Yijin Wang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Previous Organizations / Institutes
The organizations / institutes from the scholar's other articles in the past ten years
Research Subjects
Yijin Wang has a total of 11 co-authors. Here are the top co-authors by number of shared publications:
- Dong Zhang
School of Biomedical Engineering, Sun Yat-Sen University, Shenzhen, China
1 shared publication - Hao Li1 shared publication
Hongjiang Zhang
Microsoft Research Asia, Beijing, China
1 shared publication- Mingjing Li
School of Aeronautic Science and Engineering Beihang University Beijing China
1 shared publication - Peng Zhao
University of Tennessee Space Institute, University of Tennessee Knoxville
1 shared publication Qianzhi Ma
Yunnan Open University (Yunnan Technical College of Industry)
1 shared publication
Latest Publications
- Open Access
- 8 July 2026
- Published by Elsevier in SSRN Electronic Journal
Client Contribution Evaluation Considering Multi-metric for Adaptive Global Model Aggregation in Industrial Federated Learning
In Industrial Federated Learning (IFL), heterogeneous data distributions, limited computing resources, and unstable client participation make it difficult to assign appropriate aggregation weights to local models. Existing client contribution evaluation methods mainly rely on the Shapley value, which provides a game-theoretic measure of marginal contribution but suffers from high computational cost and cannot fully capture indirect factors such as data distribution and resource reliability. To address these limitations, this paper proposes a multi-metric framework for evaluating client contributions in adaptive Global Model Aggregation (GMA) in IFL. The proposed framework evaluates each client from three complementary perspectives: Shapley contribution, data contribution, and resource contribution. Guided Truncation Gradient Shapley (GTG-Shapley) is adopted to approximate marginal utility with reduced computational complexity, while Maximum Mean Discrepancy (MMD)-based data contribution and resource-related indicators are introduced to characterize client heterogeneity. Furthermore, an Adaptive Warm-Started Genetic Algorithm (AWS-GA) is designed to optimize client-specific metric weights, enabling each client to adaptively determine the relative importance of Shapley, data, and resource contributions during global training. Experiments on the GC10-DET industrial defect dataset demonstrate that the proposed method achieves a final accuracy of 0.9713 within 5 global training rounds, outperforming FedAvg, Shapley only, Shapley + Data, and Shapley + Resource in convergence speed and overall performance. Although the proposed method introduces additional aggregation overhead, it reduces the total training time to 4000s on a high-performance server, which is shorter than that of the compared methods. Additional scalability analysis with 50, 100, and 500 clients further verifies that the proposed method maintains higher accuracy, lower loss, and lower contribution evaluation overhead than the conventional Shapley-based method. These results indicate that the proposed multi-metric and client-specific contribution evaluation framework can improve the rationality of aggregation, convergence efficiency, and scalability in IFL scenarios with Non-IID data and heterogeneous resources.
- 8 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessClient Contribution Evaluation Considering Multi-metric for Adaptive Global Model Aggregation in Industrial Federated Learning
In Industrial Federated Learning (IFL), heterogeneous data distributions, limited computing resources, and unstable client participation make it difficult to assign appropriate aggregation weights to local models. Existing client contribution evaluation methods mainly rely on the Shapley value, which provides a game-theoretic measure of marginal contribution but suffers from high computational cost and cannot fully capture indirect factors such as data distribution and resource reliability. To address these limitations, this paper proposes a multi-metric framework for evaluating client contributions in adaptive Global Model Aggregation (GMA) in IFL. The proposed framework evaluates each client from three complementary perspectives: Shapley contribution, data contribution, and resource contribution. Guided Truncation Gradient Shapley (GTG-Shapley) is adopted to approximate marginal utility with reduced computational complexity, while Maximum Mean Discrepancy (MMD)-based data contribution and resource-related indicators are introduced to characterize client heterogeneity. Furthermore, an Adaptive Warm-Started Genetic Algorithm (AWS-GA) is designed to optimize client-specific metric weights, enabling each client to adaptively determine the relative importance of Shapley, data, and resource contributions during global training. Experiments on the GC10-DET industrial defect dataset demonstrate that the proposed method achieves a final accuracy of 0.9713 within 5 global training rounds, outperforming FedAvg, Shapley only, Shapley + Data, and Shapley + Resource in convergence speed and overall performance. Although the proposed method introduces additional aggregation overhead, it reduces the total training time to 4000s on a high-performance server, which is shorter than that of the compared methods. Additional scalability analysis with 50, 100, and 500 clients further verifies that the proposed method maintains higher accuracy, lower loss, and lower contribution evaluation overhead than the conventional Shapley-based method. These results indicate that the proposed multi-metric and client-specific contribution evaluation framework can improve the rationality of aggregation, convergence efficiency, and scalability in IFL scenarios with Non-IID data and heterogeneous resources.
- Open Access
- 7 July 2026
- Published by Springer Nature in Journal of Intelligent Manufacturing
Context-aware Transferability Estimation for Industrial Transfer Learning
Transfer learning (TL) has emerged as an effective approach in intelligent manufacturing by reusing knowledge from related domains to address challenges such as data scarcity, product diversity, and frequent process reconfiguration. However, the effectiveness of TL critically depends on selecting source domains with high transferability. Existing transferability metrics primarily emphasize algorithmic or data-driven aspects and often lack interpretability for manufacturing practice. This study proposes a framework that integrates context-aware transferability score (CTS), including structural alignment, monotonicity consistency, and key parameter analysis, with a data-based transferability score (DTS) that captures distributional similarity through Maximum Mean Discrepancy (MMD), Wasserstein Distance (WD), and Jensen–Shannon (JS) divergence. These components are combined into a holistic transferability score (HTS), which balances process-informed interpretability with data-driven rigor. Validation on a real-world battery anode manufacturing case and complementary synthetic stress tests demonstrates that the proposed framework provides explainable, computationally efficient, and reliable guidance for source domain selection. The proposed method establishes a systematic and interpretable foundation for transfer learning in industrial environments, mitigating negative transfer and supporting informed decision-making in intelligent manufacturing.
- 7 July 2026
- Published by Springer Nature in Journal of Intelligent Manufacturing
Open AccessContext-aware Transferability Estimation for Industrial Transfer Learning
Transfer learning (TL) has emerged as an effective approach in intelligent manufacturing by reusing knowledge from related domains to address challenges such as data scarcity, product diversity, and frequent process reconfiguration. However, the effectiveness of TL critically depends on selecting source domains with high transferability. Existing transferability metrics primarily emphasize algorithmic or data-driven aspects and often lack interpretability for manufacturing practice. This study proposes a framework that integrates context-aware transferability score (CTS), including structural alignment, monotonicity consistency, and key parameter analysis, with a data-based transferability score (DTS) that captures distributional similarity through Maximum Mean Discrepancy (MMD), Wasserstein Distance (WD), and Jensen–Shannon (JS) divergence. These components are combined into a holistic transferability score (HTS), which balances process-informed interpretability with data-driven rigor. Validation on a real-world battery anode manufacturing case and complementary synthetic stress tests demonstrates that the proposed framework provides explainable, computationally efficient, and reliable guidance for source domain selection. The proposed method establishes a systematic and interpretable foundation for transfer learning in industrial environments, mitigating negative transfer and supporting informed decision-making in intelligent manufacturing.
- 9 July 2019
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 78 (19) , 28257-28283
Improving person re-identification by multi-task learning
- 9 July 2019
- Published by Springer Nature in Multimedia Tools and Applications
- Vol. 78 (19) , 28257-28283
Improving person re-identification by multi-task learning
- 1 September 2016
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 180-187
Indoor Localization Service Based on the Data Fusion of Wi-Fi and RFID
With more and more requirements of navigation for complex indoor environments, the indoor location service has become the hotspot in the field of mobile computing. However, with only one single type of wireless signal, it is difficult to achieve ideal accuracy of positioning in the indoor environments full of indoor noises. In order to improve the performance of indoor location service, we propose a novel indoor localization mechanism, which realizes an effective data fusion of Wi-Fi and RFID signals via on-demand deployment of Wi-Fi access points and RFID tags. This mechanism can eliminate the blind areas of location so as to realize the low-cost and high-accurate indoor localization. In order to further improve the location performance, we put forward the Kalman filter algorithm based on singular value judgment (KFASVJ) and KFASVJ-based indoor localization algorithm (KILA). The KILA is adopted to judge the maximum singular value of Wi-Fi signal wave, so as to optimize the wireless signal wave. KILA can reduce the indoor noise interference with Wi-Fi signals, so to realize a high accuracy of positioning and real-time positioning stability in complex indoor environments. The experimental results and the performance analysis show that KILA achieves a better accuracy of positioning than typical Kalman-filter-based localization algorithm (KFLA), about 13% to 28% accuracy of positioning improvement in the indoor environments with the [35dB, 65dB] indoor noise. KILA has the lower time complexity, the higher location speed and the better stability, and it can maintain a good localization performance even in the indoor environment with the indoor noises changing dynamicly.
- 1 September 2016
- Published by Institute of Electrical and Electronics Engineers (IEEE)
- p. 180-187
Indoor Localization Service Based on the Data Fusion of Wi-Fi and RFID
With more and more requirements of navigation for complex indoor environments, the indoor location service has become the hotspot in the field of mobile computing. However, with only one single type of wireless signal, it is difficult to achieve ideal accuracy of positioning in the indoor environments full of indoor noises. In order to improve the performance of indoor location service, we propose a novel indoor localization mechanism, which realizes an effective data fusion of Wi-Fi and RFID signals via on-demand deployment of Wi-Fi access points and RFID tags. This mechanism can eliminate the blind areas of location so as to realize the low-cost and high-accurate indoor localization. In order to further improve the location performance, we put forward the Kalman filter algorithm based on singular value judgment (KFASVJ) and KFASVJ-based indoor localization algorithm (KILA). The KILA is adopted to judge the maximum singular value of Wi-Fi signal wave, so as to optimize the wireless signal wave. KILA can reduce the indoor noise interference with Wi-Fi signals, so to realize a high accuracy of positioning and real-time positioning stability in complex indoor environments. The experimental results and the performance analysis show that KILA achieves a better accuracy of positioning than typical Kalman-filter-based localization algorithm (KFLA), about 13% to 28% accuracy of positioning improvement in the indoor environments with the [35dB, 65dB] indoor noise. KILA has the lower time complexity, the higher location speed and the better stability, and it can maintain a good localization performance even in the indoor environment with the indoor noises changing dynamicly.
- 1 December 2002
- Published by Association for Computing Machinery (ACM)
- p. 412-413
Open AccessMyVideos