Man-Sheng Chen
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Man-Sheng Chen has a total of 80 co-authors. Here are the top co-authors by number of shared publications:
- Chang-Dong Wang
Sun Yat-sen University
33 shared publications - Jia-Qi Lin
School of Mathematics (Zhuhai), Sun Yat-sen University, Guangzhou, China
10 shared publications - Jianhuang Lai
Sun Yat-Sen University, Guangzhou, China
10 shared publications - Dong Huang
South China Agricultural University
6 shared publications - Haizhang Zhang
School of Mathematics (Zhuhai), Sun Yat-Sen University, Zhuhai 519082, China
5 shared publications - Qingyun Dai
School of Information Engineering, Guangdong University of Technology, Guangzhou, China
5 shared publications
Latest Publications
- 22 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. PP (10518215) , 1
Density Peak Anchor Selection for Multi-View Clustering
Multi-view clustering (MVC) typically enhances clustering performance by leveraging consistent and complementary information across views. With increasing data scale, many MVC methods employ few anchors to represent data, effectively reducing complexity. Traditional anchor selection methods include random sampling and k-means. Recent methods leverage variance or inter-class error to guide anchor selection. However, their anchors are selected sequentially, where each newly selected anchor focuses on divergences from the previous anchor. This local perspective leads to unstable anchor coverage and fails to fully represent complex cluster shapes. An anchor selection algorithm, termed DPAS, is proposed to solve this problem. From a novel density perspective, it selects density peaks satisfying criteria as anchors and reduces complexity based on granular balls. As a result, the efficient DPAS exhibits stronger representativeness and robustness. Furthermore, the order and number of anchors selected from different views typically need to be aligned. The assumptions of existing alignment methods are often too restrictive, compromising the quality of alignment. Thus, an alignment method FAA based on flexible assumptions is designed to improve generalization. In this paper, we propose a new fast MVC method without alternating optimization named DPAS-MC. First, DPAS is applied to select high-quality anchors for views. Then, anchor graphs are computed, aligned, and fused via FAA. Finally, a low-rank tensor is calculated to obtain clustering results. Experiments show that our methods outperform existing state-of-the-art methods in performance and efficiency. The code is available at https://github.com/ytccyw.
- 22 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. PP (10518215) , 1
Density Peak Anchor Selection for Multi-View Clustering
Multi-view clustering (MVC) typically enhances clustering performance by leveraging consistent and complementary information across views. With increasing data scale, many MVC methods employ few anchors to represent data, effectively reducing complexity. Traditional anchor selection methods include random sampling and k-means. Recent methods leverage variance or inter-class error to guide anchor selection. However, their anchors are selected sequentially, where each newly selected anchor focuses on divergences from the previous anchor. This local perspective leads to unstable anchor coverage and fails to fully represent complex cluster shapes. An anchor selection algorithm, termed DPAS, is proposed to solve this problem. From a novel density perspective, it selects density peaks satisfying criteria as anchors and reduces complexity based on granular balls. As a result, the efficient DPAS exhibits stronger representativeness and robustness. Furthermore, the order and number of anchors selected from different views typically need to be aligned. The assumptions of existing alignment methods are often too restrictive, compromising the quality of alignment. Thus, an alignment method FAA based on flexible assumptions is designed to improve generalization. In this paper, we propose a new fast MVC method without alternating optimization named DPAS-MC. First, DPAS is applied to select high-quality anchors for views. Then, anchor graphs are computed, aligned, and fused via FAA. Finally, a low-rank tensor is calculated to obtain clustering results. Experiments show that our methods outperform existing state-of-the-art methods in performance and efficiency. The code is available at https://github.com/ytccyw.
- 15 April 2026
- Published by Elsevier in Information Fusion
- Vol. 134
Multi-pattern structure refinement and fusion for heterophilic graphs
- 15 April 2026
- Published by Elsevier in Information Fusion
- Vol. 134
Multi-pattern structure refinement and fusion for heterophilic graphs
- 2 February 2026
- Published by Springer Nature in International Journal of Machine Learning and Cybernetics
- Vol. 17 (2) , 68
Multi-source knowledge enhancement for multimodal semantic representation
- 2 February 2026
- Published by Springer Nature in International Journal of Machine Learning and Cybernetics
- Vol. 17 (2) , 68
Multi-source knowledge enhancement for multimodal semantic representation
- 6 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Systems, Man, and Cybernetics: Systems
- Vol. 56 (2) , 1165-1178
Knowledge-Aware Clustering
Data clustering aims to partition the input data entities into several disjoint categories, where similar entities are grouped together while dissimilar ones are pulled apart. In general, the existing data clustering methods merely depend on the attribute information or constructed similarity graph information of the input data entities, and such one-sided cues may result in an incomplete understanding and potentially biased conclusions. How to well consider the inherent heterogeneous knowledge of data while maintaining the local semantics remains a challenging problem. To address this, we propose a novel knowledge-aware clustering (KAC) method, where the attribute information and inherent heterogeneous knowledge of data are jointly considered for better cluster structure recovery. Within this framework, there are four major modules. They are: 1) the target attribute autoencoder to capture a customizable target feature representation; 2) the meta-path aggregated knowledge graph encoder to learn an informative graph feature representation; 3) the dual information maximization to ensure the consistent semantics between two embedding representations; and 4) the self-training clustering to self-supervise the clustering learning. Experiments on several real-world datasets are conducted to validate the superiority of KAC, indicating the significance of explicitly considering the heterogeneous knowledge while maintaining local semantics.
- 6 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Systems, Man, and Cybernetics: Systems
- Vol. 56 (2) , 1165-1178
Knowledge-Aware Clustering
Data clustering aims to partition the input data entities into several disjoint categories, where similar entities are grouped together while dissimilar ones are pulled apart. In general, the existing data clustering methods merely depend on the attribute information or constructed similarity graph information of the input data entities, and such one-sided cues may result in an incomplete understanding and potentially biased conclusions. How to well consider the inherent heterogeneous knowledge of data while maintaining the local semantics remains a challenging problem. To address this, we propose a novel knowledge-aware clustering (KAC) method, where the attribute information and inherent heterogeneous knowledge of data are jointly considered for better cluster structure recovery. Within this framework, there are four major modules. They are: 1) the target attribute autoencoder to capture a customizable target feature representation; 2) the meta-path aggregated knowledge graph encoder to learn an informative graph feature representation; 3) the dual information maximization to ensure the consistent semantics between two embedding representations; and 4) the self-training clustering to self-supervise the clustering learning. Experiments on several real-world datasets are conducted to validate the superiority of KAC, indicating the significance of explicitly considering the heterogeneous knowledge while maintaining local semantics.
- 4 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Neural Networks and Learning Systems
- Vol. 37 (6) , 2563-2576
P 3 L: Patent Prediction With Prompt Learning
Patents are crucial for protecting technological innovations and fostering competitive advancements in industry. Patent prediction, a novel task in the field of patent mining, aims to forecast future technological trends, providing valuable insights for strategic planning and innovation in the industry. However, the complexity of patent data and the diversity of technological fields make effective patent prediction a significant challenge. Existing methods for predicting scientific research trends struggle to effectively model patent structures and capture dependencies between patents, resulting in suboptimal patent trend predictions. In this article, we propose a novel method, patent prediction with prompt learning (P3L), to achieve effective and accurate prediction of future patent developments based on a pretrained language model (PLM). P3L includes a patent similarity path extraction module to extract multiple patent development paths from extensive datasets. Following this, we design a patent prompt learning approach that integrates patent development paths, keywords, and patent similarities into the prompts. To mitigate potential noise introduced by this integration, we introduce an attention mask matrix for prompt denoising. Finally, we introduce three patent datasets with rich structures, and conduct extensive experiments on these datasets as well as a public dataset, demonstrating the superiority of the proposed method. The dataset and code have been made publicly available at https://github.com/AllminerLab/P3L
- 4 December 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Neural Networks and Learning Systems
- Vol. 37 (6) , 2563-2576
P 3 L: Patent Prediction With Prompt Learning
Patents are crucial for protecting technological innovations and fostering competitive advancements in industry. Patent prediction, a novel task in the field of patent mining, aims to forecast future technological trends, providing valuable insights for strategic planning and innovation in the industry. However, the complexity of patent data and the diversity of technological fields make effective patent prediction a significant challenge. Existing methods for predicting scientific research trends struggle to effectively model patent structures and capture dependencies between patents, resulting in suboptimal patent trend predictions. In this article, we propose a novel method, patent prediction with prompt learning (P3L), to achieve effective and accurate prediction of future patent developments based on a pretrained language model (PLM). P3L includes a patent similarity path extraction module to extract multiple patent development paths from extensive datasets. Following this, we design a patent prompt learning approach that integrates patent development paths, keywords, and patent similarities into the prompts. To mitigate potential noise introduced by this integration, we introduce an attention mask matrix for prompt denoising. Finally, we introduce three patent datasets with rich structures, and conduct extensive experiments on these datasets as well as a public dataset, demonstrating the superiority of the proposed method. The dataset and code have been made publicly available at https://github.com/AllminerLab/P3L
- 28 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Knowledge and Data Engineering
- Vol. 37 (10) , 6173-6188
Smoothness-Induced Efficient Incomplete Multi-View Clustering
Efficient incomplete multi-view clustering has received increasing attention due to its ability to handle large-scale and missing data. Although existing methods have promising performance, 1) they typically generate anchors directly from incomplete and noisy raw data, resulting in uncomprehensive anchor coverage and unreliable results; 2) they typically use only sparse regularization to remove noise and overlook outliers; 3) they ignore the inherent consistency of features in a view. To address these issues, we propose a smoothness-induced efficient incomplete multi-view clustering (SEIC) method. SEIC regards available data as natural anchors selected from complete data, and performs matrix decomposition only on them to obtain reliable small-size representation matrices. View-specific representation matrices are constructed as a tensor to capture consensus and guide matrix decomposition. More significantly, we enforce both smoothness and low-rank coupling on the tensor. Smoothness induces continuous variation of the tensor to further eliminate noise and enhance the relation among features. Benefiting from the noise robustness of SEIC, we design an adaptive noise balance parameter that renders SEIC parameter-free. Furthermore, by constructing a sparse anchor graph on the learned tensor, we propose the spectral clustering version SEIC-SC. Experiments on multiple datasets demonstrate the superior performance and efficiency of SEIC and SEIC-SC.
- 28 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Knowledge and Data Engineering
- Vol. 37 (10) , 6173-6188
Smoothness-Induced Efficient Incomplete Multi-View Clustering
Efficient incomplete multi-view clustering has received increasing attention due to its ability to handle large-scale and missing data. Although existing methods have promising performance, 1) they typically generate anchors directly from incomplete and noisy raw data, resulting in uncomprehensive anchor coverage and unreliable results; 2) they typically use only sparse regularization to remove noise and overlook outliers; 3) they ignore the inherent consistency of features in a view. To address these issues, we propose a smoothness-induced efficient incomplete multi-view clustering (SEIC) method. SEIC regards available data as natural anchors selected from complete data, and performs matrix decomposition only on them to obtain reliable small-size representation matrices. View-specific representation matrices are constructed as a tensor to capture consensus and guide matrix decomposition. More significantly, we enforce both smoothness and low-rank coupling on the tensor. Smoothness induces continuous variation of the tensor to further eliminate noise and enhance the relation among features. Benefiting from the noise robustness of SEIC, we design an adaptive noise balance parameter that renders SEIC parameter-free. Furthermore, by constructing a sparse anchor graph on the learned tensor, we propose the spectral clustering version SEIC-SC. Experiments on multiple datasets demonstrate the superior performance and efficiency of SEIC and SEIC-SC.
- 24 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Biology and Bioinformatics
- Vol. 22 (5) , 2267-2279
Knowledge-Aware Synergistic Discovery of Drug Combinations: A Large Language Model Perspective
Drug combination therapy with significant advantages is a well-established concept in cancer treatment. Some related efforts have been made with multiple artful deep learning techniques. However, they are usually based on data for drug synergy prediction, ignoring the professional characteristics of data and the systematic knowledge accumulation. Meanwhile, integrating the dispersed professional knowledge and effectively utilizing it in data remains a crucial technical challenge. In this study, we propose KSDDC, a novel model for knowledge-aware synergistic discovery of drug combinations from a large language model (LLM) perspective (i.e., from the continuously learnable and refined large database). Within this framework, three main modules are well-designed, i.e., knowledge-aware drug feature auto-encoding, knowledge-aware cell line feature encoding and drug-drug synergy prediction. Informative embeddings of samples are discovered and combined to make accurate drug synergy prediction. Overall, KSDDC is superior compared with the other shallow machine learning based methods and deep learning based methods on several synergistic prediction benchmarks, where about 19% F1-score improvements over the second best method on DrugComb_1 can be observed. Starting with drug synergy prediction, our studies with knowledge-enabled data mining offer valuable insights and serve as a reference method for future research in this field.
- 24 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Biology and Bioinformatics
- Vol. 22 (5) , 2267-2279
Knowledge-Aware Synergistic Discovery of Drug Combinations: A Large Language Model Perspective
Drug combination therapy with significant advantages is a well-established concept in cancer treatment. Some related efforts have been made with multiple artful deep learning techniques. However, they are usually based on data for drug synergy prediction, ignoring the professional characteristics of data and the systematic knowledge accumulation. Meanwhile, integrating the dispersed professional knowledge and effectively utilizing it in data remains a crucial technical challenge. In this study, we propose KSDDC, a novel model for knowledge-aware synergistic discovery of drug combinations from a large language model (LLM) perspective (i.e., from the continuously learnable and refined large database). Within this framework, three main modules are well-designed, i.e., knowledge-aware drug feature auto-encoding, knowledge-aware cell line feature encoding and drug-drug synergy prediction. Informative embeddings of samples are discovered and combined to make accurate drug synergy prediction. Overall, KSDDC is superior compared with the other shallow machine learning based methods and deep learning based methods on several synergistic prediction benchmarks, where about 19% F1-score improvements over the second best method on DrugComb_1 can be observed. Starting with drug synergy prediction, our studies with knowledge-enabled data mining offer valuable insights and serve as a reference method for future research in this field.
- 1 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. 29 (7) , 4847-4860
“Pre-Train, Prompt” Framework to Boost Graph Neural Networks Performance in EEG Analysis
Electroencephalography (EEG) is a vital noninvasive technique used in neuroscience research and clinical diagnosis. However, EEG data have a complex nonEuclidean structure and are often scarce, making training effective graph neural network (GNN) models difficult. We propose a "pre-train, prompt" framework in graph neural networks for EEG analysis, called GNN-based EEG Prompt Learning (GEPL). The framework first uses unsupervised contrastive learning to pre-train on a large-scale EEG dataset. It then transfers the generic EEG knowledge learned by the model to target EEG datasets through graph prompt learning, thereby enhancing the model's performance with a limited amount of EEG data from the target domain. We tested the framework on five EEG datasets, and the results showed that GEPL outperformed traditional fine-tuning methods in classification accuracy and area under the ROC curve (AUC). GEPL demonstrated improved generalization, robustness, and computational efficiency, thereby significantly reducing the overfitting risks associated with limited EEG data. Moreover, the model provided interpretable results, highlighting relevant brain regions during classification tasks. This research suggests that the "pre-train, prompt" paradigm is well-suited for EEG analysis and offers potential applications in other domains where data are limited.
- 1 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. 29 (7) , 4847-4860
“Pre-Train, Prompt” Framework to Boost Graph Neural Networks Performance in EEG Analysis
Electroencephalography (EEG) is a vital noninvasive technique used in neuroscience research and clinical diagnosis. However, EEG data have a complex nonEuclidean structure and are often scarce, making training effective graph neural network (GNN) models difficult. We propose a "pre-train, prompt" framework in graph neural networks for EEG analysis, called GNN-based EEG Prompt Learning (GEPL). The framework first uses unsupervised contrastive learning to pre-train on a large-scale EEG dataset. It then transfers the generic EEG knowledge learned by the model to target EEG datasets through graph prompt learning, thereby enhancing the model's performance with a limited amount of EEG data from the target domain. We tested the framework on five EEG datasets, and the results showed that GEPL outperformed traditional fine-tuning methods in classification accuracy and area under the ROC curve (AUC). GEPL demonstrated improved generalization, robustness, and computational efficiency, thereby significantly reducing the overfitting risks associated with limited EEG data. Moreover, the model provided interpretable results, highlighting relevant brain regions during classification tasks. This research suggests that the "pre-train, prompt" paradigm is well-suited for EEG analysis and offers potential applications in other domains where data are limited.
- 23 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 35 (10) , 10213-10224
Homophily Induced Contrastive Attributed Graph Clustering
Attributed graph clustering, aiming to discover the underlying graph structure and partition the graph nodes into several disjoint categories, is a basic task in graph data analysis. Although recent efforts over graph contrastive clustering have achieved decent performances, most of them get accustomed to construct the positive neighbor set by the generated pseudo clustering information, directly ignoring the ready-made neighbor nodes and the underlying semantics of edges in a graph. How to well deal with the graph neighbor-specific information to facilitate the performance of graph contrastive clustering is still a challenging problem. Therefore, in this work, we propose a novel Homophily Induced Contrastive Attributed Graph Clustering (HomoCAGC) method, where the power of homophily is exploited in facilitating the performance of contrastive attributed graph clustering, while the pseudo homophily in a graph is also explored and distinguished. Especially, the node feature as reliable information guidance is used to compute the underlying feature-oriented pair-wise node similarity, based on which the positive node pairs in contrastive regularizer are adjusted for better node representation characterization. According to the refined node representations, a triplet self-supervised clustering objective is well-designed to ensure the output embedding is cluster-oriented, and suitable for the downstream clustering task. Extensive experiments on seven benchmarks are conducted to demonstrate the effectiveness of HomoCAGC.
- 23 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Circuits and Systems for Video Technology
- Vol. 35 (10) , 10213-10224
Homophily Induced Contrastive Attributed Graph Clustering
Attributed graph clustering, aiming to discover the underlying graph structure and partition the graph nodes into several disjoint categories, is a basic task in graph data analysis. Although recent efforts over graph contrastive clustering have achieved decent performances, most of them get accustomed to construct the positive neighbor set by the generated pseudo clustering information, directly ignoring the ready-made neighbor nodes and the underlying semantics of edges in a graph. How to well deal with the graph neighbor-specific information to facilitate the performance of graph contrastive clustering is still a challenging problem. Therefore, in this work, we propose a novel Homophily Induced Contrastive Attributed Graph Clustering (HomoCAGC) method, where the power of homophily is exploited in facilitating the performance of contrastive attributed graph clustering, while the pseudo homophily in a graph is also explored and distinguished. Especially, the node feature as reliable information guidance is used to compute the underlying feature-oriented pair-wise node similarity, based on which the positive node pairs in contrastive regularizer are adjusted for better node representation characterization. According to the refined node representations, a triplet self-supervised clustering objective is well-designed to ensure the output embedding is cluster-oriented, and suitable for the downstream clustering task. Extensive experiments on seven benchmarks are conducted to demonstrate the effectiveness of HomoCAGC.
- 26 March 2025
- Published by Elsevier in Information Fusion
- Vol. 121
Triplets-based large-scale multi-view spectral clustering
- 26 March 2025
- Published by Elsevier in Information Fusion
- Vol. 121
Triplets-based large-scale multi-view spectral clustering