Yong Xu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Yong Xu has a total of 44 co-authors. Here are the top co-authors by number of shared publications:
- Xin He
School of Computer and Information, Anhui Normal University, Wuhu, China
6 shared publications - Jinde Cao
School of Mathematics, Southeast University, Nanjing 210096, People’s Republic of China
5 shared publications Qun Fang
School of Computer and Information, Anhui Normal University, Wuhu, China
4 shared publicationsCheng Li
Anhui Normal University
3 shared publications- Hongchen Tang
School of Computer and Information, Anhui Normal University, Wuhu, China
3 shared publications - Qun Fang
Anhui Normal University
3 shared publications
Latest Publications
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (19) , 44999-45017
A Multimodal Benchmark for Open-Pit Mine Remote Sensing Change Detection With Vision–Language Guidance
- 1 October 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. 13 (19) , 44999-45017
A Multimodal Benchmark for Open-Pit Mine Remote Sensing Change Detection With Vision–Language Guidance
- Open Access
- 4 December 2025
- Published by Springer Nature in Research Square
Behavior-Aware Aggregation and Graph Attention Fusion Network for Multi-Behavior Recommendation
In real world recommender systems, users interact with the same item through multiple behaviors—click, favorite, add-to-cart, and purchase that are heterogeneous and time dependent. Existing multi-behavior approaches suffer from three gaps: (i) static behavior categorization that ignores how frequency, intensity, and anomalies dynamically affect preference; (ii) shallow or fixed graph propagation that misses higher-order preferences and cross-graph semantics; and (iii) weak handling of temporal dependencies, which obscures interest evolution. We propose the Behavior-Aware Aggregation and Graph-Attention Fusion Network (BAGAR) to jointly model behavior differences, temporal dynamics, and semantic relations. BAGAR includes: a behavior-aware aggregation module combining behavior-intensity pooling, core-preference pooling, and anomaly filtering, fused by an adaptive graph multilayer perceptron; a multi-hop graph-attention module that forms hierarchical channels for first-order direct interests, higher-order latent preferences, and global semantic paths, reconciled by gating and a joint behavior-time dependency mechanism that embeds behavior relations and inter-event time gaps within a unified attention framework to track preference evolution. Experiments on Taobao, Tmall, and Yelp show that BAGAR outperforms strong baselines on hit rate and normalized discounted cumulative gain, validating its effectiveness for complex behavior modeling.
- 4 December 2025
- Published by Springer Nature in Research Square
Open AccessBehavior-Aware Aggregation and Graph Attention Fusion Network for Multi-Behavior Recommendation
In real world recommender systems, users interact with the same item through multiple behaviors—click, favorite, add-to-cart, and purchase that are heterogeneous and time dependent. Existing multi-behavior approaches suffer from three gaps: (i) static behavior categorization that ignores how frequency, intensity, and anomalies dynamically affect preference; (ii) shallow or fixed graph propagation that misses higher-order preferences and cross-graph semantics; and (iii) weak handling of temporal dependencies, which obscures interest evolution. We propose the Behavior-Aware Aggregation and Graph-Attention Fusion Network (BAGAR) to jointly model behavior differences, temporal dynamics, and semantic relations. BAGAR includes: a behavior-aware aggregation module combining behavior-intensity pooling, core-preference pooling, and anomaly filtering, fused by an adaptive graph multilayer perceptron; a multi-hop graph-attention module that forms hierarchical channels for first-order direct interests, higher-order latent preferences, and global semantic paths, reconciled by gating and a joint behavior-time dependency mechanism that embeds behavior relations and inter-event time gaps within a unified attention framework to track preference evolution. Experiments on Taobao, Tmall, and Yelp show that BAGAR outperforms strong baselines on hit rate and normalized discounted cumulative gain, validating its effectiveness for complex behavior modeling.
- 28 November 2025
- Published by Springer Nature in International Journal of Dynamics and Control
- Vol. 14 (1) , 15
Dynamics of an Aedes aegypti model on an asymptotically periodically evolving domain
- 28 November 2025
- Published by Springer Nature in International Journal of Dynamics and Control
- Vol. 14 (1) , 15
Dynamics of an Aedes aegypti model on an asymptotically periodically evolving domain
- 18 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Aerospace and Electronic Systems
- Vol. 62 (00189251) , 2188-2200
ST-Gesture: Spatio-Temporal Feature Extraction Network for Gesture Recognition Based on Attention Mechanism
Gesture recognition, a crucial technology in human-computer interaction, finds applications across various domains, including smart homes, automotive driver assistance systems, and more. However, existing millimeter-wave radar-based gesture recognition algorithms often face challenges in achieving high accuracy and robustness in complex environments. To address these limitations, this paper proposes a Space-Time Gesture Recognition algorithm (ST-Gesture) that leverages attention mechanisms. The approach first exploits the multi-dimensional information within mmWave radar data cubes. ST-Gesture processes the Range-Azimuth-Doppler (RAD) data to generate three types of spectrograms: Range-Azimuth Map (RAM), Azimuth-Doppler Map (ADM), and Range-Doppler Map (RDM). Spatial and temporal features are then extracted from these spectrograms using an enhanced model that combines Convolutional Neural Networks (CNNs) with Temporal Convolutional Networks (TCNs). A cross-attention mechanism is employed to fuse complementary information from the different spectrograms, thereby improving recognition accuracy. Experimental results, based on both a self-constructed dataset encompassing diverse environments and positions, as well as public dataset, demonstrate that ST-Gesture outperforms state-of-the-art methods, achieving up to 99.2% accuracy. Additionally, the system's robustness to environmental and positional variations is thoroughly validated.
- 18 November 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Aerospace and Electronic Systems
- Vol. 62 (00189251) , 2188-2200
ST-Gesture: Spatio-Temporal Feature Extraction Network for Gesture Recognition Based on Attention Mechanism
Gesture recognition, a crucial technology in human-computer interaction, finds applications across various domains, including smart homes, automotive driver assistance systems, and more. However, existing millimeter-wave radar-based gesture recognition algorithms often face challenges in achieving high accuracy and robustness in complex environments. To address these limitations, this paper proposes a Space-Time Gesture Recognition algorithm (ST-Gesture) that leverages attention mechanisms. The approach first exploits the multi-dimensional information within mmWave radar data cubes. ST-Gesture processes the Range-Azimuth-Doppler (RAD) data to generate three types of spectrograms: Range-Azimuth Map (RAM), Azimuth-Doppler Map (ADM), and Range-Doppler Map (RDM). Spatial and temporal features are then extracted from these spectrograms using an enhanced model that combines Convolutional Neural Networks (CNNs) with Temporal Convolutional Networks (TCNs). A cross-attention mechanism is employed to fuse complementary information from the different spectrograms, thereby improving recognition accuracy. Experimental results, based on both a self-constructed dataset encompassing diverse environments and positions, as well as public dataset, demonstrate that ST-Gesture outperforms state-of-the-art methods, achieving up to 99.2% accuracy. Additionally, the system's robustness to environmental and positional variations is thoroughly validated.
- 14 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Big Data
- Vol. 12 (2) , 362-375
Long-Tail Inductive Feature Transfer for Knowledge-Aware Multi-Behavior Recommendation
Multi-Behavior recommendation captures fine- grained user intents by jointly modeling multiple interaction behaviors, significantly improving recommendation accuracy and diversity. However, existing methods often overlook complex semantic relations between items and entities. Moreover, constructing separate subgraphs for different interaction types frequently results in sparse graph structures, which exacerbates the long-tail problem. To address these challenges, we propose Long-tail Inductive Feature Transfer (LIFT), a method for knowledge-aware multi-behavior recommendation. To mitigate uneven node distribution in graph structures, we introduce a knowledge transfer-based feature reconstruction mechanism. Specifically, we first drop a portion of the neighbors of head nodes to construct proxy representations for tail nodes, training a reconstructor on the tail proxy to reconstruct the original head node features. The trained reconstructor is then used to backfill missing neighbor information for tail nodes, thereby achieving a more balanced feature distribution across nodes. Furthermore, we integrate multi-behavior and semantic contrastive learning to jointly optimize the representations. Extensive experiments on four datasets demonstrate that LIFT outperforms state-of-the-art methods, with further analysis validating its uniformity in representation learning.
- 14 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Big Data
- Vol. 12 (2) , 362-375
Long-Tail Inductive Feature Transfer for Knowledge-Aware Multi-Behavior Recommendation
Multi-Behavior recommendation captures fine- grained user intents by jointly modeling multiple interaction behaviors, significantly improving recommendation accuracy and diversity. However, existing methods often overlook complex semantic relations between items and entities. Moreover, constructing separate subgraphs for different interaction types frequently results in sparse graph structures, which exacerbates the long-tail problem. To address these challenges, we propose Long-tail Inductive Feature Transfer (LIFT), a method for knowledge-aware multi-behavior recommendation. To mitigate uneven node distribution in graph structures, we introduce a knowledge transfer-based feature reconstruction mechanism. Specifically, we first drop a portion of the neighbors of head nodes to construct proxy representations for tail nodes, training a reconstructor on the tail proxy to reconstruct the original head node features. The trained reconstructor is then used to backfill missing neighbor information for tail nodes, thereby achieving a more balanced feature distribution across nodes. Furthermore, we integrate multi-behavior and semantic contrastive learning to jointly optimize the representations. Extensive experiments on four datasets demonstrate that LIFT outperforms state-of-the-art methods, with further analysis validating its uniformity in representation learning.
- 6 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Social Systems
- Vol. 13 (2) , 1960-1975
Motif-Guided Multiview Contrastive Learning for Knowledge Graph-Enhanced Recommendation
Contrastive learning (CL) has demonstrated exceptional capability in extracting supervised signals and mitigating noise, increasingly attracting interest for its application in knowledge graph-enhanced (KG) recommendation. Nevertheless, existing approaches predominantly rely on a single augmentation strategy to construct contrastive views, often failing to effectively balance feature perturbation with semantic preservation. To overcome this limitation, we propose motif-guided multiview contrastive learning (MMCL), a unified framework for contrastive augmentation. MMCL leverages diverse view generation strategies across the user–item bipartite graph and the KG, perturbing graph structures while preserving essential semantics to the greatest extent possible. Specifically, for collaborative signals, we devise a data augmentation mechanism to model the interaction dynamics between users and items. For semantic information, we introduce a novel graph augmentation technique by constructing an interactive motif KG, enabling semantic contrastive learning within localized views. Furthermore, joint enhancement of graph and interaction data allows the model to capture global topological features of nodes. MMCL integrates contrastive learning at both local and global levels, effectively embedding local semantics and global topology into the learned representations. Extensive experiments conducted on three publicly available datasets reveal that MMCL outperforms advanced methods, particularly in scenarios with sparse interactions and noisy KG, while significantly alleviating popularity bias in recommendation.
- 6 October 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Social Systems
- Vol. 13 (2) , 1960-1975
Motif-Guided Multiview Contrastive Learning for Knowledge Graph-Enhanced Recommendation
Contrastive learning (CL) has demonstrated exceptional capability in extracting supervised signals and mitigating noise, increasingly attracting interest for its application in knowledge graph-enhanced (KG) recommendation. Nevertheless, existing approaches predominantly rely on a single augmentation strategy to construct contrastive views, often failing to effectively balance feature perturbation with semantic preservation. To overcome this limitation, we propose motif-guided multiview contrastive learning (MMCL), a unified framework for contrastive augmentation. MMCL leverages diverse view generation strategies across the user–item bipartite graph and the KG, perturbing graph structures while preserving essential semantics to the greatest extent possible. Specifically, for collaborative signals, we devise a data augmentation mechanism to model the interaction dynamics between users and items. For semantic information, we introduce a novel graph augmentation technique by constructing an interactive motif KG, enabling semantic contrastive learning within localized views. Furthermore, joint enhancement of graph and interaction data allows the model to capture global topological features of nodes. MMCL integrates contrastive learning at both local and global levels, effectively embedding local semantics and global topology into the learned representations. Extensive experiments conducted on three publicly available datasets reveal that MMCL outperforms advanced methods, particularly in scenarios with sparse interactions and noisy KG, while significantly alleviating popularity bias in recommendation.
- 20 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. PP, 1
GF-DecNet: Geometry-Aware Feature Decoupling Network for Millimeter-Wave Radar Pose Estimation
Millimeter-wave (mmWave) radar is an effective sensor for human pose estimation, particularly in challenging environments where vision-based systems may struggle. However, the sparse, noisy nature of mmWave radar point clouds, along with multipath effects, presents difficulties for precise pose reconstruction. This paper introduces the Geometry-aware Feature Decoupling Network (GF-DecNet), an approach that employs a dynamic graph convolutional network (DGCN) framework to overcome these challenges. GF-DecNet combines Set Abstraction (SA) modules with Geometry-aware Feature Decoupling (GFD) blocks for effective multi-scale feature learning. This design captures both local geometric details and global context, which are crucial for accurate pose estimation. The GFD blocks are designed to disentangle local geometric features from higher-level semantic features, enhancing the network’s ability to process sparse and noisy data. A multi-scale feature aggregation module merges these learned features, providing a comprehensive representation of human pose that integrates low-level geometric details with high-level semantic information. Additionally, a learnable geometric constraint mechanism, grounded in biomechanical models, enhances robustness by ensuring consistency with human body proportions. Extensive experiments on four established public datasets (MiliPoint, MM-Fi, mRI, and mmBody) and our curated WavePose dataset demonstrate that GF-DecNet consistently achieves competitive performance, achieving significant improvements in Mean Per Joint Position Error (MPJPE). Ablation studies further validate the effectiveness of each key component of GF-DecNet, These findings establish GF-DecNet as a strong benchmark in the field of mmWave radar-based pose estimation.
- 20 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Internet of Things Journal
- Vol. PP, 1
GF-DecNet: Geometry-Aware Feature Decoupling Network for Millimeter-Wave Radar Pose Estimation
Millimeter-wave (mmWave) radar is an effective sensor for human pose estimation, particularly in challenging environments where vision-based systems may struggle. However, the sparse, noisy nature of mmWave radar point clouds, along with multipath effects, presents difficulties for precise pose reconstruction. This paper introduces the Geometry-aware Feature Decoupling Network (GF-DecNet), an approach that employs a dynamic graph convolutional network (DGCN) framework to overcome these challenges. GF-DecNet combines Set Abstraction (SA) modules with Geometry-aware Feature Decoupling (GFD) blocks for effective multi-scale feature learning. This design captures both local geometric details and global context, which are crucial for accurate pose estimation. The GFD blocks are designed to disentangle local geometric features from higher-level semantic features, enhancing the network’s ability to process sparse and noisy data. A multi-scale feature aggregation module merges these learned features, providing a comprehensive representation of human pose that integrates low-level geometric details with high-level semantic information. Additionally, a learnable geometric constraint mechanism, grounded in biomechanical models, enhances robustness by ensuring consistency with human body proportions. Extensive experiments on four established public datasets (MiliPoint, MM-Fi, mRI, and mmBody) and our curated WavePose dataset demonstrate that GF-DecNet consistently achieves competitive performance, achieving significant improvements in Mean Per Joint Position Error (MPJPE). Ablation studies further validate the effectiveness of each key component of GF-DecNet, These findings establish GF-DecNet as a strong benchmark in the field of mmWave radar-based pose estimation.
- 26 July 2025
- Published by Springer Nature
- p. 138-149
Knowledge Enhanced Global Graph Contrastive Denoising for Recommendation
- 26 July 2025
- Published by Springer Nature
- p. 138-149
Knowledge Enhanced Global Graph Contrastive Denoising for Recommendation
- 25 July 2025
- Published by Springer Nature
- p. 123-134
DAGR: DyHGSampler and AdaSGC-Based Knowledge Graph Diffusion for Recommendation
- 25 July 2025
- Published by Springer Nature
- p. 123-134
DAGR: DyHGSampler and AdaSGC-Based Knowledge Graph Diffusion for Recommendation
- Open Access
- 10 December 2024
- Published by MDPI AG in Electronics
- Vol. 13 (24) , 4861
Detection of Defective Apples Using Learnable Residual Multi-Head Attention Networks Integrated with CNNs
Many traditional fruit vendors still rely on manual sorting to pick out high-quality apples. This process is not only time-consuming but can also damage the apples. Meanwhile, automated detection technology is still in its early stage and lacks full reliability. To improve this technology, we propose a novel method, which incorporates a learnable scaling factor and residual connection to enhance the Multi-Head Attention mechanism. In our approach, a learnable scaling factor is first applied to adjust the attention weights dynamically, and then a residual connection combines the scaled attention output with the original input to preserve essential features from the initial data. By integrating Multi-Head Attention with Convolutional Neural Networks (CNNs) using this method, we propose a lightweight deep learning model called “Learnable Residual Multi-Head Attention Networks Fusion with CNNs” to detect defective apples. Compared to existing models, our proposed model has lower memory usage, shorter training time, and higher detection precision. On the test set, the model achieves an accuracy of 97.5%, a recall of 98%, and a specificity of 97%, along with the lowest detection time of 46 ms. Experimental results show that the proposed model using our method is highly promising for commercial sorting, as it reduces labor costs, increases the supply of high-quality apples, and boosts consumer satisfaction.
- 10 December 2024
- Published by MDPI AG in Electronics
- Vol. 13 (24) , 4861
Open AccessDetection of Defective Apples Using Learnable Residual Multi-Head Attention Networks Integrated with CNNs
Many traditional fruit vendors still rely on manual sorting to pick out high-quality apples. This process is not only time-consuming but can also damage the apples. Meanwhile, automated detection technology is still in its early stage and lacks full reliability. To improve this technology, we propose a novel method, which incorporates a learnable scaling factor and residual connection to enhance the Multi-Head Attention mechanism. In our approach, a learnable scaling factor is first applied to adjust the attention weights dynamically, and then a residual connection combines the scaled attention output with the original input to preserve essential features from the initial data. By integrating Multi-Head Attention with Convolutional Neural Networks (CNNs) using this method, we propose a lightweight deep learning model called “Learnable Residual Multi-Head Attention Networks Fusion with CNNs” to detect defective apples. Compared to existing models, our proposed model has lower memory usage, shorter training time, and higher detection precision. On the test set, the model achieves an accuracy of 97.5%, a recall of 98%, and a specificity of 97%, along with the lowest detection time of 46 ms. Experimental results show that the proposed model using our method is highly promising for commercial sorting, as it reduces labor costs, increases the supply of high-quality apples, and boosts consumer satisfaction.