Yanmei Zhang
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Yanmei Zhang has a total of 76 co-authors. Here are the top co-authors by number of shared publications:
- Guan Yuan
China University of Mining and Technology
18 shared publications Shujuan Jiang
9 shared publications- Shujuan Jiang
School of Computer Science and Technology, Engineering Research Center of Mine Digitalization Ministry of Education, China University of Mining and Technology, Xuzhou, China
7 shared publications Yanru Ding
School of Computer Science and Technology, China University of Mining and Technology, Xuzhou, China
7 shared publicationsGuan Yuan
School of Computer Science and Technolog, China University of Mining and Technology
6 shared publications- Guixian Zhang
China University of Mining and Technology
4 shared publications
Latest Publications
- 26 September 2026
- Published by World Scientific Pub Co Pte Ltd in International Journal of Software Engineering and Knowledge Engineering
- p. 1-34
Applying Cost-Sensitive Learning to Mitigate the Effect of Class Imbalance for Deep Learning-Based Fault Localization
Automated software fault localization has become a research focus with the growing software scale. Deep learning-based fault localization (DLFL) has been widely studied due to its excellent performance. Most existing DLFL methods use the execution traces of the test cases as features. However, the number of passing test cases is far more than that of failed test cases in the practical test suite. The severe class imbalance problem is ignored in existing DLFL methods. To alleviate the impact of the class imbalance on DLFL, we propose applying cost-sensitive learning to set different weights for each sample to optimize the performance of fault localization. We reduce the number of passing test cases and compute the weight for each sample. The sample weight is combined with a loss function to optimize the accuracy of the DLFL method. To evaluate our approach, we conduct within-project and cross-project fault localization experiments on the Defects4J benchmark. Experimental results show our approach outperforms the traditional DLFL and existing data augmentation methods in fault localization (e.g. localizing 133 faults within Top-5). The experimental results show that applying cost-sensitive learning can effectively alleviate the class imbalance problem in DLFL.
- 26 September 2026
- Published by World Scientific Pub Co Pte Ltd in International Journal of Software Engineering and Knowledge Engineering
- p. 1-34
Applying Cost-Sensitive Learning to Mitigate the Effect of Class Imbalance for Deep Learning-Based Fault Localization
Automated software fault localization has become a research focus with the growing software scale. Deep learning-based fault localization (DLFL) has been widely studied due to its excellent performance. Most existing DLFL methods use the execution traces of the test cases as features. However, the number of passing test cases is far more than that of failed test cases in the practical test suite. The severe class imbalance problem is ignored in existing DLFL methods. To alleviate the impact of the class imbalance on DLFL, we propose applying cost-sensitive learning to set different weights for each sample to optimize the performance of fault localization. We reduce the number of passing test cases and compute the weight for each sample. The sample weight is combined with a loss function to optimize the accuracy of the DLFL method. To evaluate our approach, we conduct within-project and cross-project fault localization experiments on the Defects4J benchmark. Experimental results show our approach outperforms the traditional DLFL and existing data augmentation methods in fault localization (e.g. localizing 133 faults within Top-5). The experimental results show that applying cost-sensitive learning can effectively alleviate the class imbalance problem in DLFL.
- Open Access
- 9 June 2026
- Published by Elsevier in SSRN Electronic Journal
Beyond Conceptual Path Bias: A Unified Hierarchical Framework for Learning Path Recommendation in Online Educational Systems
Learning Path Recommendation (LPR) aims to generate personalized sequences of concepts and exercises that improve learner mastery in online educational systems. Since effective learning path recommendation depends on modeling prerequisite relations among concepts, existing methods usually infer knowledge structures from learner interaction logs and restrict recommendation to the current target concept. As a result, they may confound observed learning transitions with true prerequisite relations, leading to \textit{Conceptual Path Bias}, and may also cause prolonged stagnation when learners fail to progress on the current concept. To address these limitations, we propose \textbf{UniSys-LPR}, a framework for structure-aware and adaptively recoverable learning path recommendation. UniSys-LPR first constructs prerequisite and semantic similarity graphs with the assistance of Large Language Models (LLMs), and models directional dependencies using Directed Graph Neural Networks (DGCNs) to provide concept-level structural priors. It then tracks the evolution of learner proficiency with a bidirectional Mamba-based sequence encoder. Building on these components, we introduce a Dual-Pool Adaptive Selection (DPAS) mechanism that expands the candidate space from the current target concept to related concepts when progress stalls. In this way, UniSys-LPR couples concept planning and exercise recommendation through shared structural and learner-state representations, supporting both prerequisite-consistent progression and adaptive detours. Experiments on three real-world datasets show that UniSys-LPR consistently outperforms strong baselines in learning gain.
- 9 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessBeyond Conceptual Path Bias: A Unified Hierarchical Framework for Learning Path Recommendation in Online Educational Systems
Learning Path Recommendation (LPR) aims to generate personalized sequences of concepts and exercises that improve learner mastery in online educational systems. Since effective learning path recommendation depends on modeling prerequisite relations among concepts, existing methods usually infer knowledge structures from learner interaction logs and restrict recommendation to the current target concept. As a result, they may confound observed learning transitions with true prerequisite relations, leading to \textit{Conceptual Path Bias}, and may also cause prolonged stagnation when learners fail to progress on the current concept. To address these limitations, we propose \textbf{UniSys-LPR}, a framework for structure-aware and adaptively recoverable learning path recommendation. UniSys-LPR first constructs prerequisite and semantic similarity graphs with the assistance of Large Language Models (LLMs), and models directional dependencies using Directed Graph Neural Networks (DGCNs) to provide concept-level structural priors. It then tracks the evolution of learner proficiency with a bidirectional Mamba-based sequence encoder. Building on these components, we introduce a Dual-Pool Adaptive Selection (DPAS) mechanism that expands the candidate space from the current target concept to related concepts when progress stalls. In this way, UniSys-LPR couples concept planning and exercise recommendation through shared structural and learner-state representations, supporting both prerequisite-consistent progression and adaptive detours. Experiments on three real-world datasets show that UniSys-LPR consistently outperforms strong baselines in learning gain.
- 27 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 75 (00189456) , 1-11
Adaptive Cross-Domain Fault Diagnosis of Bearings Using Momentum Contrastive Learning for Nonstationary Data
To address the challenges of bearing fault diagnosis under variable operating conditions and limited labeled data in the target domain, we propose a domain adaptation model based on momentum contrastive learning. The model first uses momentum contrastive learning to extract robust contextual representations, and then applies nearest-neighbor contrastive learning to align these representations between source and target domains. To further ensure spatial structure consistency across domains, we incorporate maximum mean discrepancy and domain-adversarial learning, effectively mitigating the impact of diverse working conditions and data scarcity. In addition, we introduce a balanced dynamic adaptive threshold to improve the model’s resistance to noise, and employ wavelet-based weight initialization to enhance its extrapolation capability. Extensive experiments, including comparative studies, parameter analysis, and ablation studies under various noise levels, are conducted on datasets to validate our model’s effectiveness. This work significantly advances the practical application of bearing fault diagnosis in industrial settings by enhancing noise resistance and robustness.
- 27 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 75 (00189456) , 1-11
Adaptive Cross-Domain Fault Diagnosis of Bearings Using Momentum Contrastive Learning for Nonstationary Data
To address the challenges of bearing fault diagnosis under variable operating conditions and limited labeled data in the target domain, we propose a domain adaptation model based on momentum contrastive learning. The model first uses momentum contrastive learning to extract robust contextual representations, and then applies nearest-neighbor contrastive learning to align these representations between source and target domains. To further ensure spatial structure consistency across domains, we incorporate maximum mean discrepancy and domain-adversarial learning, effectively mitigating the impact of diverse working conditions and data scarcity. In addition, we introduce a balanced dynamic adaptive threshold to improve the model’s resistance to noise, and employ wavelet-based weight initialization to enhance its extrapolation capability. Extensive experiments, including comparative studies, parameter analysis, and ablation studies under various noise levels, are conducted on datasets to validate our model’s effectiveness. This work significantly advances the practical application of bearing fault diagnosis in industrial settings by enhancing noise resistance and robustness.
- Open Access
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 4093-4104
Harnessing LLM for Noise-Robust Cognitive Diagnosis in Web-Based Intelligent Education Systems
Cognitive diagnostics in the Web-based Intelligent Education System (WIES) aims to assess students' mastery of knowledge concepts from heterogeneous, noisy interactions. Recent work has tried to utilize Large Language Models (LLMs) for cognitive diagnosis, yet LLMs struggle with structured data and are prone to noise-induced misjudgments. Specially, WIES's open environment continuously attracts new students and produces vast amounts of response logs, exacerbating the data imbalance and noise issues inherent in traditional educational systems. To address these challenges, we propose DLLM, a Diffusion-based LLM framework for noise-robust cognitive diagnosis. DLLM first constructs independent subgraphs based on response correctness, then applies relation augmentation alignment module to mitigate data imbalance. The two subgraph representations are then fused and aligned with LLM-derived, semantically augmented representations. Importantly, before each alignment step, DLLM employs a two-stage denoising diffusion module to eliminate intrinsic noise while assisting structural representation alignment. Specifically, unconditional denoising diffusion first removes erroneous information, followed by conditional denoising diffusion based on graph signal to eliminate misleading information. Finally, the noise-robust representation that integrates semantic knowledge and structural information is fed into existing cognitive diagnosis models for prediction. Experimental results on three publicly available web-based educational platform datasets demonstrate that our DLLM achieves optimal predictive performance across varying noise levels, which demonstrates that DLLM achieves noise robustness while effectively leveraging semantic knowledge from LLM.
- 12 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 4093-4104
Open AccessHarnessing LLM for Noise-Robust Cognitive Diagnosis in Web-Based Intelligent Education Systems
Cognitive diagnostics in the Web-based Intelligent Education System (WIES) aims to assess students' mastery of knowledge concepts from heterogeneous, noisy interactions. Recent work has tried to utilize Large Language Models (LLMs) for cognitive diagnosis, yet LLMs struggle with structured data and are prone to noise-induced misjudgments. Specially, WIES's open environment continuously attracts new students and produces vast amounts of response logs, exacerbating the data imbalance and noise issues inherent in traditional educational systems. To address these challenges, we propose DLLM, a Diffusion-based LLM framework for noise-robust cognitive diagnosis. DLLM first constructs independent subgraphs based on response correctness, then applies relation augmentation alignment module to mitigate data imbalance. The two subgraph representations are then fused and aligned with LLM-derived, semantically augmented representations. Importantly, before each alignment step, DLLM employs a two-stage denoising diffusion module to eliminate intrinsic noise while assisting structural representation alignment. Specifically, unconditional denoising diffusion first removes erroneous information, followed by conditional denoising diffusion based on graph signal to eliminate misleading information. Finally, the noise-robust representation that integrates semantic knowledge and structural information is fed into existing cognitive diagnosis models for prediction. Experimental results on three publicly available web-based educational platform datasets demonstrate that our DLLM achieves optimal predictive performance across varying noise levels, which demonstrates that DLLM achieves noise robustness while effectively leveraging semantic knowledge from LLM.
- 3 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Software Engineering
- Vol. 52 (5) , 1688-1703
DyCITO+: Scalable Deep Reinforcement Learning for Generating Class Integration Test Orders of Java Programs
Class Integration Test Order (CITO) generation is essential to minimize testing cost in object-oriented software.Traditional methods based on static dependencies often producesuboptimal results, while recent approaches that incorporatedynamic dependencies typically neglect accurate stubbing costestimation and face scalability challenges. We propose DyCITO+,an extension of DyCITO, which originally modeled CITO generationas a Reinforcement Learning (RL) problem using Qlearning.However, DyCITO relies on tabular methods, and thislimits its scalability. DyCITO+ addresses this by introducingthree Deep Reinforcement Learning (DRL) algorithms: DeepQ-Network (DQN), Proximal Policy Optimization (PPO), andAdvantage Actor-Critic (A2C), to handle the complexity oflarge-scale systems more effectively. DyCITO+ builds on thedynamic dependency analysis mechanism from DyCITO, whichcaptures more accurate runtime relationships, including interfaceimplementation, abstract class inheritance, method overriding,and multilevel inheritance. We evaluated DyCITO+ on eightJava programs of varying sizes. The results show that DyCITO+significantly improves the scalability and effectiveness of CITOgeneration. Among the three DRL methods, A2C consistentlyproduces the lowest overall stubbing complexity, particularly inmedium- and large-scale systems.
- 3 April 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Software Engineering
- Vol. 52 (5) , 1688-1703
DyCITO+: Scalable Deep Reinforcement Learning for Generating Class Integration Test Orders of Java Programs
Class Integration Test Order (CITO) generation is essential to minimize testing cost in object-oriented software.Traditional methods based on static dependencies often producesuboptimal results, while recent approaches that incorporatedynamic dependencies typically neglect accurate stubbing costestimation and face scalability challenges. We propose DyCITO+,an extension of DyCITO, which originally modeled CITO generationas a Reinforcement Learning (RL) problem using Qlearning.However, DyCITO relies on tabular methods, and thislimits its scalability. DyCITO+ addresses this by introducingthree Deep Reinforcement Learning (DRL) algorithms: DeepQ-Network (DQN), Proximal Policy Optimization (PPO), andAdvantage Actor-Critic (A2C), to handle the complexity oflarge-scale systems more effectively. DyCITO+ builds on thedynamic dependency analysis mechanism from DyCITO, whichcaptures more accurate runtime relationships, including interfaceimplementation, abstract class inheritance, method overriding,and multilevel inheritance. We evaluated DyCITO+ on eightJava programs of varying sizes. The results show that DyCITO+significantly improves the scalability and effectiveness of CITOgeneration. Among the three DRL methods, A2C consistentlyproduces the lowest overall stubbing complexity, particularly inmedium- and large-scale systems.
- Open Access
- 23 March 2026
- Published by Elsevier in SSRN Electronic Journal
Applying Cost-sensitive Learning to Mitigate the Effect of Class Imbalance for Deep Learning-based Fault Localization
Context: Automated software fault localization has become a research focus with the growing software scale. Deep learning-based fault localization (DLFL) has been widely studied due to its excellent performance. Most existing DLFL methods use the execution traces of the test cases as features. However, the number of passing test cases is far more than that of failed test cases in the practical test suite. The severe class imbalance problem is ignored in existing DLFL methods.Objective: To alleviate the impact of the class imbalance on DLFL, we propose applying cost-sensitive learning to set different weights for each sample to optimize the performance of fault localization.Method: We reduce the number of passing test cases and compute the weight for each sample. The sample weight is combined with a loss function to optimize the accuracy of the DLFL method.Results: To evaluate our approach, we conduct within-project and cross-project fault localization experiments on the Defects4J benchmark. Experimental results show our approach outperforms the traditional DLFL and existing data augmentation methods in fault localization (e.g., localizing 133 faults within Top-5).Conclusion: The experimental results show that applying cost-sensitive learning can effectively alleviate the class imbalance problem in DLFL and improve the effectiveness of fault localization.
- 23 March 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessApplying Cost-sensitive Learning to Mitigate the Effect of Class Imbalance for Deep Learning-based Fault Localization
Context: Automated software fault localization has become a research focus with the growing software scale. Deep learning-based fault localization (DLFL) has been widely studied due to its excellent performance. Most existing DLFL methods use the execution traces of the test cases as features. However, the number of passing test cases is far more than that of failed test cases in the practical test suite. The severe class imbalance problem is ignored in existing DLFL methods.Objective: To alleviate the impact of the class imbalance on DLFL, we propose applying cost-sensitive learning to set different weights for each sample to optimize the performance of fault localization.Method: We reduce the number of passing test cases and compute the weight for each sample. The sample weight is combined with a loss function to optimize the accuracy of the DLFL method.Results: To evaluate our approach, we conduct within-project and cross-project fault localization experiments on the Defects4J benchmark. Experimental results show our approach outperforms the traditional DLFL and existing data augmentation methods in fault localization (e.g., localizing 133 faults within Top-5).Conclusion: The experimental results show that applying cost-sensitive learning can effectively alleviate the class imbalance problem in DLFL and improve the effectiveness of fault localization.
- Open Access
Harnessing LLM for Noise-Robust Cognitive Diagnosis in Web-Based Intelligent Education Systems
Cognitive diagnostics in the Web-based Intelligent Education System (WIES) aims to assess students' mastery of knowledge concepts from heterogeneous, noisy interactions. Recent work has tried to utilize Large Language Models (LLMs) for cognitive diagnosis, yet LLMs struggle with structured data and are prone to noise-induced misjudgments. Specially, WIES's open environment continuously attracts new students and produces vast amounts of response logs, exacerbating the data imbalance and noise issues inherent in traditional educational systems. To address these challenges, we propose DLLM, a Diffusion-based LLM framework for noise-robust cognitive diagnosis. DLLM first constructs independent subgraphs based on response correctness, then applies relation augmentation alignment module to mitigate data imbalance. The two subgraph representations are then fused and aligned with LLM-derived, semantically augmented representations. Importantly, before each alignment step, DLLM employs a two-stage denoising diffusion module to eliminate intrinsic noise while assisting structural representation alignment. Specifically, unconditional denoising diffusion first removes erroneous information, followed by conditional denoising diffusion based on graph-guided to eliminate misleading information. Finally, the noise-robust representation that integrates semantic knowledge and structural information is fed into existing cognitive diagnosis models for prediction. Experimental results on three publicly available web-based educational platform datasets demonstrate that our DLLM achieves optimal predictive performance across varying noise levels, which demonstrates that DLLM achieves noise robustness while effectively leveraging semantic knowledge from LLM.
Open AccessHarnessing LLM for Noise-Robust Cognitive Diagnosis in Web-Based Intelligent Education Systems
Cognitive diagnostics in the Web-based Intelligent Education System (WIES) aims to assess students' mastery of knowledge concepts from heterogeneous, noisy interactions. Recent work has tried to utilize Large Language Models (LLMs) for cognitive diagnosis, yet LLMs struggle with structured data and are prone to noise-induced misjudgments. Specially, WIES's open environment continuously attracts new students and produces vast amounts of response logs, exacerbating the data imbalance and noise issues inherent in traditional educational systems. To address these challenges, we propose DLLM, a Diffusion-based LLM framework for noise-robust cognitive diagnosis. DLLM first constructs independent subgraphs based on response correctness, then applies relation augmentation alignment module to mitigate data imbalance. The two subgraph representations are then fused and aligned with LLM-derived, semantically augmented representations. Importantly, before each alignment step, DLLM employs a two-stage denoising diffusion module to eliminate intrinsic noise while assisting structural representation alignment. Specifically, unconditional denoising diffusion first removes erroneous information, followed by conditional denoising diffusion based on graph-guided to eliminate misleading information. Finally, the noise-robust representation that integrates semantic knowledge and structural information is fed into existing cognitive diagnosis models for prediction. Experimental results on three publicly available web-based educational platform datasets demonstrate that our DLLM achieves optimal predictive performance across varying noise levels, which demonstrates that DLLM achieves noise robustness while effectively leveraging semantic knowledge from LLM.
- Open Access
- 1 September 2025
- Published by International Joint Conferences on Artificial Intelligence
- Vol. 10, 8563-8571
RobustHAR: Multi-scale Spatial-temporal Masked Self-supervised Pre-training for Robust Human Activity Recognition
Human activity recognition (HAR) is prone to performance degradation in real-world applications due to data missing between intra-sensor and inter-sensor channels. Masked modeling, as one mainstream paradigm of self-supervised pre-training, can learn robust representations across sensors in the data missing scenario by reconstructing the masked content based on the unmasked part. However, the existing methods predominantly emphasize the temporal dynamics of human activities, which limits their ability to effectively capture the spatial interdependencies among multiple sensors. Besides, different human activities often span across various spatial-temporal scales, which results in activity recognizer failing to capture intricate spatial-temporal semantic information. To address these issues, we propose RobustHAR, a new HAR model with multi-scale spatial-temporal masked self-supervised pre-training designed to improve model performance on the data missing context. RobustHAR involves three main steps: (1) RobustHAR constructs location-inspired spatial-temporal 3D-variation modeling to capture spatial-temporal correlated information in human activity data. (2) RobustHAR then designs multi-scale spatial-temporal masked self-supervised pre-training with semantic-consistent multi-scale feature co-learning for learning robust features at different scales. (3) Finally, RobustHAR fine-tunes the pretraining model with adaptive multi-scale feature fusion for human activity recognition. Extensive experiments on three public multi-sensor datasets demonstrate that RobustHAR outperforms existing state-of-the-art methods.
- 1 September 2025
- Published by International Joint Conferences on Artificial Intelligence
- Vol. 10, 8563-8571
Open AccessRobustHAR: Multi-scale Spatial-temporal Masked Self-supervised Pre-training for Robust Human Activity Recognition
Human activity recognition (HAR) is prone to performance degradation in real-world applications due to data missing between intra-sensor and inter-sensor channels. Masked modeling, as one mainstream paradigm of self-supervised pre-training, can learn robust representations across sensors in the data missing scenario by reconstructing the masked content based on the unmasked part. However, the existing methods predominantly emphasize the temporal dynamics of human activities, which limits their ability to effectively capture the spatial interdependencies among multiple sensors. Besides, different human activities often span across various spatial-temporal scales, which results in activity recognizer failing to capture intricate spatial-temporal semantic information. To address these issues, we propose RobustHAR, a new HAR model with multi-scale spatial-temporal masked self-supervised pre-training designed to improve model performance on the data missing context. RobustHAR involves three main steps: (1) RobustHAR constructs location-inspired spatial-temporal 3D-variation modeling to capture spatial-temporal correlated information in human activity data. (2) RobustHAR then designs multi-scale spatial-temporal masked self-supervised pre-training with semantic-consistent multi-scale feature co-learning for learning robust features at different scales. (3) Finally, RobustHAR fine-tunes the pretraining model with adaptive multi-scale feature fusion for human activity recognition. Extensive experiments on three public multi-sensor datasets demonstrate that RobustHAR outperforms existing state-of-the-art methods.
- 15 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-17
Enhancing Multimodal Action Recognition With Semantic-Guided Feature Disentangled and Multichannel Adaptive Feature Fusion
Multimodal action recognition exploits the consistency and complementarity among different modalities to mitigate measurement biases inherent in single-modality data. However, existing methods primarily focus on inter-modal correlation, often at the expense of modality-specific semantic features. This limits them to discerning ambiguous actions when different actions with similar sequences, as crucial modality-specific nuances are lost in the fused representation. To address the aforementioned issue, we propose Feature Disentangled Representation Learning for Multimodal Action Recognition (FDRL-MAR). Unlike previous approaches that blend multimodal information indiscriminately, FDRL-MAR explicitly disentangles action features into modality-common and modality-specific representations. This ensures that crucial modality-specific cues, such as skeleton posture details or inertial motion signatures, remain distinguishable while preserving cross-modal consistency. Furthermore, since action recognition is inherently dynamic, we introduce a multi-channel adaptive fusion mechanism that dynamically adjusts the contribution of different semantic features. Finally, extensive experiments on three benchmark datasets demonstrate the effectiveness of FDRL-MAR over state-of-the-art methods. Our code and data will be released on the GitHub: https://github.com/YooLiuXiao/FDRL-MHAD.
- 15 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Instrumentation and Measurement
- Vol. 74 (00189456) , 1-17
Enhancing Multimodal Action Recognition With Semantic-Guided Feature Disentangled and Multichannel Adaptive Feature Fusion
Multimodal action recognition exploits the consistency and complementarity among different modalities to mitigate measurement biases inherent in single-modality data. However, existing methods primarily focus on inter-modal correlation, often at the expense of modality-specific semantic features. This limits them to discerning ambiguous actions when different actions with similar sequences, as crucial modality-specific nuances are lost in the fused representation. To address the aforementioned issue, we propose Feature Disentangled Representation Learning for Multimodal Action Recognition (FDRL-MAR). Unlike previous approaches that blend multimodal information indiscriminately, FDRL-MAR explicitly disentangles action features into modality-common and modality-specific representations. This ensures that crucial modality-specific cues, such as skeleton posture details or inertial motion signatures, remain distinguishable while preserving cross-modal consistency. Furthermore, since action recognition is inherently dynamic, we introduce a multi-channel adaptive fusion mechanism that dynamically adjusts the contribution of different semantic features. Finally, extensive experiments on three benchmark datasets demonstrate the effectiveness of FDRL-MAR over state-of-the-art methods. Our code and data will be released on the GitHub: https://github.com/YooLiuXiao/FDRL-MHAD.