Xiaofeng Liu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Xiaofeng Liu has a total of 26 co-authors. Here are the top co-authors by number of shared publications:
- Yinglong Ma
North China Electric Power University
7 shared publications - Chenqi Guo
School of Control and Computer Engineering, North China Electric Power University, Beijing, China
3 shared publications - Bin Wang
College of Computer Science and Technology, Taiyuan University of Technology, Jinzhong, China
2 shared publications Dandan Li
2 shared publications- Huili Liu
School of Control and Computer Engineering, North China Electric Power University, Beijing, China
2 shared publications Jiayu Lu
2 shared publications
Latest Publications
- 1 September 2026
- Published by Elsevier in Medical Image Analysis
- Vol. 114
HAHN-SGCL: Hierarchical Attention and Hard Negatives-aware State Graph Contrastive Learning for functional connectome fingerprinting
- 1 September 2026
- Published by Elsevier in Medical Image Analysis
- Vol. 114
HAHN-SGCL: Hierarchical Attention and Hard Negatives-aware State Graph Contrastive Learning for functional connectome fingerprinting
- 4 February 2026
- Published by Elsevier in Information Sciences
- Vol. 740
Post-hoc explainability of graph neural networks: A comprehensive survey
- 4 February 2026
- Published by Elsevier in Information Sciences
- Vol. 740
Post-hoc explainability of graph neural networks: A comprehensive survey
- 21 November 2025
- Published by Elsevier in Biomedical Signal Processing and Control
- Vol. 113
Intra- and Inter-Modality Interactions Region Alignment Fusion for Alzheimer’s disease diagnosis
- 21 November 2025
- Published by Elsevier in Biomedical Signal Processing and Control
- Vol. 113
Intra- and Inter-Modality Interactions Region Alignment Fusion for Alzheimer’s disease diagnosis
- 15 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Network and Service Management
- Vol. 22 (3) , 2853-2865
MOHFL: Multi-Level One-Shot Hierarchical Federated Learning With Enhanced Model Aggregation Over Non-IID Data
Hierarchical federated learning (HFL) is a privacy-preserving distributed machine learning framework with a client-edge-cloud hierarchy, where multiple edge servers perform partial model aggregation to reduce costly communication with the cloud server. Nevertheless, most existing HFL methods require extensive iterative communication and public datasets, which not only increase communication overhead but also raise privacy and security concerns. Moreover, non-independent and identically distributed (non-IID) data among devices can significantly impact the accuracy of the global model in HFL. To address these challenges, we propose a multi-level one-shot HFL framework (MOHFL), which aims to improve the performance of the global model in a single communication round. Specifically, we employ conditional variational autoencoders (CVAEs) as local models and use the aggregated decoders to generate an IID training set for the global model, thereby mitigating the negative impact of non-IID data. We improve the performance of CVAEs under different levels of data heterogeneity through a dominant class-based data selection method. Subsequently, an edge aggregation scheme based on multi-teacher knowledge distillation and contrastive learning is proposed to aggregate the knowledge from local decoders to edge decoders. Extensive experiments on four real-world datasets demonstrate that MOHFL is very competitive against four state-of-the-art baselines under various settings.
- 15 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Network and Service Management
- Vol. 22 (3) , 2853-2865
MOHFL: Multi-Level One-Shot Hierarchical Federated Learning With Enhanced Model Aggregation Over Non-IID Data
Hierarchical federated learning (HFL) is a privacy-preserving distributed machine learning framework with a client-edge-cloud hierarchy, where multiple edge servers perform partial model aggregation to reduce costly communication with the cloud server. Nevertheless, most existing HFL methods require extensive iterative communication and public datasets, which not only increase communication overhead but also raise privacy and security concerns. Moreover, non-independent and identically distributed (non-IID) data among devices can significantly impact the accuracy of the global model in HFL. To address these challenges, we propose a multi-level one-shot HFL framework (MOHFL), which aims to improve the performance of the global model in a single communication round. Specifically, we employ conditional variational autoencoders (CVAEs) as local models and use the aggregated decoders to generate an IID training set for the global model, thereby mitigating the negative impact of non-IID data. We improve the performance of CVAEs under different levels of data heterogeneity through a dominant class-based data selection method. Subsequently, an edge aggregation scheme based on multi-teacher knowledge distillation and contrastive learning is proposed to aggregate the knowledge from local decoders to edge decoders. Extensive experiments on four real-world datasets demonstrate that MOHFL is very competitive against four state-of-the-art baselines under various settings.
- 11 December 2024
- Published by Springer Nature in Applied Intelligence
- Vol. 55 (2) , 1-19
KnowGNN: a knowledge-aware and structure-sensitive model-level explainer for graph neural networks
- 11 December 2024
- Published by Springer Nature in Applied Intelligence
- Vol. 55 (2) , 1-19
KnowGNN: a knowledge-aware and structure-sensitive model-level explainer for graph neural networks
- 21 October 2024
- Published by Elsevier in Expert Systems with Applications
- Vol. 262
Why does Knowledge Distillation work? Rethink its attention and fidelity mechanism
- 21 October 2024
- Published by Elsevier in Expert Systems with Applications
- Vol. 262
Why does Knowledge Distillation work? Rethink its attention and fidelity mechanism
- 27 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Fuzzy Systems
- Vol. 32 (12) , 6951-6964
Toward Embedding Ambiguity-Sensitive Graph Neural Network Explainability
Recently, many post-hoc GNN explanation methods have been explored to uncover GNNs' predictive behaviors by analyzing the embeddings produced by the GNN models. However, these methods suffer from explanation ambiguity inherent in learned graph embeddings because aggregation-based embeddings can lead to the loss of unique identifiers for individual graph components and thus allow non-causal nodes that are adjacent to true causal patterns to unintentionally embody causal information in their embeddings, hindering the explanations from faithfully representing the true insights of GNNs' predictive reasoning. In this paper, we present an Embedding AmbiguitySensitive GNN Explanation framework (EAGX). EAGX can effectively mitigate the impact of embedding-induced explanation ambiguity by creating edges' ambiguity feature extractor, exploring edges' predictive relevance, and integrating them into the explanation process, thereby capturing each graph component's contribution to the predictions. Specifically, we first propose a Centroid-constrained Fuzzy c-Means (C-FCM) algorithm to construct an ambiguity feature extractor. Then, we leverage the ambiguity features for edges to develop the ambiguity-based edge attribution module for assigning a prediction relevance score to each edge. Finally, instead of focusing only on the edges with high influence to the GNN prediction, we introduce a joint optimization strategy to refine the learning process of our edge attribution module, empowering EAGX to capture the subtle interplay of both causal and non-causal subgraphs on model predictions, which further improve the explainability of GNN predictions. Experimental results demonstrate that EAGX outperforms the leading explainers on most evaluation metrics, underscoring its effectiveness in generating reliable and precise explanations for GNNs
- 27 September 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Fuzzy Systems
- Vol. 32 (12) , 6951-6964
Toward Embedding Ambiguity-Sensitive Graph Neural Network Explainability
Recently, many post-hoc GNN explanation methods have been explored to uncover GNNs' predictive behaviors by analyzing the embeddings produced by the GNN models. However, these methods suffer from explanation ambiguity inherent in learned graph embeddings because aggregation-based embeddings can lead to the loss of unique identifiers for individual graph components and thus allow non-causal nodes that are adjacent to true causal patterns to unintentionally embody causal information in their embeddings, hindering the explanations from faithfully representing the true insights of GNNs' predictive reasoning. In this paper, we present an Embedding AmbiguitySensitive GNN Explanation framework (EAGX). EAGX can effectively mitigate the impact of embedding-induced explanation ambiguity by creating edges' ambiguity feature extractor, exploring edges' predictive relevance, and integrating them into the explanation process, thereby capturing each graph component's contribution to the predictions. Specifically, we first propose a Centroid-constrained Fuzzy c-Means (C-FCM) algorithm to construct an ambiguity feature extractor. Then, we leverage the ambiguity features for edges to develop the ambiguity-based edge attribution module for assigning a prediction relevance score to each edge. Finally, instead of focusing only on the edges with high influence to the GNN prediction, we introduce a joint optimization strategy to refine the learning process of our edge attribution module, empowering EAGX to capture the subtle interplay of both causal and non-causal subgraphs on model predictions, which further improve the explainability of GNN predictions. Experimental results demonstrate that EAGX outperforms the leading explainers on most evaluation metrics, underscoring its effectiveness in generating reliable and precise explanations for GNNs
- 30 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Network Science and Engineering
- Vol. 11 (6) , 6007-6023
GAXG: A Global and Self-Adaptive Optimal Graph Topology Generation Framework for Explaining Graph Neural Networks
Numerous explainability techniques have been developed to reveal the prediction principles of Graph Neural Networks (GNNs) across diverse domains. However, many existing approaches, particularly those concentrating on model-level explanations, tend to grapple with the tunnel vision problem, leading to less-than-optimal outcomes and constraining users' comprehensive understanding of GNNs. Furthermore, these methods typically require hyperparameters to mold the explanations, introducing unintended human biases. In response, we present GAXG, a global and self-adaptive optimal graph topology generation framework for explaining GNNs' prediction principles at model-level. GAXG addresses the challenges of tunnel vision and hyperparameter reliance by integrating a strategically tailored Monte Carlo Tree Search (MCTS) algorithm. Notably, our tailored MCTS algorithm is modified to incorporate an Edge Mask Learning and Simulated Annealing-based subgraph screening strategy during the expansion phase, resolving the inherent time-consuming challenges of the tailored MCTS and enhancing the quality of the generated explanatory graph topologies. Experimental results underscore GAXG's effectiveness in discovering global explanations for GNNs, outperforming leading explainers on most evaluation metrics.
- 30 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Network Science and Engineering
- Vol. 11 (6) , 6007-6023
GAXG: A Global and Self-Adaptive Optimal Graph Topology Generation Framework for Explaining Graph Neural Networks
Numerous explainability techniques have been developed to reveal the prediction principles of Graph Neural Networks (GNNs) across diverse domains. However, many existing approaches, particularly those concentrating on model-level explanations, tend to grapple with the tunnel vision problem, leading to less-than-optimal outcomes and constraining users' comprehensive understanding of GNNs. Furthermore, these methods typically require hyperparameters to mold the explanations, introducing unintended human biases. In response, we present GAXG, a global and self-adaptive optimal graph topology generation framework for explaining GNNs' prediction principles at model-level. GAXG addresses the challenges of tunnel vision and hyperparameter reliance by integrating a strategically tailored Monte Carlo Tree Search (MCTS) algorithm. Notably, our tailored MCTS algorithm is modified to incorporate an Edge Mask Learning and Simulated Annealing-based subgraph screening strategy during the expansion phase, resolving the inherent time-consuming challenges of the tailored MCTS and enhancing the quality of the generated explanatory graph topologies. Experimental results underscore GAXG's effectiveness in discovering global explanations for GNNs, outperforming leading explainers on most evaluation metrics.
- Open Access
Why does Knowledge Distillation Work? Rethink its Attention and Fidelity Mechanism
Does Knowledge Distillation (KD) really work? Conventional wisdom viewed it as a knowledge transfer procedure where a perfect mimicry of the student to its teacher is desired. However, paradoxical studies indicate that closely replicating the teacher's behavior does not consistently improve student generalization, posing questions on its possible causes. Confronted with this gap, we hypothesize that diverse attentions in teachers contribute to better student generalization at the expense of reduced fidelity in ensemble KD setups. By increasing data augmentation strengths, our key findings reveal a decrease in the Intersection over Union (IoU) of attentions between teacher models, leading to reduced student overfitting and decreased fidelity. We propose this low-fidelity phenomenon as an underlying characteristic rather than a pathology when training KD. This suggests that stronger data augmentation fosters a broader perspective provided by the divergent teacher ensemble and lower student-teacher mutual information, benefiting generalization performance. These insights clarify the mechanism on low-fidelity phenomenon in KD. Thus, we offer new perspectives on optimizing student model performance, by emphasizing increased diversity in teacher attentions and reduced mimicry behavior between teachers and student.
Open AccessWhy does Knowledge Distillation Work? Rethink its Attention and Fidelity Mechanism
Does Knowledge Distillation (KD) really work? Conventional wisdom viewed it as a knowledge transfer procedure where a perfect mimicry of the student to its teacher is desired. However, paradoxical studies indicate that closely replicating the teacher's behavior does not consistently improve student generalization, posing questions on its possible causes. Confronted with this gap, we hypothesize that diverse attentions in teachers contribute to better student generalization at the expense of reduced fidelity in ensemble KD setups. By increasing data augmentation strengths, our key findings reveal a decrease in the Intersection over Union (IoU) of attentions between teacher models, leading to reduced student overfitting and decreased fidelity. We propose this low-fidelity phenomenon as an underlying characteristic rather than a pathology when training KD. This suggests that stronger data augmentation fosters a broader perspective provided by the divergent teacher ensemble and lower student-teacher mutual information, benefiting generalization performance. These insights clarify the mechanism on low-fidelity phenomenon in KD. Thus, we offer new perspectives on optimizing student model performance, by emphasizing increased diversity in teacher attentions and reduced mimicry behavior between teachers and student.
- Open Access
- 27 July 2022
- Published by Elsevier in Software Impacts
- Vol. 14
HE-HMTC: A hybrid embedding-based text representation for Hierarchical multi-label text classification
- 27 July 2022
- Published by Elsevier in Software Impacts
- Vol. 14
Open AccessHE-HMTC: A hybrid embedding-based text representation for Hierarchical multi-label text classification