Man-Fai Leung
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
Man-Fai Leung has a total of 194 co-authors. Here are the top co-authors by number of shared publications:
- Hangjun Che
College of Electronic and Information Engineering, Southwest University, Chongqing, China
28 shared publications - Sin-Chun Ng
School of Computing and Information Science, Faculty of Science and Engineering, Anglia Ruskin University, Cambridge CB1 1PT, UK
12 shared publications - Cheng Liu
College of Computer Science and Technology, Huaqiao University, Xiamen, China
9 shared publications Hangjun Che
8 shared publications- Jun Wang
Department of Physics, Southeast University, Nanjing 210096, China
8 shared publications - Britto Pari James
Vel Tech Rangarajan Dr. Sagunthala R&D Institute of Science and Technology, Chennai 600062, India
7 shared publications
Latest Publications
- Open Access
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 5320-5329
Self-representation Tensor Tri-Factorization with View Consistency for Multi-view Clustering
- 9 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 5320-5329
Open AccessSelf-representation Tensor Tri-Factorization with View Consistency for Multi-view Clustering
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Systems, Man, and Cybernetics: Systems
- Vol. 56 (11) , 6575-6588
Robust Subspace Learning-Based Orthogonal Nonnegative Tucker Decomposition for Incomplete Multiview Clustering
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Systems, Man, and Cybernetics: Systems
- Vol. 56 (11) , 6575-6588
Robust Subspace Learning-Based Orthogonal Nonnegative Tucker Decomposition for Incomplete Multiview Clustering
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 72 (4) , 10986-10995
Bridging Domain Gap and Preserving Privacy for Battery Diagnosis in Data Centers
Data centers, the cornerstone of the digital economy, depend on Uninterruptible Power Supply (UPS) systems, where battery health is critical for operational continuity. However, the practical application of data-driven anomaly detection is severely hampered by two interconnected challenges: distribution shifts between different battery batches degrade model generalization, while data silos arising from commercial and privacy constraints prevent cross-institutional data aggregation. To overcome these barriers, we systematically formulate this industrial challenge as a Federated Unsupervised Domain Adaptation (FUDA) task, which enables collaborative model training without sharing raw data. The core contribution of this paper is the proposal of a novel scenario-driven adaptation strategy, which posits that no single algorithm is universally optimal. We demonstrate that for homogeneous scenarios with small domain gaps, a lightweight discrepancy-based method (CORAL) paired with a powerful feature extractor (Transformer) is most effective. Conversely, for heterogeneous scenarios with large domain gaps, a stable, decoupled adversarial approach (ADDA) combined with a capacity-constrained feature extractor (DeeperMLP) is superior as it ensures convergence and promotes generalization. Experiments on a real-world data center dataset validate our FUDA approach (F1-score of 0.872) not only significantly outperforms local non-adapted models but also achieves performance highly comparable to, and in some cases exceeding, that of privacy-violating centralized models (0.908).
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 72 (4) , 10986-10995
Bridging Domain Gap and Preserving Privacy for Battery Diagnosis in Data Centers
Data centers, the cornerstone of the digital economy, depend on Uninterruptible Power Supply (UPS) systems, where battery health is critical for operational continuity. However, the practical application of data-driven anomaly detection is severely hampered by two interconnected challenges: distribution shifts between different battery batches degrade model generalization, while data silos arising from commercial and privacy constraints prevent cross-institutional data aggregation. To overcome these barriers, we systematically formulate this industrial challenge as a Federated Unsupervised Domain Adaptation (FUDA) task, which enables collaborative model training without sharing raw data. The core contribution of this paper is the proposal of a novel scenario-driven adaptation strategy, which posits that no single algorithm is universally optimal. We demonstrate that for homogeneous scenarios with small domain gaps, a lightweight discrepancy-based method (CORAL) paired with a powerful feature extractor (Transformer) is most effective. Conversely, for heterogeneous scenarios with large domain gaps, a stable, decoupled adversarial approach (ADDA) combined with a capacity-constrained feature extractor (DeeperMLP) is superior as it ensures convergence and promotes generalization. Experiments on a real-world data center dataset validate our FUDA approach (F1-score of 0.872) not only significantly outperforms local non-adapted models but also achieves performance highly comparable to, and in some cases exceeding, that of privacy-violating centralized models (0.908).
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 72 (4) , 12737-12748
Multimodal Weather-Embedding-Assisted Framework for Traffic Flow Forecasting
The integration of generative artificial intelligence (GAI) in traffic forecasting is transforming urban mobility systems, particularly through multimodal data fusion. While prior research has advanced the modeling of spatial and temporal dependencies in traffic systems, the role of weather information has been largely underexplored. To address this gap, we propose a weather-aware dynamic graph neural network (WADGNet) that leverages GAI to enhance traffic flow prediction by integrating multisource weather data, improving both accuracy and robustness. The framework consists of two key modules: 1) The WAEformer utilizes generative representations to capture latent meteorological patterns. 2) The SaGCN, a semi-autoregressive spatiotemporal graph convolution network, learns dynamic spatial and temporal dependencies in traffic flow. By fusing multimodal data from traffic and weather systems, WADGNet enables more accurate and reliable traffic flow forecasting in diverse urban environments. Extensive experiments on several benchmark datasets show that WADGNet consistently outperforms state-of-the-art models, demonstrating strong generalizability. These findings highlight the potential of GAI to power smart city infrastructures through improved traffic management and prediction systems.
- 1 November 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 72 (4) , 12737-12748
Multimodal Weather-Embedding-Assisted Framework for Traffic Flow Forecasting
The integration of generative artificial intelligence (GAI) in traffic forecasting is transforming urban mobility systems, particularly through multimodal data fusion. While prior research has advanced the modeling of spatial and temporal dependencies in traffic systems, the role of weather information has been largely underexplored. To address this gap, we propose a weather-aware dynamic graph neural network (WADGNet) that leverages GAI to enhance traffic flow prediction by integrating multisource weather data, improving both accuracy and robustness. The framework consists of two key modules: 1) The WAEformer utilizes generative representations to capture latent meteorological patterns. 2) The SaGCN, a semi-autoregressive spatiotemporal graph convolution network, learns dynamic spatial and temporal dependencies in traffic flow. By fusing multimodal data from traffic and weather systems, WADGNet enables more accurate and reliable traffic flow forecasting in diverse urban environments. Extensive experiments on several benchmark datasets show that WADGNet consistently outperforms state-of-the-art models, demonstrating strong generalizability. These findings highlight the potential of GAI to power smart city infrastructures through improved traffic management and prediction systems.
- Open Access
- 10 October 2026
- Published by MDPI AG in Journal of Risk and Financial Management
- Vol. 19 (10) , 801
An Unsupervised Intelligent System for Adaptive Decision Control in Non-Stationary Financial Environments
Intelligent decision systems assume a stable relationship between observations and outcomes. Financial markets violate that assumption, since shifts in volatility, trend and liquidity erode models trained on historical data. This study proposes an unsupervised decision system that restricts trading according to market conditions. Features from multiple time horizons are partitioned by walk-forward clustering into distinct market states, and execution is permitted only when the state satisfies predefined conditions, so that clustering governs when predictions are acted upon rather than how they are formed. The framework is evaluated on gold, Bitcoin, the S&P 500, Nasdaq 100, Dow Jones and the US dollar–Japanese yen exchange rate between January 2024 and December 2025, against a non-gated baseline with identical signals, risk controls and position sizing. Maximum drawdown falls on every asset, by half in several cases, whilst trade frequency declines by a third to two thirds; on the S&P 500 the Sortino ratio rises from 2.57 to 6.63 as maximum drawdown falls from 42.27% to 22.51%. A random-drop ablation shows the gain is not explained by trading less often, and bootstrap resampling indicates higher Sharpe ratios across all six markets, with a combined p-value of 0.0074. Results are gross of transaction costs.
- 10 October 2026
- Published by MDPI AG in Journal of Risk and Financial Management
- Vol. 19 (10) , 801
Open AccessAn Unsupervised Intelligent System for Adaptive Decision Control in Non-Stationary Financial Environments
Intelligent decision systems assume a stable relationship between observations and outcomes. Financial markets violate that assumption, since shifts in volatility, trend and liquidity erode models trained on historical data. This study proposes an unsupervised decision system that restricts trading according to market conditions. Features from multiple time horizons are partitioned by walk-forward clustering into distinct market states, and execution is permitted only when the state satisfies predefined conditions, so that clustering governs when predictions are acted upon rather than how they are formed. The framework is evaluated on gold, Bitcoin, the S&P 500, Nasdaq 100, Dow Jones and the US dollar–Japanese yen exchange rate between January 2024 and December 2025, against a non-gated baseline with identical signals, risk controls and position sizing. Maximum drawdown falls on every asset, by half in several cases, whilst trade frequency declines by a third to two thirds; on the S&P 500 the Sortino ratio rises from 2.57 to 6.63 as maximum drawdown falls from 42.27% to 22.51%. A random-drop ablation shows the gain is not explained by trading less often, and bootstrap resampling indicates higher Sharpe ratios across all six markets, with a combined p-value of 0.0074. Results are gross of transaction costs.
- Open Access
- 1 October 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (19) , 9751
Design of Low-Complexity and High-Speed Reconfigurable FIR Filter Using a Modified Distributed Arithmetic Structure
The implementation of Software-Defined Radio (SDR) requires various reconfigurable modules. One of the important modules is the reconfigurable finite impulse response (FIR) filter, which SDR has to support different wireless standards by reconfiguring the single filter structure. The complexity of implementing these filters depends on the method of changing coefficients, which can be resolved using the coefficient decimation method. This paper presents a Distributed Arithmetic (DA)-based reconfigurable FIR filter employing the Coefficient Decimation method (CDM) to achieve a low-complexity and scalable hardware architecture for Software Defined Radio (SDR) applications. FPGA synthesis results demonstrate approximately proportional resource growth with increasing filter order, requiring 400, 950, 2200, 4680, and 9800 LUTs for the 32-, 64-, 128-, 256-, and 512-tap implementations, respectively. The design operates at maximum frequencies of 165 MHz and 150 MHz, while consuming only 145 mW and 295 mW of power for the 128-tap and 256-tap configurations, respectively, without the use of dedicated DSP blocks. In addition, the proposed architecture achieves the lowest LUT-per-tap ratio among the compared designs, highlighting its efficient utilization of FPGA resources. Overall, the results demonstrate that the proposed CDM–DA architecture provides an effective trade-off among hardware resource utilization, operating frequency, power consumption, and scalability, making it a promising solution for high-order reconfigurable FIR filters in SDR channelization, multirate signal processing, and adaptive filtering applications.
- 1 October 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (19) , 9751
Open AccessDesign of Low-Complexity and High-Speed Reconfigurable FIR Filter Using a Modified Distributed Arithmetic Structure
The implementation of Software-Defined Radio (SDR) requires various reconfigurable modules. One of the important modules is the reconfigurable finite impulse response (FIR) filter, which SDR has to support different wireless standards by reconfiguring the single filter structure. The complexity of implementing these filters depends on the method of changing coefficients, which can be resolved using the coefficient decimation method. This paper presents a Distributed Arithmetic (DA)-based reconfigurable FIR filter employing the Coefficient Decimation method (CDM) to achieve a low-complexity and scalable hardware architecture for Software Defined Radio (SDR) applications. FPGA synthesis results demonstrate approximately proportional resource growth with increasing filter order, requiring 400, 950, 2200, 4680, and 9800 LUTs for the 32-, 64-, 128-, 256-, and 512-tap implementations, respectively. The design operates at maximum frequencies of 165 MHz and 150 MHz, while consuming only 145 mW and 295 mW of power for the 128-tap and 256-tap configurations, respectively, without the use of dedicated DSP blocks. In addition, the proposed architecture achieves the lowest LUT-per-tap ratio among the compared designs, highlighting its efficient utilization of FPGA resources. Overall, the results demonstrate that the proposed CDM–DA architecture provides an effective trade-off among hardware resource utilization, operating frequency, power consumption, and scalability, making it a promising solution for high-order reconfigurable FIR filters in SDR channelization, multirate signal processing, and adaptive filtering applications.
- Open Access
- 25 September 2026
- Published by MDPI AG in Biomolecules
- Vol. 16 (10) , 1402
Topology-Aware Multi-Omics Tensor Graph Clustering for Cancer Subtype Identification
Cancer subtype identification from multi-omics profiles is important for understanding tumor heterogeneity and advancing computational precision oncology. Various subtyping methods have been proposed, among which graph-based approaches have become a widely adopted paradigm by capturing patient-to-patient relationships from heterogeneous molecular profiles. However, the reliability of learned patient similarity graphs remains a critical challenge, as noisy or spurious edges induced by high-dimensional and heterogeneous omics measurements can impair subtype discrimination. To address this issue, we propose topology-aware tensor graph learning (TaTGL), an unsupervised multi-omics clustering framework for cancer subtyping. TaTGL first constructs omics-specific patient graphs using a topology-aware distance metric that integrates molecular-profile distance with shared-neighborhood consistency. Furthermore, to exploit both complementary and consistent information across different omics views, the resulting graphs are integrated through frequency-weighted t-SVD-based tensor graph learning, which captures high-order cross-omics structural consistency. Experimental studies on several cancer subtype datasets show that TaTGL achieves competitive or superior clustering performance on labeled datasets and identifies clinically relevant patient subgroups on survival-based datasets. These results demonstrate that the proposed method can discover molecularly coherent and clinically meaningful cancer subtypes from heterogeneous multi-omics profiles.
- 25 September 2026
- Published by MDPI AG in Biomolecules
- Vol. 16 (10) , 1402
Open AccessTopology-Aware Multi-Omics Tensor Graph Clustering for Cancer Subtype Identification
Cancer subtype identification from multi-omics profiles is important for understanding tumor heterogeneity and advancing computational precision oncology. Various subtyping methods have been proposed, among which graph-based approaches have become a widely adopted paradigm by capturing patient-to-patient relationships from heterogeneous molecular profiles. However, the reliability of learned patient similarity graphs remains a critical challenge, as noisy or spurious edges induced by high-dimensional and heterogeneous omics measurements can impair subtype discrimination. To address this issue, we propose topology-aware tensor graph learning (TaTGL), an unsupervised multi-omics clustering framework for cancer subtyping. TaTGL first constructs omics-specific patient graphs using a topology-aware distance metric that integrates molecular-profile distance with shared-neighborhood consistency. Furthermore, to exploit both complementary and consistent information across different omics views, the resulting graphs are integrated through frequency-weighted t-SVD-based tensor graph learning, which captures high-order cross-omics structural consistency. Experimental studies on several cancer subtype datasets show that TaTGL achieves competitive or superior clustering performance on labeled datasets and identifies clinically relevant patient subgroups on survival-based datasets. These results demonstrate that the proposed method can discover molecularly coherent and clinically meaningful cancer subtypes from heterogeneous multi-omics profiles.
- 21 September 2026
- Published by Royal Society of Chemistry (RSC) in Food & Function
- Vol. 17 (18) , 7908-7941
Artificial intelligence in food and nutrition science: a paradigm-centric review of computational frameworks and system-level integration
AI integrates recognition, mapping, search, synchronization, and reasoning to transform multimodal food data into actionable intelligence for precision nutrition, food safety, smart manufacturing, sustainability, and adaptive food systems.
- 21 September 2026
- Published by Royal Society of Chemistry (RSC) in Food & Function
- Vol. 17 (18) , 7908-7941
Artificial intelligence in food and nutrition science: a paradigm-centric review of computational frameworks and system-level integration
AI integrates recognition, mapping, search, synchronization, and reasoning to transform multimodal food data into actionable intelligence for precision nutrition, food safety, smart manufacturing, sustainability, and adaptive food systems.
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Knowledge and Data Engineering
- Vol. 38 (7) , 4627-4640
SAFA: Sequential Recommendation With Adaptive Sparse Attention and Frequency-Aware Encoding
- 1 July 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Knowledge and Data Engineering
- Vol. 38 (7) , 4627-4640
SAFA: Sequential Recommendation With Adaptive Sparse Attention and Frequency-Aware Encoding
- 17 April 2026
- Published by Elsevier in Neural Networks
- Vol. 202, 108973
Self-supervised semantic graph propagation for multi-view clustering
- 17 April 2026
- Published by Elsevier in Neural Networks
- Vol. 202, 108973
Self-supervised semantic graph propagation for multi-view clustering