Understanding the complex internal relationships of climate elements and revealing causal relationships in climate phenomena has attracted attention in the field of climate science and data mining. However, current causal discovery methods in climate science primarily rely on single-modal data, which limits the available information and complicates the accurate identification of climate causal relationships. In this paper, we propose a novel climate causal discovery method based on cross-modal Transformer and reinforcement learning, called CTRL. The method consists of three modules: unimodal feature representation, cross-modal feature fusion, and reinforcement learning for causal discovery. Specifically, we first design a novel unimodal feature representation method that uses the seasonal-trend decomposition based on LOESS (STL) strategy to decompose air quality and meteorological features into trend and seasonal components, with the aim of effectively capturing temporal data patterns. Next, we develop a cross-modal feature fusion method using cross-attention to capture cross-modal interactions and adaptively fuse bimodal features. Finally, we employ reinforcement learning to search for the optimal climate causal network with the best Bayesian information criterion (BIC) score. Numerous experiments show that CTRL outperforms singlemodal approaches in climate causal discovery, achieving even better results when predicting a target climate variable using only the causal features identified by CTRL. Code and data are available at https://anonymous.4open.science/r/CTRL-2622/.

Static functional brain network (sFBN) and dynamic functional brain network (dFBN) characterize brain func tional states from different scales, respectively. Integrating their complementary information has become an emerging direction for brain network classification. However, most existing methods often overlook the inherent variability within the two types of brain networks, thereby failing to fully exploit the complementary information during the process of static-dynamic interaction and fusion. To address this issue, we propose a static-dynamic variability-aware brain network classification method utilizing variability-modulated attention mechanisms to perform static-dynamic interaction and fusion, which consists of three main stages. In the first stage, based on the constructed sFBN and dFBN, we propose a variability calculation module that obtains node variability and graph variability by quantifying node activity and window fluctuations, respectively. In the second stage, follow ing node-level feature extraction, we propose a node-level static-dynamic feature interaction module that utilizes node variability to modulate the cross-attention weights, emphasizing complementary information in high-variability nodes. In the third stage, after graph-level feature extraction, we propose a graph-level static-dynamic feature fusion module that utilizes graph variability to modulate the self-attention weights, focusing on discrimi native information in high-variability windows. Experiments on ABIDE I and ABIDE II datasets demonstrate that the proposed method achieves superior classification accuracies of 74.68% and 73.98%, respectively.

Brain effective connectivity (EC) characterizes directional causal interactions among brain regions. However, learning stable and directionally explicit EC networks from multimodal data remains challenging. In practice, functional magnetic resonance imaging (fMRI) and electroencephalography (EEG) differ substantially in spatio-temporal resolution and noise characteristics. Existing methods often rely on manual spatio-temporal alignment and tend to recover only partial causal structures under high noise and Bayesian equivalence classes. To address these challenges, we propose a spatio-temporal constrained Bayesian causal network for multimodal brain effective connectivity learning (STCBN-EC). First, STCBN-EC constructs an anatomically guided EEG–fMRI spatial mapping and derives a unified spatio-temporal representation through slice-level alignment and adaptive modality fusion. Then, a Bayesian causal network is employed to model nonlinear inter-regional dependencies, where uncertainty-driven surrogate scoring is used to evaluate candidate structures. Finally, multimodal representation learning and EC structure estimation are jointly optimized via a gradient-free global optimization strategy. Experiments on simulated and real EEG–fMRI datasets demonstrate that STCBN-EC outperforms state-of-the-art methods and effectively captures state-dependent directional interactions among brain regions.

Graph neural networks have achieved tremendous success in graph classification tasks recently. The essential idea is to combine local aggregation and global readout operations to learn effective graph features so that the classification tasks can be easily launched. However, most current methods pay less attention to the subgraph structures which are usually the key indicators for discriminating the categories of graphs. In this paper, we aim to fully explore the potential in subgraphs and build a framework based on subgraph neural networks and multi-scale contrastive learning so that the patterns in subgraphs can be leveraged to formulate discriminative graph features. Beyond the conventional graph classification pipeline, we first propose a novel subgraph generation method that utilizes feature clustering and modularity to formulate subgraphs with high-order cluster-wise structures. Then, to leverage subgraph structures, on the one hand, we design subgraph neural networks to further extract node features with subgraph information. On the other hand, we propose a multi-scale contrastive learning framework that separately considers node-subgraph and graph-graph contrastive learning to improve the discriminative ability of graph features. Extensive experimental results on several benchmark datasets demonstrate the important roles of subgraphs and the effectiveness of the proposed method in graph classification.

With the progression of Neural Architecture Search (NAS), it has increasingly caught the attention of researchers from various domains. Up to now, researchers have developed numerous Demand-Driven NAS (DD-NAS) methods to alleviate difficulties of constructing deep neural network architectures across various practical applications. However, there lacks a review that summarizes the technical characteristics of NAS methods from a demand-driven perspective. Thus, in this paper, we provide a comprehensive survey of the latest advancements in DD-NAS. Specifically, first of all, we outline the fundamental process and core components of DD-NAS. Then, we summarize the unique characteristics of DD-NAS methods used in four main domains. Following this, we take representative DD-NAS methods based on electroencephalogram as an example to conduct empirical researches. Finally, we explore potential directions for the future development of DD-NAS. Note to Practitioners-This paper aims to provide a comprehensive survey of Neural Architecture Search (NAS) methods for interdisciplinary practitioners outside the artificial intelligence field. Specifically, our survey delves into practical applications and potential value of NAS methodologies across four critical inter-disciplinary domains: smart industry, intelligent transportation systems, remote sensing, and smart healthcare. While constrained by computational resources and article length limitations, our analysis primarily focuses on NAS implementations in these selected domains. Nevertheless, the proposed framework and summarized insights offer valuable References for professionals in various other vertical fields, including smart agriculture and environmental monitoring. Future work will expand the research scope to incorporate NAS applications in both emerging and traditional domains, thereby more comprehensively demonstrating its universality and practical utility.

Cell segmentation is a research hotspot in biology and clinical medicine for its role in assisting medical diagnosis. Although deep learning has been widely applied for this task, existing methods often use the same feature extraction mode for all regions. This ignores the similar features within regions and the density differences between them, easily leading to poor performance. To address these problems, this paper proposes a Cell Region-aware segmentation method (CellRa). Specifically, CellRa mainly includes three modules: regional feature extraction module, regional feature enhancement module, and regional feature fusion module. The regional feature extraction module employs a novel feature clustering attention mechanism to distinguish cell regions of varying densities and adaptively extract detailed cellular features. Subsequently, the regional feature enhancement module utilizes three parallel decoders to refine the extracted features from distinct perspectives. Connections among these decoders are established via shared separable attention, enabling comprehensive capture of similar cellular features within the same region. Finally, the regional feature fusion module incorporates a joint loss function weighting mechanism to balance the contributions from the different decoders. Experimental results on two public datasets demonstrate that CellRa achieves state-of-the-art performance, surpassing strong benchmark models. Specifically, CellRa achieves Dice similarity coefficients (Dice) of 91.58% and 97.14%, and Intersection over Union (IoU) scores of 84.47% and 94.44%. These results demonstrate significant and quantifiable improvements in segmentation accuracy.

Traffic prediction provides vital support for urban traffic management and has received extensive research interest. By virtue of the ability to effectively learn spatial and temporal dependencies from a global view, Transformers have achieved superior performance in long-term traffic prediction. However, existing methods usually underrate the complex spatio-temporal entanglement in long-range sequences. Compared with purely temporal entanglement, spatio-temporal data emphasizes the entangled dynamics under the restrictions of traffic networks, which brings additional difficulties. Moreover, the computational costs of spatio-temporal Transformers scale quadratically as the sequence length grows, limiting their applications on long-range and large-scale scenarios. To address these problems, we propose a decomposed spatio-temporal Mamba (DST-Mamba) for traffic prediction. We aim to apply temporal decomposition to the entangled sequences and obtain the seasonal and trend parts. Shifting from the temporal view to the spatial view, we leverage Mamba, a state space model with near-linear complexity, to capture seasonal variations in a node-centric manner. Meanwhile, multi-scale trend information is extracted and aggregated by simple linear layers. Such combination equips DST-Mamba with superior capability to model long-range spatio-temporal dependencies while remaining efficient compared with Transformers. Experimental results across five real-world datasets demonstrate that DST-Mamba can capture both local fluctuations and global trends within traffic patterns, achieving state-of-the-art performance with favorable efficiency.

The high-dimensional decision space of large-scale multi-objective optimization poses challenges to evolutionary algorithms, which leads to potential trapping in local optima. This paper proposes a brain information decomposition mechanism inspired large-scale multi-objective evolutionary algorithm (IDLMEA). Three strategies are developed. 1) The redundant information subpopulation division strategy selects high-quality solutions using shift-based density estimation, which determines the global optimal information. 2) The synergistic information subpopulation division strategy identifies exploration direction information by dividing solutions using local sensitive hashing. 3) The redundant and synergistic information-based reproduction strategy generates offspring for effective exploration of the search space. The Friedman test values of IDLMEA for inverted generational distance on LSMOP and IMF benchmarks outperform ten competitors by at least 25.64

Evolutionary multitasking optimization (EMTO) aims to make use of evolutionary algorithms to solve multiple optimization tasks simultaneously through exploiting the relevant information between different tasks. However, the existing EMTO algorithms (EMTOAs) usually begin the optimization process from scratch without considering the prior information between different tasks, which would restrict the solving performance. This paper proposes a pre-communication mechanism (PCM) for EMTO, which takes the distribution information of the initial population corresponding to each task as the prior information, and uses the correlation of the prior information to provide refined solutions for each task in the each generation of the early evolution. Firstly, after generating initial individual solutions, PCM constructs a Gaussian distribution model on the individual solutions of each task as the prior information. Next, in the each generation of the early evolution, PCM takes each task as the target task in turn and learns the similarity information between the target task and other tasks by constructing a Gaussian mixture model. Mixture coefficients represent the learned similarity information between the target task and other tasks. Finally, the individual solutions sampled from the obtained Gaussian mixture model of each task compete with original individual solutions of each generation of early evolution to obtain refined individual solutions. The experimental results show that PCM can help the existing EMTOAs to solve multitasking optimization problems more effectively and efficiently.

For the latest two years, relation classification-based surrogate assisted evolutionary algorithms show good potential for solving expensive multi-objective optimization problems (EMOPs). However, the existing studies are still at the initial stage and lack specific research on the dominance relation. This paper proposes a difference vector angle dominance relation for EMOPs, which uses an angle threshold phi to control the selection pressure and is called DVAD-phi. The proposed DVAD-phi has adaptive selection pressure and considers the convergence and diversity of solutions when picking out superior solutions, which makes it beneficial to pick out promising solutions for expensive real FEs and reduce expensive real FEs. To be specific, we firstly give the definition of DVAD-phi that measures the superiority from one solution to another solution according to the angle threshold phi. Then, we deduce that there is monotonicity between the angle threshold phi and the number of non-dominated solutions in the sense of DVAD-phi. At last, we propose an adaptive determination strategy of angle threshold based on bisection to set proper pressure for picking out promising solutions for expensive real FEs. Experiments have been conducted on 23 test functions from two benchmark sets and one real-world problem. The experimental results have verified the effectiveness of DVAD-phi.

Estimating brain effective connectivity (EC) from functional magnetic resonance imaging (fMRI) data can aid in comprehending the neural mechanisms underlying human behavior and cognition, providing a foundation for disease diagnosis. However, current spatiotemporal attention modules handle temporal and spatial attention separately, extracting temporal and spatial features either sequentially or in parallel. These approach overlooks the inherent spatiotemporal correlations present in real world fMRI data. Additionally, the presence of noise in fMRI data further limits the performance of existing methods. In this paper, we propose a novel brain effective connectivity estimation method based on Fourier spatiotemporal attention (FSTA-EC), which combines Fourier attention and spatiotemporal attention to simultaneously capture inter-series (spatial) dynamics and intra-series (temporal) dependencies from high-noise fMRI data. Specifically, Fourier attention is designed to convert the high-noise fMRI data to frequency domain, and map the denoised fMRI data back to physical domain, and spatiotemporal attention is crafted to simultaneously learn spatiotemporal dynamics. Furthermore, through a series of proofs, we demonstrate that incorporating learnable filters into fast Fourier transform and inverse fast Fourier transform processes is mathematically equivalent to performing cyclic convolution. The experimental results on simulated and real-resting-state fMRI datasets demonstrate that the proposed method exhibits superior performance when compared to state-of-the-art methods. The code is available at https://github.com/XiongWenXww/FSTA.

Brain effective connectivity (EC) is key to understanding causal neural interactions and brain organization. However, learning EC from single-modal brain data, such as functional magnetic resonance imaging (fMRI) or electroencephalography (EEG), is limited by the inability to simultaneously capture sparse temporal and spatial information. This paper proposes a novel multimodal sparse generative flow network (MSGFlowNet), which integrates fMRI and EEG data through an attention-guided encoder and employs a multi-head self-attention sparse Transformer to extract features from the fused data. These features are then processed by two output heads of the generative flow network: one computes state transition probabilities and updates the mask, while the other determines the probability of generating a termination state. Experiments on synthetic and real-world datasets demonstrate that MSGFlowNet significantly outperforms state-of-the-art methods.

The brain connectivity patterns, such as functional connectivity (FC) and effective connectivity (EC), describe the complex interaction patterns in brain networks. By combining FC and EC information, essential biomarkers for characterizing brain abnormalities can be further extracted. However, existing FC-EC fusion methods often process FC and EC networks independently, neglecting the intrinsic interrelationships between them. In addition, current FC-EC fusion methods rely on single segmentation atlas data, whereas multi-atlas data provides a more comprehensive display of brain information. To address these issues, we propose a novel multi atlas dual branch graph convolution functional effective connectivity attention fusion (MaFECAF) method. It comprises three modules: multi-atlas data fusion, functional and effective connectivity construction, and functional and effective connectivity attention fusion. First, two atlases are fused using the cosine similarity matrix as weights. Next, we use the fused data to construct the FC and EC networks via spatio-temporal graph convolutional network and variational autoencoder parameterized by graph neural network architecture, respectively. Subsequently, the FC and EC networks are merged via a dynamic graph convolutional network and an adaptive self-attention feature extraction unit, followed by the application of the obtained features to autism spectrum disorder (ASD) classification. The experimental results have demonstrated the effectiveness of the proposed method and the benefits of integrating FC and EC. The proposed method achieved a 74.7% classification accuracy for ASD and outperformed state-of-the-art methods.

Deep models are increasingly used to analyze brain graphs for the diagnosis and understanding of brain diseases. However, due to the multi-site data aggregation and individual differences, brain graph datasets exhibit widespread distribution shifts, which impair the model’s generalization ability to the test set, thereby limiting the performance of existing methods. To address these issues, we propose a Causally Invariance-aware Augmentation for brain Graph Contrastive Learning, called CIA-GCL. This method first generates a brain graph by extracting node features based on the topological structure. Then, a learnable brain invariant subgraph is identified based on a causal decoupling approach to capture the maximum label-related invariant information with invariant learning. Around this invariant subgraph, we design a novel invariance-aware augmentation strategy to generate meaningful augmented samples for graph contrast learning. Finally, the extracted invariant subgraph is utilized for brain disease classification, effectively mitigating distribution shifts while also identifying critical local graph structures, enhancing the model’s interpretability. Experiments on three real-world brain disease datasets demonstrate that our method achieves state-of-the-art performance, effectively generalizes to multi-site brain datasets, and provides certain interpretability.

Inferring causal protein signaling networks from human immune system cellular data is an important approach to reveal underlying tissue signaling biology and dysfunction in diseased cells. In recent years, reinforcement learning (RL) methods have shown excellent performance in the field of causal protein signaling network inference. However, the complexity of RL models and the need for manual hyperparameter tuning can hinder performance. In this paper, we propose a actor-critic RL model via artificial bee colony (ABC) neural architecture search, called ABCNAS-RL. Specifically, the entire method is divided into two phases: ABC neural architecture search and actor-critic RL search. In phase one, we represent each bee as a set of hyperparameter, utilizing the ABC algorithm searching for optimal hyperparameters of the actor-critic RL model on the training set. In phase two, we use the actor-critic RL model to infer the causal protein signaling network on the test set. The actor network consists of an encoder-decoder architecture, composed of a transformer and a bidirectional gated recurrent unit (BiGRU) with an integrated attention mechanism. The critic network consists of a fully connected neural network that estimates the output state of the actor network. By maximizing cumulative rewards, we ultimately derive the causal protein signaling network. Extensive experimental results on simulated and real datasets verify that ABCNAS-RL outperforms the comparison methods and has superior performance.

Functional Brain Network (FBN) classification methods based on Neural Architecture Search (NAS) have been increasingly emerging, with their core advantage being the ability to automatically construct high-quality network architectures. However, existing methods exhibit poor robustness when dealing with FBNs that have inherent high-noise characteristics. To address these issues, we propose a robust NAS with progressive-enhanced strategies for FBN classification. Specifically, this method adopts Particle Swarm Optimization as the search method, while treating candidate architectures as individuals, and proposes two progressive-enhanced (PE) strategies to optimize the critical stages of population sampling and fitness evaluation. In the population sampling stage, we first utilize Latin Hypercube Sampling to initialize a small-scale population, ensuring a broad search range. Subsequently, to reduce random fluctuations in searches, we propose a PE supplementary sampling strategy that identifies advantageous regions of the solution space, and performs precise supplementary sampling of the population. In the fitness evaluation stage, to enhance the noise resistance of the searched architectures, we propose a PE fitness evaluation strategy. This strategy first evaluates individual fitness separately using both original data and artificially constructed noise-augmented data, then combines the two fitness scores through a novel progressive formula to determine the final individual fitness. Experiments were conducted on two public datasets: the ABIDE I dataset (1,112 subjects, 17 sites), and ADHD-200 (776 subjects, 8 sites), using AAL/CC200 atlases. Results demonstrate that PE-RBNAS achieves state-of-the-art performance, with 72.61% accuracy on clean ABIDE I data (vs. 71.05% for MC-APSONAS) and 71.82% accuracy under 0.2 noise (vs. 68.15% for PSO-BNAS). The results indicate that, compared to other methods, the proposed method demonstrates better model performance and superior noise resistance.

The analysis of electrocardiogram (ECG) signals is of great significance in predicting arrhythmias and preventing cardiovascular diseases. However, ECG data has high noise and highly temporally correlated temporal characteristics, which challenges machine learning algorithms in classification tasks. This paper proposes an ECG signal classification method based on the bidirectional long short-term memory (BiLSTM) network and multi-task temporal attention, named BiLMTA, which combines a time series encoding module and a multi-task decoding module. Specifically, the first step of the method involves feeding the filtered ECG signals into the encoder module, in which BiLSTM is utilized to model temporal dependencies between preceding and succeeding time steps. Subsequently, the multi-task decoding module handles two key tasks: one is responsible for the main task, the classification task of ECG signals; the other serves as an auxiliary task to predict normal and abnormal heartbeats. To better capture the key information in ECG signals and improve sensitivity to abnormal patterns, a temporal attention mechanism is introduced. In addition, in order to promote the interactive learning between two tasks, a new joint loss function is proposed. This loss function comprehensively considers the cross entropy loss of the main task and auxiliary task to ensure that the two tasks can jointly optimize the model and improve the accuracy of classification results. The experimental results show that the BiLMTA method performs well on two publicly available datasets, surpassing traditional methods and state-of-the-art deep learning methods.

The automatic generation of brain CT reports has gained widespread attention, given its potential to assist radiologists in diagnosing cranial diseases. However, brain CT scans involve extensive medical entities, such as diverse anatomy regions and lesions, exhibiting highly inconsistent spatial patterns in 3D volumetric space. This leads to biased learning of medical entities in existing methods, resulting in repetitiveness and inaccuracy in generated reports. To this end, we propose a Medical Entity-balanced Prompting Network (MEPNet), which harnesses the large language model (LLM) to fairly interpret various entities for accurate brain CT report generation. By introducing the visual embedding and the learning status of medical entities as enriched clues, our method prompts the LLM to balance the learning of diverse entities, thereby enhancing reports with comprehensive findings. First, to extract visual embedding of entities, we propose Knowledge-driven Joint Attention to explore and distill entity patterns using both explicit and implicit medical knowledge. Then, a Learning Status Scorer is designed to evaluate the learning of entity visual embeddings, resulting in unique learning status for individual entities. Finally, these entity visual embeddings and status are elaborately integrated into multi-modal prompts, to guide the text generation of LLM. This process allows LLM to self-adapt the learning process for biased-fitted entities, thereby covering detailed findings in generated reports. We conduct experiments on two brain CT report generation benchmarks, showing the effectiveness in clinical accuracy and text coherence.

Solving dynamic multiobjective optimization problems (DMOPs) is extremely challenging due to the need to address multiple conflicting objectives that change over time. Transfer prediction- based strategies typically leverage solutions from historical environments to generate an initial population for a new environment. However, these strategies often overlook the similarity between the historical and new environments, which can negatively impact the quality of the initial population. To address this issue, we propose a similar environment transfer strategy. Firstly, we select Pareto-optimal solutions from a randomly generated population in the new environment to form a prior Pareto set (PS). The prior PS is expand by oversampling sparse solutions. Then, we apply the maximum mean discrepancy (MMD) to measure the discrepancy between the prior PS and the PS from each historical environment. The historical environment with the smallest MMD is identified as the similar environment. Finally, we use solutions from this similar environment to establish a kernelized easy transfer learning model, which is employed to predict the quality of random solutions in the new environment. The initial population is formed by combining excellent solutions predicted by the model with the prior PS. Experimental results demonstrate that the proposed strategy significantly outperforms several state-of-the-art strategies.

Multi-head attention (MA), which allows the model to jointly attend to crucial information from diverse representation subspaces through its heads, has yielded remarkable achievement in image captioning. However, there is no explicit mechanism to ensure MA attends to appropriate positions in diverse subspaces, resulting in overfocused attention for each head and redundancy between heads. In this paper, we propose a novel Intra- and Inter-Head Orthogonal Attention (I2OA) to efficiently improve MA in image captioning by introducing a concise orthogonal regularization to heads. Specifically, Intra-Head Orthogonal Attention enhances the attention learning of MA by introducing orthogonal constraint to each head, which decentralizes the object-centric attention to more comprehensive content-aware attention. Inter-Head Orthogonal Attention reduces the heads redundancy by applying orthogonal constraint between heads, which enlarges the diversity of representation subspaces and improves the representation ability for MA. Moreover, the proposed I2OA is flexible to combine with various multi-head attention based image captioning methods and improve the performances without increasing model complexity and parameters. Experiments on the MS COCO dataset demonstrate the effectiveness of the proposed model.