Jin Cao
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
Jin Cao has a total of 32 co-authors. Here are the top co-authors by number of shared publications:
- Ta Zhou
School of Computing, Jiangsu University of Science and Technology, Zhenjiang, China
3 shared publications - Jing Cai
The Hong Kong Polytechnic University
2 shared publications Defang He
1 shared publication- Defeng Sun
Department of Applied Mathematics, The Hong Kong Polytechnic University, Hong Kong, China
1 shared publication Ge Ren
1 shared publicationHaifeng Fang
1 shared publication
Latest Publications
- Open Access
- 21 July 2026
- Published by Elsevier in SSRN Electronic Journal
STR-BNet: Joint Spatiotemporal Representation Learning and Adaptive Brain Network Modeling for EEG Recognition
Multichannel electroencephalography (EEG) reflects the spatiotemporal orchestration of brain activity. However, conventional methods typically decouple raw temporal feature mining from spatial brain network modeling, failing to capture the intrinsic synergy between neural timing and topological organization. This disjointed approach creates a critical information bottleneck that obstructs the recognition of holistic brain patterns. To bridge this gap, we propose STR-BNet, an end-to-end framework designed for integrated spatiotemporal dynamic brain pattern analysis. The framework reconciles signal-level representation learning with topology-level network generation through two synergistic modules. First, Latent-Space Spatiotemporal Regularization (STR) employs a Spatiotemporal Transformer and contrastive learning to mine robust semantic features from raw EEG, mitigating the low-level reconstruction bias of traditional self-supervised learning. Second, Prior-Coupled Semantic Mixture-of-Experts Graph (PCSM-G) adaptively constructs global brain networks by dynamically fusing these semantic features with neurophysiological Phase Locking Value priors. Through the unified end-to-end optimization of both feature encoding and adaptive graph construction, STR-BNet facilitates a seamless information flow that preserves the spatiotemporal integrity of brain dynamics. Extensive experiments on two public datasets demonstrate that STR-BNet establishes new state-of-the-art performance in EEG-based emotion recognition, achieving 96.93% accuracy on the SEED dataset and 97.79% accuracy with 97.77% F1-score on the MuSer dataset, while reducing cross-subject F1-score standard deviation to only 1.23%, underscoring its exceptional robustness.
- 21 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessSTR-BNet: Joint Spatiotemporal Representation Learning and Adaptive Brain Network Modeling for EEG Recognition
Multichannel electroencephalography (EEG) reflects the spatiotemporal orchestration of brain activity. However, conventional methods typically decouple raw temporal feature mining from spatial brain network modeling, failing to capture the intrinsic synergy between neural timing and topological organization. This disjointed approach creates a critical information bottleneck that obstructs the recognition of holistic brain patterns. To bridge this gap, we propose STR-BNet, an end-to-end framework designed for integrated spatiotemporal dynamic brain pattern analysis. The framework reconciles signal-level representation learning with topology-level network generation through two synergistic modules. First, Latent-Space Spatiotemporal Regularization (STR) employs a Spatiotemporal Transformer and contrastive learning to mine robust semantic features from raw EEG, mitigating the low-level reconstruction bias of traditional self-supervised learning. Second, Prior-Coupled Semantic Mixture-of-Experts Graph (PCSM-G) adaptively constructs global brain networks by dynamically fusing these semantic features with neurophysiological Phase Locking Value priors. Through the unified end-to-end optimization of both feature encoding and adaptive graph construction, STR-BNet facilitates a seamless information flow that preserves the spatiotemporal integrity of brain dynamics. Extensive experiments on two public datasets demonstrate that STR-BNet establishes new state-of-the-art performance in EEG-based emotion recognition, achieving 96.93% accuracy on the SEED dataset and 97.79% accuracy with 97.77% F1-score on the MuSer dataset, while reducing cross-subject F1-score standard deviation to only 1.23%, underscoring its exceptional robustness.
- 30 August 2025
- Published by World Scientific Pub Co Pte Ltd in International Journal of Pattern Recognition and Artificial Intelligence
- Vol. 39 (15)
A Hierarchical Fuzzy Classifier for Radiation Dermatitis Diagnosis Based on Rule-Feature Reconstruction and Specific Approximation
Radiation dermatitis (RD) represents the primary adverse outcome of ionizing radiation associated with chemotherapy in breast, head and neck, anal cancers and other medical conditions. For nasopharyngeal cancer (NPC) recognition, how to build an interpretable lightweight recognition model is a serious challenge. Traditional medical diagnostic models typically exhibit opaque decision-making processes with limited interpretability, and achieving high classification accuracy often necessitates increased model complexity. In order to address these limitations of traditional methods, the study proposes an innovative stacked hierarchical Takagi–Sugeno–Kang (TSK) fuzzy classifier for diagnosing the severity of RD in NPC patients. The classifier incorporates a low-rank rule feature structure and bidirectional approximation consequents, with each sub-classifier constructed from first-order TSK fuzzy systems. Its antecedent part employs a rule-feature matrix to characterize relationships between fuzzy rules and training features, where matrix decomposition techniques are applied to simplify rule-feature associations and generate representative discriminative rules. The consequent part integrates outputs from preceding sub-classifiers into the objective function, achieving approximation of both prior sub-classifier outputs and target outputs through independent training mechanisms. Finally, comparative experiments demonstrate that each sub-classifier in LRR-TSK requires only 3–15 fuzzy rules yet achieves an average accuracy exceeding 99.6%, highlighting its excellent performance. These improvements stem from the low-rank rule-feature matrix, which generates clear rules via matrix decomposition, and the bidirectional approximation method, which optimizes prediction accuracy, providing a transparent and efficient solution for handling complex medical data. However, the current training models still face the predicament that it is difficult to obtain the hyperparameter combinations of the models and the number of fuzzy rules is hard to determine. In future work, we aim to address these issues by integrating global optimization algorithms to achieve a global optimum solution, and extending the model into a multi-view joint learning framework to enhance diagnostic comprehensiveness.
- 30 August 2025
- Published by World Scientific Pub Co Pte Ltd in International Journal of Pattern Recognition and Artificial Intelligence
- Vol. 39 (15)
A Hierarchical Fuzzy Classifier for Radiation Dermatitis Diagnosis Based on Rule-Feature Reconstruction and Specific Approximation
Radiation dermatitis (RD) represents the primary adverse outcome of ionizing radiation associated with chemotherapy in breast, head and neck, anal cancers and other medical conditions. For nasopharyngeal cancer (NPC) recognition, how to build an interpretable lightweight recognition model is a serious challenge. Traditional medical diagnostic models typically exhibit opaque decision-making processes with limited interpretability, and achieving high classification accuracy often necessitates increased model complexity. In order to address these limitations of traditional methods, the study proposes an innovative stacked hierarchical Takagi–Sugeno–Kang (TSK) fuzzy classifier for diagnosing the severity of RD in NPC patients. The classifier incorporates a low-rank rule feature structure and bidirectional approximation consequents, with each sub-classifier constructed from first-order TSK fuzzy systems. Its antecedent part employs a rule-feature matrix to characterize relationships between fuzzy rules and training features, where matrix decomposition techniques are applied to simplify rule-feature associations and generate representative discriminative rules. The consequent part integrates outputs from preceding sub-classifiers into the objective function, achieving approximation of both prior sub-classifier outputs and target outputs through independent training mechanisms. Finally, comparative experiments demonstrate that each sub-classifier in LRR-TSK requires only 3–15 fuzzy rules yet achieves an average accuracy exceeding 99.6%, highlighting its excellent performance. These improvements stem from the low-rank rule-feature matrix, which generates clear rules via matrix decomposition, and the bidirectional approximation method, which optimizes prediction accuracy, providing a transparent and efficient solution for handling complex medical data. However, the current training models still face the predicament that it is difficult to obtain the hyperparameter combinations of the models and the number of fuzzy rules is hard to determine. In future work, we aim to address these issues by integrating global optimization algorithms to achieve a global optimum solution, and extending the model into a multi-view joint learning framework to enhance diagnostic comprehensiveness.
- 6 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Fuzzy Systems
- Vol. 33 (8) , 2805-2817
A Lightweight TSK Fuzzy Classifier With Quantitative Equivalent Fuzzy Rules via Adaptive Weighting
The first-order Takagi-Sugeno-Kang (TSK) fuzzy classifier with a fully combined fuzzy rule base (FuCo-FRB) is a potent and interpretable classifier for multiple input and multiple output (MIMO) tasks. However, FuCo-FRB possess an exponential increase in the number of fuzzy rules, and it poses challenges for efficient identification of the parameter matrix in MIMO tasks. A lightweight TSK fuzzy classifier (LW-TSK-FC) is proposed to achieve the balance of training efficiency and predictive performance especially on MIMO tasks. It has the following advantages: (1) An adaptive weighting method based on directly connected FRB enables the efficient generation of fuzzy rules, meanwhile retaining the distribution characteristic of FuCo-FRB to reduce information loss. (2) The consequent network of first-order TSK is optimized to a novel series structure by using matrix factorization. This new series structure increases the depth of consequent network and enable the implementation of kernel function and least learning machine (LLM), enhancing the calculation efficiency and predictive performance. (3) There are only weight parameters in consequent network of LW-TSK-FC and these parameters can be identified quantitatively by LLM, leading to a great training efficiency. The comparison experiments with other 14 power classifiers on 13 public UCI datasets shown comparable predictive performance and outstanding training and testing efficiency in both MISO and MIMO tasks. The experiments on a real-world clinical task demonstrated the significant capability of LW-TSK-FC on handling imbalanced small data, and can provide the semantic interpretability for reasoning process.
- 6 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Fuzzy Systems
- Vol. 33 (8) , 2805-2817
A Lightweight TSK Fuzzy Classifier With Quantitative Equivalent Fuzzy Rules via Adaptive Weighting
The first-order Takagi-Sugeno-Kang (TSK) fuzzy classifier with a fully combined fuzzy rule base (FuCo-FRB) is a potent and interpretable classifier for multiple input and multiple output (MIMO) tasks. However, FuCo-FRB possess an exponential increase in the number of fuzzy rules, and it poses challenges for efficient identification of the parameter matrix in MIMO tasks. A lightweight TSK fuzzy classifier (LW-TSK-FC) is proposed to achieve the balance of training efficiency and predictive performance especially on MIMO tasks. It has the following advantages: (1) An adaptive weighting method based on directly connected FRB enables the efficient generation of fuzzy rules, meanwhile retaining the distribution characteristic of FuCo-FRB to reduce information loss. (2) The consequent network of first-order TSK is optimized to a novel series structure by using matrix factorization. This new series structure increases the depth of consequent network and enable the implementation of kernel function and least learning machine (LLM), enhancing the calculation efficiency and predictive performance. (3) There are only weight parameters in consequent network of LW-TSK-FC and these parameters can be identified quantitatively by LLM, leading to a great training efficiency. The comparison experiments with other 14 power classifiers on 13 public UCI datasets shown comparable predictive performance and outstanding training and testing efficiency in both MISO and MIMO tasks. The experiments on a real-world clinical task demonstrated the significant capability of LW-TSK-FC on handling imbalanced small data, and can provide the semantic interpretability for reasoning process.
- 19 January 2025
- Published by Elsevier in Journal of Cleaner Production
- Vol. 492
A lightweight model for plastic classification based on data augmentation
- 19 January 2025
- Published by Elsevier in Journal of Cleaner Production
- Vol. 492
A lightweight model for plastic classification based on data augmentation
- Open Access
- 26 January 2024
- Published by Elsevier in Information Sciences
- Vol. 662
Fuzzy inference system with interpretable fuzzy rules: Advancing explainable artificial intelligence for disease diagnosis—A comprehensive review
- 26 January 2024
- Published by Elsevier in Information Sciences
- Vol. 662
Open AccessFuzzy inference system with interpretable fuzzy rules: Advancing explainable artificial intelligence for disease diagnosis—A comprehensive review
- Open Access
- 18 March 2022
- Published by Wiley in Computational Intelligence and Neuroscience
- Vol. 2022, 1-10
A Small Network MicronNet-BF of Traffic Sign Classification
One of a very significant computer vision task in many real-world applications is traffic sign recognition. With the development of deep neural networks, state-of-art performance traffic sign recognition has been provided in recent five years. Getting very high accuracy in object classification is not a dream any more. However, one of the key challenges is becoming making the deep neural network suitable for an embedded system. As a result, a small neural network with as less parameters as possible and high accuracy needs to be explored. In this paper, the MicronNet which is a small but powerful convolutional neural network is improved by batch normalization and factorization, and the proposed MicronNet-BN-Factorization (MicronNet-BF) takes advantages about reducing parameters and improving accuracy. The effect of image brightness is reduced for feature recognition by the elimination of mean and variance of each input layer in MicronNet via BN. A lower number of parameters are realized with the replacement of convolutional layers in MicronNet, which is the inspiration of factorization. In addition, data augmentation is also been changed to get higher accuracy. Most important, the experiment shows that the accuracy of MicronNet-BF is 99.383% on German traffic sign recognition benchmark (GTSRB) which is much higher than the original MicronNet (98.9%), and the most influence factor is batch normalization after the confirmation of orthogonal experimental. Furthermore, the handsome training efficiency and generality of MicronNet-BF indicate the wide application in embedded scenarios.
- 18 March 2022
- Published by Wiley in Computational Intelligence and Neuroscience
- Vol. 2022, 1-10
Open AccessA Small Network MicronNet-BF of Traffic Sign Classification
One of a very significant computer vision task in many real-world applications is traffic sign recognition. With the development of deep neural networks, state-of-art performance traffic sign recognition has been provided in recent five years. Getting very high accuracy in object classification is not a dream any more. However, one of the key challenges is becoming making the deep neural network suitable for an embedded system. As a result, a small neural network with as less parameters as possible and high accuracy needs to be explored. In this paper, the MicronNet which is a small but powerful convolutional neural network is improved by batch normalization and factorization, and the proposed MicronNet-BN-Factorization (MicronNet-BF) takes advantages about reducing parameters and improving accuracy. The effect of image brightness is reduced for feature recognition by the elimination of mean and variance of each input layer in MicronNet via BN. A lower number of parameters are realized with the replacement of convolutional layers in MicronNet, which is the inspiration of factorization. In addition, data augmentation is also been changed to get higher accuracy. Most important, the experiment shows that the accuracy of MicronNet-BF is 99.383% on German traffic sign recognition benchmark (GTSRB) which is much higher than the original MicronNet (98.9%), and the most influence factor is batch normalization after the confirmation of orthogonal experimental. Furthermore, the handsome training efficiency and generality of MicronNet-BF indicate the wide application in embedded scenarios.
- Open Access
Style Attuned Pre-training and Parameter Efficient Fine-tuning for Spoken Language Understanding
Neural models have yielded state-of-the-art results in deciphering spoken language understanding (SLU) problems; however, these models require a significant amount of domain-specific labeled examples for training, which is prohibitively expensive. While pre-trained language models like BERT have been shown to capture a massive amount of knowledge by learning from unlabeled corpora and solve SLU using fewer labeled examples for adaption, the encoding of knowledge is implicit and agnostic to downstream tasks. Such encoding results in model inefficiencies in parameter usage: an entirely new model is required for every domain. To address these challenges, we introduce a novel SLU framework, comprising a conversational language modeling (CLM) pre-training task and a light encoder architecture. The CLM pre-training enables networks to capture the representation of the language in conversation style with the presence of ASR errors. The light encoder architecture separates the shared pre-trained networks from the mappings of generally encoded knowledge to specific domains of SLU, allowing for the domain adaptation to be performed solely at the light encoder and thus increasing efficiency. With the framework, we match the performance of state-of-the-art SLU results on Alexa internal datasets and on two public ones (ATIS, SNIPS), adding only 4.4% parameters per task.
Open AccessStyle Attuned Pre-training and Parameter Efficient Fine-tuning for Spoken Language Understanding
Neural models have yielded state-of-the-art results in deciphering spoken language understanding (SLU) problems; however, these models require a significant amount of domain-specific labeled examples for training, which is prohibitively expensive. While pre-trained language models like BERT have been shown to capture a massive amount of knowledge by learning from unlabeled corpora and solve SLU using fewer labeled examples for adaption, the encoding of knowledge is implicit and agnostic to downstream tasks. Such encoding results in model inefficiencies in parameter usage: an entirely new model is required for every domain. To address these challenges, we introduce a novel SLU framework, comprising a conversational language modeling (CLM) pre-training task and a light encoder architecture. The CLM pre-training enables networks to capture the representation of the language in conversation style with the presence of ASR errors. The light encoder architecture separates the shared pre-trained networks from the mappings of generally encoded knowledge to specific domains of SLU, allowing for the domain adaptation to be performed solely at the light encoder and thus increasing efficiency. With the framework, we match the performance of state-of-the-art SLU results on Alexa internal datasets and on two public ones (ATIS, SNIPS), adding only 4.4% parameters per task.