Infrared–visible image fusion leverages the complementary strengths of both modalities to enhance visual perception in challenging environments. While image-level fusion has achieved promising results, extending it to video remains challenging due to temporal illumination variations, brightness flickering, and visual inconsistency caused by motion under non-uniform illumination. To address these issues, we propose Seq-IF, a sequential fusion framework that ensures visual consistency and structural fidelity when processing video sequences. The framework comprises a static–dynamic decoupling module for robust foreground–background separation. For the background, fusion is performed by selecting the frame with the highest contrast to ensure clarity and stability. For the dynamic objects, the pixel intensity is adaptively adjusted across consecutive frames by a lightweight MLP-based illumination-consistency fine-tuning module that performs online adaptation and dynamically optimizes brightness in response to scene changes. Later we introduce a spatial–frequency fusion module integrating multi-scale encoder and edge-guided decoder to ensure structural consistency. Extensive experiments demonstrate that the fusion results produced by Seq-IF outperform baseline methods in terms of clarity, detail preservation, and illumination stability, achieving smoother temporal transitions and enhanced perceptual quality. Furthermore, the effectiveness of Seq-IF is validated on downstream tasks, including object detection and optical flow estimation, highlighting its applicability to real-world scenarios.

Enhancing model performance on low-capacity devices remains a significant challenge in the Internet of Models (IoM). Inspired by the efficiency of disentangled representation learning, this study proposes a model decomposition approach to improve response time on low-capacity devices while maintaining accuracy, through collaborative learning with models deployed on higher-capacity devices. First, we introduce a divide-and-conquer strategy that decomposes and learns knowledge within each data sample. This enables lightweight sub-models, tailored to specific knowledge components, to respond efficiently on low-capacity devices in the IoM system. Second, we design a progressive crossfusion mechanism to promote mutual enhancement among these knowledge-specific sub-models. Third, we optimize the learning process by aligning the updates of these sub-models with those of the models on higher-capacity devices. This collaborative optimization is guided via knowledge distillation during model aggregation. Experimental results on image classification tasks demonstrate that our approach reduces response time by 9.87 % to 60.31 % without compromising accuracy.

Anomaly detection is a fundamental component of intelligent monitoring in the Internet of Things (IoT), where accuracy, efficiency, and interpretability are critical requirements. However, existing methods often overlook the unique characteristics of IoT signals such as seasonality, trends, and irregular residual components, as well as the complex interactions among them. This oversight can lead to anomaly masking, increased false positives, and reduced interpretability in anomaly identification. Motivated by the effectiveness of disentangled representation learning, we propose TRAdetector, a novel disentangled reconstruction-based framework for IoT signals anomaly detection. TRAdetector explicitly models recurrent and consistent patterns, as well as irregular variations in the latent space by leveraging variational inference strategies, thereby enhancing probabilistic guidance in learning both regular and irregular temporal representations. A sparse coding strategy is incorporated within the latent space of the residual component to directly model inconsistent temporal fluctuations. Finally, a multihead cross-attention mechanism and a gated, decomposition-aware reconstruction strategy are designed to effectively model the complex interactions among different components. Extensive experiments show that our model achieves state-of-the-art performance on multiple benchmark datasets in terms of accuracy, efficiency, and interpretability.

Accurate classification of electroencephalogram (EEG) signals is crucial for brain activity understanding. However, EEG signals are characterized by data heterogeneity and label scarcity, which present a challenging low-data learning regime when building machine learning models. Existing methods tend to suffer from overfitting problem. To this end, we propose an adaptive Bayesian meta-learning framework for instance-specific learning and inference in EEG classification tasks. Specifically, first, a query set-driven dynamic parameter-based support set selection strategy is designed to adaptively fit the query set when constructing a meta-training task. Second, we employ an amortized variational inference network to generate task-specific adapted parameters given the support set, thereby achieving rapid model adaption for the inference of the data in the query set. Especially, a time- and frequency-aware representation learning encoder is leveraged to extract more task-relevant information guided by information bottleneck principle from time and frequency views, respectively, alleviating the low signal-to-noise ratio issue. Extensive experimental results on three public datasets demonstrate the superior effectiveness of our method.

Recent advances in graph representation learning provide new opportunities for computational drug-target interaction (DTI) prediction. However, it still suffers from deficiencies of dependence on manual labels and vulnerability to attacks. Inspired by the success of self-supervised learning (SSL) algorithms, which can leverage input data itself as supervision,we propose SupDTI, a SSL-enhanced drug-target interaction prediction framework based on a heterogeneous network (i.e., drug-protein, drug-drug, and protein-protein interaction network; drug-disease, drug-side-effect, and protein-disease association network; drug-structure and protein-sequence similarity network). Specifically, SupDTI is an end-to-end learning framework consisting of five components. First, localized and globalized graph convolutions are designed to capture the nodes' information from both local and global perspectives, respectively. Then, we develop a variational autoencoder to constrain the nodes' representation to have desired statistical characteristics. Finally, a unified self-supervised learning strategy is leveraged to enhance the nodes' representation, namely, a contrastive learning module is employed to enable the nodes' representation to fit the graph-level representation, followed by a generative learning module which further maximizes the node-level agreement across the global and local views by learning the probabilistic connectivity distribution of the original heterogeneous network. Experimental results show that our model can achieve better prediction performance than state-of-the-art methods.

Predicting treatment medicines is a key task in many intelligent healthcare systems. Prediction of treatment medicines can assist doctors in making informed prescription decisions for patients according to their Electronic Health Records (EHRs). However, predicting treatment medicines is a challenging task due to the following reasons: (1) heterogeneous nature of EHR data that typically includes laboratory results, treatment records, disease conditions, and demographic information; (2) complex correlations among EHR sequences, including inter-correlations between sequences and temporal intra-correlations within each sequence; (3) temporal dynamics of these correlations changing with disease progression. In this paper, we predict treatment medicines for patients with dual adaptive sequential networks (DASNet). Specifically, DASNet is designed with three components. First, a decomposed adaptive long short-term memory network (DA-LSTM) is designed to capture the intra- and inter-correlations in multiple heterogeneous temporal sequences. Then, we develop an attentive meta learning network (AT-MetaNet) to learn dynamic weight parameters for DA-LSTM, thus enabling it to model various correlation structures. Finally, we employ an attentive fusion network (AT-FuNet) to incorporate historical information and collectively fuse representation embeddings of heterogeneous data to predict treatment medicines. Our results on the public MIMIC-III dataset covering 11 medical conditions demonstrate that the proposed end-to-end model can achieve the state-of-the-art prediction performance while providing clinically useful insights.

Graph Neural Networks (GNNs) are prone to adversarial attacks and discriminatory biases. The cutting-edge studies usually adopt a perturbation-invariant consistency regularization strategy without considering the inherent prediction uncertainties, which can lead to unsatisfactory overconfidence for incorrect prediction under intent graph topology or node features attacks. Besides, operating on the complete graph structure is biased towards global level graph noise and brings severe computational issues. In this work, we develop a multi-view confidence-calibrated framework, called MCCNIFTY, for unified fair and stable graph representation learning. At its core is a multi-view uncertainty-aware node embedding learning module derived from evidential theory, including an intra-view evidence calibration, an inter-view evidence fusion, and an uncertainty-aware message passing process in a GNN architecture, which simultaneously optimizes for counterfactual fairness and stability at the sub-graph level. Experimental results on three real-world datasets demonstrate that our method is capable of adequately capturing inherent uncertainties while improving the fairness and stability via subgraph-induced multiview confidence calibration.

Recent advances in graph representation learning provide new opportunities for computational drug-target interaction (DTI) prediction. Inspired by the emerging graph mutual information-based algorithms, we propose MMIDTI, a multi-level mutual information-aware DTI prediction framework based on a heterogeneous network (i.e., drug-protein, drug-drug and protein-protein interaction network; drug-disease, drug-side-effect, and protein-disease association network; drug-structure and protein-sequence similarity network). More specifically, MMIDTI leverages an encoder-decoder framework that can learn the type-aware and meta-path augmented node representations by following a contrastive learning paradigm. The encoder part is a Graph Convolutional Network (GCN) and the decoder is an inner product of the learned representations to recover the original heterogeneous network. Meanwhile, MMIDTI exploits two levels of mutual information: (1) maximizing local mutual information, to obtain node representations that capture the global information content of the entire heterogeneous graph. (2) maximizing the global mutual information, to constrain the node representation to have desired statistical characteristics. Experimental results show that our model can achieve better prediction performance than state-of-the-art methods.

Social recommendation tasks exploit social connections to enhance recommendation performance. To fully utilize each user's first-order and high-order neighborhood preferences, recent approaches incorporate influence diffusion process for better user preference modeling. Despite the superior performance of these models, they either neglect the latent individual interests hidden in the user-item interactions or rely on computationally expensive graph attention models to uncover the item-induced sub-relations, which essentially determine the influence propagation passages. Considering the sparse substructures are derived from original social network, we name them as partial relationships between users. We argue such relationships can be directly modeled such that both personal interests and shared interests can propagate along a few channels (or dimensions) of latent users' embeddings. To this end, we propose a partial relationship aware influence diffusion structure via a computationally efficient multi-channel encoding scheme. Specifically, the encoding scheme first simplifies graph attention operation based on a channel-wise sparsity assumption, and then adds an InfluenceNorm function to maintain such sparsity. Moreover, ChannelNorm is designed to alleviate the oversmoothing problem in graph neural network models. Extensive experiments on two benchmark datasets show that our method is comparable to state-of-the-art graph attention-based social recommendation models while capturing user interests according to partial relationships more efficiently.

Partial correlation-based connectivity networks can describe the direct connectivity between features while avoiding spurious effects, and hence they can be implemented in diagnosing complex dynamic multivariate systems. However, existing studies mainly focus on single systems that are ill-equipped for incremental learning. Moreover, related methods estimate temporal connectivity network by imposing only sparse regularization without integrating pattern priors (e.g., inter-system shared pattern and intra-system intrinsic pattern), which have been proven effective in limiting noise interference. To this end, we develop an adaptive connectivity estimation model that incorporates prior patterns, namely Sparse Adaptive Meta-Learning Connectivity Network (SAMCN). Specifically, our model extends ideas of the gradient-based meta-learning to capture inter-system shared prior information by generating fast adaptive initialization parameters for the connectivity matrix. Then, a sparse variational autoencoder is proposed to generate a weight matrix for sparse regularization penalty in reweighted LASSO, which helps extract intra-system intrinsic patterns (local manifold structure). Experimental results on both synthetic data and real-world datasets demonstrate that our method is capable of adequately capturing the aforementioned pattern priors. Further, experiments from corresponding classification tasks validate the strength of the prior pattern-aware features connectivity network in resulting in better classification performance.

Following the recent advances of artificial intelligence, financial text mining has gained new potential to benefit theoretical research with practice impacts. An essential research question for financial text mining is how to accurately identify the actual financial opinions (e.g., bullish or bearish) behind words in plain text. Traditional methods mainly consider this task as a text classification problem with solutions based on machine learning algorithms. However, most of them rely heavily on the hand-crafted features extracted from the text. Indeed, a critical issue along this line is that the latent global and local contexts of the financial opinions usually cannot be fully captured. To this end, we propose a context-aware deep embedding network for financial text mining, named CADEN, by jointly encoding the global and local contextual information. Especially, we capture and include an attitude-aware user embedding to enhance the performance of our model. We validate our method with extensive experiments based on a real-world dataset and several state-of-the-art baselines for investor sentiment recognition. Our results show a consistently superior performance of our approach for identifying the financial opinions from texts of different formats.