Zhiyi Song
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Zhiyi Song has a total of 14 co-authors. Here are the top co-authors by number of shared publications:
- Aoqi Wang
School of Automation, Beijing Information Science and Technology University, Beijing 102206, China
1 shared publication - Bingnan Liu
School of Business, Macau University of Science and Technology, Taipa, Macau 999078, China
1 shared publication - Guo Chen
College of Earth Science and Technology, Southwest Petroleum University, Chengdu 610500, China
1 shared publication - Huiji Wang
School of Humanities and Public Administration, Jiangxi Agricultural University, Nanchang 330045, China
1 shared publication Jiaen Lin
Beijing Information Science and Technology University
1 shared publication- Jiaen Lin
School of Automation, Beijing Information Science and Technology University, Beijing 102206, China
1 shared publication
Latest Publications
- Open Access
- 29 July 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (15) , 7535
Chaos-Driven Evolutionary Optimization of a Hybrid Machine Learning Model for Accurate Heart Disease Prediction
In this paper, a hybrid machine learning approach is proposed for accurate prediction of heart disease. The features are categorized into numerical and categorical types, which are extracted independently by a deep neural network (DNN) and CatBoost, respectively, before being fused via logistic regression to produce the output; simultaneously, an improved chaotic evolutionary algorithm (CEO) is employed to optimise hyperparameters, and adaptive chaotic perturbations and a restart mechanism are introduced to enhance global optimisation capabilities. The integration of these components facilitates the successful development of a hybrid model, referred to as the CEO-Enhanced Hybrid CatBoost and Deep Learning Model (CEO-HCDLM), for effective heart disease classification. To assess model stability and generalizability, multi-seed validation, Gaussian noise injection, and cross-dataset validation using the Kaggle heart disease dataset were conducted. The results demonstrate the efficacy and practicality of this approach in supporting decision-making for cardiovascular risk assessment.
- 29 July 2026
- Published by MDPI AG in Applied Sciences
- Vol. 16 (15) , 7535
Open AccessChaos-Driven Evolutionary Optimization of a Hybrid Machine Learning Model for Accurate Heart Disease Prediction
In this paper, a hybrid machine learning approach is proposed for accurate prediction of heart disease. The features are categorized into numerical and categorical types, which are extracted independently by a deep neural network (DNN) and CatBoost, respectively, before being fused via logistic regression to produce the output; simultaneously, an improved chaotic evolutionary algorithm (CEO) is employed to optimise hyperparameters, and adaptive chaotic perturbations and a restart mechanism are introduced to enhance global optimisation capabilities. The integration of these components facilitates the successful development of a hybrid model, referred to as the CEO-Enhanced Hybrid CatBoost and Deep Learning Model (CEO-HCDLM), for effective heart disease classification. To assess model stability and generalizability, multi-seed validation, Gaussian noise injection, and cross-dataset validation using the Kaggle heart disease dataset were conducted. The results demonstrate the efficacy and practicality of this approach in supporting decision-making for cardiovascular risk assessment.
- 19 June 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (13)
A Kolmogorov-Arnold Network enhanced Tabular Parameter-Efficient Ensembling Model with SHAP Analysis for Athlete Overtraining Binary Classification
In this paper, an interpretable hybrid machine learning framework is proposed for the binary classification of athlete overtraining status. Using the publicly available Kaggle athlete overtraining dataset, the input features are first preprocessed through a standardised workflow. Since all input variables are represented as numerical or binary numerical features, no additional categorical feature encoding is required. Subsequently, a dual-layer Kolmogorov-Arnold Network (KAN) module is employed to enhance the nonlinear representation of the standardised tabular features. The original standardised features and the KAN-enhanced features are then concatenated and fed into TabM, a parameter-efficient ensemble model designed for tabular data classification.To improve model interpretability, Shapley Additive Explanations (SHAP) are introduced to quantify the contribution of each input feature to the prediction outcomes. By integrating KAN-based nonlinear feature enhancement, TabM-based parameter-efficient ensemble classification, and SHAP-based interpretability analysis, the proposed KAN-TabM framework aims to provide an effective and interpretable approach for athlete overtraining identification. To evaluate the model, stratified five-fold cross-validation and ablation experiments are conducted, and the proposed framework is compared with baseline models including Random Forest, Multilayer Perceptron (MLP), XGBoost, CatBoost, LightGBM, standalone KAN, and standalone TabM. The results indicate that the proposed framework can support athlete overtraining risk assessment, training load regulation, and recovery management in an interpretable decision-support manner.
- 19 June 2026
- Published by IOP Publishing in Engineering Research Express
- Vol. 8 (13)
A Kolmogorov-Arnold Network enhanced Tabular Parameter-Efficient Ensembling Model with SHAP Analysis for Athlete Overtraining Binary Classification
In this paper, an interpretable hybrid machine learning framework is proposed for the binary classification of athlete overtraining status. Using the publicly available Kaggle athlete overtraining dataset, the input features are first preprocessed through a standardised workflow. Since all input variables are represented as numerical or binary numerical features, no additional categorical feature encoding is required. Subsequently, a dual-layer Kolmogorov-Arnold Network (KAN) module is employed to enhance the nonlinear representation of the standardised tabular features. The original standardised features and the KAN-enhanced features are then concatenated and fed into TabM, a parameter-efficient ensemble model designed for tabular data classification.To improve model interpretability, Shapley Additive Explanations (SHAP) are introduced to quantify the contribution of each input feature to the prediction outcomes. By integrating KAN-based nonlinear feature enhancement, TabM-based parameter-efficient ensemble classification, and SHAP-based interpretability analysis, the proposed KAN-TabM framework aims to provide an effective and interpretable approach for athlete overtraining identification. To evaluate the model, stratified five-fold cross-validation and ablation experiments are conducted, and the proposed framework is compared with baseline models including Random Forest, Multilayer Perceptron (MLP), XGBoost, CatBoost, LightGBM, standalone KAN, and standalone TabM. The results indicate that the proposed framework can support athlete overtraining risk assessment, training load regulation, and recovery management in an interpretable decision-support manner.
- Open Access
- 2 November 2025
- Published by MDPI AG in Biomimetics
- Vol. 10 (11) , 735
An Economic Scheduling Management Method for Microgrids Using Multi-Strategy Improved Sand Cat Swarm Optimization
With the rise of the digital economy, energy management has become increasingly intelligent and data-driven. Environmental, Social, and Governance (ESG) considerations have emerged as a key driver of corporate competitiveness, while microgrid scheduling serves as an essential pathway for enterprises to achieve carbon reduction, attract green investment, and meet low-carbon development goals. However, traditional microgrid economic dispatch algorithms often suffer from low optimization efficiency, limited scalability, and poor flexibility. To address these challenges, this paper proposes a multi-strategy improved sand cat swarm optimization (MISCSO) algorithm for the economic scheduling of microgrids. First, a distribution-optimized initialization method based on adaptive diversity guidance is developed to enhance the quality of the initial population. This approach improves algorithmic performance by generating individuals in high-potential regions to ensure solution quality while maintaining population diversity through the inclusion of individuals from low-potential regions. Subsequently, an elite-centered global random movement strategy is introduced to balance elite guidance and global exploration, thereby improving both convergence speed and optimization accuracy. In addition, an adaptive elastic boundary mapping mechanism is proposed to effectively handle boundary violations, striking a balance between boundary constraints and global search capability. To evaluate the effectiveness of MISCSO, it is compared with 11 state-of-the-art algorithms using the IEEE CEC2017 benchmark set, and statistical analyses are conducted to assess performance differences. Experimental results demonstrate that MISCSO achieves superior optimization accuracy, convergence performance, and robustness. Finally, the applicability of MISCSO is verified through its implementation in microgrid economic scheduling, where it achieves outstanding optimization results.
- 2 November 2025
- Published by MDPI AG in Biomimetics
- Vol. 10 (11) , 735
Open AccessAn Economic Scheduling Management Method for Microgrids Using Multi-Strategy Improved Sand Cat Swarm Optimization
With the rise of the digital economy, energy management has become increasingly intelligent and data-driven. Environmental, Social, and Governance (ESG) considerations have emerged as a key driver of corporate competitiveness, while microgrid scheduling serves as an essential pathway for enterprises to achieve carbon reduction, attract green investment, and meet low-carbon development goals. However, traditional microgrid economic dispatch algorithms often suffer from low optimization efficiency, limited scalability, and poor flexibility. To address these challenges, this paper proposes a multi-strategy improved sand cat swarm optimization (MISCSO) algorithm for the economic scheduling of microgrids. First, a distribution-optimized initialization method based on adaptive diversity guidance is developed to enhance the quality of the initial population. This approach improves algorithmic performance by generating individuals in high-potential regions to ensure solution quality while maintaining population diversity through the inclusion of individuals from low-potential regions. Subsequently, an elite-centered global random movement strategy is introduced to balance elite guidance and global exploration, thereby improving both convergence speed and optimization accuracy. In addition, an adaptive elastic boundary mapping mechanism is proposed to effectively handle boundary violations, striking a balance between boundary constraints and global search capability. To evaluate the effectiveness of MISCSO, it is compared with 11 state-of-the-art algorithms using the IEEE CEC2017 benchmark set, and statistical analyses are conducted to assess performance differences. Experimental results demonstrate that MISCSO achieves superior optimization accuracy, convergence performance, and robustness. Finally, the applicability of MISCSO is verified through its implementation in microgrid economic scheduling, where it achieves outstanding optimization results.