Zhuodong Liu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Zhuodong Liu has a total of 22 co-authors. Here are the top co-authors by number of shared publications:
- Xiangyu Li
Department of Electronic Engineering, Shanghai Jiao Tong University, Shanghai, China
6 shared publications - Zhaoen Qu
School of Economics and Management, Beijing Jiaotong University, Beijing 100044, China
2 shared publications - Bodong Shang
Zhejiang Key Laboratory of Industrial Intelligence and Digital Twin, Eastern Institute of Technology, Ningbo, Zhejiang, China
1 shared publication - Bohan Guo
University of Malaya, Kuala Lumpur, Malaysia
1 shared publication - Chunhong Yuan
Faculty of Control Systems and Robotics, ITMO University, St. Petersburg, Russia
1 shared publication - Hongyang Du
The University of Hong Kong
1 shared publication
Latest Publications
- Open Access
- 6 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 3362-3373
FedCGR: Federated Cross-Domain Generative Recommendation
- 6 November 2026
- Published by Association for Computing Machinery (ACM)
- p. 3362-3373
Open AccessFedCGR: Federated Cross-Domain Generative Recommendation
- Open Access
RECAP-RL: Symmetry-Guided Retention Shaping and Mirrored Credit Assignment for Personalized Instrumental-Practice Recommendation
Personalized instrumental-practice planning requires selecting and ordering exercises, allocating duration and difficulty, and accounting for prerequisites, fatigue, spacing, and forgetting across multiple sessions. This paper presents RECAP-RL, a symmetry-guided reinforcement-learning framework that post-trains a structured large language model (LLM) policy to generate budget-feasible, teacher-editable practice plans optimized for delayed retention. Symmetry enters at two levels. At the objective level, retention-anchored dense analytic reward shaping (RADAR) re-expresses terminal retention-adjusted learning gain (RALG) as telescoping differences in a retention potential based on the learner’s projected retained mastery; this re-expression preserves core-objective policy ordering and yields an action-independent retention-state baseline whose variance effect is characterized analytically. At the intervention level, matched-intervention replay with residualized order-graph rewards (MIRROR) pairs each exercise with an equal-duration null-practice twin under a common continuation; exchanging the twins reverses the signed retained and prerequisite-unlock contrast, and zero-mean residualization preserves the mean exercise-block advantage when credit is assigned to exercise-token blocks. We instantiate the framework for piano practice in a mechanistic environment with 72 skills and 520 exercises. Against eleven comparison planners on held-out and structurally shifted learner families, RECAP-RL reaches a RALG of , exceeding the strongest observable-state model-based planner by 0.034 and terminal-reward group-relative policy optimization by 0.027, while reducing mean cross-family degradation from 26.6% to 18.8% and improving prerequisite repair and spacing alignment. These simulation results support symmetry-guided reward and credit design for delayed educational planning; longitudinal human studies remain necessary to establish educational effectiveness.Open AccessRECAP-RL: Symmetry-Guided Retention Shaping and Mirrored Credit Assignment for Personalized Instrumental-Practice Recommendation
Personalized instrumental-practice planning requires selecting and ordering exercises, allocating duration and difficulty, and accounting for prerequisites, fatigue, spacing, and forgetting across multiple sessions. This paper presents RECAP-RL, a symmetry-guided reinforcement-learning framework that post-trains a structured large language model (LLM) policy to generate budget-feasible, teacher-editable practice plans optimized for delayed retention. Symmetry enters at two levels. At the objective level, retention-anchored dense analytic reward shaping (RADAR) re-expresses terminal retention-adjusted learning gain (RALG) as telescoping differences in a retention potential based on the learner’s projected retained mastery; this re-expression preserves core-objective policy ordering and yields an action-independent retention-state baseline whose variance effect is characterized analytically. At the intervention level, matched-intervention replay with residualized order-graph rewards (MIRROR) pairs each exercise with an equal-duration null-practice twin under a common continuation; exchanging the twins reverses the signed retained and prerequisite-unlock contrast, and zero-mean residualization preserves the mean exercise-block advantage when credit is assigned to exercise-token blocks. We instantiate the framework for piano practice in a mechanistic environment with 72 skills and 520 exercises. Against eleven comparison planners on held-out and structurally shifted learner families, RECAP-RL reaches a RALG of , exceeding the strongest observable-state model-based planner by 0.034 and terminal-reward group-relative policy optimization by 0.027, while reducing mean cross-family degradation from 26.6% to 18.8% and improving prerequisite repair and spacing alignment. These simulation results support symmetry-guided reward and credit design for delayed educational planning; longitudinal human studies remain necessary to establish educational effectiveness. - Open Access
- 1 January 2026
- Published by Tech Science Press in Computers, Materials & Continua
- Vol. 88 (2) , 1-10
HiFraud: Hierarchical Privacy-Preserving Federated Learning with Star-Chain Knowledge Transfer for Cross-Institutional Fraud Detection
Financial fraud detection across institutions faces a fundamental tension between the need for diverse training data and regulatory prohibitions on sharing sensitive records. Existing federated learning approaches suffer from performance degradation under non-IID distributions and substantial utility losses when uniform differential privacy is applied to inherently sparse fraud signals. To this end, this paper proposes HiFraud, a hierarchical federated framework featuring three key components: fraud-aware dynamic clustering with complementarity regularization to group institutions by fraud pattern similarity while preserving rare-type representation; star-chain knowledge transfer augmented by not-true-class distillation to propagate novel fraud patterns rapidly within clusters while mitigating catastrophic forgetting; and privacy-adaptive aggregation via Rényi differential privacy composition, calibrating noise intensity to distributional divergence and fraud rarity. Experiments on IEEE-CIS, PaySim, and Worldline datasets show that HiFraud achieves an area under the receiver operating characteristic curve (AUC-ROC) of 0.935 under ε=2.3, outperforming DP-FedAvg by 10.5% while reducing convergence from 49 to 30 rounds. The framework also suppresses membership inference attack success to 10.2%, detects emerging fraud patterns within 3 h inside clusters, and improves rare fraud type detection by 23.0% over uniform privacy baselines. These results demonstrate that hierarchical architectures can effectively reconcile detection performance, formal privacy guarantees, and rapid threat response in collaborative fraud detection.
- 1 January 2026
- Published by Tech Science Press in Computers, Materials & Continua
- Vol. 88 (2) , 1-10
Open AccessHiFraud: Hierarchical Privacy-Preserving Federated Learning with Star-Chain Knowledge Transfer for Cross-Institutional Fraud Detection
Financial fraud detection across institutions faces a fundamental tension between the need for diverse training data and regulatory prohibitions on sharing sensitive records. Existing federated learning approaches suffer from performance degradation under non-IID distributions and substantial utility losses when uniform differential privacy is applied to inherently sparse fraud signals. To this end, this paper proposes HiFraud, a hierarchical federated framework featuring three key components: fraud-aware dynamic clustering with complementarity regularization to group institutions by fraud pattern similarity while preserving rare-type representation; star-chain knowledge transfer augmented by not-true-class distillation to propagate novel fraud patterns rapidly within clusters while mitigating catastrophic forgetting; and privacy-adaptive aggregation via Rényi differential privacy composition, calibrating noise intensity to distributional divergence and fraud rarity. Experiments on IEEE-CIS, PaySim, and Worldline datasets show that HiFraud achieves an area under the receiver operating characteristic curve (AUC-ROC) of 0.935 under ε=2.3, outperforming DP-FedAvg by 10.5% while reducing convergence from 49 to 30 rounds. The framework also suppresses membership inference attack success to 10.2%, detects emerging fraud patterns within 3 h inside clusters, and improves rare fraud type detection by 23.0% over uniform privacy baselines. These results demonstrate that hierarchical architectures can effectively reconcile detection performance, formal privacy guarantees, and rapid threat response in collaborative fraud detection.
- Top Cited PapersOpen Access
- 6 August 2025
- Published by MDPI AG in Mathematics
- Vol. 13 (15) , 2526
Interpretable Machine Learning Framework for Corporate Financialization Prediction: A SHAP-Based Analysis of High-Dimensional Data
High-dimensional prediction problems with complex non-linear feature interactions present significant algorithmic challenges in machine learning, particularly when dealing with imbalanced datasets and multicollinearity issues. This study proposes an innovative Shapley Additive Explanations (SHAP)-enhanced machine learning framework that integrates SHAP with advanced ensemble methods for interpretable financialization prediction. The methodology simultaneously addresses high-dimensional feature selection using 40 independent variables (19 CSR-related and 21 financialization-related), multicollinearity issues, and model interpretability requirements. Using a comprehensive dataset of 25,642 observations from 3776 Chinese A-share companies (2011–2022), we implement nine optimized machine learning algorithms with hyperparameter tuning via the Hippopotamus Optimization algorithm and five-fold cross-validation. XGBoost demonstrates superior performance with 99.34% explained variance, achieving an RMSE of 0.082 and R2 of 0.299. SHAP analysis reveals non-linear U-shaped relationships between key predictors and financialization outcomes, with critical thresholds at approximately 10 for CSR_SocR, 1.5 for CSR_S, and 5 for CSR_CV. SOE status, EPU, ownership concentration, firm size, and housing prices emerge as the most influential predictors. Notable shifts in factor importance occur during the COVID-19 pandemic period (2020–2022). This work contributes a scalable, interpretable machine learning architecture for high-dimensional financial prediction problems, with applications in risk assessment, portfolio optimization, and regulatory monitoring systems.
- 6 August 2025
- Published by MDPI AG in Mathematics
- Vol. 13 (15) , 2526
Top Cited PapersOpen AccessInterpretable Machine Learning Framework for Corporate Financialization Prediction: A SHAP-Based Analysis of High-Dimensional Data
High-dimensional prediction problems with complex non-linear feature interactions present significant algorithmic challenges in machine learning, particularly when dealing with imbalanced datasets and multicollinearity issues. This study proposes an innovative Shapley Additive Explanations (SHAP)-enhanced machine learning framework that integrates SHAP with advanced ensemble methods for interpretable financialization prediction. The methodology simultaneously addresses high-dimensional feature selection using 40 independent variables (19 CSR-related and 21 financialization-related), multicollinearity issues, and model interpretability requirements. Using a comprehensive dataset of 25,642 observations from 3776 Chinese A-share companies (2011–2022), we implement nine optimized machine learning algorithms with hyperparameter tuning via the Hippopotamus Optimization algorithm and five-fold cross-validation. XGBoost demonstrates superior performance with 99.34% explained variance, achieving an RMSE of 0.082 and R2 of 0.299. SHAP analysis reveals non-linear U-shaped relationships between key predictors and financialization outcomes, with critical thresholds at approximately 10 for CSR_SocR, 1.5 for CSR_S, and 5 for CSR_CV. SOE status, EPU, ownership concentration, firm size, and housing prices emerge as the most influential predictors. Notable shifts in factor importance occur during the COVID-19 pandemic period (2020–2022). This work contributes a scalable, interpretable machine learning architecture for high-dimensional financial prediction problems, with applications in risk assessment, portfolio optimization, and regulatory monitoring systems.
- Open Access
- 9 June 2025
- Published by MDPI AG in Mathematics
- Vol. 13 (12) , 1924
Hierarchical Multi-Scale Decomposition and Deep Learning Ensemble Framework for Enhanced Carbon Emission Prediction
Carbon emission prediction is critical for climate change mitigation across industrial, transportation, and urban sectors. Traditional statistical and machine learning methods struggle to capture complex multi-scale temporal patterns and long-range dependencies in emission data. This paper proposes a hierarchical multi-scale decomposition and deep learning ensemble framework that addresses these limitations. We integrate complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to decompose carbon emission time series into intrinsic mode functions (IMFs) capturing different frequency bands. Each IMF is processed through a hybrid convolutional neural network (CNN)–Transformer architecture: CNNs extract local features and transformers model long-range dependencies via multi-head attention. An adaptive ensemble mechanism dynamically weights component predictions based on stability and performance metrics. Experiments on four real-world datasets (133,225 observations) demonstrate that our CEEMDAN–CNN–Transformer framework outperforms 12 state-of-the-art methods, achieving a 13.3% reduction in root mean square error (RMSE) to 0.117, 12.7% improvement in mean absolute error (MAE) to 0.088, and 13.0% improvement in continuous ranked probability score (CRPS) to 0.060. The proposed framework not only improves predictive accuracy, but also enhances interpretability by revealing emission patterns across multiple temporal scales, supporting both operational and strategic carbon management decisions.
- 9 June 2025
- Published by MDPI AG in Mathematics
- Vol. 13 (12) , 1924
Open AccessHierarchical Multi-Scale Decomposition and Deep Learning Ensemble Framework for Enhanced Carbon Emission Prediction
Carbon emission prediction is critical for climate change mitigation across industrial, transportation, and urban sectors. Traditional statistical and machine learning methods struggle to capture complex multi-scale temporal patterns and long-range dependencies in emission data. This paper proposes a hierarchical multi-scale decomposition and deep learning ensemble framework that addresses these limitations. We integrate complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) to decompose carbon emission time series into intrinsic mode functions (IMFs) capturing different frequency bands. Each IMF is processed through a hybrid convolutional neural network (CNN)–Transformer architecture: CNNs extract local features and transformers model long-range dependencies via multi-head attention. An adaptive ensemble mechanism dynamically weights component predictions based on stability and performance metrics. Experiments on four real-world datasets (133,225 observations) demonstrate that our CEEMDAN–CNN–Transformer framework outperforms 12 state-of-the-art methods, achieving a 13.3% reduction in root mean square error (RMSE) to 0.117, 12.7% improvement in mean absolute error (MAE) to 0.088, and 13.0% improvement in continuous ranked probability score (CRPS) to 0.060. The proposed framework not only improves predictive accuracy, but also enhances interpretability by revealing emission patterns across multiple temporal scales, supporting both operational and strategic carbon management decisions.
- Open Access
- 6 May 2025
- Published by MDPI AG in Mathematics
- Vol. 13 (9) , 1529
Multi-Domain Controversial Text Detection Based on a Machine Learning and Deep Learning Stacked Ensemble
Due to the rapid proliferation of social media and online reviews, the accurate identification and classification of controversial texts has emerged as a significant challenge in the field of natural language processing. However, traditional text-classification methodologies frequently encounter critical limitations, such as feature sensitivity and inadequate generalization capabilities. This results in a notably suboptimal performance when confronted with diverse controversial content. To address these substantial limitations, this paper proposes a novel controversial text-detection framework based on stacked ensemble learning to enhance the accuracy and robustness of text classification. Firstly, considering the multidimensional complexity of textual features, we integrate comprehensive feature engineering, i.e., encompassing word frequency, statistical metrics, sentiment analysis, and comment tree structure features, as well as advanced feature selection methodologies, particularly lassonet, i.e., a neural network with feature sparsity, to effectively address dimensionality challenges while enhancing model interpretability and computational efficiency. Secondly, we design a two-tier stacked ensemble architecture, which not only combines the strengths of multiple machine learning algorithms, e.g., gradient-boosted decision tree (GBDT), random forest (RF), and extreme gradient boosting (XGBoost), with deep learning models, e.g., gated recurrent unit (GRU) and long short-term memory (LSTM), but also implements the support vector machine (SVM) for efficient meta-learning. Furthermore, we systematically compare three hyperparameter optimization algorithms, including the sparrow search algorithm (SSA), particle swarm optimization (PSO), and Bayesian optimization (BO). The experimental results demonstrate that the SSA exhibits a superior performance in exploring high-dimensional parameter spaces. Extensive experimentation across diverse topics and domains also confirms that our proposed methodology significantly outperforms the state-of-the-art approaches.
- 6 May 2025
- Published by MDPI AG in Mathematics
- Vol. 13 (9) , 1529
Open AccessMulti-Domain Controversial Text Detection Based on a Machine Learning and Deep Learning Stacked Ensemble
Due to the rapid proliferation of social media and online reviews, the accurate identification and classification of controversial texts has emerged as a significant challenge in the field of natural language processing. However, traditional text-classification methodologies frequently encounter critical limitations, such as feature sensitivity and inadequate generalization capabilities. This results in a notably suboptimal performance when confronted with diverse controversial content. To address these substantial limitations, this paper proposes a novel controversial text-detection framework based on stacked ensemble learning to enhance the accuracy and robustness of text classification. Firstly, considering the multidimensional complexity of textual features, we integrate comprehensive feature engineering, i.e., encompassing word frequency, statistical metrics, sentiment analysis, and comment tree structure features, as well as advanced feature selection methodologies, particularly lassonet, i.e., a neural network with feature sparsity, to effectively address dimensionality challenges while enhancing model interpretability and computational efficiency. Secondly, we design a two-tier stacked ensemble architecture, which not only combines the strengths of multiple machine learning algorithms, e.g., gradient-boosted decision tree (GBDT), random forest (RF), and extreme gradient boosting (XGBoost), with deep learning models, e.g., gated recurrent unit (GRU) and long short-term memory (LSTM), but also implements the support vector machine (SVM) for efficient meta-learning. Furthermore, we systematically compare three hyperparameter optimization algorithms, including the sparrow search algorithm (SSA), particle swarm optimization (PSO), and Bayesian optimization (BO). The experimental results demonstrate that the SSA exhibits a superior performance in exploring high-dimensional parameter spaces. Extensive experimentation across diverse topics and domains also confirms that our proposed methodology significantly outperforms the state-of-the-art approaches.