Applied Intelligence最新文献

With the rapid advancement of deep learning, remote sensing image change detection (CD) has made significant progress. The complementary strengths of convolutional neural networks (CNNs) and Transformers have attracted considerable attention. This has prompted researchers to explore CNN-Transformer parallel architectures for CD tasks. However, existing methods often rely on spatial-domain operations, such as convolution and pooling, to integrate local and global features, which can result in the loss of fine-grained details, leading to incomplete detection of small changes and blurred boundaries in change regions. Additionally, many methods employ a single strategy for difference extraction, limiting their ability to model temporal dependencies and making them susceptible to irrelevant environmental variations, which can cause pseudo-changes. To address these challenges, we propose a Hybrid CNN-Transformer Network with Difference Enhancement and Frequency Fusion (HCTFNet). HCTFNet introduces a Temporal Feature Fusion Module (TFFM) that efficiently extracts difference features via a dual-branch operation, while integrating an attention mechanism to highlight actual changes and suppress irrelevant noise. Furthermore, the Hybrid CNN-Transformer Fusion (HCTF) module extracts both local and global features and applies frequency-domain processing to enhance the interaction between local and global features, thereby preserving fine-grained spatial details more effectively. Extensive experiments conducted on three publicly available benchmark datasets demonstrate that HCTFNet achieves superior CD performance compared to existing mainstream methods.

In this paper, we investigate optimization strategies for predicting high-mortality diseases in response to the challenges of unstable data quality and missing labels in healthcare big data. Firstly, a systematic review of the current state of research and application of machine learning algorithms for predicting heart disease, cardiovascular and cerebrovascular disease, lung cancer, diabetes, and breast cancer is presented. The core contributions are: 1) Semi-supervised self-training experiments were conducted on 14 datasets with labeling ratios of 30(%)-95(%), and the performance of classifiers such as SVM, KNN, LR, AB, DT, and RF was evaluated, and it was found that data balance is crucial for the accuracy improvement of most of the classifiers, and that the sensitivities to labeling ratios varied significantly between different classifiers, e.g., SVM robustness and KNN dependency. 2) In-depth analysis of feature importance identifies the key attributes of each dataset, whose absence significantly degrades the model performance, while RF exhibits optimal robustness due to integration properties. The study provides methodological references and empirical evidence for precision medicine prediction under limited labeling conditions.

Graph meta-learning models can fast adapt to new tasks with extremely limited labeled data by learning transferable meta knowledge and inductive bias on graph. Existing methods construct meta-training tasks with abundant labeled nodes from base classes, which limit the application scenarios of graph meta-learning. Therefore, we propose an unsupervised graph meta-learning framework via local subgraph augmentation (UMLGA). Specifically, we firstly propose a graph clustering-based sampling method to sample anchor nodes from different natural classes and extract corresponding local subgraphs. Then, supposing that the generated augmentation samples share the same labels, we design structure-wise and feature-wise graph augmentation strategies to generate diverse augmentation subgraphs while keeping the semantics unchanged. Finally, we perform meta-training on the unsupervised constructed tasks with weighted meta-loss, which can extract cross-tasks knowledge for fast adaption to novel classes. To evaluate the effectiveness of UMLGA, series of experiments are conducted on four real-world graph datasets. Experiment results show that, even without relying on extensive labeled data, UMLGA can achieve comparable and even better few-shot node classification performance comparing with the supervised graph meta-learning backbone models. With GPN as the backbone model, the improvements of UMLGA are respectively 3.0(sim )9.3%, 4.4(sim )11.6%, -1.2(sim )9.3%, and 1.8(sim )15.1% on Amazon-Clothing, Amazon-Electronics, DBLP, and ogbn-products datasets.

Multimodal biometric systems provide advantages over unimodal systems, such as improved accuracy, spoofing resistance, and broader population coverage. However, challenges related to privacy and template protection remain. To address these issues, we propose a integrated framework for multimodal biometric recognition with cancelable template protection, a privacy-preserving mechanism, using deep learning and an optimized bloom filter to significantly enhance recognition performance while ensuring robust security and privacy. First, a key mapping and management system is designed to generate secure keys that support the entire framework. Second, the Dynamic Attention and Hash Network (DAHNet) is employed to extract discriminative palmprint features through a hybrid attention mechanism and a deep hashing network. Third, a quantized fingerprint feature mapping technique is used to generate the corresponding binary fingerprint vector. Finally, the system applies an XOR operation to fuse DAHNet-extracted palmprint features and quantized fingerprint features, followed by an optimized bloom filter to generate secure cancelable templates, ensuring cancelability, irreversibility, and protection against template reconstruction. Experimental evaluations on the TJU and PolyU palmprint datasets, as well as the FVC2002 fingerprint dataset, demonstrate the outstanding accuracy of our state-of-the-art approach, achieving a remarkably low Equal Error Rate (EER). Comparative analysis further shows that the proposed multimodal system significantly outperforms unimodal systems in both recognition accuracy and security. Moreover, security analysis confirms that the framework satisfies all critical requirements for cancelable biometric template protection, including irreversibility, unlinkability, revocability, and robust privacy against various attack scenarios.

Time series forecasting has broad applications in fields such as finance, energy management, and weather prediction. PatchTST, a widely recognized model, has significantly improved the input length and accuracy of time series predictions. However, it uniformly treats all patches without considering their varying impacts on future predictions and relies heavily on global attention, often neglecting local information. To address these issues, we have developed an enhanced model: Dynamic Feature Weighting PatchTST (DynamicPatchTST). This model features three key enhancements: (i) a dynamic feature weighting strategy that assigns weights to each patch based on the characteristics of individual time series, emphasizing more impactful patches; (ii) a linear embedding method to replace positional encoding, preventing the over-adjustment or redundant modification of patch weights; (iii) a dual-pathway strategy that integrates local and global information, with the local pathway refined through gated adaptive adjustments to enhance the model’s ability to capture local details. Extensive experiments across multiple standard time series datasets (e.g., ETT, Exchanges, Weather) have demonstrated that our DynamicPatchTST significantly outperforms existing state-of-the-art models, with an average improvement of 4.4% in Mean Squared Error (MSE). This work provides a novel perspective on time series forecasting and paves the way for future advancements in the field.

Due to the complementary strengths of language models and knowledge graphs, question-answering systems that combine both techniques have emerged. However, existing models suffer from issues such as introducing excessive misleading information and an unreasonable distribution of heterogeneous representations, which negatively affect answer accuracy and reasoning efficiency. To address these problems, we propose a question-answering model based on split-hop information propagation and heterogeneous representation alignment. First, we train a graph attention network using the split-hop information propagation method along multi-hop reasoning paths, enhancing the relevance between extracted sub-graphs and the question context while allowing the network to learn differences across reasoning hops. Next, we align the heterogeneous representations of text and knowledge graph through post-training, ensuring the pre-trained representations fall into a reasonable parameter distribution. We validate the effectiveness and generalizability of our model on the CommonsenseQA and OpenBookQA datasets in the commonsense domain, as well as the MedQA-UMSLE dataset in the biomedical domain, achieving accuracy improvements of 0.5%, 0.6%, and 1.2% over baseline models of the same type, respectively.

Diabetic retinopathy (DR) is one of the most common ocular complications in diabetic patients. Therefore, early DR screening is crucial for preventing disease deterioration and timely diagnosis. However, the challenge of DR grading arose due to the tiny and unevenly distributed lesions and the complex pathological relationships, which are difficult to capture. This paper proposes the pathological relationship feature perception and dual attention guided network (PRANet) for DR grading. Firstly, a pathological relationship feature perception module (PRFM) is introduced to capture the pathological relationships between different types of lesions. A lesion detection network is used to locate the rough regions of the lesions, and K-Means clustering is performed on the lesion features to generate lesion nodes. The co-occurrence relationships between lesion nodes are used to construct an adjacency matrix, which is then input into a graph convolutional network to explore the complex relationships among the lesion nodes. Moreover, the EfficientNetV2-M backbone is employed to obtain global information of DR images and input them into a dual attention module (DAM). The DAM utilizes spatial and channel attention to emphasize the features of suspicious lesion regions. Pathological relationship features, channel features and spatial features are integrated to achieve precise DR grading. Extensive experiments were carried out on the DDR, APTOS2019 and FGADR datasets. Results show that our method exceeds most existing approaches in classification accuracy and robustness, yielding superior classification performance.

Personalized dialogue systems represent an innovative application in the field of conversational AI, aiming to endow chatbots with distinct personas to address the lack of individuality and specificity in traditional human-computer inter- actions.Current approaches often fail to incorporate rich external knowledge, making it difficult to maintain coherence and depth in long-term personal- ized interactions.Thus, to bridge the gap between static persona design and dynamic, knowledge-enhanced personalized dialogue generation, we did the fol- lowing work:1) We propose a novel Knowledge-expanded Personalized Dialogue Generation (KPDG) model to extend predefined personas using a commonsense knowledge graph of related personas. During the decoding of generated responses, this method adaptively integrates the most relevant personas that are optimally selected and partitioned with the dialogue history. 2) We design a two-stage prompting approach that leverages large language models (LLMs) for personal- ized dialogue generation. In the first stage, LLMs are used to enhance and expand the persona, enriching the persona’s intrinsic characteristics and emotional state. All in all, the main contributions of this work are as follows: (a) identification of the key limitations of current persona-based dialogue systems and formulation of a knowledge-enhanced framework to address them; (b) a novel persona expan- sion approach that combines structured common sense knowledge with adaptive selection; and (c) a two-stage LLM prompting paradigm that achieves state- of-the-art results without fine-tuning. Experiments on the PERSONA-CHAT dataset demonstrate that our approach outperforms strong baselines in both. 1. automatic metrics and human evaluations, validating its effectiveness in enhanc- ing persona diversity, contextual consistency, and conversational engagement. The novel contribution of this work is that we propose the KPDG framework, which first employs a knowledge graph–driven persona expansion module to enrich persona attributes and adaptively select context-relevant traits during decoding. Then, a two-stage LLM prompting strategy is applied: the first stage enhances and diversifies persona characteristics, while the second stage gener- ates responses via ICL without model retraining. This design not only addresses persona sparsity and consistency issues but also provides a scalable solution adaptable to different dialogue settings.

In real decision-making scenarios involving multi-agent games, such as intelligent unmanned systems, military confrontations, and autonomous navigation, the coupling of participant strategies and the incompleteness of perceived information make the accurate inference of dynamic game trajectories a critical and challenging task. To address this problem, this paper proposes a game trajectory modeling method that integrates a cross-attention mechanism with a moving diffusion process, termed CARM-Diff (Cross-Attention Autoregressive Moving Diffusion Model). The model combines autoregressive structures with cross-attention to capture the temporal evolution of sequences while explicitly modeling strategic interactions between agents. We design a lightweight feature extraction module and, leveraging the Markov property of diffusion models, introduce a deterministic evolution process of historical states to simulate noise, thereby enhancing the model’s capability to learn local temporal patterns. Meanwhile, a cross-attention mechanism is introduced in the reverse diffusion stage to guide the model in focusing on the opponent’s historical sequential behavior, enabling more precise capture of inter-agent influences. Furthermore, we design a residual gated trajectory modeling structure that fuses the agent’s own behavioral evolution with interaction effects induced by opponents. Gating factors are dynamically generated through multilayer perceptrons to achieve adaptive information fusion. We construct a dynamic trajectory dataset based on an underwater pursuit-evasion game to validate our approach, and the proposed CARM Diff framework is generalizable to a wide range of multi-agent interactive systems. Experimental results show that CARM-Diff outperforms mainstream baseline methods in both prediction accuracy and dynamic interaction modeling, demonstrating the effectiveness and practical potential of the proposed model.