Selected work

Representative Papers

DYNA : Dynamic Episodic Memory Networks for Augmenting Large Language Models with Temporal Knowledge Graphs in Continuous Learning

Jun 14, 2026

This work addresses the challenges of catastrophic forgetting and high retraining costs in continual learning with large language models (LLMs). The authors propose modeling episodic memory as a dynamic temporal knowledge graph—where events serve as nodes and timestamped directed edges represent relations—as an external, updatable memory for a frozen LLM. During inference, retrieval-augmented generation (RAG) is enhanced by combining random walks with centrality-based retrieval from this graph. This approach effectively incorporates new knowledge without requiring model fine-tuning and, for the first time, demonstrates that structural properties of the graph—such as the clustering coefficient—can predict retrieval performance. Experiments show that, across three temporal recall tasks, the method reduces forgetting by approximately 7% compared to fine-tuning baselines and improves temporal ordering accuracy by about 5% over standard RAG.

0 citationsRead paper

MultiCNKG: Integrating Cognitive Neuroscience, Gene, and Disease Knowledge Graphs Using Large Language Models

Oct 08, 2025

This study addresses the limitation of conventional machine learning in modeling complex semantic associations among genes, diseases, and cognitive processes. We propose the first cross-domain alignment framework integrating cognitive neuroscience, genomics, and disease knowledge graphs. Methodologically, we innovatively leverage a large language model (GPT-4) to perform entity alignment, semantic enrichment, and structural completion across three heterogeneous knowledge graphs, followed by link prediction evaluation using TransE and RotatE. We construct a multi-scale knowledge graph spanning molecular to behavioral levels, comprising 6.9K nodes and 11.3K edges. Our framework achieves 85.20% precision and 87.30% recall on entity alignment, with 89.50% expert validation agreement; link prediction performance matches state-of-the-art baselines. The resulting knowledge infrastructure is interpretable, scalable, and supports mechanistic analysis of cognitive disorders and personalized diagnosis and treatment.

0 citationsRead paper

ExKG-LLM: Leveraging Large Language Models for Automated Expansion of Cognitive Neuroscience Knowledge Graphs

Mar 09, 2025

To address low accuracy, narrow coverage, and infrequent updates in Cognitive Neuroscience Knowledge Graph (CNKG) expansion, this paper proposes the first LLM-driven incremental construction framework tailored for CNKGs. The method integrates large language model–based joint entity-relation extraction, semantic consistency–constrained relation inference, knowledge graph embedding optimization, and dynamic graph structure integration—enabling efficient, scalable, and semantically faithful graph updates. Experimental results demonstrate significant improvements: precision reaches 0.80 (+6.67%), recall 0.81 (+15.71%), and F1-score increases by 11.81%. Node and edge counts grow by over 21%, while user interaction rates in clinical decision support scenarios rise by 20%. This work establishes a reusable methodology and technical paradigm for automated, iterative evolution of domain-specific knowledge graphs.

0 citationsRead paper

SKG-LLM: Developing a Mathematical Model for Stroke Knowledge Graph Construction Using Large Language Models

Mar 09, 2025

Existing stroke-domain knowledge graphs (KGs) suffer from insufficient accuracy and semantic depth. Method: This paper proposes a novel KG construction framework integrating large language models (LLMs) with mathematical modeling. GPT-4 is systematically embedded across all KG construction stages—literature preprocessing, entity-relation extraction, and embedding generation—complemented by expert validation to ensure quality. A mathematical optimization model refines graph topology and enforces semantic consistency. Contribution/Results: The approach achieves state-of-the-art performance with precision (0.923) and recall (0.918), outperforming Wikidata and WN18RR benchmarks. The resulting high-quality stroke-domain KG comprises 2,692 nodes (13 types) and 5,012 edges (24 types). In the fully automated construction phase, it attains precision of 0.906 and recall of 0.923.

0 citationsRead paper

Predicting Drive Test Results in Mobile Networks Using Optimization Techniques

Feb 13, 2025

Mobile network operators traditionally rely on drive testing to collect signal data for coverage optimization—a method that is costly, non-reproducible, and environmentally constrained. To address these limitations, this paper proposes a novel received signal strength (RSS) prediction model integrating sparse regression, geographically weighted optimization, and physics-informed signal propagation constraints. It is the first to jointly formulate spatial signal interpolation as an optimization problem, enabling high-fidelity, full-area radio map reconstruction from limited measurements. The model incorporates location-aware regularization and multi-source environmental feature embedding to significantly enhance generalization. Evaluated on real-world 4G/5G networks, it reduces prediction error by 37%, decreases drive-test frequency by 60%, and achieves over 95% accuracy in critical-area coverage assessment—effectively overcoming the cost and reproducibility bottlenecks inherent in conventional drive testing.

0 citationsRead paper

Recent publications

Latest Papers

DYNA : Dynamic Episodic Memory Networks for Augmenting Large Language Models with Temporal Knowledge Graphs in Continuous Learning

Jun 14, 2026

This work addresses the challenges of catastrophic forgetting and high retraining costs in continual learning with large language models (LLMs). The authors propose modeling episodic memory as a dynamic temporal knowledge graph—where events serve as nodes and timestamped directed edges represent relations—as an external, updatable memory for a frozen LLM. During inference, retrieval-augmented generation (RAG) is enhanced by combining random walks with centrality-based retrieval from this graph. This approach effectively incorporates new knowledge without requiring model fine-tuning and, for the first time, demonstrates that structural properties of the graph—such as the clustering coefficient—can predict retrieval performance. Experiments show that, across three temporal recall tasks, the method reduces forgetting by approximately 7% compared to fine-tuning baselines and improves temporal ordering accuracy by about 5% over standard RAG.

0 citationsRead paper

MultiCNKG: Integrating Cognitive Neuroscience, Gene, and Disease Knowledge Graphs Using Large Language Models

Oct 08, 2025

This study addresses the limitation of conventional machine learning in modeling complex semantic associations among genes, diseases, and cognitive processes. We propose the first cross-domain alignment framework integrating cognitive neuroscience, genomics, and disease knowledge graphs. Methodologically, we innovatively leverage a large language model (GPT-4) to perform entity alignment, semantic enrichment, and structural completion across three heterogeneous knowledge graphs, followed by link prediction evaluation using TransE and RotatE. We construct a multi-scale knowledge graph spanning molecular to behavioral levels, comprising 6.9K nodes and 11.3K edges. Our framework achieves 85.20% precision and 87.30% recall on entity alignment, with 89.50% expert validation agreement; link prediction performance matches state-of-the-art baselines. The resulting knowledge infrastructure is interpretable, scalable, and supports mechanistic analysis of cognitive disorders and personalized diagnosis and treatment.

0 citationsRead paper

ExKG-LLM: Leveraging Large Language Models for Automated Expansion of Cognitive Neuroscience Knowledge Graphs

Mar 09, 2025

To address low accuracy, narrow coverage, and infrequent updates in Cognitive Neuroscience Knowledge Graph (CNKG) expansion, this paper proposes the first LLM-driven incremental construction framework tailored for CNKGs. The method integrates large language model–based joint entity-relation extraction, semantic consistency–constrained relation inference, knowledge graph embedding optimization, and dynamic graph structure integration—enabling efficient, scalable, and semantically faithful graph updates. Experimental results demonstrate significant improvements: precision reaches 0.80 (+6.67%), recall 0.81 (+15.71%), and F1-score increases by 11.81%. Node and edge counts grow by over 21%, while user interaction rates in clinical decision support scenarios rise by 20%. This work establishes a reusable methodology and technical paradigm for automated, iterative evolution of domain-specific knowledge graphs.

0 citationsRead paper

SKG-LLM: Developing a Mathematical Model for Stroke Knowledge Graph Construction Using Large Language Models

Mar 09, 2025

Existing stroke-domain knowledge graphs (KGs) suffer from insufficient accuracy and semantic depth. Method: This paper proposes a novel KG construction framework integrating large language models (LLMs) with mathematical modeling. GPT-4 is systematically embedded across all KG construction stages—literature preprocessing, entity-relation extraction, and embedding generation—complemented by expert validation to ensure quality. A mathematical optimization model refines graph topology and enforces semantic consistency. Contribution/Results: The approach achieves state-of-the-art performance with precision (0.923) and recall (0.918), outperforming Wikidata and WN18RR benchmarks. The resulting high-quality stroke-domain KG comprises 2,692 nodes (13 types) and 5,012 edges (24 types). In the fully automated construction phase, it attains precision of 0.906 and recall of 0.923.

0 citationsRead paper

Predicting Drive Test Results in Mobile Networks Using Optimization Techniques

Feb 13, 2025

Mobile network operators traditionally rely on drive testing to collect signal data for coverage optimization—a method that is costly, non-reproducible, and environmentally constrained. To address these limitations, this paper proposes a novel received signal strength (RSS) prediction model integrating sparse regression, geographically weighted optimization, and physics-informed signal propagation constraints. It is the first to jointly formulate spatial signal interpolation as an optimization problem, enabling high-fidelity, full-area radio map reconstruction from limited measurements. The model incorporates location-aware regularization and multi-source environmental feature embedding to significantly enhance generalization. Evaluated on real-world 4G/5G networks, it reduces prediction error by 37%, decreases drive-test frequency by 60%, and achieves over 95% accuracy in critical-area coverage assessment—effectively overcoming the cost and reproducibility bottlenecks inherent in conventional drive testing.

0 citationsRead paper