As the green economy continuously integrates into the digital financial system, the structure of credit risk in consumer finance has changed accordingly. Traditional models gradually exhibit limitations in characterizing complex feature interactions and nonlinear relationships. To tackle this problem, this study establishes an attention-enhanced multi-head self-attention deep factorization machine. The attention mechanism is introduced into low-order feature interactions, and the multi-head self-attention mechanism is fused with deep neural networks in high-order interaction parts to achieve multi-level joint feature modeling. Experiments are carried out using the Lending Club dataset. This study verifies that the proposed method has strong adaptability and stability in consumer finance credit risk identification under the background of digital economy.
This study applies a two-stage data envelopment analysis (DEA)-Tobit framework to evaluate logistics efficiency and its determinants in 12 provinces of Western China from 2014 to 2023. Logistics efficiency is assessed using DEA, which yields measures of comprehensive, pure technical, and scale efficiency. Subsequently, a Tobit regression model examines the causal mechanisms underlying efficiency variations. The results reveal a general upward trend in regional logistics efficiency, alongside persistent inter-provincial disparities. Infrastructure quality, industrial structure upgrading, information technology development, and the scale of specialized logistics personnel exhibit significant positive effects. Based on these findings, policy recommendations are proposed, focusing on transport network optimization, industrial transformation, logistics informatization, and talent cultivation to support high-quality logistics development in the region.
Handwritten mathematical expression recognition aims to recover symbol sequences and hierarchical relations from two-dimensional layouts. Existing methods often rely on linear sequence modeling or local structural composition, making it difficult to jointly capture the main chain and nested dependencies, leading to structural entanglement and unstable parsing. To address this issue, this paper proposes a structural re-indexing and nested decoding network. A structural re-indexing module separates the main chain and hierarchical relations on shared features and uses mutual constraints to model their dependency. An offset field rearranges features into a disentangled representation, reducing relational ambiguity. Based on this, a nested decoder progressively constructs expressions. The main chain provides the global framework, and local structures are expanded at key positions according to relation predictions. Experiments show that the structural re-indexing and nested decoding network achieves higher accuracy, robustness, and structural consistency, especially for complex expressions and long sequences.
This study explored whether digital trace data could effectively predict the process of low-carbon urbanization and promote data-driven decisions for sustainable urban development. The study integrated 12 types of digital traces, including mobile phone signaling, point of interest check-ins, traffic checkpoint records, shared bicycle trajectories, nighttime light remote sensing, and enterprise electricity consumption data. Meanwhile, it established a multidimensional indicator system based on population mobility, energy intensity, and spatial interaction. A hybrid model combining a spatiotemporal graph convolutional network and long short-term memory was proposed. The study demonstrated that digital traces could reveal carbon emission drivers that traditional methods failed to capture. Unstructured data, such as point of interest check-ins in social platforms and food delivery heatmaps, could serve as leading indicators for low-carbon governance. Therefore, this study opened new pathways for global urban low-carbon transitions based on digital footprints.
This study examined the influence of psychological capital on regional innovation. Based on a theoretical analysis of psychological capital, intrinsic motivation, and regional innovation, a model analysis of their mediating effect was carried out. It was found that psychological capital is significantly positively correlated with regional innovation, intrinsic motivation is significantly positively correlated with regional innovation, and intrinsic motivation plays a mediating role between psychological capital and regional innovation. The study distributed questionnaires to employees of multiple companies in different regions to collect data on employees' psychological capital and intrinsic motivation. The collected data were used for correlation analysis, regression analysis, and other methods to test the hypothesis. The empirical results show that the correlation coefficients between psychological capital and intrinsic motivation and the number of patents, taken as an indicator of regional innovation, are 0.743 and 0.705, respectively.
Ensuring teaching quality in higher education increasingly requires standardized, systemic, and intelligent methods aligned with advances in information and communication technologies. The authors propose an artificial intelligence-driven systemic evaluation and knowledge-management framework that operationalizes emerging information and communication technology-aligned practices in educational informatics. The framework combines a three-layer backpropagation neural network with a structured taxonomy that covers teaching design, classroom delivery, student engagement, assessment practices, and faculty development, unifying heterogeneous indicators into a single, interpretable score. A teaching knowledge base and organizational memory layer support interoperability, recursive feedback, and data-informed decisions by administrators. Using normalized data from 120 faculty members, the model achieves a strong correlation with expert-panel ratings (R2 = 0.89) and reduces mean absolute error by 18% against a principal-component-analysis baseline.
As large-scale electric vehicle clusters become increasingly integrated into grid-wide collaborative scheduling, the cross-domain flow of massive user data introduces serious privacy risks. To address these challenges, this study proposed a distributed data privacy protection framework tailored for multiscale collaborative optimization of electric vehicle clusters. The framework mitigated single-point failures and trust issues commonly found in centralized scheduling systems. The proposed approach combined hierarchical federated learning with adaptive differential privacy to establish a three-tier collaborative architecture—vehicle, station, and cloud. At the charging station level, local models were trained with perturbed gradients, where an adaptive noise injection mechanism enforced (ε,δ)-differential privacy. At the cloud level, a multitimescale optimization model was employed: in the day-ahead stage, the globally aggregated hierarchical-federated-learning model predicted the schedulable capacity of electric vehicle clusters.
Aiming at the difficult problem of transformer mechanical fault detection, this article aims to build an efficient and accurate identification model to ensure the stable operation of an electric power system. First, the principle of transformer mechanical fault and the characteristics of voiceprint are described in detail. On this basis, sparse component analysis is combined with deep learning technology to realize sparse component-driven deep voiceprint feature modeling. Specifically, extensive transformer voiceprint data are collected by building an experimental platform. Using sparse component analysis, the voiceprint features are accurately extracted, and fault identification is carried out with the help of a mixed deep learning model combining a convolutional neural network and long short-term memory network. After many experiments, the model effectively suppressed the overfitting phenomenon in the training stage.
Handwritten mathematical expression recognition converts two-dimensional formula images into structured LaTeX sequences. However, existing vision-to-sequence methods struggle to explicitly model complex spatial structures and symbol-level ambiguities. To address this issue, the authors propose the Structure-Role Refined Network (SR2-Net). The model incorporates explicit structural modeling in the encoder. The Explicit Structural Parsing Module (ESP) aggregates spatial features into symbol-level representations and captures geometric relationships among symbols, producing structure-enhanced features. The Semantic–Role Decoupling Module (SRD) further separates visual semantics from structural roles and integrates them through structure-aware fusion, improving consistency under complex layouts. Finally, a structure-aware decoder generates LaTeX sequences by jointly exploiting structural context and semantic constraints. Experiments show that SR2-Net improves recognition accuracy and structural consistency, especially for formulas with complex nesting and superscript–subscript structures.
This work aims to enhance the performance of sequence generation models for popular music recognition and recommendation: an optimized generalized regression neural network combined with a genetic algorithm (GRNN-GA) evaluated through a thorough training and testing process. By tuning key hyperparameters such as learning rate and number of iterations, the model's training performance significantly improved. Experimental results show that the GRNN-GA model achieved increases in accuracy, recall, and F1 score of greater than 25%, 27%, and 31%, respectively, resulting in an overall performance gain exceeding 29%. The proposed GRNN-GA optimization approach contributes to the development of more accurate and efficient sequence generation models in the music domain. The training and testing results collectively confirm the model's robustness and generalization capabilities, providing a valuable reference for future research in music recognition and recommendation systems.
In the context of smart tourism, modern theme parks operate as integrated information technology (IT) ecosystems. This study is the first to apply a structural equation model to examine the impact of self-congruity and functional congruity on tourists' loyalty within such an IT-enabled environment. The aim was to construct and test a conceptual framework integrating self-congruity, functional congruity, tourists' involvement, and tourists' loyalty in theme parks conceptualized as complex service systems. A sample of 370 tourists was collected via questionnaires from four major theme parks in Wuxi. Data were analyzed using SPSS 22 and AMOS 24, yielding five key findings: (1) self-congruity positively influences functional congruity; (2) functional congruity positively affects tourists' loyalty; (3) self-congruity does not directly influence tourists' loyalty; (4) tourists' involvement moderates the relationship between self-congruity and functional congruity; and (5) tourists' involvement does not moderate the relationship between functional congruity and tourists' loyalty. The study adopted a systems modeling approach using SEM and moderation analysis to delineate the psychological and behavioral pathways underlying the interaction between tourists and theme parks as IT-enabled systems. Implications for information systems design are twofold: first, incorporating self-congruity and functional congruity metrics into user-profiling algorithms can enhance personalization; second, system interfaces and customer relationship management (CRM) strategies should be tailored to users' involvement levels to foster sustained engagement and loyalty.
Achievement motivation theory has been applied across a wide range of academic contexts, and use of achievement motivation as a theoretical foundation for examining individual sport consumption behavior aimed at improving physical condition has received increasing attention. This study employed the preferred reporting items for systematic reviews and meta-analyses framework for systematic review and meta-analysis to synthesize research trends and developments in this area. The Web of Science database was used for data collection due to its strong reputation for scientific indexing and comprehensive coverage. An exhaustive keyword search was conducted, yielding 1,143 publications that met the selection criteria. For data processing and visualization, CiteSpace-based knowledge graphs were extensively applied. The identification of influential publications, key researchers, and leading institutions supporting the findings reflects the rapid growth of academic output on achievement motivation in sport consumption behavior in recent years and demonstrates the field's expanding global influence.
As an essential core component in mechanical equipment, rolling bearings have a straight impact on the reliability, safety, and operational efficiency of the equipment. In actual industrial scenarios, equipment often operates in a normal state, and the cost of obtaining and annotating fault data is enormous, which cannot meet the requirements of model training. Based on this, this study designed a self-supervised learning Transformer model (SSFormer) for bearing fault diagnosis under small sample conditions. The SSFormer effectively utilized a large number of unlabeled samples to mine deep representations of fault signals and potential dependencies between subsequences through self-supervised learning and constructed a pretrained model with good generalization ability. Subsequently, with only a small number of labeled samples for supervised finetuning, the SSFormer's ability to distinguish multiple types of bearing faults was significantly improved.
To investigate the educational impact of domestic animated films incorporating human-computer interaction technologies and artificial intelligence on adolescents' cognitive development, this study conducted an empirical analysis. A total of 120 students were selected as participants and divided into an experimental group and a control group for a 12-week instructional intervention. The experimental group engaged with domestically produced animations based on virtual reality technology, while the control group received conventional multimedia instruction. The study primarily examined the effects of domestic animations on adolescents' deductive, inductive, and analogical reasoning abilities. The findings demonstrate the effectiveness of high-tech-enabled domestic animations in promoting the development of adolescents' cognitive thinking and provide a reference for the development of related educational resources.
This study explores the application of the deep temporal model within the internet of things to enhance rural revitalization and development through digital finance, highlighting how digital finance addresses challenges in rural development. The integration of deep temporal models within the internet of things, the incorporation of digital finance into rural development strategies, and the framework for empowering rural development are outlined. The proposed model outperformed traditional models at predicting time series data indicators, with a root mean squared error of 0.012, a mean absolute error of 0.008, and an R2 score of 0.95. Additionally, the deep temporal model showed shorter training times (36 h) and fewer parameters (2.5 million) compared to eXtreme Gradient Boosting and random forest, emphasizing its advantages in performance and efficiency. In conclusion, the deep temporal model, with its strong support for practical implementation, offers significant benefits for empowering rural revitalization through digital finance.
This study proposed an intelligent classification and content recognition framework for enterprise financial archives based on a multimodal deep-learning architecture. The model integrated image, audio-video, and text modalities, using DenseNet-a 3D densely connected convolutional neural network with openSMILE software features-and a bidirectional-encoder-representations-fr om-transformers encoder. An attention-guided bidirectional gated-recurrent-unit fusion module modeled intramodal and intermodal dependencies, and a grid-structured classification head preserved spatial and temporal consistency. A class-aware readout with additive-margin softmax strengthened category separation and improved interpretability. Experiments on a dataset of 38,420 multimodal financial records showed that the proposed framework achieved an accuracy of 95.8%, an F1-score of 95.3%, and an area under the curve of 0.981, surpassing unimodal and transformer-based baselines. The results indicated that the multimodal attention-gated-recurrent-unit grid framework effectively captured cross-modal semantics and localized features, offering an efficient approach for financial document analysis and audit automation.
In the field of sports training, the hardware constraints of physical cables and the lack of robustness of visual estimation algorithms in complex backgrounds severely restrict the real-time performance and accuracy of motion monitoring. This study developed a lightweight high-resolution network algorithm optimized based on the coordinate attention mechanism. By encoding spatial location information into feature channels, it enhanced the sensitivity of key point localization in dynamic scenes. The core logic utilized depthwise separable convolution to reconstruct multiscale feature fusion branches and introduced a spatiotemporal constraint mechanism in the interaction stage. It also utilized biomechanical topological correlation to repair the skeletal logic of occluded parts. Combined with an efficient wireless communication protocol, an auxiliary system integrating data acquisition, pose calculation, and feedback output was constructed. Experimental results show that the system achieved a positioning accuracy of 94.8% of correct keypoints with a normalized distance threshold of 0.5. The number of parameters was compressed to 1.46M, the single-frame end-to-end inference latency was stable at 14 ms, and the data packet loss rate was less than 5% under complex wireless channel interference. This indicates that the system maintains high-precision attitude calculation while possessing excellent real-time performance and long-term monitoring stability.
To address the challenges of lacking real-time feedback and insufficient personalization in traditional art therapy, this study proposes a Smart Emotional Healing-Virtual Reality (SEH-VR) model comprising three modules: multimodal emotion recognition, personalized content generation, and adaptive interaction optimization. The study contributions are as follows: (1) A cross-modal fusion algorithm based on the transformer architecture with dynamic allocation of visual-speech attention weights improves emotion recognition accuracy to 0.92 in VR scenarios, reduces processing latency to 150 ms, and optimizes memory usage to 1500 MB. (2) An enhanced Style Generative Adversarial Network 3 (StyleGAN3) model achieves quality and personalization scores of 0.85 and 0.80, respectively, for generating art therapy content, while shortening generation time to 45 seconds. (3) The Context-Aware Particle Swarm Optimization (CA-PSO) algorithm optimizes system response time to 120 ms, stabilizes frame rate at 50 fps, and reduces system load to 55%.
This study constructs a model suitable for studying the performance of a Japanese speech recognition translation system based on a long short-term memory (LSTM) neural network. The impact of the matrix on the accuracy and recognition ratio of the speech recognition system is considered. This study provides a general introduction to the LSTM neural network, gives an overview of a Japanese speech recognition translation system, and presents experimental data to evaluate its performance. Then, the LSTM neural network is recommended for studying the Japanese speech translation recognition system, and a suitable model is established to conduct experiments on the system. By comparing example analyses, the experimental results show that the optimized PLSTM (Part-Aware Long Short-Term Memory) achieved an 8.09% improvement over an unoptimized LSTM network while maintaining recognition accuracy.
Handwritten mathematical expression recognition converts two-dimensional formula images into structured LaTeX sequences. However, existing vision-to-sequence methods struggle to explicitly model complex spatial structures and symbol-level ambiguities. To address this issue, the authors propose the Structure-Role Refined Network (SR2-Net). The model incorporates explicit structural modeling in the encoder. The Explicit Structural Parsing Module (ESP) aggregates spatial features into symbol-level representations and captures geometric relationships among symbols, producing structure-enhanced features. The Semantic-Role Decoupling Module (SRD) further separates visual semantics from structural roles and integrates them through structure-aware fusion, improving consistency under complex layouts. Finally, a structure-aware decoder generates LaTeX sequences by jointly exploiting structural context and semantic constraints. Experiments show that SR2-Net improves recognition accuracy and structural consistency, especially for formulas with complex nesting and superscript-subscript structures.
