Lejun Ai
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Lejun Ai has a total of 41 co-authors. Here are the top co-authors by number of shared publications:
- Min Chen
Pazhou Laboratory, Guangzhou 510640, China
7 shared publications - Xiao-Kun Wu
School of Journalism and Communication, Renmin University of China, Beijing 100872, China
3 shared publications - Yixue Hao
Huazhong University of Science and Technology, Wuhan, China
3 shared publications - Yue Wang
School of Computer Science and Engineering, South China University of Technology, Guangzhou, China
3 shared publications - Delu Zeng
School of Electronic and Information Engineering, South China University of Technology, Guangzhou, China
2 shared publications - Dusit Niyato
Nanyang technological university, Singapore, Singapore
2 shared publications
Latest Publications
- Open Access
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 3198-3209
Mining Point-of-No-Return Boundaries in Constrained Dynamical Systems via Counterfactual Auditing
- 8 August 2026
- Published by Association for Computing Machinery (ACM)
- p. 3198-3209
Open AccessMining Point-of-No-Return Boundaries in Constrained Dynamical Systems via Counterfactual Auditing
- 15 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Social Systems
- Vol. PP (4) , 1-14
HateMediator: Fine-Tuning Large Language Models for Counter-Hate Speech via Multiturn Mediations
The proliferation of hate speech on social media presents an escalating threat to both public discourse and individual mental well-being. Traditional strategies that prioritize detection and removal often neglect to engage directly with hate speakers or address the underlying causes of their hostility. This article proposes HateMediator, a dialogue-based intervention framework that fine-tunes large language models (LLMs) to generate persuasive, context-aware counter-hate speech. The framework emphasizes two core aspects: the generation of effective counter-hate responses and their evaluation through multiturn dialogues. Our fine-tuning approach integrates tutorial-based learning with critical token guidance, enabling LLMs to recognize and reproduce strategic rhetorical patterns observed in expert interventions. To support training and evaluation, we introduce the MedHate dataset, grounded in social science theory, comprising complete dialogue records from 85 real-world hate incidents (including 255 dialogues), expert-crafted counter-responses, and feedback from the original hate speakers. Experimental results show that HateMediator consistently outperforms baseline LLMs across multiple evaluation dimensions. This study advances both the technical frontier of hate speech intervention and the ethical deployment of LLMs in addressing complex social issues.
- 15 May 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computational Social Systems
- Vol. PP (4) , 1-14
HateMediator: Fine-Tuning Large Language Models for Counter-Hate Speech via Multiturn Mediations
The proliferation of hate speech on social media presents an escalating threat to both public discourse and individual mental well-being. Traditional strategies that prioritize detection and removal often neglect to engage directly with hate speakers or address the underlying causes of their hostility. This article proposes HateMediator, a dialogue-based intervention framework that fine-tunes large language models (LLMs) to generate persuasive, context-aware counter-hate speech. The framework emphasizes two core aspects: the generation of effective counter-hate responses and their evaluation through multiturn dialogues. Our fine-tuning approach integrates tutorial-based learning with critical token guidance, enabling LLMs to recognize and reproduce strategic rhetorical patterns observed in expert interventions. To support training and evaluation, we introduce the MedHate dataset, grounded in social science theory, comprising complete dialogue records from 85 real-world hate incidents (including 255 dialogues), expert-crafted counter-responses, and feedback from the original hate speakers. Experimental results show that HateMediator consistently outperforms baseline LLMs across multiple evaluation dimensions. This study advances both the technical frontier of hate speech intervention and the ethical deployment of LLMs in addressing complex social issues.
- 13 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 72 (1) , 1948-1957
A Hydro-Intelligence Body: Cognitive Architecture for Project-Oriented Autonomous Water Systems Driven by Multi-Modal Sensing and Water Utility Panel
The advancement of emerging technologies such as unmanned aerial vehicles and distributed sensing devices has accelerated the development of Autonomous Intelligent Systems (AIS) with integrated perception, communication, and computation capabilities. However, in mission-critical consumer applications such as smart cities, a major challenge lies in designing a computational architecture that combines deployment flexibility with domain-specific expertise, enabling continuous self-evolution and collaborative decision-making based on dynamic sensing data. To address this challenge, we propose a novel hierarchical cognitive architecture. The framework adopts a project-based methodology, where project engineers can autonomously construct systems by selecting utilities from a utility panel that integrates physical models, data-driven models, and AI models. The system’s intelligence is powered by federated knowledge engine, which coordinates a set of domain-specific fine-tuned large language models (LLMs), and orchestrates their collaboration through a multi-agent workflow to execute complex tasks. We conduct extensive experiments demonstrating that, compared with several general-purpose LLMs, our fine-tuned model under our approach exhibits a significantly lower frequency of moderate-to-severe hallucinations in the water resources engineering domain. These findings validate the feasibility of constructing a federated knowledge engine. Furthermore, this work provides a robust and scalable platform for developing next-generation consumer-grade AIS that require high reliability and physical interpretability.
- 13 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Consumer Electronics
- Vol. 72 (1) , 1948-1957
A Hydro-Intelligence Body: Cognitive Architecture for Project-Oriented Autonomous Water Systems Driven by Multi-Modal Sensing and Water Utility Panel
The advancement of emerging technologies such as unmanned aerial vehicles and distributed sensing devices has accelerated the development of Autonomous Intelligent Systems (AIS) with integrated perception, communication, and computation capabilities. However, in mission-critical consumer applications such as smart cities, a major challenge lies in designing a computational architecture that combines deployment flexibility with domain-specific expertise, enabling continuous self-evolution and collaborative decision-making based on dynamic sensing data. To address this challenge, we propose a novel hierarchical cognitive architecture. The framework adopts a project-based methodology, where project engineers can autonomously construct systems by selecting utilities from a utility panel that integrates physical models, data-driven models, and AI models. The system’s intelligence is powered by federated knowledge engine, which coordinates a set of domain-specific fine-tuned large language models (LLMs), and orchestrates their collaboration through a multi-agent workflow to execute complex tasks. We conduct extensive experiments demonstrating that, compared with several general-purpose LLMs, our fine-tuned model under our approach exhibits a significantly lower frequency of moderate-to-severe hallucinations in the water resources engineering domain. These findings validate the feasibility of constructing a federated knowledge engine. Furthermore, this work provides a robust and scalable platform for developing next-generation consumer-grade AIS that require high reliability and physical interpretability.
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Network Science and Engineering
- Vol. 13, 10965-10979
HongAvg: Hierarchical On-Demand Cognitive Split Federated Learning for Twin Agent Orchestration
- 1 January 2026
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Network Science and Engineering
- Vol. 13, 10965-10979
HongAvg: Hierarchical On-Demand Cognitive Split Federated Learning for Twin Agent Orchestration
- 29 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 74 (11) , 3763-3777
Information Sharing in Multi-Tenant Metaverse via Intent-Driven Multicasting
A multi-tenant metaverse enables multiple users in a common virtual world to interact with each other online. Information sharing will occur when interactions between a user and the environment are multicast to other users by an interactive metaverse (IM) service. However, ineffective information-sharing strategies intensify competitions among users for limited resources in networks, and fail to interpret optimization intent prompts conveyed in high-level natural languages, ultimately diminishing user immersion. In this paper, we explore reliable information sharing in a multi-tenant metaverse with time-varying resource capacities and costs, where IM services are unreliable and alter the volumes of data processed by them, while the service provider dynamically adjusts global intent to minimize multicast delays and costs. To this end, we first formulate the information sharing problem as a Markov decision process and show its NP-hardness. Then, we propose a learning-based system GTP, which combines the proximal policy optimization reinforcement learning with feature extraction networks, including graph attention network and gated recurrent unit, and a Transformer encoder for multi-feature comparison to process a sequence of incoming multicast requests without the knowledge of future arrival information. The GTP operates through three modules: a deployer that allocates primary and backup IM services across the network to minimize a weighted goal of server computation costs and communication distances between users and services, an intent extractor that dynamically infers provider intent conveyed in natural language, and a router that constructs on-demand multicast routing trees adhering to users, the provider, and network constraints. We finally conduct theoretical and empirical analysis on the proposed algorithms for the system. Experimental results show that the proposed algorithms are promising, and superior to their comparison baseline algorithms.
- 29 August 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Computers
- Vol. 74 (11) , 3763-3777
Information Sharing in Multi-Tenant Metaverse via Intent-Driven Multicasting
A multi-tenant metaverse enables multiple users in a common virtual world to interact with each other online. Information sharing will occur when interactions between a user and the environment are multicast to other users by an interactive metaverse (IM) service. However, ineffective information-sharing strategies intensify competitions among users for limited resources in networks, and fail to interpret optimization intent prompts conveyed in high-level natural languages, ultimately diminishing user immersion. In this paper, we explore reliable information sharing in a multi-tenant metaverse with time-varying resource capacities and costs, where IM services are unreliable and alter the volumes of data processed by them, while the service provider dynamically adjusts global intent to minimize multicast delays and costs. To this end, we first formulate the information sharing problem as a Markov decision process and show its NP-hardness. Then, we propose a learning-based system GTP, which combines the proximal policy optimization reinforcement learning with feature extraction networks, including graph attention network and gated recurrent unit, and a Transformer encoder for multi-feature comparison to process a sequence of incoming multicast requests without the knowledge of future arrival information. The GTP operates through three modules: a deployer that allocates primary and backup IM services across the network to minimize a weighted goal of server computation costs and communication distances between users and services, an intent extractor that dynamically infers provider intent conveyed in natural language, and a router that constructs on-demand multicast routing trees adhering to users, the provider, and network constraints. We finally conduct theoretical and empirical analysis on the proposed algorithms for the system. Experimental results show that the proposed algorithms are promising, and superior to their comparison baseline algorithms.
- 23 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. 30 (1) , 230-242
Joint Fine-Grained Representation Learning and Masked Relational Modeling for EEG-Based Automatic Sleep Staging in Fabric Space
Sleep staging is a crucial method for the evaluation of sleep quality and the diagnosis of sleep disorders. In recent years, rapid progress has been made in sleep research through the application of fabric computing and neural networks. Flexible fabric sensors introduced by fabric computing minimize the discomfort of data collection devices on individuals, while neural networkbased algorithms can automatically perform sleep staging based on the collected signals. However, there are two key challenges hinder the integration of automatic sleep staging networks with fabric computing: (1) signals in fabric-based environments exhibit strong heterogeneity due to the wide range of individuals, and (2) interactions between individuals and the fabric space introduce behavioral dynamics to the system. In this paper, we propose a masked autoencoder-based sleep staging neural networks (MAESleepNet), designed to integrate automatic sleep staging algorithm with fabric space. Specifically, MAESleepNet addresses the challenge of signal heterogeneity by learning fine-grained representations from local signals. Furthermore, MAESleepNet tackle the challenge of behavioral dynamics through stochastic masking and reconstruction pre-training. Experiments were conducted on three public datasets: (1) Sleep-EDF-20, (2) Sleep-EDF-78 and (3) SHHS. MAESleepNet achieves overall accuracies of 88.9%, 85.5%, and 87.3%, respectively, outperforming other state-of-theart models. Furthermore, feature visualization and reconstruction visualization experiments were also conducted. The results demonstrates that MAESleepNet is an effective solution to the aforementioned challenges, paving the way for seamless integration into the fabric space.
- 23 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Biomedical and Health Informatics
- Vol. 30 (1) , 230-242
Joint Fine-Grained Representation Learning and Masked Relational Modeling for EEG-Based Automatic Sleep Staging in Fabric Space
Sleep staging is a crucial method for the evaluation of sleep quality and the diagnosis of sleep disorders. In recent years, rapid progress has been made in sleep research through the application of fabric computing and neural networks. Flexible fabric sensors introduced by fabric computing minimize the discomfort of data collection devices on individuals, while neural networkbased algorithms can automatically perform sleep staging based on the collected signals. However, there are two key challenges hinder the integration of automatic sleep staging networks with fabric computing: (1) signals in fabric-based environments exhibit strong heterogeneity due to the wide range of individuals, and (2) interactions between individuals and the fabric space introduce behavioral dynamics to the system. In this paper, we propose a masked autoencoder-based sleep staging neural networks (MAESleepNet), designed to integrate automatic sleep staging algorithm with fabric space. Specifically, MAESleepNet addresses the challenge of signal heterogeneity by learning fine-grained representations from local signals. Furthermore, MAESleepNet tackle the challenge of behavioral dynamics through stochastic masking and reconstruction pre-training. Experiments were conducted on three public datasets: (1) Sleep-EDF-20, (2) Sleep-EDF-78 and (3) SHHS. MAESleepNet achieves overall accuracies of 88.9%, 85.5%, and 87.3%, respectively, outperforming other state-of-theart models. Furthermore, feature visualization and reconstruction visualization experiments were also conducted. The results demonstrates that MAESleepNet is an effective solution to the aforementioned challenges, paving the way for seamless integration into the fabric space.
- 23 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 24 (11) , 11731-11747
Privacy-Enhanced Healthcare Monitoring Service Refreshment in Human Digital Twin-Assisted Fabric Metaverse
Human digital twin bridges humans with digital avatars in the fabric metaverse, assisting users and healthcare professionals with real-time visualization, analysis, and prediction of personal data sensed by fabric sensors. The human digital twin-assisted healthcare monitoring (HHM) service refreshment refers to sending personal health data to corresponding services hosted on nearby edge servers and receiving the results to update local digital avatars continuously. However, the malicious nature and resource limitations of edge servers may lead to user privacy leaks and refreshment timeout, thereby impacting diagnostics. In this paper, we investigate a novel privacy-enhanced HHM service refreshment maximization problem in the fabric metaverse by considering privacy data encryption, model compression, and personalized user requirements. To this end, we first formulate the above issue as an Integer Linear Programming (ILP) problem, and prove its NP-hardness. Then, a resource scheduler named Wiper is designed, consisting of a shallow-deep distiller and an agile refresher library. To enable efficient inference while preserving user privacy, the former replaces violation modules in existing models with approximations and conducts shallow distillation on model layers to meet operation type and depth limits of homomorphic encryption, and then deep distillation on model parameters to decrease end-to-end refreshment delay. Finally, to satisfy user requirements on accuracy and delay during encrypted refreshments while maximizing the throughput of HHM services in offline and online situations with different problem scales, a series of HHM service refreshment algorithms are merged into the latter, including exact, performance-guaranteed approximation, and residual diffusion reinforcement learning algorithms. Theoretical analyses and experiments demonstrate that our algorithms are promising compared with baseline algorithms.
- 23 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Mobile Computing
- Vol. 24 (11) , 11731-11747
Privacy-Enhanced Healthcare Monitoring Service Refreshment in Human Digital Twin-Assisted Fabric Metaverse
Human digital twin bridges humans with digital avatars in the fabric metaverse, assisting users and healthcare professionals with real-time visualization, analysis, and prediction of personal data sensed by fabric sensors. The human digital twin-assisted healthcare monitoring (HHM) service refreshment refers to sending personal health data to corresponding services hosted on nearby edge servers and receiving the results to update local digital avatars continuously. However, the malicious nature and resource limitations of edge servers may lead to user privacy leaks and refreshment timeout, thereby impacting diagnostics. In this paper, we investigate a novel privacy-enhanced HHM service refreshment maximization problem in the fabric metaverse by considering privacy data encryption, model compression, and personalized user requirements. To this end, we first formulate the above issue as an Integer Linear Programming (ILP) problem, and prove its NP-hardness. Then, a resource scheduler named Wiper is designed, consisting of a shallow-deep distiller and an agile refresher library. To enable efficient inference while preserving user privacy, the former replaces violation modules in existing models with approximations and conducts shallow distillation on model layers to meet operation type and depth limits of homomorphic encryption, and then deep distillation on model parameters to decrease end-to-end refreshment delay. Finally, to satisfy user requirements on accuracy and delay during encrypted refreshments while maximizing the throughput of HHM services in offline and online situations with different problem scales, a series of HHM service refreshment algorithms are merged into the latter, including exact, performance-guaranteed approximation, and residual diffusion reinforcement learning algorithms. Theoretical analyses and experiments demonstrate that our algorithms are promising compared with baseline algorithms.