Aming Wu
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Aming Wu has a total of 19 co-authors. Here are the top co-authors by number of shared publications:
- Jianji Ren
School of Software, Henan Polytechnic University, Jiaozuo, China
4 shared publications - Yanan Li
School of Software, Henan Polytechnic University, Jiaozuo, China
4 shared publications - Young-Woo Kwon
Kyungpook National University, Daegu, Republic of Korea
4 shared publications - Jingzhe Kang3 shared publications
- Yongliang Yuan
School of Mechanical and Power Engineering, Henan Polytechnic University, Jiaozuo, China
3 shared publications Yun Xin
3 shared publications
Latest Publications
- Open Access
- 16 August 2026
- Published by Elsevier in SSRN Electronic Journal
FedPOD-LoRA: Federated Fine-Tuning Based on Proximal Optimization and Dual-LoRA Collaboration
Fine-tuning methods that combine federated learning with low-rank adaptation (LoRA) enable efficient collaborative training of large language models under resource constraints. Existing approaches widely adopt single-LoRA or multi-LoRA architectures. However, non-IID client data contain both global and personalized knowledge. Single-LoRA architectures tend to entangle them, while, multi-LoRA architectures lack persistent constraints between the global and private branches, causing the progressive discrepancy between local knowledge and aggregation knowledge. To address these limitations, this paper proposes FedPOD-LoRA, a federated fine-tuning method based on proximal optimization and dual-LoRA collaboration. First, we introduce a targeted proximal constraint on the global LoRA matrix B at the client side to reduce representation divergence between LoRA branches, thereby mitigating knowledge drift and fusion conflicts under non-IID data. Second, we further optimize the aggregated global Bmatrix using public data at the server side to enhance its global knowledge representation capability. Compared with five state-of-the-art (SOTA) methods, FedPOD-LoRA achieves better understanding performance, faster convergence. Across five natural language understanding tasks, including MNLI, SST-2, QNLI, QQP, and RTE, it improves accuracy by up to 2.5% and reduces the communication rounds required to reach the target performance on RTE and QNLI by 25.6% and 63.9%.
- 16 August 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessFedPOD-LoRA: Federated Fine-Tuning Based on Proximal Optimization and Dual-LoRA Collaboration
Fine-tuning methods that combine federated learning with low-rank adaptation (LoRA) enable efficient collaborative training of large language models under resource constraints. Existing approaches widely adopt single-LoRA or multi-LoRA architectures. However, non-IID client data contain both global and personalized knowledge. Single-LoRA architectures tend to entangle them, while, multi-LoRA architectures lack persistent constraints between the global and private branches, causing the progressive discrepancy between local knowledge and aggregation knowledge. To address these limitations, this paper proposes FedPOD-LoRA, a federated fine-tuning method based on proximal optimization and dual-LoRA collaboration. First, we introduce a targeted proximal constraint on the global LoRA matrix B at the client side to reduce representation divergence between LoRA branches, thereby mitigating knowledge drift and fusion conflicts under non-IID data. Second, we further optimize the aggregated global Bmatrix using public data at the server side to enhance its global knowledge representation capability. Compared with five state-of-the-art (SOTA) methods, FedPOD-LoRA achieves better understanding performance, faster convergence. Across five natural language understanding tasks, including MNLI, SST-2, QNLI, QQP, and RTE, it improves accuracy by up to 2.5% and reduces the communication rounds required to reach the target performance on RTE and QNLI by 25.6% and 63.9%.
- Open Access
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
Prompt-Guided Causal Log Reasoning with Self-Evolving Semantic Memory for Robust Anomaly Detection in Noisy Distributed Systems
Log anomaly detection is critical for ensuring the reliability of large-scale distributed systems. Existing methods often fail to balance adaptability to concept drift with the interpretability required for effective root cause analysis. Deep learning models generally lack transparency, while Large Language Model (LLM)-based approaches are prone to hallucination and exhibit high latency. To overcome these challenges, this paper proposes a neuro-symbolic framework, PGCLR+SESAM, which combines two complementary modules for robust log analysis. The first module, Prompt-Guided Causal Log Reasoning (PGCLR), uses an LLM (LLaMA-3.3-70B) and prompt-based causal reconstruction to infer multi-hop dependencies among log events and generate explicit root-cause reasoning chains. The second module, Self-Evolving Semantic Anomaly Memory (SESAM), implements an active memory-guided prompting mechanism. In contrast to passive storage, SESAM dynamically retrieves and injects semantic summaries of long-tail anomalies to refine the reasoning context and enhance generalisation under distribution drift. Extensive experiments on two large-scale datasets, LogHub~2.0 (15M entries) and LogLAB (9M entries), demonstrate that this approach outperforms previous methods, including specialized LLMs such as LLM-LADE, achieving an F1-score of 0.961, maintaining real-time inference efficiency with less than 150 ms per sequence. Automated interpretability metrics, including Graph Edit Distance and Semantic Consistency Score, as well as human evaluation, are used to quantitatively assess the structural accuracy of the generated explanations. The findings demonstrate that the proposed framework enables transparent causal reasoning and offers a practical solution for trustworthy, self-adaptive system reliability management.
- 29 July 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessPrompt-Guided Causal Log Reasoning with Self-Evolving Semantic Memory for Robust Anomaly Detection in Noisy Distributed Systems
Log anomaly detection is critical for ensuring the reliability of large-scale distributed systems. Existing methods often fail to balance adaptability to concept drift with the interpretability required for effective root cause analysis. Deep learning models generally lack transparency, while Large Language Model (LLM)-based approaches are prone to hallucination and exhibit high latency. To overcome these challenges, this paper proposes a neuro-symbolic framework, PGCLR+SESAM, which combines two complementary modules for robust log analysis. The first module, Prompt-Guided Causal Log Reasoning (PGCLR), uses an LLM (LLaMA-3.3-70B) and prompt-based causal reconstruction to infer multi-hop dependencies among log events and generate explicit root-cause reasoning chains. The second module, Self-Evolving Semantic Anomaly Memory (SESAM), implements an active memory-guided prompting mechanism. In contrast to passive storage, SESAM dynamically retrieves and injects semantic summaries of long-tail anomalies to refine the reasoning context and enhance generalisation under distribution drift. Extensive experiments on two large-scale datasets, LogHub~2.0 (15M entries) and LogLAB (9M entries), demonstrate that this approach outperforms previous methods, including specialized LLMs such as LLM-LADE, achieving an F1-score of 0.961, maintaining real-time inference efficiency with less than 150 ms per sequence. Automated interpretability metrics, including Graph Edit Distance and Semantic Consistency Score, as well as human evaluation, are used to quantitatively assess the structural accuracy of the generated explanations. The findings demonstrate that the proposed framework enables transparent causal reasoning and offers a practical solution for trustworthy, self-adaptive system reliability management.
- Open Access
- 24 June 2026
- Published by Elsevier in SSRN Electronic Journal
Fat-Cat: Document-Driven Metacognitive Multi-Agent System for Complex Reasoning
The efficacy of Large Language Model (LLM) agents is often constrained by the inefficiency of context utilization during runtime. Traditional ”log-centric” frameworks rely on syntax-heavy state representations like nested JSON, which impose a syntactic tax that diverts finite attention from semantic reasoning to structural parsing, leading to context dilution in long-horizon tasks. We propose Fat-Cat, a document-driven metacognitive architecture that transitions agent state management from passive logging to active semantic orchestration. Fat-Cat introduces two synergistic mechanisms: (1) Document-Driven State Modeling. By utilizing a markdown-based global workspace aligned with pre-training priors, we achieve representational alignment that collapses context growth from a quadratic 𝑂(𝑇 2) trajectory into a linear 𝑂(𝑇 ) one. (2) Closed-Loop Semantic Watcher. An independent auditor performs pre-commit verification to detect invalid state transitions and prevent error cascades. Additionally, a textual strategy evolution module facilitates parameter-free learning through methodological distillation. Extensive benchmarking suggests that document-driven state orchestration improves the reliability and context efficiency of LLM agents across multiple backbone models, especially in long-horizon reasoning and domain-grounded verification tasks.
- 24 June 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessFat-Cat: Document-Driven Metacognitive Multi-Agent System for Complex Reasoning
The efficacy of Large Language Model (LLM) agents is often constrained by the inefficiency of context utilization during runtime. Traditional ”log-centric” frameworks rely on syntax-heavy state representations like nested JSON, which impose a syntactic tax that diverts finite attention from semantic reasoning to structural parsing, leading to context dilution in long-horizon tasks. We propose Fat-Cat, a document-driven metacognitive architecture that transitions agent state management from passive logging to active semantic orchestration. Fat-Cat introduces two synergistic mechanisms: (1) Document-Driven State Modeling. By utilizing a markdown-based global workspace aligned with pre-training priors, we achieve representational alignment that collapses context growth from a quadratic 𝑂(𝑇 2) trajectory into a linear 𝑂(𝑇 ) one. (2) Closed-Loop Semantic Watcher. An independent auditor performs pre-commit verification to detect invalid state transitions and prevent error cascades. Additionally, a textual strategy evolution module facilitates parameter-free learning through methodological distillation. Extensive benchmarking suggests that document-driven state orchestration improves the reliability and context efficiency of LLM agents across multiple backbone models, especially in long-horizon reasoning and domain-grounded verification tasks.
- Open Access
- 10 April 2026
- Published by Elsevier in SSRN Electronic Journal
Fat-Cat: Document-Driven Metacognitive Multi-Agent System for Complex Reasoning
The efficacy of Large Language Model (LLM) agents is often constrained by the inefficiency of context utilization during runtime. Traditional “log-centric” frameworks rely on syntax-heavy state representations like nested JSON, which impose a syntactic tax that diverts finite attention from semantic reasoning to structural parsing, leading to context dilution in long-horizon tasks. We propose Fat-Cat, a document-driven metacognitive architecture that transitions agent state management from passive logging to active semantic orchestration. Fat-Cat introduces two synergistic mechanisms: (1) Document-Driven State Modeling. By utilizing a Markdown-based semantic file system aligned with pre-training priors, we achieve representational alignment that collapses context growth from a quadratic 𝑂(𝑇 ²) trajectory into a linear 𝑂(𝑇 ) one. (2) Closed-Loop Semantic Watcher. An independent auditor performs pre-emptive audits, functioning as a semantic firewall to intercept hallucinations and prevent error cascades. Additionally, a textual strategy evolution module facilitates parameter-free learning through methodological distillation. Extensive benchmarking demonstrates that architectural-level representational alignment effectively enables open-weight models to rival or surpass proprietary systems, providing a robust framework for efficient autonomous agents.
- 10 April 2026
- Published by Elsevier in SSRN Electronic Journal
Open AccessFat-Cat: Document-Driven Metacognitive Multi-Agent System for Complex Reasoning
The efficacy of Large Language Model (LLM) agents is often constrained by the inefficiency of context utilization during runtime. Traditional “log-centric” frameworks rely on syntax-heavy state representations like nested JSON, which impose a syntactic tax that diverts finite attention from semantic reasoning to structural parsing, leading to context dilution in long-horizon tasks. We propose Fat-Cat, a document-driven metacognitive architecture that transitions agent state management from passive logging to active semantic orchestration. Fat-Cat introduces two synergistic mechanisms: (1) Document-Driven State Modeling. By utilizing a Markdown-based semantic file system aligned with pre-training priors, we achieve representational alignment that collapses context growth from a quadratic 𝑂(𝑇 ²) trajectory into a linear 𝑂(𝑇 ) one. (2) Closed-Loop Semantic Watcher. An independent auditor performs pre-emptive audits, functioning as a semantic firewall to intercept hallucinations and prevent error cascades. Additionally, a textual strategy evolution module facilitates parameter-free learning through methodological distillation. Extensive benchmarking demonstrates that architectural-level representational alignment effectively enables open-weight models to rival or surpass proprietary systems, providing a robust framework for efficient autonomous agents.
- Open Access
- 6 November 2023
- Published by Association for Computing Machinery (ACM)
- p. 229-236
AD-TIN: Edge Anomaly Detection for Temporal Interaction Networks using Multi-representation Attention
Anomaly detection in temporal interaction networks (TINs) has become critical in network security, digital finance, and social networks. While recent studies based on Graph Neural Networks (GNNs) have yielded promising results, the existing methods are still limited by insufficient labels and noisy data, often ignoring the information filtering for unrelated user interactions. Therefore, this paper proposes a dynamic edge anomaly detection framework, AD-TIN, to address these challenges based on a multi-representation attention mechanism. It encodes graph structural information using a network information propagation module with neighbor sampling and graph diffusion. Furthermore, the network update module combines past node states with current structural features to capture the temporal information in potential user relationships, effectively mitigating the impact of noisy data. Extensive experiments on three real-world datasets demonstrate the robustness and efficacy of AD-TIN in addressing noise and unrelated interactions for edge anomaly detection.
- 6 November 2023
- Published by Association for Computing Machinery (ACM)
- p. 229-236
Open AccessAD-TIN: Edge Anomaly Detection for Temporal Interaction Networks using Multi-representation Attention
Anomaly detection in temporal interaction networks (TINs) has become critical in network security, digital finance, and social networks. While recent studies based on Graph Neural Networks (GNNs) have yielded promising results, the existing methods are still limited by insufficient labels and noisy data, often ignoring the information filtering for unrelated user interactions. Therefore, this paper proposes a dynamic edge anomaly detection framework, AD-TIN, to address these challenges based on a multi-representation attention mechanism. It encodes graph structural information using a network information propagation module with neighbor sampling and graph diffusion. Furthermore, the network update module combines past node states with current structural features to capture the temporal information in potential user relationships, effectively mitigating the impact of noisy data. Extensive experiments on three real-world datasets demonstrate the robustness and efficacy of AD-TIN in addressing noise and unrelated interactions for edge anomaly detection.
- Open Access
- 1 May 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 45850-45861
Enhancing Recommendation Capabilities Using Multi-Head Attention-Based Federated Knowledge Distillation
As the internet and mobile computing have advanced, recommendation algorithms are used to manage large amounts of data. However, traditional recommendation systems usually require collecting user data on a central server, which may expose user privacy. Furthermore, data and models from different organizations may be proprietary and cannot be shared directly, leading to data isolation. To address these challenges, we propose a method that combines federated learning (FL) with the recommendation system using a federated knowledge distillation algorithm based on a multi-head attention mechanism. In the proposed approach, knowledge distillation is introduced on the basis of FL to induce the training of the local network and facilitate knowledge transfer. Further, to address the non-independent identical distribution of training samples in FL, Wasserstein distance and regularization terms are incorporated into the objective function of federated knowledge distillation to reduce the distribution difference between server and client networks. A multi-head attention mechanism is used to enhance user encoding information. A combined adaptive learning rate is adopted to further improve the convergence. Compared to the benchmark model, this approach demonstrates significant improvements, with accuracy enhanced up to 10%, model training time shortened by approximately 45%, and average error and NDCG values decreased by 10%.
- 1 May 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 11, 45850-45861
Open AccessEnhancing Recommendation Capabilities Using Multi-Head Attention-Based Federated Knowledge Distillation
As the internet and mobile computing have advanced, recommendation algorithms are used to manage large amounts of data. However, traditional recommendation systems usually require collecting user data on a central server, which may expose user privacy. Furthermore, data and models from different organizations may be proprietary and cannot be shared directly, leading to data isolation. To address these challenges, we propose a method that combines federated learning (FL) with the recommendation system using a federated knowledge distillation algorithm based on a multi-head attention mechanism. In the proposed approach, knowledge distillation is introduced on the basis of FL to induce the training of the local network and facilitate knowledge transfer. Further, to address the non-independent identical distribution of training samples in FL, Wasserstein distance and regularization terms are incorporated into the objective function of federated knowledge distillation to reduce the distribution difference between server and client networks. A multi-head attention mechanism is used to enhance user encoding information. A combined adaptive learning rate is adopted to further improve the convergence. Compared to the benchmark model, this approach demonstrates significant improvements, with accuracy enhanced up to 10%, model training time shortened by approximately 45%, and average error and NDCG values decreased by 10%.
- Open Access
- 27 March 2023
- Published by Association for Computing Machinery (ACM)
- p. 1221-1224
MAFD: A Federated Distillation Approach with Multi-head Attention for Recommendation Tasks
The key challenges that recommendation systems must overcome are data isolation and privacy protection issues. Federated learning can efficiently train global models using decentralized data while preserving privacy. In real-world applications, however, it is difficult to achieve high prediction accuracy due to the heterogeneity of devices, the lack of data, and the limited generalization capacity of models. In this research, we introduce a personalized federated knowledge distillation model for a recommendation system based on a multi-head attention mechanism for recommendation systems. Specifically, we first employ federated distillation to improve the performance of student models and introduce a multi-head attention mechanism to enhance user encoding information. Next, we incorporate Wasserstein distance into the objective function of combined distillation to reduce the distribution gap between teacher and student networks and also use an adaptive learning rate technique to enhance convergence. We show that the proposed approach achieves better effectiveness and robustness through benchmarks.
- 27 March 2023
- Published by Association for Computing Machinery (ACM)
- p. 1221-1224
Open AccessMAFD: A Federated Distillation Approach with Multi-head Attention for Recommendation Tasks
The key challenges that recommendation systems must overcome are data isolation and privacy protection issues. Federated learning can efficiently train global models using decentralized data while preserving privacy. In real-world applications, however, it is difficult to achieve high prediction accuracy due to the heterogeneity of devices, the lack of data, and the limited generalization capacity of models. In this research, we introduce a personalized federated knowledge distillation model for a recommendation system based on a multi-head attention mechanism for recommendation systems. Specifically, we first employ federated distillation to improve the performance of student models and introduce a multi-head attention mechanism to enhance user encoding information. Next, we incorporate Wasserstein distance into the objective function of combined distillation to reduce the distribution gap between teacher and student networks and also use an adaptive learning rate technique to enhance convergence. We show that the proposed approach achieves better effectiveness and robustness through benchmarks.
- Open Access
- 6 December 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 9, 167140-167153
Augmenting Seismic Data Using Generative Adversarial Network for Low-Cost MEMS Sensors
The performance of a deep learning (DL) model depends on sufficient training datasets and its algorithmic structure. Even though seismological research using low-cost micro-electro-mechanical systems (MEMS) sensor received much attention recently, because of the lack of data recorded by such MEMS sensors whose data are usually polluted by different types of noise. Therefore, increasing seismic datasets is required by intelligently generating seismic data through data-augmentation techniques. However, it is difficult to characterize and measure the evolution process of seismic sequences, making the feature extraction and data generation of seismic sequences still a significant challenge. By combining the framework of Generative Adversarial Network (GAN) with long short-term memory (LSTM), attention mechanism and neural network (NN), a novel deep generation model (DGM) named EQGAN is developed to overcome the challenges, which can automatically capture the different time histories and dimension characteristics of seismic sequences, meanwhile stably generating high-quality seismic data. The reality of generated data is qualitatively clarified through the analysis of frequency domain and data autocorrelation distribution. Based on the High-throughput Screening (HTS) Theory, the quantitative evaluation index of statistical metrics is designed, and the generation performance of different machine learning models (standard GAN, LSTM, NN) is compared to prove the stability and effectiveness of EQGAN. The experimental results denote that the EQGAN has excellent stability and performance (up to 81%, much higher than that of other generation models), which provides a suitable data expansion approach for the field of seismological research.
- 6 December 2021
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 9, 167140-167153
Open AccessAugmenting Seismic Data Using Generative Adversarial Network for Low-Cost MEMS Sensors
The performance of a deep learning (DL) model depends on sufficient training datasets and its algorithmic structure. Even though seismological research using low-cost micro-electro-mechanical systems (MEMS) sensor received much attention recently, because of the lack of data recorded by such MEMS sensors whose data are usually polluted by different types of noise. Therefore, increasing seismic datasets is required by intelligently generating seismic data through data-augmentation techniques. However, it is difficult to characterize and measure the evolution process of seismic sequences, making the feature extraction and data generation of seismic sequences still a significant challenge. By combining the framework of Generative Adversarial Network (GAN) with long short-term memory (LSTM), attention mechanism and neural network (NN), a novel deep generation model (DGM) named EQGAN is developed to overcome the challenges, which can automatically capture the different time histories and dimension characteristics of seismic sequences, meanwhile stably generating high-quality seismic data. The reality of generated data is qualitatively clarified through the analysis of frequency domain and data autocorrelation distribution. Based on the High-throughput Screening (HTS) Theory, the quantitative evaluation index of statistical metrics is designed, and the generation performance of different machine learning models (standard GAN, LSTM, NN) is compared to prove the stability and effectiveness of EQGAN. The experimental results denote that the EQGAN has excellent stability and performance (up to 81%, much higher than that of other generation models), which provides a suitable data expansion approach for the field of seismological research.