Graph Structure Research Articles

Graph-Based Similarity of Deep Neural Networks

Understanding the enigmatic black-box representations within Deep Neural Networks (DNNs) is an essential problem in the community of deep learning. An initial step towards tackling this conundrum lies in quantifying the degree of similarity between these representations. Various approaches have been proposed in prior research, however, as the field of representation similarity continues to develop, existing metrics are not compatible with each other and struggling to meet the evolving demands. To address this, we propose a comprehensive similarity measurement framework inspired by the natural graph structure formed by samples and their corresponding features within the neural network. Our novel Graph-Based Similarity (GBS) framework gauges the similarity of DNN representations by constructing a weighted, undirected graph based on the output of hidden layers. In this graph, each node represents an input sample, and the edges are weighted in accordance with the similarity between pairs of nodes. Consequently, the measure of representational similarity can be derived through graph similarity metrics, such as layer similarity. We observe that input samples belonging to the same category exhibit dense interconnections within the deep layers of the DNN. To quantify this phenomenon, we employ a motif-based approach to gauge the extent of these interconnections. This serves as a metric to evaluate whether the representation derived from one model can be accurately classified by another. Experimental results show that GBS gets state-of-the-art performance in the sanity check. We also extensively evaluate GBS on downstream tasks to demonstrate its effectiveness, including measuring the transferability of pretrained models and model pruning.

Properties of Cubic Fuzzy Competition Graphs with Applications

This study extends the concept of competition graphs to cubic fuzzy competition graphs by introducing additional variations including cubic fuzzy out-neighbourhoods, cubic fuzzy in-neighbourhoods, open neighbourhood cubic fuzzy graphs, closed neighbourhood cubic fuzzy graphs, cubic fuzzy (k) neighbourhood graphs and cubic fuzzy [k]-neighbourhood graphs. These variations provide further insights into the relationships and competition within the graph structure, each with its own defined characteristics and examples. These cubic fuzzy CMGs are further classified as cubic fuzzy k-competition graphs that show competition in the k th order between components, p -competition cubic fuzzy graphs that concentrate on competition in terms of membership degrees, and m -step cubic fuzzy competition graphs that analyze competition in terms of steps. Further, some related results about independent strong vertices and edges have been obtained for these cubic fuzzy competition graph classes. Finally, the proposed concept of cubic fuzzy competition graphs is supported by a numerical example. This example showcases how the framework of cubic fuzzy competition graphs can be practically applied to the predator-prey model to illustrate the representation and analysis of ambiguous information within the graph structures.

Towards attributed graph clustering using enhanced graph and reconstructed graph structure

Attributed graph clustering, leveraging both structural and attribute information, is crucial in various real-world applications. However, current approaches face challenges stemming from the sparsity of graphs and sensitivity to noise in Graph Convolutional Networks (GCNs). Moreover, GCN-based methods are often designed based on the assumption of homophilic graph and ignore heterophilic graph. To address these, we propose a graph clustering method that consists of four phases: graph enhance, graph reconstruction, graph refine, and dual-guidance supervisor module. An enhanced graph module is defined by an auxiliary graph to consider distant relationships in the topology structure to alleviate the limitations of sparse graphs. The graph reconstruction phase includes the creation and integration of homophily and heterophily graphs to achieve graph-agnostic. In graph refine, the auxiliary graph is iteratively improved to enhance the generalization of the representation. In this phase, a subspace clustering module is applied to convert attribute-based embeddings into relationship-based representations. Finally, the extracted graphs are fed to a dual-guidance supervisor module to obtain the final clustering result. Experimental validation on several benchmark datasets demonstrates the efficiency of our model. Meanwhile, the findings offer significant advancements in attributed graph clustering, promising improved applicability in various domains.

Topological coindices and QSPR analysis for some potential drugs used in lung cancer treatment via CoM and CoNM-polynomials

Topological indices are used to convert a chemical structure into a real number, usually to study the physicochemical and biological properties of molecules. The groundwork is prepared for the interpretation of the obtained data by processing with Quantitative Structure Property/Activity Relationship (QSPR/QSAR). In this study, the drugs lorlatinib, gefitinib, sotorasib, pralsetinib, crizotinib, adagrasib, alectinib, brigatinib, dacomitinib and entrectinib, which are potential to be used in the treatment of lung cancer, are discussed. Topological coindices are calculated with the help of CoM and CoNM polynomials obtained with the graph structures of these drugs. The relationship between topological coindices and physicochemical properties such as evaporation enthalpy, flash point, molar refraction, polarisation, surface tension, molar volume are investigated by QSPR analysis. At this stage, linear, logarithmic and quadratic regression methods have been used. The results show that the values of these topological indices are highly correlated with certain physicochemical properties of the used some drugs in the treatment of lung cancer. In addition, using comparative analysis, the actual values and the values calculated with the help of topological indices have been examined in terms of predictive ability. The findings of this search demonstrate topological indices’ potential as tools for cancer drug discovery and design.

Efficient Task Scheduling for Large-scale Graph Data Processing in Cloud Computing: A Particle Swarm Optimization Approach

To fulfil the requirements of task scheduling for processing massive amounts of graph data in cloud computing environments, this thesis offers the LGPPSO method, which is based on Particle Swarm Optimisation. The LGPPSO algorithm considers the task’s overall structure when initialising numerous particles in order to broaden the search range and raise the likelihood of finding an approximation optimal solution. We evaluated it in large-scale simulation trials with 100 performance-heterogeneous virtual machines (VMs) using both randomly generated and real application graphs, and evaluated its effectiveness against the CCSH and HEFT algorithms. The experimental findings demonstrate that, in both randomly generated graphs and real graph structure applications, significantly reducing the scheduling duration of large-scale graph data is the LGPPSO algorithm. For randomly generated 200 and 400 node tasks, respectively, the scheduling length is shortened by approximately 8.3% and 9.7% on average when compared to the HEFT algorithm. The LGPPSO technique minimises the scheduling length for actual graphical structure applications by an average of 14.6% and 16.9% for the Gaussian 100 and 200 node examples, respectively.

VulTR: Software vulnerability detection model based on multi-layer key feature enhancement

Software vulnerabilities pose a huge threat to current network security, which continues to lead to data leaks and system damage. In order to effectively identify and patch these vulnerabilities, researchers have proposed automated detection methods based on deep learning. However, most of the existing methods only rely on single-dimensional data representation and fail to fully explore the composite characteristics of the code. Among them, the sequence embedding method fails to effectively capture the structural characteristics of the code, while the graph embedding method focuses more on the global characteristics of the overall graph structure and is still insufficient in optimizing the representation of nodes. In view of this, this paper constructs the VulTR model, which incorporates an importance assessment mechanism to strengthen the key syntax levels of the source code (from lexical elements to nodes and graph-level structures), significantly improving the importance of key vulnerability features in classification decisions. At the same time, a relationship connection diagram is constructed to describe the spatial characteristics of the correlations between functions. Experimentally verified, VulTR's F1 scores on both synthetic and real data sets exceed those of the compared models (VulDeePecker, SySeVR, Devign, VulCNN, IVDetect, and mVulPreter).

A hierarchical GNN across semantic and topological domains for predicting circRNA-microRNA interactions

Identifying circRNA-microRNA interactions (CMI) is a significant biomedical issue in recent years. This problem provides insights into using circRNA as biomarkers, developing cancer therapies and producing cancer vaccines. Using computational methods for identification is a more time-efficient and cost-effective approach. In computational methods, using graphs to represent and explore the CMI is a mainstream approach. However, existing relevant methods do not achieve optimal results by utilizing both the semantic information extracted from sequences and the topological information extracted from graph structures. To address this issue, we propose HGLMALLM, a graph contrastive learning method that learns node representation crossing both the semantic domain generated via motif-aware pre-trained LLMs and the topological domain extracted from hierarchical graph structures. Our method effectively addresses the issue in existing Message Passing Neural Network (MPNN) method that edge components losing heterogeneity after multiple iterations. Moreover, this method utilizes the heterogeneity of graph which is extended from the traditional bipartite graph to heterogeneous through the semantic domain. Two commonly used datasets were partitioned based on the distribution of node degrees. Then, we benchmarked our method against existing methods. In the independent testing set evaluation, it achieved a 3 % and 1 % improvement on two datasets. Our method demonstrated the best stability in ten-fold cross-validation on the training set. A test conducted on the peripheral components reveals robust performance of our model. A dataset collected from real scenarios was used to demonstrate the strong predictive ability of our method for identifying unidentified CMI.

Multi-Objective Combinatorial Optimization Algorithm Based on Asynchronous Advantage Actor–Critic and Graph Transformer Networks

Multi-objective combinatorial optimization problems (MOCOPs) are designed to identify solution sets that optimally balance multiple competing objectives. Addressing the challenges inherent in applying deep reinforcement learning (DRL) to solve MOCOPs, such as model non-convergence, lengthy training periods, and insufficient diversity of solutions, this study introduces a novel multi-objective combinatorial optimization algorithm based on DRL. The proposed algorithm employs a uniform weight decomposition method to simplify complex multi-objective scenarios into single-objective problems and uses asynchronous advantage actor–critic (A3C) instead of conventional REINFORCE methods for model training. This approach effectively reduces variance and prevents the entrapment in local optima. Furthermore, the algorithm incorporates an architecture based on graph transformer networks (GTNs), which extends to edge feature representations, thus accurately capturing the topological features of graph structures and the latent inter-node relationships. By integrating a weight vector layer at the encoding stage, the algorithm can flexibly manage issues involving arbitrary weights. Experimental evaluations on the bi-objective traveling salesman problem demonstrate that this algorithm significantly outperforms recent similar efforts in terms of training efficiency and solution diversity.

IMCSN: An improved neighborhood aggregation interaction strategy for multi-scale contrastive Siamese networks

Graph contrastive learning is an effective method for enhancing graph representation by maximizing the similarity of representations between analogous graphs while minimizing the similarity between dissimilar ones. A common approach to improve the representation capabilities of graphs involves data augmentation, which generates additional training samples through transformations applied to the original graphs. However, traditional data augmentation techniques, such as random deletion of nodes or edges, often compromise the structural integrity of graphs, result in loss of crucial information, and lead to representation instability. To overcome these drawbacks, this study introduces a novel data augmentation paradigm, the integration of directed random noise (IDR), which incorporates controlled random noise to optimize the augmentation objectives. IDR enhances the diversity and robustness of representations without sacrificing the structural integrity of the graphs. This approach significantly improves the performance of graph contrastive learning, avoiding the common issues of graph structure damage and information loss associated with conventional methods. To further refine the model, this paper proposes an improved multiscale contrastive Siamese network framework that employs three Siamese networks to process different views of input graphs. It utilizes cross-network and cross-view contrastive learning objectives to optimize graph representations, leveraging the complementary information between different views to maximize the consistency of representations among similar graphs. This enhancement improves the quality and generalizability of graph representations. Additionally, the introduction of a self-supervised loss function based on graph reconstruction within the loss framework capitalizes on the structural similarity between the original and reconstructed graphs to further refine graph representations. This loss function ensures that the global view retains more structural information from the original graph, thus enhancing the complementarity between the global and local views. The effectiveness and stability of this research framework are demonstrated through node classification tasks and visualizations on six real-world datasets, comparing favorably against existing methods such as MERIT and GraphVAT. The proposed model achieves accuracy improvements of 3.21 % and 3.71 % on the Cora dataset and 1.12 % and 1.42 % on the CiteSeer dataset. Additionally, it ranks first in an average accuracy comparison across six datasets, with scores of 80.6 % and 52.67 %. These results underscore the robustness and efficacy of the proposed model in self-supervised graph representation learning, offering substantial advancements over existing techniques.

CPMKG: a condition-based knowledge graph for precision medicine.

Personalized medicine tailors treatments and dosages based on a patient's unique characteristics, particularly its genetic profile. Over the decades, stratified research and clinical trials have uncovered crucial drug-related information-such as dosage, effectiveness, and side effects-affecting specific individuals with particular genetic backgrounds. This genetic-specific knowledge, characterized by complex multirelationships and conditions, cannot be adequately represented or stored in conventional knowledge systems. To address these challenges, we developed CPMKG, a condition-based platform that enables comprehensive knowledge representation. Through information extraction and meticulous curation, we compiled 307 614 knowledge entries, encompassing thousands of drugs, diseases, phenotypes (complications/side effects), genes, and genomic variations across four key categories: drug side effects, drug sensitivity, drug mechanisms, and drug indications. CPMKG facilitates drug-centric exploration and enables condition-based multiknowledge inference, accelerating knowledge discovery through three pivotal applications. To enhance user experience, we seamlessly integrated a sophisticated large language model that provides textual interpretations for each subgraph, bridging the gap between structured graphs and language expressions. With its comprehensive knowledge graph and user-centric applications, CPMKG serves as a valuable resource for clinical research, offering drug information tailored to personalized genetic profiles, syndromes, and phenotypes. Database URL: https://www.biosino.org/cpmkg/.

Combining graph neural networks and transformers for few-shot nuclear receptor binding activity prediction

Nuclear receptors (NRs) play a crucial role as biological targets in drug discovery. However, determining which compounds can act as endocrine disruptors and modulate the function of NRs with a reduced amount of candidate drugs is a challenging task. Moreover, the computational methods for NR-binding activity prediction mostly focus on a single receptor at a time, which may limit their effectiveness. Hence, the transfer of learned knowledge among multiple NRs can improve the performance of molecular predictors and lead to the development of more effective drugs. In this research, we integrate graph neural networks (GNNs) and Transformers to introduce a few-shot GNN-Transformer, Meta-GTNRP to predict the binding activity of compounds using the combined information of different NRs and identify potential NR-modulators with limited data. The Meta-GTNRP model captures the local information in graph-structured data and preserves the global-semantic structure of molecular graph embeddings for NR-binding activity prediction. Furthermore, a few-shot meta-learning approach is proposed to optimize model parameters for different NR-binding tasks and leverage the complementarity among multiple NR-specific tasks to predict binding activity of compounds for each NR with just a few labeled molecules. Experiments with a compound database containing annotations on the binding activity for 11 NRs shows that Meta-GTNRP outperforms other graph-based approaches. The data and code are available at: https://github.com/ltorres97/Meta-GTNRP.Scientific contributionThe proposed few-shot GNN-Transformer model, Meta-GTNRP captures the local structure of molecular graphs and preserves the global-semantic information of graph embeddings to predict the NR-binding activity of compounds with limited available data; A few-shot meta-learning framework adapts model parameters across NR-specific tasks for different NRs in a joint learning procedure to predict the binding activity of compounds for each NR with just a few labeled molecules in highly imbalanced data scenarios; Meta-GTNRP is a data-efficient approach that combines the strengths of GNNs and Transformers to predict the NR-binding properties of compounds through an optimized meta-learning procedure and deliver robust results valuable to identify potential NR-based drug candidates.

SGIR-Tree: Integrating R-Tree Spatial Indexing as Subgraphs in Graph Database Management Systems

Efficient spatial query processing in Graph Database Management Systems (GDBMSs) has become increasingly important owing to the prevalence of spatial graph data. However, current GDBMSs lack effective spatial indexing, causing performance issues with complex spatial graph queries. This study proposes a spatial index called Subgraph Integrated R-Tree (SGIR-Tree) for efficient spatial query processing in GDBMSs. The SGIR-Tree integrates the hierarchical R-Tree structure with the graph structure of GDBMSs by converting R-Tree elements into graph components like nodes and edges. The Minimum Bounding Rectangle (MBR) information of spatial objects and R-Tree nodes is stored as properties of these graph elements, and the leaf nodes are directly connected to the spatial nodes. This approach combines the efficiency of spatial indexing with the flexibility of graph databases, thereby allowing spatial query results to be directly utilized in graph traversal. Experiments using OpenStreetMap datasets demonstrate that the SGIR-Tree outperforms the previous approaches in terms of query overhead and index overhead. The results are expected to improve spatial graph data processing in various fields, including location-based service and urban planning, significantly advancing spatial data management in GDBMSs.

Constructing the graphical structure of expert-based Bayesian networks in the context of software engineering: A systematic mapping study

Context:In scenarios where data availability issues hinder the applications of statistical causal modeling in software engineering (SE), Bayesian networks (BNs) have been widely used due to their flexibility in incorporating expert knowledge. However, the general understanding of how the graphical structure, i.e., the directed acyclic graph (DAG), of these models is built from domain experts is still insufficient. Objective:This study aims to characterize the SE landscape of constructing the graphical structure of BNs, including their potential for causal modeling. Method:We conducted a systematic mapping study employing a hybrid search strategy that combines a database search with parallel backward and forward snowballing. Results:Our mapping included a total of 106 studies. Different methods are commonly combined to construct expert-based BN structures. These methods span across data gathering & analysis (e.g., interviews, focus groups, literature research, grounded theory, and statistical analysis) and reasoning mechanisms (e.g., using idioms combined with the adoption of lifecycle models, risk-centric modeling, and other frameworks to guide BN construction). We found a lack of consensus regarding validation procedures, particularly critical when modeling cause–effect relationships from knowledge. Additionally, expert-based BNs are mainly applied at the tactical level to address problems related to software engineering management and software quality. Challenges in creating expert-based structures include validation procedures, experts’ availability, expertise level, and structure complexity handling. Key recommendations involve empirical validation, participatory involvement, and balance between adaptation to organizational constraints and model construction requirements. Conclusion:The construction of expert-based BN structures in SE varies in rigor, with some methods being systematic while others appear ad hoc. To enhance BN application, reducing expert knowledge subjectivity, enhancing methodological rigor, and clearly articulating the construction rationale is essential. Addressing these challenges is crucial for improving the reliability of causal inferences drawn from these models, ultimately leading to better-informed decisions in SE practices.

Dual channel visible graph convolutional neural network for microleakage monitoring of pipeline weld homalographic cracks

When using a single sensor to monitor early microleakage of nuclear power pressure pipeline leakage, there are problems such as low monitoring accuracy and poor early warning reliability due to the limitations of the monitoring range and weak difference between the leakage signals. To address these challenges, this paper proposes a dual channel visible graph convolutional neural network (DCV-GCN). Firstly, the acoustic emission time-series data of each channel are truncated and divided, and the significant frequency bands are selected based on the envelope spectrum. On this basis, the sequence group is averaged to obtain the graph structure sequence. Then, the limited penetrable visibility (LPV) graph construction algorithm is used to calculate the adjacency matrix, and the important nodes is reserved according to the eigenvector centrality. Furthermore, the inverse ratio of the distance from the sensor in each single channel to the center of the crack is used as the fusion weight, and the adjacency matrices are merged after normalization to transform the construction of the graph structure dataset. Finally, the dataset is input into the graph convolutional neural network, and the effectiveness of the method is verified by carefully designing three homalographic cracks. The results show that the proposed method can effectively extract the distinguishing features with similar frequency components and similar leakage rates, and the recognition accuracy of different leakage states can reach 98.56 %. In addition, through ablation experiments and different parameter strategy settings, the operating mechanism is explained, which can provide a reference for monitoring and analysis by industrial technicians.

Graph Agent Transformer Network With Contrast Learning for Cross-Domain Recommendation of E-Commerce

Cross-domain recommendation(CDR) system is a kind of recommendation system based on multiple data fields, its core idea is to integrate data from different fields, so as to provide users with more accurate and personalized recommendation services. This paper constructs a hybrid contrast learning graph agent transformer network for CDR of e-commerce. In this method, data in target domain and source domain are first constructed into a multi-part graph structure, and then data mining is carried out by double-branch graph agent transformer. The final embedding of each node in the graph is obtained by training the model through mixed supervised learning and comparative learning. This method not only increases the recommended performance of the model through comparative learning, but also improves the transformer structure through agent and reduces the calculation cost of the model. Simulation results on two real cross-domain recommendation datasets of e-commerce demonstrate the effectiveness of the proposed method.

Interaction-knowledge semantic alignment for recommendation

In order to alleviate the issue of data sparsity, knowledge graphs are introduced into recommender systems because they contain diverse information about items. The existing knowledge graph enhanced recommender systems utilize both user–item interaction data and knowledge graph, but those methods ignore the semantic difference between interaction data and knowledge graph. On the other hand, for the item representations obtained from two kinds of graph structure data respectively, the existing methods of fusing representations only consider the item representations themselves, without considering the personalized preference of users. In order to overcome the limitations mentioned above, we present a recommendation method named Interaction-Knowledge Semantic Alignment for Recommendation (IKSAR). By introducing a semantic alignment module, the semantic difference between the interaction bipartite graph and the knowledge graph is reduced. The representation of user is integrated during the fusion of representations of item, which improves the quality of the fused representation of item. To validate the efficacy of the proposed approach, we perform comprehensive experiments on three datasets. The experimental results demonstrate that the IKSAR is superior to the existing methods, showcasing notable improvement.

A graph representation framework for encrypted network traffic classification

Network Traffic Classification (NTC) is crucial for ensuring internet security, but encryption presents significant challenges to this task. While Machine Learning (ML) and Deep Learning (DL) methods have shown promise, issues such as limited representativeness leading to sub-optimal generalizations and performance remain prevalent. These problems become more pronounced with advanced obfuscation, network security, and privacy technologies, indicating a need for improved model robustness. To address these issues, we focus on feature extraction and representation in NTC by leveraging the expressive power of graphs to represent network traffic at various granularity levels. By modeling network traffic as interconnected graphs, we can analyze both flow-level and packet-level data. Our graph representation method for encrypted NTC effectively preserves crucial information despite encryption and obfuscation. We enhance the robustness of our approach by using cosine similarity to exploit correlations between encrypted network flows and packets, defining relationships between abstract entities. This graph structure enables the creation of structural embeddings that accurately define network traffic across different encryption levels. Our end-to-end process demonstrates significant improvements where traditional NTC methods struggle, such as in Tor classification, which employs anonymization to further obfuscate traffic. Our packet-level classification approach consistently outperforms existing methods, achieving accuracies exceeding 96%.

On the Independent Uphill Domination Polynomial of a Graph

Objectives: The concept of independent uphill domination polynomial is newly defined to deal with the present graph theory in this paper. It is characterized and analyzed within various contexts and types of graphs, its roots are discovered, its relation to different graph transforms is discussed and the study of this graph can help reveal new structures of graphs. Methods: The independent uphill domination polynomial is defined as is an independent uphill dominating set of size in . The uphill paths are used in identifying the criterion of independent uphill dominating sets. The computation of specific values of the polynomial is performed for basic graphs, and the polynomial’s response to various flip operations is observed. Findings: Checking the standard graphs for some properties of the independent uphill dominance polynomial shows the features of the independent uphill dominance polynomial as an uphill dominance polynomial of a certain type of digraph. Families of polynomials can be best understood when it comes to stability and characteristics by analyzing their roots on different graphical planes. They demonstrate how the structure of the graph and the polynomial change as a result of the different graph operations. Novelty: Incorporating uphill paths into independent dominating sets extends common domination polynomials and provides a distinct perspective distinct from previous investigations on graph dominance. Hence, this work enriches the subject and suggests new avenues for studying related polynomials and their uses by encompassing more structural properties of graphs. Keywords: Domination, Polynomial, Uphill domination, Independent domination, Root, Book graph, Independent uphill domination polynomial

Human hippocampal and entorhinal neurons encode the temporal structure of experience.

Extracting the underlying temporal structure of experience is a fundamental aspect of learning and memory that allows us to predict what is likely to happen next. Current knowledge about the neural underpinnings of this cognitive process in humans stems from functional neuroimaging research1-5. As these methods lack direct access to the neuronal level, it remains unknown how this process is computed by neurons in the human brain. Here we record from single neurons in individuals who have been implanted with intracranial electrodes for clinical reasons, and show that human hippocampal and entorhinal neurons gradually modify their activity to encode the temporal structure of a complex image presentation sequence. This representation was formed rapidly, without providing specific instructions to the participants, and persisted when the prescribed experience was no longer present. Furthermore, the structure recovered from the population activity of hippocampal-entorhinal neurons closely resembled the structural graph defining the sequence, but at the same time, also reflected the probability of upcoming stimuli. Finally, learning of the sequence graph was related to spontaneous, time-compressed replay of individual neurons' activity corresponding to previously experienced graph trajectories. These findings demonstrate that neurons in the hippocampus and entorhinal cortex integrate the 'what' and 'when' information to extract durable and predictive representations of the temporal structure of human experience.

Large-language-model empowered 3D dose prediction for intensity-modulated radiotherapy.

Treatment planning is currently a patient specific, time-consuming, and resource demanding task in radiotherapy. Dose-volume histogram (DVH) prediction plays a critical role in automating this process. The geometric relationship between DVHs in radiotherapy plans and organs-at-risk (OAR) and planning target volume (PTV) has been well established. This study explores the potential of deep learning models for predicting DVHs using images and subsequent human intervention facilitated by a large-language model (LLM) to enhance the planning quality. We propose a pipeline to convert unstructured images to a structured graph consisting of image-patch nodes and dose nodes. A novel Dose Graph Neural Network (DoseGNN) model is developed for predicting DVHs from the structured graph. The proposed DoseGNN is enhanced with the LLM to encode massive knowledge from prescriptions and interactive instructions from clinicians. In this study, we introduced an online human-AI collaboration (OHAC) system as a practical implementation of the concept proposed for the automation of intensity-modulated radiotherapy (IMRT) planning. The proposed DoseGNN model was compared to widely employed DL models used in radiotherapy, including Swin Transformer, 3D U-Net CNN, and vanilla MLP. For PTV, DoseGNN achieved the mean absolute error (MAE) of , , , and between true plans and predicted plans that were 64%, 53%, 64%, 61% of the best baseline model. For the worst case among OARs (left lung, right lung, chest wall, heart, spinal cord), DoseGNN achieved the mean absolute error of , , that were 85%, 91%, 80% of the best baseline model. Moreover, the LLM-empowered DoseGNN model facilitates seamless adjustment to treatment plans through interaction with clinicians using natural language. We developed DoseGNN, a novel deep learning model for predicting delivered radiation doses from medical images, enhanced by LLM to allow adjustment through seamless interaction with clinicians. The preliminary results confirm DoseGNN's superior accuracy in DVH prediction relative to typical DL methods, highlighting its potential to facilitate an online clinician-AI collaboration system for streamlined treatment planning automation.