling frequency. This article proposes a discriminative dictionary pair learning method constrained by ordinal locality for mixed frequency data classification (shorted by DPLOL-MF). This method integrates synthesis dictionary and analysis dictionary into a dictionary pair, which not only improves computational cost caused by the${\ell _0}$or${\ell _1}$-norm constraint, but also can deal with the sampling frequency inconsistency. The DPLOL-MF utilizes a synthesis dictionary to learn class-specified reconstruction information and employs an analysis dictionary to generate coding coefficients by analyzing samples. Particularly, the ordinal locality preserving term is leveraged to constrain the atoms of dictionaries pair to further facilitate the learned dictionary pair to be more discriminative. Besides, we design a specific classification scheme for the inconsistent sample size of mixed frequency data. This paper illustrates a novel idea to solve the classification task of mixed frequency data and the experimental results demonstrate the effectiveness of the proposed method.

The existing multi-view clustering algorithms require that a sample in a view is completely or partially mapped onto one or more samples in a different corresponding view. However, this requirement could not be satisfied in many practical applications. Fortunately, there is a common cognition that the graph structure formed from each view should be as consistent as possible. Thus, this paper proposes a novel multi-view clustering method for unknown mapping relationships between cross-view samples based on the framework of non-negative matrix factorization, as an attempt to solve this problem. The objective function is designed by effectively building reconstruction error terms, local structural constraint terms, and cross-view mapping loss terms by exploring cross-view relationships. The experimental results show that the proposed method not only performs well to reveal the real mapping relationships between cross-view samples but also outperforms the comparison algorithms on the obtained clustering results.

The problem of initialization of active learning is considered in this paper. Especially, this paper studies the problem in an imbalanced data scenario, which is called as class-imbalance active learning cold-start. The novel method is two-stage clustering-based active learning cold-start (ALCS). In the first stage, to separate the instances of minority class from that of majority class, a multi-center clustering is constructed based on a new inter-cluster tightness measure, thus the data is grouped into multiple clusters. Then, in the second stage, the initial training instances are selected from each cluster based on an adaptive candidate representative instances determination mechanism and a clusters-cyclic instance query mechanism. The comprehensive experiments demonstrate the effectiveness of the proposed method from the aspects of class coverage, classification performance, and impact on active learning.

Aluminum alloy is widely used in transportation, catering, industry, sports, health and other fields, because of their excellent high specific strength and corrosion resistance. Aluminum industry has been an important mainstay industry of a national economy. In the process of electrolytic aluminum, the superheat degree is a very important production target. When an electrolysis cell is working in the appropriate superheat degree state, the life of cell is prolonged and the amounts of aluminum released will enhanced. However, to measure the superheat degree is very difficult and the measured results cannot timely feedback to the process of production. To address the problem, a soft sensing prediction model of superheat degree is proposed in this paper, in which some new concepts such as the decay function of data weight, the credibility of a rule and the rule tree are introduced. Basically speaking, the processing of the soft sensing prediction model is mainly based on the rough set data analysis method and a tree data structure. The static rules are obtained from the history data by using the attribute reduction and value reduction method in the rough sets. The rule tree is updated timely based on the incremental data set accordingly. The effectiveness of the proposed model is verified with the aluminum production data provided by Shandong Weiqiao Aluminum Electrolysis limited company in China.

In the human cognition learning, the noise self-filtering filters the noise data by itself, making the human brain very robust. Although k-Nearest Neighbors (kNN) can decrease the affection of noise data by optimizing the parameter k, noise data still may deteriorate the learning results to an extent on various data. Therefore, this paper proposes the strategy for kNNs. Here, we formalize such training strategies in the context of kNNs, and they are called with “Noise Self-filtering k-Nearest Neighbors” (NSF-kNNs). As its name suggests, it is determined by the model itself based on what it has already learned, as oppose to some predefined heuristic criteria. We derive the mathematical model of the NSF-kNNs, which is an algorithm framework. The NSF-kNNs are compared with the exact kNNs and as far as we know the most efficient approximate kNN [1]. The results show the achievement of improvements in the robustness and generalizability on various data sets.

This special issue of Fundamenta Informaticae (FI) contains a selection of papers presented initially at the 5th International Conference on Rough Sets and Knowledge Technology (RSKT

In many applications, clusters tend to have vague or imprecise boundaries. It is desirable that clustering techniques should consider such an issue. The decision-theoretic rough set (DTRS) model is a typical probabilistic rough set model, which has t