Applied Intelligence最新文献

Illicit financial activities in cryptocurrency transaction networks are becoming increasingly sophisticated, often involving multi-hop and high-order transaction paths that obscure their origins and complicate detection. Traditional rule-based or machine learning methods typically rely on static features and limited local structures, making them insufficient for uncovering deeply embedded anomalous behaviors. Graph Neural Networks (GNNs) have recently shown promise in modeling such relational data; however, most existing GNN-based approaches struggle to effectively capture high-order dependencies and often lack interpretability, which is critical in financial security applications. To address these challenges, this study proposes an explainable graph neural network framework for illicit transaction detection. The framework is built upon the Topology Adaptive Graph Convolutional Network (TAGCN), which allows flexible integration of higher-order neighborhood information to capture complex propagation patterns within transaction graphs. We have evaluated the model using the publicly available Elliptic dataset. Experimental results demonstrate that our method achieves an accuracy of 98.14%, a recall of 86.22%, a precision of 94.23%, an F1-score of 90.05%, and a Matthews correlation coefficient (MCC) of 0.8913, outperforming several baseline models. Furthermore, SHapley Additive exPlanations (SHAP) are employed to provide post hoc interpretability, offering transparent insights into model predictions and enhancing trustworthiness for regulatory decision-making. The proposed framework not only significantly improves detection performance but also enhances model transparency through interpretability, providing important theoretical value and practical potential for anti-money laundering and financial risk management applications.

Accurate prediction of corn futures prices is critical for global food security and agricultural risk management, yet existing deep learning models struggle to fully exploit residual patterns and integrate macroeconomic uncertainty factors effectively. We propose a three-stage hierarchical ensemble framework that synergizes Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) networks for complementary temporal modeling, followed by Extreme Gradient Boosting (XGBoost) for residual correction with multi-source feature fusion. Hyperparameters were systematically optimized via Optuna framework. Empirical evaluation on 1,196 trading days of Chinese corn futures data (March 2018 to February 2023) demonstrates that our approach achieves R² = 0.9286, representing a 2.69% improvement over simple averaging (R² = 0.9018) and 1.28% over the best single model (R² = 0.9167). Statistical significance is confirmed via Diebold-Mariano tests (p < 0.001), with RMSE reduced by 14.77%. Ablation studies reveal that macroeconomic factors, particularly interest rates (44.94%) and geopolitical risk (15.15%), contribute over 60% to predictive power. This framework provides a generalizable methodology for financial time series forecasting and demonstrates the critical role of uncertainty quantification in ensemble learning.

With the increasing prevalence of incomplete multi-view data in practical situations, incomplete multi-view clustering (IMVC) has gained prominence as an essential methodology for processing incomplete and unlabeled multi-view data in real-world scenarios. Nevertheless, some irrelevant and insignificant data leads to sub-optimal solutions when directly modeling the raw data. And previous methods failed to fully utilize cross-view interactions, neglecting the consistency and heterogeneity among different views. Furthermore, the previous methods ignored multi-level data structures, thus limiting the exploration of hierarchical relationships. In response to the mentioned limitations, a novel method HSIMTC is proposed, which introduces a hierarchical tensor designed to harness multi-level data structures and intricate high-order correlations across multiple layers. Specifically, an orthogonal projection operator is optimized to filter out irrelevant content and extract the most critical and distinguishing features from the raw data for further analysis. In addition, dynamic cross-view cooperation facilitates information interaction and sufficiently explores consistency from different views. Meanwhile, a hierarchical tensor is first proposed which evolves bottom-up, unlike conventional flat modeling. Each layer not only consolidates internal structures from the other layers but also improves layer-to-layer interactions, enabling a deeper insight into multi-level data. Additionally, our method adopts an information simplification strategy to eliminate extraneous data and deeply analyzes the angular information of its principal direction, resulting in a more discriminative affinity matrix for spectral clustering. Experimental results on various benchmark datasets verify that HSIMTC achieves superior outcomes compared to current advanced algorithms.

Object detection in military operations faces critical challenges, including camouflaged targets and occlusion, where traditional RGB-based systems often fail. This study proposes a systematic framework for optimizing multimodal late fusion in object detection by integrating color space transformations with depth information. We contribute three key elements: (1) the development of the “militar-VALID” dataset, a specialized, defense-oriented collection of 6,054 images curated for challenging detection scenarios; (2) a comprehensive statistical evaluation framework comparing four late-fusion algorithms across 247 unique configurations; and (3) the hyperparameter optimization of the optimal fusion configuration through Bayesian search. Leveraging the YOLOv8-small architecture trained on eight parallel color representations (RGB, BGR, Grayscale, HSV, CIELab, YUV, YCrCb, and Depth), we establish that Weighted Boxes Fusion combining RGB, Depth, and HSV modalities delivers statistically significant improvements ((p le 0.001), Cliff’s (delta ge 0.8)). Specifically, this configuration achieves a 1.30% increase in mean Average Precision (mAP@50-95), a 3.20% improvement in precision, and a 3.22% enhancement in Average Precision (AP@50) for small objects compared to RGB-only baselines. Statistical analysis highlights the depth maps as the most impactful modality and HSV as the optimal color space complement. This work provides a quantitative framework for multimodal color space fusion in military object detection and discusses its implications for deployment in high-stakes operational environments.

Flooding is a devastating natural disaster that often causes inestimable human and economic loss. With historical data, flood forecasting with machine learning methods is considered an effective way to evaluate the potential risks in the concerned region. In this work, the recursive feature elimination (RFE) using SHapley Additive exPlanations (SHAP) is employed to select informative features, which iteratively evaluates feature importance and eliminates the least important features. With the least significant feature discarded, a one-dimension-reduced feature vector is formed for the next iteration, the proposed machine learning model is trained with the input samples of updated features. After this recursive feature elimination procedure, the selected features are finalized by finding the iteration where the model’s trained performance is optimal or dramatically drops. With the proposed recursive feature elimination, the training time of the conventional machine learning models, such as XGB, RF, and CatBoost, can be reduced by 17.8%, 56.0%, 30.6% while keeping the loss of prediction accuracy below 1%, respectively. Furthermore, a hybrid machine learning model integrating the conventional Catboost model and Random Forest model is proposed to evaluate flood susceptibility in flood-prone areas of Malawi, where numerical experiments showed its superiority over the individual models.

Performing node classification in text-attributed graphs (TAGs) has become a critical area of research. While two-stage methods effectively address scalability by decoupling feature extraction from GNN training, they often suffer from a “structural gap” where global topology is ignored during text encoding, and local aggregation leads to feature oversmoothing. To bridge this gap, we propose DCD-CAWA, a framework that integrates community structures into both feature learning and message passing. Specifically, in the first stage, we employ Decoupled Community Detection (DCD) to generate structural priors. These are integrated with text attributes through an auxiliary multi-task fine-tuning strategy, forcing the language model to capture the correlations between semantic content and community membership. In the second stage, we introduce the Community-Aware Weighted Aggregation (CAWA) module. The workflow of CAWA involves computing dynamic attention weights between nodes and their respective community prototypes, allowing the model to adaptively refine node representations by emphasizing global community contexts. This approach not only eliminates discrepancies between static topology and dynamic features but also effectively mitigates oversmoothing by preserving community-level distinctiveness. Experimental results across multiple datasets demonstrate that DCD-CAWA significantly outperforms state-of-the-art baselines.

Salient object detection (SOD) aims to imitate the human visual system (HVS) to accurately locate and segment salient objects in the scenes. In particular, it has been very successful detection results in fully supervised salient object detection (FSOD) methods, but they rely on great labor costs to provide pixel-by-pixel labels. The unsupervised salient object detection (USOD) methods do not utilize artificial labels to detect the salient objects in the scenes to avoid the drawback of FSOD. The initial pseudo-labels required by current deep learning-based USOD inevitably introduce some non-salient noise and ignore some fine-grained information of salient objects. In this paper, we propose EFNet, a novel two-stage USOD method. Its core lies in an innovative pseudo-label initialization stage. This stage employs an Enhanced Activation Strategy (EAS) based on a hybrid attention mechanism to increase the weight of salient regions, reduce the weight of non-salient regions, and guide the model to efficiently extract salient knowledge from shallow to deep layers. Meanwhile, a Fine-grained Information Auxiliary Strategy (FIAS) based on upper and lower branches is introduced to capture the boundary and texture information of salient objects, thereby generating cleaner and more accurate pseudo-labels for training in the second stage. Extensive experiments conducted on five benchmark datasets prove that our method achieves state-of-the-art USOD performance.

Multiple object tracking in dense scenes presents a significant challenge because of mutual occlusion and redundant detections, which can result in feature loss and cumulative error. To address these challenges, joint-detection-and-tracking frameworks based on the bipartite graph model have emerged as a popular paradigm. However, the bipartite graph model is often constrained by König’s minimax theorem, which postulates the need for one-to-one relationships in matching and equates the maximum number of matches with the minimum number of points covered. This restriction easily causes tracking failure for frequently appearing similarity targets. To overcome these limitations, this paper proposes a hypergraph random field model that uses domain hypernodes and trajectory hypernodes to differentiate targets on the basis of domain space features and trajectory fragments, respectively. This approach avoids cumulative error and enables one-to-many relationships for efficient tracking. Additionally, this paper presents an approximate solution that reduces the original complexity from (O(n^2)) to O(n). The experimental results on MOTChallenge demonstrate competitive performance, with a 1-2% improvement in MOTA compared with the baseline, particularly on MOT20, by almost 2%.

Data-driven machine learning models have increasingly been applied to survival analysis in recent years. However, these models require sufficient training samples, which is often impractical due to privacy, security, and legal constraints. As a result, survival data is typically distributed across multiple institutions and cannot be directly aggregated. In this study, we propose DA-FedSurX, a federated survival analysis framework that integrates gradient boosted survival trees (GBST) with a data augmentation strategy to effectively analyze distributed and private survival data. Unlike previous deep learning-based federated survival models, DA-FedSurX employs histogram-based GBST, which improves computational efficiency while maintaining strong interpretability. Furthermore, our data augmentation component enhances model robustness under Non-IID data distributions, a common challenge in federated learning. Extensive experiments on both simulated and real-world survival datasets demonstrate that DA-FedSurX significantly outperforms state-of-the-art deep survival models, including DeepSurv, DeepHit, CoxCC, and NnetSurvival, in terms of predictive accuracy. These results confirm the potential of DA-FedSurX as an effective and interpretable solution for federated survival analysis.

Label distribution learning is a popular learning paradigm for dealing with label polysemy scenarios, providing rich semantic information and being applied in many practical tasks. However, annotating training label distributions is challenging, which inevitably introduces uncertainty, noise, and bias into the label distribution. To mitigate this issue, this paper proposes a label distribution learning method based on the estimation of label distribution robust intervals. Specifically, we first establish robust interval estimation methods for the label distribution, which captures the uncertainty in the label distribution. Subsequently, we propose a loss function and a corresponding label distribution learning algorithm, which measure the training error of the robust interval and predict the label distribution. Extensive experiments are conducted to validate the effectiveness of the proposed algorithm. The results demonstrate that the algorithm presented in this paper statistically outperforms comparison algorithms and achieves optimal performance in (varvec{90.63%}) of cases.