Xi Xiong
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Xi Xiong has a total of 64 co-authors. Here are the top co-authors by number of shared publications:
- Ling He
Sichuan University
6 shared publications Yuanyuan Li
6 shared publications- Fei Xiong
School of Electronic and Information Engineering, Beijing Jiaotong University, Beijing, China
5 shared publications - Jing Zhang
Sichuan University
4 shared publications - Fei He
School of Information and Communication Engineering, University of Electronic Science and Technology of China (UESTC), Chengdu, China
3 shared publications Nan Han
3 shared publications
Latest Publications
- 27 October 2026
- Published by Elsevier in Information Systems
- Vol. 144
Dynamic graph rewiring for fraud detection based on one-class homophily
- 27 October 2026
- Published by Elsevier in Information Systems
- Vol. 144
Dynamic graph rewiring for fraud detection based on one-class homophily
- Open Access
- 27 March 2025
- Published by Darcy & Roy Press Co. Ltd. in Journal of Computer Science and Artificial Intelligence
- Vol. 2 (3) , 50-54
Efficient Differentially Private Fine-Tuning with QLoRA and Prefix Tuning for Large Language Models
Large language models (LLMs) have achieved remarkable success in natural language processing (NLP) tasks. However, fine-tuning LLMs using private datasets raises significant privacy concerns, as models can inadvertently memorize sensitive information. Differentially Private Stochastic Gradient Descent (DP-SGD) provides a mathematically rigorous solution but suffers from high computational overhead, slow convergence, and excessive privacy budget consumption, making it impractical for large-scale models. To address these challenges, we propose an efficient differentially private fine-tuning method that combines Quantized Low-Rank Adaptation (QLoRA) and Prefix Tuning. QLoRA employs 4-bit NormalFloat quantization and low-rank adaptation, significantly reducing memory consumption and improving computational efficiency. Prefix Tuning optimizes a small set of prefix vectors without modifying the model’s main parameters, further reducing the impact of DP noise. Additionally, we introduce a hybrid adaptive gradient clipping strategy, which applies sample-wise adaptive clipping for Prefix Tuning and group-wise clipping for QLoRA, effectively balancing privacy protection and model utility. We evaluate our approach on GPT-2 using benchmark datasets including E2E NLG Challenge, XSum, SST-2, and DART, measuring performance using BLEU, ROUGE, and F1-score. Results demonstrate that QLoRA + Prefix Tuning achieves up to 75% memory reduction while maintaining over 95% of the original model performance under a moderate privacy budget (ε=3), outperforming traditional DP fine-tuning methods. Our work provides a practical and scalable solution for privacy-preserving LLM fine-tuning in resource-constrained environments.
- 27 March 2025
- Published by Darcy & Roy Press Co. Ltd. in Journal of Computer Science and Artificial Intelligence
- Vol. 2 (3) , 50-54
Open AccessEfficient Differentially Private Fine-Tuning with QLoRA and Prefix Tuning for Large Language Models
Large language models (LLMs) have achieved remarkable success in natural language processing (NLP) tasks. However, fine-tuning LLMs using private datasets raises significant privacy concerns, as models can inadvertently memorize sensitive information. Differentially Private Stochastic Gradient Descent (DP-SGD) provides a mathematically rigorous solution but suffers from high computational overhead, slow convergence, and excessive privacy budget consumption, making it impractical for large-scale models. To address these challenges, we propose an efficient differentially private fine-tuning method that combines Quantized Low-Rank Adaptation (QLoRA) and Prefix Tuning. QLoRA employs 4-bit NormalFloat quantization and low-rank adaptation, significantly reducing memory consumption and improving computational efficiency. Prefix Tuning optimizes a small set of prefix vectors without modifying the model’s main parameters, further reducing the impact of DP noise. Additionally, we introduce a hybrid adaptive gradient clipping strategy, which applies sample-wise adaptive clipping for Prefix Tuning and group-wise clipping for QLoRA, effectively balancing privacy protection and model utility. We evaluate our approach on GPT-2 using benchmark datasets including E2E NLG Challenge, XSum, SST-2, and DART, measuring performance using BLEU, ROUGE, and F1-score. Results demonstrate that QLoRA + Prefix Tuning achieves up to 75% memory reduction while maintaining over 95% of the original model performance under a moderate privacy budget (ε=3), outperforming traditional DP fine-tuning methods. Our work provides a practical and scalable solution for privacy-preserving LLM fine-tuning in resource-constrained environments.
- Open Access
- 27 December 2024
- Published by Association for Computing Machinery (ACM)
- p. 90-95
Keyword Fusion Optimization: A Simple and Efficient Query Rewriting Framework
- 27 December 2024
- Published by Association for Computing Machinery (ACM)
- p. 90-95
Open AccessKeyword Fusion Optimization: A Simple and Efficient Query Rewriting Framework
- 12 June 2024
- Published by Elsevier in Neural Networks
- Vol. 178, 106461
HyGloadAttack: Hard-label black-box textual adversarial attacks via hybrid optimization
- 12 June 2024
- Published by Elsevier in Neural Networks
- Vol. 178, 106461
HyGloadAttack: Hard-label black-box textual adversarial attacks via hybrid optimization
- 13 May 2024
- Published by Elsevier in Information Sciences
- Vol. 674
SCSQ: A sample cooperation optimization method with sample quality for recurrent neural networks
- 13 May 2024
- Published by Elsevier in Information Sciences
- Vol. 674
SCSQ: A sample cooperation optimization method with sample quality for recurrent neural networks
- Open Access
- 8 January 2024
- Published by Association for Computing Machinery (ACM) in ACM Transactions on the Web
- Vol. 18 (2) , 1-26
Incorporating a Triple Graph Neural Network with Multiple Implicit Feedback for Social Recommendation
Graph neural networks have been clearly proven to be powerful in recommendation tasks since they can capture high-order user-item interactions and integrate them with rich attributes. However, they are still limited by the cold-start problem and data sparsity. Using social relationships to assist recommendation is an effective practice, but it can only moderately alleviate these problems. In addition, rich attributes are often unavailable, which prevents graph neural networks from being fully effective. Hence, we propose to enrich the model by mining multiple implicit feedback and constructing a triple GCN component. We have noticed that users may be influenced not only by their trusted friends but also by the ratings that already exist. The implicit influence spreads among the item’s previous and potential raters, and makes a difference on future ratings. The implicit influence is analyzed on the mechanism of information propagation, and fused with the user’s binary implicit attitude, since negative influence propagates as well as the positive one. Furthermore, we leverage explicit feedback, social relationships, and multiple implicit feedback in the triple GCN component. Abundant experiments on real-world datasets reveal that our model has improved significantly in the rating prediction task compared with other state-of-the-art methods.
- 8 January 2024
- Published by Association for Computing Machinery (ACM) in ACM Transactions on the Web
- Vol. 18 (2) , 1-26
Open AccessIncorporating a Triple Graph Neural Network with Multiple Implicit Feedback for Social Recommendation
Graph neural networks have been clearly proven to be powerful in recommendation tasks since they can capture high-order user-item interactions and integrate them with rich attributes. However, they are still limited by the cold-start problem and data sparsity. Using social relationships to assist recommendation is an effective practice, but it can only moderately alleviate these problems. In addition, rich attributes are often unavailable, which prevents graph neural networks from being fully effective. Hence, we propose to enrich the model by mining multiple implicit feedback and constructing a triple GCN component. We have noticed that users may be influenced not only by their trusted friends but also by the ratings that already exist. The implicit influence spreads among the item’s previous and potential raters, and makes a difference on future ratings. The implicit influence is analyzed on the mechanism of information propagation, and fused with the user’s binary implicit attitude, since negative influence propagates as well as the positive one. Furthermore, we leverage explicit feedback, social relationships, and multiple implicit feedback in the triple GCN component. Abundant experiments on real-world datasets reveal that our model has improved significantly in the rating prediction task compared with other state-of-the-art methods.
- Open Access
- 17 November 2023
- Published by Association for Computing Machinery (ACM)
- p. 535-541
Joint Learning Framework of Semantics and Knowledge Graph Reasoning
The existing embedded Knowledge Graph Question Answering (KGQA) methods based on relationship chain reasoning primarily rely on explicit relationship chains in natural language questions and implicit relationship chains in the knowledge graph (KG) for reasoning. However, these methods overlook the semantic contributions of entity semantics and question context to the relationship chain, merely focusing on the order of relationships within the chain. To address this limitation, a joint learning model is proposed that incorporates both text and KG reasoning using the Graph Attention Network (GAT). In this approach, the semantics of entities are integrated into the reasoning process of multi-hop KGQA, thereby enhancing the semantic expression within the relationship chain. Additionally, a fusion of the question context with the KG is performed to achieve a comprehensive understanding of the abundant semantic information contained in the question. The experimental results, obtained from three benchmark datasets, clearly demonstrate the significant superiority of the proposed model over the previous state-of-the-art methods, including GraftNet, EmbedKGQA, and Rce-KGQA. This validation verifies that the integration of question context and entity semantics effectively enhances the expression ability of incomplete KGs, leading to improved KGQA performance.
- 17 November 2023
- Published by Association for Computing Machinery (ACM)
- p. 535-541
Open AccessJoint Learning Framework of Semantics and Knowledge Graph Reasoning
The existing embedded Knowledge Graph Question Answering (KGQA) methods based on relationship chain reasoning primarily rely on explicit relationship chains in natural language questions and implicit relationship chains in the knowledge graph (KG) for reasoning. However, these methods overlook the semantic contributions of entity semantics and question context to the relationship chain, merely focusing on the order of relationships within the chain. To address this limitation, a joint learning model is proposed that incorporates both text and KG reasoning using the Graph Attention Network (GAT). In this approach, the semantics of entities are integrated into the reasoning process of multi-hop KGQA, thereby enhancing the semantic expression within the relationship chain. Additionally, a fusion of the question context with the KG is performed to achieve a comprehensive understanding of the abundant semantic information contained in the question. The experimental results, obtained from three benchmark datasets, clearly demonstrate the significant superiority of the proposed model over the previous state-of-the-art methods, including GraftNet, EmbedKGQA, and Rce-KGQA. This validation verifies that the integration of question context and entity semantics effectively enhances the expression ability of incomplete KGs, leading to improved KGQA performance.
- 20 May 2023
- Published by Elsevier in Expert Systems with Applications
- Vol. 229
HCURec: Hierarchical candidate-aware user modeling for news recommendation
- 20 May 2023
- Published by Elsevier in Expert Systems with Applications
- Vol. 229
HCURec: Hierarchical candidate-aware user modeling for news recommendation
- 26 September 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Audio, Speech, and Language Processing
- Vol. 31 (23299290) , 721-733
WNSA-Net: An Axial-Attention-Based Network for Schizophrenia Detection Using Wideband and Narrowband Spectrograms
Schizophrenia is a severe mental disease that affects patients' thoughts, feelings, and behaviors. Speech signal has proven to be a biomarker in the early diagnosis of schizophrenia. Previous studies on schizophrenic speech detection are mainly based on manual feature extraction engineering, which requires domain knowledge for researchers and has difficulties extracting effective features. This work proposes an end-to-end architecture, called Axial-attention-based Network using Wideband and Narrowband Spectrograms (WNSA-Net), to detect schizophrenia. Specifically, we adopt both wideband and narrowband spectrograms as inputs to represent speech signals using fine time and frequency structures. Then dilated convolution blocks are employed to capture detailed and long-range information in spectrograms. Axial-attention blocks are introduced to augment the information in feature maps along the time and frequency axes. In addition, we employ a gate mechanism to fuse the output feature maps from all channels. Experimental results on the Schizophrenia dataset and its subdatasets show that schizophrenic patients have difficulties in expressing emotions. To validate the performance of our WNSA-Net, experiments are conducted on Schizophrenia dataset and open-access TORGO database, achieving 97.37% and 98.16% accuracy in detecting schizophrenia and dysarthria, respectively. The results show promise for the proposed method in the diagnosis of disordered speech.
- 26 September 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE/ACM Transactions on Audio, Speech, and Language Processing
- Vol. 31 (23299290) , 721-733
WNSA-Net: An Axial-Attention-Based Network for Schizophrenia Detection Using Wideband and Narrowband Spectrograms
Schizophrenia is a severe mental disease that affects patients' thoughts, feelings, and behaviors. Speech signal has proven to be a biomarker in the early diagnosis of schizophrenia. Previous studies on schizophrenic speech detection are mainly based on manual feature extraction engineering, which requires domain knowledge for researchers and has difficulties extracting effective features. This work proposes an end-to-end architecture, called Axial-attention-based Network using Wideband and Narrowband Spectrograms (WNSA-Net), to detect schizophrenia. Specifically, we adopt both wideband and narrowband spectrograms as inputs to represent speech signals using fine time and frequency structures. Then dilated convolution blocks are employed to capture detailed and long-range information in spectrograms. Axial-attention blocks are introduced to augment the information in feature maps along the time and frequency axes. In addition, we employ a gate mechanism to fuse the output feature maps from all channels. Experimental results on the Schizophrenia dataset and its subdatasets show that schizophrenic patients have difficulties in expressing emotions. To validate the performance of our WNSA-Net, experiments are conducted on Schizophrenia dataset and open-access TORGO database, achieving 97.37% and 98.16% accuracy in detecting schizophrenia and dysarthria, respectively. The results show promise for the proposed method in the diagnosis of disordered speech.
- 13 July 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 31 (10577149) , 5025-5037
A Robust Movement Quantification Algorithm of Hyperactivity Detection for ADHD Children Based on 3D Depth Images
Attention deficit hyperactivity disorder (ADHD) is one of the most common childhood mental disorders. Hyperactivity is a typical symptom of ADHD in children. Clinicians diagnose this symptom by evaluating the children’s activities based on subjective rating scales and clinical experience. In this work, an objective system is proposed to quantify the movements of children with ADHD automatically. This system presents a new movement detection and quantification method based on depth images. A novel salient object extraction method is proposed to segment body regions. In movement detection, we explore a new local search algorithm to detect any potential motions of children based on three newly designed evaluation metrics. In the movement quantification, two parameters are investigated to quantify the participation degree and the displacements of each body part in the movements. This system is tested by a depth dataset of children with ADHD. The movement detection results of this dataset mainly range from 91.0% to 95.0%. The movement quantification results of children are consistent with the clinical observations. The public MSR Action 3D dataset is tested to validate the performance of this system.
- 13 July 2022
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Image Processing
- Vol. 31 (10577149) , 5025-5037
A Robust Movement Quantification Algorithm of Hyperactivity Detection for ADHD Children Based on 3D Depth Images
Attention deficit hyperactivity disorder (ADHD) is one of the most common childhood mental disorders. Hyperactivity is a typical symptom of ADHD in children. Clinicians diagnose this symptom by evaluating the children’s activities based on subjective rating scales and clinical experience. In this work, an objective system is proposed to quantify the movements of children with ADHD automatically. This system presents a new movement detection and quantification method based on depth images. A novel salient object extraction method is proposed to segment body regions. In movement detection, we explore a new local search algorithm to detect any potential motions of children based on three newly designed evaluation metrics. In the movement quantification, two parameters are investigated to quantify the participation degree and the displacements of each body part in the movements. This system is tested by a depth dataset of children with ADHD. The movement detection results of this dataset mainly range from 91.0% to 95.0%. The movement quantification results of children are consistent with the clinical observations. The public MSR Action 3D dataset is tested to validate the performance of this system.