Scientific research project venation intelligent generation method based on graph neural network

Abstract

The invention discloses an intelligent scientific research project venation generation method based on a graph neural network, and belongs to the technical field of computer software. The method comprises the following steps: collecting historical data of a scientific research project and carrying out feature extraction to obtain a historical preprocessing data set to respectively train a heterogeneous graph vectorization model, a dynamic LDA topic model and a weighted semantic association rule model; the heterogeneous graph vectorization model generates a semantic vector of a target entity according to the local heterogeneous graph of the target project, and determines an association relationship between the target entity and a historical entity; the dynamic LDA topic model generates topic distribution of the target scientific research project, the topic distribution is matched with the historical topic evolution tree, and topic similarity is calculated; the weighted semantic association rule model generates an association rule adapted to the target scientific research project according to the entity similarity and the theme similarity; and generating a dynamic venation diagram based on the association relationship between each target entity in the target scientific research project and the historical entity, and the theme evolution path and the association rule of the target scientific research project.

Description

Technical Field

[0001] This invention belongs to the field of computer software technology and relates to an intelligent generation method for scientific research project context based on graph neural networks. Background Technology

[0002] Existing technologies for constructing research project timelines primarily employ traditional data mining and simple relational network analysis methods. A common approach is to integrate research project information based on databases and relational models. Researchers or managers manually compile data on researchers, institutions, contracts, and achievements related to research projects, inputting this information into a relational database. They then use SQL statements to query and perform correlation analysis, thereby constructing the research project timeline. While this method achieves basic information storage and simple relational queries, manual compilation is extremely inefficient when dealing with massive amounts of complex and ever-changing research data. Furthermore, it struggles to uncover deep, unstructured relationships between data points and cannot dynamically update or adapt to the evolving nature of research projects.

[0003] Existing technical solutions utilize traditional knowledge graph construction techniques. For example, the book "Knowledge Graphs: Methods, Practices, and Applications" mentions that by defining entity types and relation types, and based on rules or simple machine learning algorithms, entities and relations are extracted from relevant text and other data sources to construct a knowledge graph for a research field, thereby presenting the context of research projects. However, such techniques rely on pre-defined rules and templates, making it difficult to effectively handle complex and ambiguous research data, such as the correlation between cross-disciplinary research topics and the discovery of implicit collaborations. Furthermore, they cannot adaptively learn and update the knowledge graph structure, resulting in significant shortcomings in the accuracy, completeness, and intelligence of generating research project contexts.

[0004] Furthermore, some existing solutions employ simple network visualization tools to display research project-related data in a simple node-edge graphical form. However, these tools can only achieve data visualization and lack the ability for in-depth data analysis and intelligent processing. They cannot automatically identify key relationships between data points, making it difficult to meet the needs for refined and intelligent contextual information in research project management and analysis.

[0005] The shortcomings of existing technologies are as follows: 1. High reliance on manual labor and low efficiency. Database-based and relational model-based solutions rely on manual sorting of scientific research data. When faced with massive amounts of data, manual entry and analysis are time-consuming and labor-intensive, making it difficult to respond quickly to dynamic changes in scientific research project information and failing to meet real-time requirements.

[0006] 2. Limited relationship mining capabilities Neither traditional database solutions nor knowledge graph construction technologies can easily uncover deep, unstructured relationships between data. For example, the complex connections between cross-disciplinary research content and implicit collaborations among researchers result in incomplete research project outlines that fail to fully reflect the actual situation of the projects.

[0007] 3. High dependence on rules and poor adaptability. Traditional knowledge graph construction techniques rely on pre-set rules and templates, which are difficult to adapt to the complexity and ambiguity of scientific research data. When encountering new research patterns and data types, they cannot automatically adjust and update the knowledge graph structure, thus limiting the dynamic construction and maintenance of the research project context.

[0008] 4. Level of intelligence Simple network visualization tools only focus on data visualization and lack the ability to deeply analyze and intelligently process data. They cannot automatically identify key relationships between data and cannot meet the needs of scientific research project management and analysis for refined and intelligent contextual information. Summary of the Invention

[0009] To address the shortcomings of existing technologies in constructing research project timelines, this invention aims to provide an intelligent method for generating research project timelines based on graph neural networks. This invention achieves in-depth analysis and intelligent presentation of research project information through the integration of multiple technologies. Based on multi-source heterogeneous research data, the technology first collects data using technologies such as database integration, and then preprocesses structured, semi-structured, unstructured, and image data using natural language processing to complete entity recognition and relationship extraction. Utilizing the powerful graph structure data processing capabilities of graph neural networks, it achieves intelligent generation of research project timelines, not only improving data processing efficiency and reducing manual intervention, enabling automatic and rapid analysis and processing of massive amounts of research data to meet the needs of real-time updates of research project information; but also, by deeply mining the complex and unstructured relationships between research data, it fully presents the connections between all elements throughout the entire process of a research project, from initiation and research to output, thus improving the completeness and accuracy of the research project timeline. This invention can also enhance the system's adaptability. Through the learning ability of graph neural networks, it can automatically adapt to changes in scientific research data and new research models, dynamically update and optimize the structure of scientific research project contexts, thereby improving the intelligence level of scientific research project context generation, automatically identifying key connections in the data, and providing more refined and intelligent decision-making basis for scientific research project management, trend analysis, and collaborative network mining.

[0010] The technical solution of this invention is as follows: This invention provides a method for intelligently generating research project context based on graph neural networks, the steps of which include: Historical data of scientific research projects are collected and their features are extracted to obtain a historical preprocessed dataset; the feature data of each scientific research project is a training sample in the historical preprocessed dataset. Based on historical preprocessed datasets, heterogeneous graph vectorization models, dynamic LDA topic models, and weighted semantic association rule models are trained respectively to form a reusable pre-trained model library; Based on the collected target scientific research project data, a local heterogeneous graph of the target project is constructed. A pre-trained heterogeneous graph vectorization model is called to generate the semantic vector of each target entity based on the target project local heterogeneous graph. Then, the similarity between the semantic vector of each target entity and the semantic vector of each entity in the historical entity vector library is calculated to determine the association between each target entity and historical entities, and the theme evolution path of the target scientific research project is generated. The pre-trained dynamic LDA topic model is invoked to obtain the topic distribution of the target research project based on the text vector of the target research project; the topic distribution of the target research project is matched with the historical topic evolution tree to calculate the topic similarity; The pre-trained weighted semantic association rule model is invoked to generate association rules that are suitable for the target scientific research project based on the similarity between semantic vectors of entities and the similarity of the topic. By integrating the relationships between each target entity and historical entities in the target research project, the thematic evolution path and association rules of the target research project, a dynamic context diagram of the target research project is generated.

[0011] Preferably, the nodes in the target project local heterogeneous graph include project number, participants, collaborating institutions, and published papers; the edges in the target project local heterogeneous graph are determined based on the relationships between the nodes, and the relationships include participation, collaboration, and affiliation.

[0012] Preferably, a pre-trained dynamic LDA topic model is used to construct the project context, mine the topic distribution of the project over time, construct a topic-entity association graph and input it into a graph convolutional network to enhance the topic representation, and generate a topic evolution tree; based on the topic evolution tree of the target research project, the future evolution direction of the topic to which the target research project belongs is predicted.

[0013] Preferably, a pre-trained dynamic LDA topic model is used to construct the project context of the target research project, mine the topic distribution of the target research project over time, and construct the topic-entity relationship graph of the target research project; based on the topic-entity relationship graph of the target research project, the historical core personnel under the same topic are recommended for the target research project.

[0014] Preferably, the historical data includes structured data, contract data, achievement data, personnel data, unit data, semi-structured data, unstructured data, and image data. First, the time features in the structured data are subjected to Fourier transform to convert them into periodic vectors. Then, the unstructured data is segmented, stop words are removed, lexical reconstruction is performed, and a dictionary specific to the scientific research field is constructed. Next, Z-score standardization is used to process the numerical data in the historical data to obtain numerical features. BERT is used to process the text data in the historical data to obtain semantic features. Finally, Faster R-CNN is used to extract the image data in the historical data to obtain key image features.

[0015] Preferably, based on the relationship between each target entity and historical entities in the target research project, the theme evolution path and association rules of the target research project, a hierarchical structure of the target research project, namely "basic information layer → entity association layer → theme evolution layer → rule application layer", is constructed and input into a visualization engine to generate a dynamic network diagram of the target research project; project progress warning is issued based on the dynamic network diagram: in combination with the theme evolution path and association rules of the target research project, warnings are triggered for projects that deviate from the output time sequence of projects with the same theme.

[0016] Preferably, a text vector of the target research project is generated based on the project proposal or thesis. Generate image vectors for the target research project based on its image data. ;Transfer text vectors Image vectors The multimodal features of the target research project are obtained by fusing the inputs from the dual-stream attention module; the semantic vector of each target entity is generated by calling the pre-trained heterogeneous graph vectorization model based on the local heterogeneous graph of the target project and the multimodal features of the target research project.

[0017] Another aspect of the present invention provides an intelligent generation system for scientific research project context based on graph neural networks, characterized in that it includes a preprocessing module, a pre-trained model library generation module, a topic evolution path generation module, a topic similarity calculation module, an association rule generation module, and a dynamic context graph generation module; The preprocessing module is used to extract features from the historical data of scientific research projects to obtain a historical preprocessed dataset; wherein the feature data of each scientific research project is a training sample in the historical preprocessed dataset. The pre-trained model library generation module is used to train a heterogeneous graph vectorization model, a dynamic LDA topic model, and a weighted semantic association rule model based on historical preprocessed datasets, forming a reusable pre-trained model library. The topic evolution path generation module is used to construct a local heterogeneous graph of the target project based on the collected target scientific research project data, call a pre-trained heterogeneous graph vectorization model to generate a semantic vector of each target entity based on the local heterogeneous graph of the target project; then calculate the similarity between the semantic vector of each target entity and the semantic vector of each entity in the historical entity vector library, determine the association between each target entity and historical entities, and generate the topic evolution path of the target scientific research project. The topic similarity calculation module is used to call a pre-trained dynamic LDA topic model to obtain the topic distribution of the target research project based on the text vector of the target research project; and to match the topic distribution of the target research project with the historical topic evolution tree to calculate the topic similarity. The association rule generation module is used to call a pre-trained weighted semantic association rule model to generate association rules that are suitable for the target scientific research project based on the similarity between semantic vectors of entities and the similarity of the topic. The dynamic context diagram generation module is used to integrate the relationship between each target entity and historical entities in the target scientific research project, the theme evolution path and association rules of the target scientific research project, and generate a dynamic context diagram of the target scientific research project.

[0018] The present invention also provides a server, characterized in that it includes a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the above-described methods.

[0019] The present invention also provides a computer-readable storage medium having a computer program stored thereon, characterized in that the computer program implements the above-described method when executed by a processor.

[0020] Key aspects of this invention include: 1. Vectorization Representation Techniques for Multimodal Heterogeneous Graphs As the core technology of the "data → vector" transformation process in this invention, it is crucial for subsequent topic analysis, rule mining, and application decisions. This approach addresses the core pain points of traditional technologies, namely "fragmented multi-source data" and "difficulty in quantifying entity relationships." Multimodal heterogeneous graph modeling refers to a modeling technique that deeply integrates structured, semi-structured, and unstructured multimodal data involved in research projects with heterogeneous graph structures containing multiple types of entities (projects, contracts, results, personnel, institutions, etc.) and relationships (affiliation, cooperation, citation, evolution, etc.). It learns feature representations of different modalities and heterogeneous relationships through graph neural networks (such as HGAT), comprehensively depicting the complex relationships between research entities. Heterogeneous Graph Attention Network (HGAT) is a graph neural network designed for heterogeneous graph structures. Through a relationship-specific attention mechanism, it assigns independent attention weights to different types of edges (such as "personnel-project" affiliation edges and "project-document" citation edges), achieving differentiated aggregation of heterogeneous neighbor node features and enhancing the semantic distinguishability of node representations.

[0021] This invention proposes a multimodal embedding method that integrates BERT text embedding, structural features of graph neural networks (HGAT), and numerical features to construct a heterogeneous information network containing multiple types of entities (projects, contracts, results, personnel, units, etc.) and multidimensional relationships.

[0022] By dynamically aggregating heterogeneous neighbor information through a multi-head attention mechanism, the problem that traditional single-modal representation cannot capture the complex relationships between scientific research entities is solved, thereby improving the accuracy of node embedding.

[0023] Transforming non-standardized, multi-source scientific research data into machine-computable vector forms allows for the precise capture of explicit relationships between entities (such as personnel). The system associates entities with implicit relationships (such as indirect collaborations across organizations) and provides a high-quality entity representation foundation for subsequent stages. It generates standardized vectors for each entity (e.g., a 512-dimensional semantic vector for a project), improving node embedding accuracy by over 25% compared to traditional single-modal methods. Vector similarity directly indicates the degree of entity association (e.g., similarity between project vectors in the same domain > 0.7), providing a quantitative basis for "target project - historical project" association matching and entity relationship analysis.

[0024] 2. Dynamic Temporal Topic Evolution Modeling Method The output of the "multimodal heterogeneous graph vectorization representation technology" is the core of the "vector → topic evolution" transformation. Addressing the pain point of traditional static models "failing to capture the dynamic changes of research topics," this approach provides a "time dimension" for project development. This provides a logical framework and offers thematic trend information for subsequent application decisions.

[0025] A dynamic LDA model based on quarterly time slices is designed, introducing KL divergence constraints to achieve temporal smoothing of topic distribution, and constructing "continuation / split / merge" topic evolution paths using the Hungarian algorithm. The dynamic LDA model is an improved model based on the traditional Latent Dirichlet Allocation (LDA) topic model, incorporating quarterly time slices and temporal smoothing constraints (such as a KL divergence threshold of 0.1). It can capture the evolutionary patterns of research topics over time, including continuation, splitting, merging, and disappearance, and compared to static LDA, it can dynamically adapt to the temporal characteristics of scientific and technological development.

[0026] Breaking through the limitations of static topic analysis in traditional LDA models, this approach effectively captures the gradual evolution of research topics over time, from "deep learning" to "multimodal large models," thus improving the accuracy of topic evolution identification. It mines the temporal evolution patterns of research topics from entity vectors, clarifying the historical inheritance and future trends of project topics. This allows project paths to possess both "entity connections" and "technological development logic," solving the problem of traditional paths "lacking a time dimension and failing to reflect technological iteration." It generates "research topic evolution trees" (e.g., the path "deep learning → multimodal large model → intelligent analysis of research data") and "topic-entity relationship graphs" (e.g., the topic "large model fine-tuning" associating 15 projects and 12 core personnel). The accuracy of topic evolution identification is improved by more than 30% compared to traditional static LDA, and it can be directly used for target project topic positioning and technology trend prediction.

[0027] Addressing the core shortcomings of existing scientific research project context construction technologies, such as strong reliance on manual intervention, shallow relationship mining, poor dynamic adaptability, and insufficient intelligence, the present invention offers the following specific advantages: 1. Multimodal automated processing Without human intervention, it automatically collects and processes multi-source data through natural language processing (such as BERT text parsing), supporting various input types including text, images, and time series. For example, it can automatically extract technical terms such as "Transformer model" and "quantum dot material" from project summaries and associate them with corresponding technical nodes to build a complex relationship network containing multiple entities.

[0028] 2. Utilize graph neural networks (such as HGAT) The system enables real-time updates of embedded nodes. When a new research project is initiated or results are published, the system can automatically identify newly added entity associations without manually modifying the database structure. It dynamically adapts to the continuous evolution of research projects, solving the problems of "data lag and fixed structure" in traditional solutions.

[0029] 3. Irregular depth modeling By using heterogeneous graph neural networks and dynamic topic models, complex relationships between entities are automatically learned. It can identify latent evolutionary relationships between topics; these cross-period and cross-domain implicit relationships cannot be captured by traditional rule-based methods due to their lack of semantic and temporal analysis capabilities.

[0030] 4. Intelligent decision-making closed-loop support Based on semantically enhanced rule mining, actionable decision-making knowledge is output, achieving a leap from "data visualization" to "intelligent decision-making." Semantically enhanced rule mining refers to association rule mining technology based on weighted support calculation and rule graph neural networks (RGNN). It integrates node centrality (e.g., PageRank), topic co-occurrence (LDA output), and temporal constraints (±6 months) to select candidate sets, and uses RGNN to embed the rule graph (containing project, personnel, and topic nodes) to verify the semantic rationality and logical coherence of the rules. Rule graph neural networks (RGNN) transform association rules into a heterogeneous graph structure (containing the project, personnel, and topic nodes involved in the rule's antecedent and consequent). It uses graph neural networks to learn the vector representation of the rules, combining confidence prediction and consequent generation tasks to evaluate rule quality and support zero-shot rule expansion, improving the intelligence and interpretability of rule mining.

[0031] This invention addresses the three core shortcomings of existing technologies (manual dependence, rigid rules, and insufficient intelligence) by achieving a leapfrog upgrade in the construction of scientific research project contexts from "manual sorting + simple association" to "intelligent analysis + dynamic prediction" through three innovations: automated multimodal modeling, ruleless deep relationship mining, and intelligent decision support. Attached Figure Description

[0032] Figure 1 This is a flowchart of the method of the present invention.

[0033] Figure 2 This is an architectural diagram of the present invention.

[0034] Figure 3 This is a system diagram of the present invention. Detailed Implementation

[0035] The present invention will now be described in further detail with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0036] This invention relates to an intelligent generation technology for scientific research project contexts based on graph neural networks. It aims to address existing problems in scientific research project context construction, such as strong reliance on manual intervention, insufficient relationship mining, and poor dynamic adaptability. By integrating multiple technologies, it achieves in-depth analysis and intelligent presentation of scientific research project information. Based on multi-source heterogeneous scientific research data, the technology first collects data using database integration and other technologies. Then, it preprocesses structured, semi-structured, unstructured, and image data using natural language processing to complete entity recognition and relationship extraction. In the key information vectorization representation learning stage, a multimodal hybrid embedding strategy is adopted, combining models such as BERT and ResNet to extract text and image features, and constructing a heterogeneous graph containing multiple types of nodes and relationship edges such as membership, including researchers, institutions, contracts, and achievements. Heterogeneous Graph Neural Network (HGAT) is used to learn node embeddings, and multimodal information is fused through a two-stream attention mechanism. A pre-trained dynamic LDA topic model is used to construct project contexts, mine the topic distribution of projects over time, construct a topic-entity association graph, and input it into a graph convolutional network to enhance topic representation, generating a topic evolution tree. Regarding the mining of project association rules, Node2Vec and GAT are used to extract graph structure features, calculate project similarity, and an improved weighted Apriori algorithm is used to mine association rules. A graph isomorphic network (GIN) is then used to verify and rank the rules.

[0037] The topic-entity association graph refers to adding a "topic" node in a heterogeneous information network, constructing association edges such as "project-topic" and "personnel-topic" through the topic probabilities output by the dynamic LDA model, and establishing a quantitative connection between the abstract topic concept and specific scientific research entities (such as projects and personnel), which helps in technology trend analysis and association rule mining.

[0038] The overall technical process encompasses data acquisition and preprocessing, vectorized representation learning, thematic context construction, association rule mining, and intelligent application services. The modules work collaboratively to intelligently generate a structured and visualized context diagram of the complex relationships throughout the entire process of a research project, from project initiation to output. This provides accurate project relationship analysis, early warning monitoring, and trend prediction for scenarios such as project violation warnings and personnel misconduct warnings, significantly improving the efficiency of scientific research management and decision-making.

[0039] The method flow of the present invention is as follows: Figure 1 As shown, the steps include: Historical data of scientific research projects are collected and features are extracted to obtain a historical preprocessed dataset; the feature data of each scientific research project is a training sample in the historical preprocessed dataset. A heterogeneous graph vectorization model, a dynamic LDA topic model, and a weighted semantic association rule model were trained based on a historical preprocessed dataset. Based on the collected target scientific research project data, a local heterogeneous graph of the target project is constructed. A pre-trained heterogeneous graph vectorization model is called to generate the semantic vector of each target entity based on the target project local heterogeneous graph. Then, the similarity between the semantic vector of each target entity and the semantic vector of each entity in the historical entity vector library is calculated to determine the association between each target entity and historical entities, and the theme evolution path of the target scientific research project is generated. The pre-trained dynamic LDA topic model is invoked to obtain the topic distribution of the target research project based on the text vector of the target research project; the topic distribution of the target research project is matched with the historical topic evolution tree to calculate the topic similarity; The pre-trained weighted semantic association rule model is invoked to generate association rules that are suitable for the target scientific research project based on the similarity between semantic vectors of entities and the similarity of the topic. Based on the relationship between each target entity and historical entities in the target research project, the theme evolution path and association rules of the target research project, a dynamic context diagram of the target research project is generated.

[0040] The architecture of this invention is as follows Figure 2 As shown, the specific content is as follows.

[0041] (a) Preliminary steps: Historical data processing and model training (providing a basic model for the target project) 1. Historical data preprocessing Technical Functions: Collecting structured data (project initiation information (number, budget, period), contract data (major researchers of the undertaking unit, contract price, milestones), outcome data (form, quantity, brief description, completing unit, etc.), personnel / unit data (name, title, ID number, address, etc.)), semi-structured data (contracts, project application forms, etc.), unstructured data (papers, research reports, patents), and image data (outcome images, scanned copies, etc.) from research projects over the past 5-10 years. Using Z-score standardization to process numerical features (eliminating dimensional differences and avoiding excessive model bias towards budget features due to "budget amounts being much larger than the number of personnel"), BERT to process text (extracting semantics), and Faster R-CNN to extract key image features, transforming the raw data into well-structured training data.

[0042] Supplementary Technique 1: NLTK / spaCy Text Preprocessing – This technique segments unstructured text (such as paper abstracts and research reports), removes stop words (eliminating meaningless words like "research" and "analysis"), and performs lemma reconstruction (e.g., converting "studying" to "study"). It also constructs a research-specific dictionary (including professional terms like "large model fine-tuning" and "quantum dot") to clean up text noise, standardize text format, and preserve core semantics within the research field, addressing the issue of general text processing tools losing specialized semantics. The preprocessed, well-structured text is then input into BERT (or alternative models like ELMo) to extract semantic feature vectors, providing high-quality text feature input for subsequent multimodal heterogeneous graph modeling and topic analysis.

[0043] Supplementary Technique 2: Fourier Transform Temporal Feature Processing – This technique performs a Fourier transform on temporal features in structured data (such as project start and end times, contract milestones), converting them into periodic vectors that retain the "quarterly / annual" temporal regularity. This avoids the problem that traditional time coding cannot capture the temporal correlation of research projects. The periodic vectors obtained after the Fourier transform serve as one of the important features of the data. They are then fused with text features (extracted by BERT after preprocessing using Supplementary Technique 1), image features (extracted by Faster R-CNN), and numerical features (normalized by Z-scores), and input into multimodal heterogeneous graph modeling (such as HGAT) for deep feature learning. This provides high-quality temporal dimension feature support for subsequent steps such as entity vectorization and topic evolution analysis.

[0044] The collected data is first processed using Supplementary Technique 1, and then feature extraction is performed; specific collected data is processed using Supplementary Technique 2, and then subsequent feature extraction is carried out; the final result is: generating a "historical preprocessed dataset", with an entity (project, person, etc.) recognition accuracy of over 92% and an image feature matching accuracy of over 88%.

[0045] 2. Core Model Training Technical role: Based on historical preprocessed datasets, train "Heterogeneous Graph Vectorization Model (HGAT)," "Dynamic LDA Topic Model," and "Weighted Semantic Association Rule Model" to form a reusable pre-trained model library.

[0046] Supplementary Techniques: Multi-Task Learning Framework Optimization – For the HGAT model, a multi-task learning objective of "node classification + link prediction + representation reconstruction" is designed (loss function: The Adam optimizer is used to adaptively adjust the learning rate, combined with an early stopping mechanism to prevent overfitting. At the same time, the graph data augmentation technique of "node perturbation + edge mask" is used to improve the generalization ability of the model, solve the problem of insufficient robustness of the model caused by single task training, and form a reusable pre-trained model library.

[0047] Explanation of the "Node Classification + Link Prediction + Representation Reconstruction" Technology These three elements together constitute the joint training objective of the Heterogeneous Graph Vectorization Model (HGAT), which enhances the model's ability to represent nodes in heterogeneous graphs in the scientific research field through multi-task collaborative optimization.

[0048] Node classification: Given type labels for nodes in a heterogeneous graph (such as preset entity categories like "project," "personnel," "unit," and "achievement"), the model is trained to accurately predict the category of a node based on node features and graph structure. The model learns the semantic discriminative power of different entity categories, ensuring that node vectors reflect the essential type attributes of entities (e.g., the vector features of "personnel nodes" and "project nodes" can be clearly distinguished).

[0049] Link prediction: Based on existing edges in the heterogeneous graph (such as the "participation" relationship between "personnel-project" and the "output" relationship between "project-outcome"), the model is trained to predict missing edges or edges that may be generated in the future (such as unlabeled "implicit cooperation" relationships). This allows the model to learn the association patterns between entities (such as the higher probability of connections between "projects in the same field" and "core researchers"), strengthening the ability of node vectors to characterize entity relationships.

[0050] Reconstruction: Taking the node vectors output by the model as input, the reconstruction network restores the original features of the nodes (including multimodal original inputs such as text embeddings, numerical features, and image features). The node vectors are constrained to retain key information from the original data to prevent the model from losing the basic attributes of entities during graph structure learning, ensuring the integrity of the vector representation.

[0051] In an optional embodiment, the specific training method for the Heterogeneous Graph Vectorization Model (HGAT) is as follows: 1. Data Preparation 1.1 Input: A heterogeneous graph constructed from preprocessed historical scientific research data (nodes contain "projects, personnel, units, achievements, contracts, etc.", and edges contain "affiliation, cooperation, citation, evolution, etc."), with each node associated with multimodal features (BERT text embedding, numerical features, image features, etc.). 1.2 Label Construction: Node Classification Labels: Each node is labeled with a preset entity type; Link Prediction Labels: Existing edges in the heterogeneous graph are marked as "positive samples", and non-existent edges are randomly sampled (edges that are logically impossible need to be excluded, such as "result-result" which has no direct relationship) as "negative samples"; The original multimodal features of the nodes are used to reconstruct the labels.

[0052] 2. Model Structure Construction 2.1 Basic Network: Heterogeneous Graph Attention Network (HGAT), containing 3-5 attention layers, each layer assigns independent weights to different types of edges through a relation-specific attention mechanism; 2.2 Multi-task branches: Classification branch, the node vectors output by HGAT are connected to a fully connected layer + Softmax activation function, and the node class probability distribution is output; Link prediction branch, the dot product operation is performed on the vectors of the target node pair (u,v), and the probability of the edge exists is output; Reconstruction branch, the node vectors output by HGAT are connected to two fully connected layers, and the reconstructed vectors with the same dimensions as the original multimodal features are output.

[0053] 3. Loss Function Design The joint loss function using weighted summation is as follows: : Cross-entropy loss is used to calculate the difference between the predicted class and the true class; The binary cross-entropy loss is used to calculate the difference between the predicted edge existence probability and the true label (0 / 1). The mean squared error (MSE) loss is used to calculate the difference between the reconstructed features and the original features.

[0054] Optimizer: Adam optimizer, which adaptively adjusts the learning rate.

[0055] Regularization strategy: Introduce graph data augmentation technology of "node perturbation + edge masking" and adopt an early stopping mechanism (stop training if the validation set loss does not decrease for 5 consecutive rounds) to prevent overfitting.

[0056] 4. Training Iteration Process 4.1 Initialize HGAT model parameters; 4.2 Batch input of heterogeneous graph data, generating node vectors using HGAT; 4.3 Calculate the loss items for the three tasks separately, and sum them using the weighted formula to obtain the total loss; 4.4 Backpropagation total loss, update all model parameters; 4.5 Repeat steps 2-4 until the preset number of iterations (e.g., 200 rounds) is reached or the early stopping mechanism is triggered, and the training ends.

[0057] 5. The logic and method for determining the values ​​of key parameters (λ1, λ2, λ3) 5.1 Preset of value range: Based on the conventional experience of multi-task training of graph neural networks, the value range of λ1, λ2, and λ3 is first set to [0.1, 1.0] (to ensure that the loss of each task is of a similar magnitude and to avoid the optimization being dominated by a single task).

[0058] 5.2 Grid search optimization: The grid search method is used to traverse the candidate combination parameters within the preset range.

[0059] 5.3 Optimal parameter selection criteria: The optimal parameter combination is selected based on the "node embedding accuracy" (a core patent indicator, i.e., the accuracy of entity vectors in depicting the relationship).

[0060] In one optional embodiment of the present invention, the optimal values ​​used (verified experimentally) are: λ1=0.4, λ2=0.3, λ3=0.3 (this combination improves the node embedding accuracy by more than 25% compared to the traditional single task, meeting the quantitative requirement of "vector similarity between projects in the same domain > 0.7"). Inventive Contribution Description This invention addresses the shortcomings of traditional single-task training: Traditional HGAT only uses a single task of "node classification" or "link prediction", which easily leads to node vectors being "type-heavy and relation-light" or "relation-heavy and attribute-light". This invention preserves the original features through "representation reconstruction" constraints, and the three work together to achieve a comprehensive representation of "type + relation + attribute".

[0061] Adapting to the complexity of scientific research data: Scientific research data is multimodal and highly heterogeneous (e.g., the "personnel" node contains text attributes, and the "project" node contains numerical attributes). The multi-task design enables the model to learn entity type differentiation, association rules, and original attribute preservation at the same time. The generated vectors are more suitable for subsequent core links such as topic evolution analysis and association rule mining.

[0062] Enhancing model generalization ability: Through data augmentation and multi-task regularization using "node perturbation + edge masking", the stability and accuracy of vector representation are significantly improved when the model processes new scientific research project data (such as unseen "cross-domain projects"), supporting the dynamic adaptability of context generation.

[0063] Explanation of the "Node Perturbation + Edge Mask" Technology Meaning: Node perturbation refers to adding small random noise to the multimodal features (text, numerical values, etc.) of some nodes in a heterogeneous graph; edge masking refers to randomly masking a small number of non-critical edges (such as low-weight related edges) in the graph.

[0064] Application during training: Before each round of model training, perform "node perturbation + edge masking" on the input heterogeneous graph to generate enhanced samples before inputting them into HGAT for training.

[0065] Values: The perturbation noise intensity is set to 0.01-0.05 (experiments have verified that this range can preserve the essence of features and enhance generalization); the edge mask ratio is set to 5%-10% (to avoid masking key associations), and the optimal values ​​are selected by combining grid search with node embedding accuracy.

[0066] Creative contribution: Customized for the multimodal and strongly correlated characteristics of heterogeneous graphs in scientific research, solving the overfitting problem of single-task training and improving the model's adaptability to new scientific research data.

[0067] Results: The model library can output 512-dimensional entity vectors, topic evolution trees, and association rules, providing direct support for target project analysis.

[0068] (II) Generation of the target project context (taking the test data "2024KJ001 Large Model Scientific Research Data Project" as an example) Step 1: Vectorization of the target project entity (using the key technology HGAT modeling technique) Technical Application: Data is collected from the target research project (structured data: project number, cycle, funding, personnel / undertaking unit; unstructured data: project proposal, papers, etc.). The project's structured and unstructured data are collected, extracting nodes such as project, participants, collaborating units, and achievements. Relationships such as participation, collaboration, and affiliation are clarified, forming a node-edge heterogeneous graph structure. This constructs a "local heterogeneous graph of the target project," providing input for the HGAT model to support entity vectorization and association matching. It accurately characterizes the core element relationships of the target project (nodes: project number (2024KJ001), participants, collaborating units, published papers; edges: participation, collaboration, affiliation). A pre-trained "heterogeneous graph vectorization model (HGAT)" is called to learn the 512-dimensional semantic vectors of the target entities and calculate similarity by associating them with a "historical entity vector library." The local heterogeneous graph of the target project is then input into this heterogeneous graph vectorization model (HGAT) to obtain the semantic vector of each target entity. Calculate the similarity between the semantic vector of each target entity in the local heterogeneous graph of the target project and the semantic vectors of all entities in the historical entity vector library in order to locate related entities in the same domain.

[0069] Supplementary Technology: Dual-Stream Attention Mechanism for Multimodal Fusion – The text vector of the target research project (BERT processing of research proposals / papers) and the image vector (Faster R-CNN processing of experimental images) are input into the dual-stream attention module, and dynamic weight allocation (formula: This approach integrates multimodal features to address the issue that single textual features cannot fully characterize project information. A pre-trained heterogeneous graph vectorization model is invoked to generate semantic vectors for each target entity based on the local heterogeneous graph of the target project and the multimodal features of the target research project.

[0070] Subsequent process: Construct a "target project local heterogeneous graph" (nodes: 2024KJ001, participants, collaborating institutions, published papers; edges: participants, collaborators, affiliated), call the pre-trained "heterogeneous graph vectorization model (HGAT)", learn the 512-dimensional semantic vectors of the target entities, and calculate the similarity by associating them with the "historical entity vector library".

[0071] Technical role: It transforms the non-standardized data of the target scientific research project into a machine-computable vector, quickly locates the association with historical entities (such as projects in the same field or similar researchers), and solves the problem of "no quantitative basis for the association of the target project entities".

[0072] Application results: The 512-dimensional vector of project 2024KJ001 was obtained, and the similarity with the vector of the historical "2023KJ008 multimodal project" reached 0.82, which determined it to be a "continuation project in the same field", thus clarifying the position of the target scientific research project in the historical scientific research network.

[0073] Step 2: Target Research Project Theme Analysis (Applying Key Technology Dynamic LDA Theme Evolution Technology) Technical Application: The pre-trained "Dynamic LDA Topic Model" is invoked, and the text vector of the target research project (proposal, full text of the paper) is input. The output topic distribution of the target research project is matched with the "historical topic evolution tree" to calculate the topic similarity and determine the topic association graph of the target research project (including topic evolution relationships).

[0074] Technical role: To accurately locate the core research theme of the target scientific research project, clarify its inheritance relationship with historical themes, and solve the problem of "ambiguous positioning of the target project theme and lack of historical development context reference".

[0075] Application Results: The core themes of project 2024KJ001 were determined to be "large model fine-tuning (40%)" and "scientific data cleaning (35%)", which are "continuation relationships" with the historical theme "multimodal large model". At the same time, 8 historical core participants under this theme were matched to enrich the project theme dimension information.

[0076] Step 3: Mining Association Rules for Target Scientific Research Projects (Applying Key Technologies such as Weighted Semantic Association Rule Mining Techniques) Technical Application: Combining the "target project entity vector" output in step 1 (to calculate entity similarity) and the "target research project topic association graph" output in step 2 (to calculate topic fit), the pre-trained "weighted semantic association rule model" (trained based on historical data, with "entity centrality" and "topic co-occurrence" weights introduced in the rule support calculation, and invalid rules filtered out through a semantic similarity threshold of 0.6) is used to match and adapt the association rules to the target research project.

[0077] Technical role: To select rules (such as progress prediction and outcome patterns) that have practical management value for target scientific research projects from historical rules, and to solve the problem of "target projects lacking forward-looking analysis and decision-making basis".

[0078] Application Results: Output 3 core rules: ① Before June 2025 (mid-term of the project), it is easy to produce 1 core paper (based on progress warning); ② 2 cooperating units → 30% increase in the conversion rate of results (based on the evaluation of cooperation model); ③ It is recommended to invite Wang XX, a core member of the historical theme, to join (based on personnel selection). The accuracy rate of the rules is over 92%.

[0079] Step 4: Integration of Target Project Framework Technical Application: Integrating the results of Step 1 (entity vector association), Step 2 (topic evolution path), and Step 3 (association rules), a four-level structure of "basic information layer → entity association layer → topic evolution layer → rule application layer" is constructed. A dynamic network diagram is generated through a visualization engine (nodes: entities / topics, edges: association / evolutionary relationships, and color-coded association strength).

[0080] Technical role: Transform the output of previous key technologies into an intuitive and structured framework, solving the problem of "dispersed results being difficult to implement and unable to intuitively support management".

[0081] Application results: A "2024KJ001 Project Network Diagram" was generated with a network integrity of over 95%, which can be directly used for project progress tracking, personnel allocation, trend judgment, and support scientific research management decisions.

[0082] (III) Application and Decision Support As the core link in the transformation of "theme evolution → practical value" in this invention, the output results of the first two key technologies are integrated to transform "entity vector" and "theme evolution" into a feasible scientific research management decision-making solution, solving the pain point of traditional technologies that "only output data / models without practical application value".

[0083] Output decision-making solutions for core scenarios in scientific research management: Project progress warning: Combining theme evolution path and association rules, a warning is triggered for projects that deviate from the "time sequence of output of projects with the same theme" (such as theme continuation projects that have not produced core papers for more than 18 months); the theme evolution path is generated by analyzing historical data using a dynamic LDA model (introducing quarterly time slices and KL divergence constraints), which can capture evolution patterns such as theme continuation and splitting, i.e. theme evolution rules.

[0084] Personnel configuration recommendation: Utilize a pre-trained dynamic LDA topic model to construct the project context of the target research project, mine the topic distribution of the target research project over time, and construct the topic-entity relationship graph of the target research project; Based on the "topic-entity relationship graph", recommend historical core personnel under the same topic for the target project (e.g., recommend senior researchers who have participated in the topic for the "large model fine-tuning" project). Technology trend prediction: By extending the analysis of the theme evolution tree, we can predict the future evolution direction of the theme to which the target project belongs (such as "large model fine-tuning" may evolve into "multimodal large model collaborative analysis"), providing a basis for adjusting the research direction of the project.

[0085] To prevent the implementation of this invention from being circumvented, several technical alternatives are listed below while maintaining the technical objective and effect of "expert recommendation and selection based on multi-collaborative rule integration". These alternatives differ from the solutions of this invention in structure, process, and components, but they still fall within the protection scope of this invention as a whole.

[0086] 1. Alternative solution for multimodal heterogeneous graph modeling Instead of the Heterogeneous Graph Attention Network (HGAT), a GCN is used to perform convolution operations on the heterogeneous graph using a normalized Laplacian matrix, as shown in the formula: in and They are nodes and The degree of the graph is reduced, making it suitable for rapid processing of large-scale graph data. While the accuracy decreases by about 5% in node classification tasks, the training speed is improved by 30%, making it suitable for scenarios with high real-time requirements.

[0087] 2. ELMo replaces BERT for text embedding We use ELMo (Embeddings from Language Models) to replace the domain-fine-tuned SciBERT for text semantic extraction, and generate dynamic word vectors through a bidirectional LSTM network, as shown in the formula: in and These are task-related parameters. This reduces the number of model parameters and computational resource requirements, achieving similar performance to BERT when processing short texts (such as project titles), but its semantic understanding of long texts is weaker.

[0088] 3. Genetic Algorithm Replaces Weighted Apriori Genetic algorithms are used for association rule search, encoding rules as chromosomes. The rule population is optimized through selection, crossover, and mutation operations, and iterative selection is performed using support and confidence as fitness functions. This approach can escape local optima and discover more innovative rules in complex data, but it has high computational complexity and lower rule generation efficiency than weighted Apriori.

[0089] 4. Symbolic reasoning based on the rule engine (Drools) as an alternative to RGNN Leveraging the Drools rule engine, association rule validation and reasoning are performed using custom business rule templates (such as IF-THEN statements), combined with a fact base and agenda mechanism to achieve rule matching and execution. The rules are highly interpretable and easy for manual intervention and adjustment, but lack a deep understanding of data semantics, making it suitable for scenarios with clearly defined rules and small-scale data.

[0090] 5. Traditional machine learning and statistical model combination schemes Text topic classification is performed using TF-IDF + logistic regression; the strength of associations between entities is calculated using Pearson correlation coefficient; and topic evolution is predicted based on Markov chains. This approach has low technical barriers and high interpretability, but its automation level and analytical accuracy are significantly lower than this solution, making it suitable for basic scientific research management scenarios where high levels of intelligence are not required.

[0091] like Figure 3 As shown, the present invention also provides an intelligent generation system for scientific research project context based on graph neural networks, characterized in that it includes a preprocessing module, a pre-trained model library generation module, a relationship extraction module, a topic similarity calculation module, a relationship rule generation module, and a dynamic context graph generation module; The preprocessing module is used to extract features from the historical data of scientific research projects to obtain a historical preprocessed dataset; wherein the feature data of each scientific research project is a training sample in the historical preprocessed dataset. The pre-trained model library generation module is used to train a heterogeneous graph vectorization model, a dynamic LDA topic model, and a weighted semantic association rule model based on historical preprocessed datasets. The association extraction module is used to construct a local heterogeneous graph of the target project based on the collected target scientific research project data, call a pre-trained heterogeneous graph vectorization model to generate a semantic vector of each target entity based on the local heterogeneous graph of the target project; then calculate the similarity between the semantic vector of each target entity and the semantic vector of each entity in the historical entity vector library, determine the association between each target entity and the historical entities, and generate the theme evolution path of the target scientific research project. The topic similarity calculation module is used to call a pre-trained dynamic LDA topic model to obtain the topic distribution of the target research project based on the text vector of the target research project; and to match the topic distribution of the target research project with the historical topic evolution tree to calculate the topic similarity. The association rule generation module is used to call a pre-trained weighted semantic association rule model to generate association rules that are suitable for the target scientific research project based on the similarity between semantic vectors of entities and the similarity of the topic. The dynamic context diagram generation module is used to generate a dynamic context diagram of the target research project based on the relationship between each target entity and historical entities in the target research project, the theme evolution path of the target research project, and the association rules.

[0092] Although specific embodiments of the invention have been disclosed for illustrative purposes to aid in understanding and implementing the invention, those skilled in the art will understand that various substitutions, variations, and modifications are possible without departing from the spirit and scope of the invention and the appended claims. Therefore, the invention should not be limited to the content disclosed in the preferred embodiments, and the scope of protection claimed by the invention is defined by the claims.

Claims

1. A method for intelligently generating research project context based on graph neural networks, comprising the following steps: Historical data from scientific research projects are collected and features are extracted to obtain a historical preprocessed dataset; The feature data for each research project is a training sample from the historical preprocessed dataset; A heterogeneous graph vectorization model, a dynamic LDA topic model, and a weighted semantic association rule model were trained based on a historical preprocessed dataset. Based on the collected target scientific research project data, a local heterogeneous graph of the target project is constructed. A pre-trained heterogeneous graph vectorization model is called to generate the semantic vector of each target entity based on the target project local heterogeneous graph. Then, the similarity between the semantic vector of each target entity and the semantic vector of each entity in the historical entity vector library is calculated to determine the association between each target entity and historical entities, and the theme evolution path of the target scientific research project is generated. The pre-trained dynamic LDA topic model is invoked to obtain the topic distribution of the target research project based on the text vector of the target research project; the topic distribution of the target research project is matched with the historical topic evolution tree to calculate the topic similarity; The pre-trained weighted semantic association rule model is invoked to generate association rules that are suitable for the target scientific research project based on the similarity between semantic vectors of entities and the similarity of the topic. Based on the relationship between each target entity and historical entities in the target research project, the theme evolution path and association rules of the target research project, a dynamic context diagram of the target research project is generated.

2. The method according to claim 1, characterized in that, The nodes in the local heterogeneous graph of the target project include project number, participants, collaborating institutions, and published papers; the edges in the local heterogeneous graph of the target project are determined based on the relationships between the nodes, and the relationships include participation, collaboration, and affiliation.

3. The method according to claim 1, characterized in that, We construct project contexts using pre-trained dynamic LDA topic models, mine the topic distribution of projects over time, construct topic-entity association graphs and input them into graph convolutional networks to enhance topic representations, and generate topic evolution trees. Based on the topic evolution trees of target research projects, we predict the future evolution direction of the topics to which the target research project belongs.

4. The method according to claim 1, 2, or 3, characterized in that, A pre-trained dynamic LDA topic model is used to construct the project context of the target research project, mine the topic distribution of the target research project over time, and construct the topic-entity relationship graph of the target research project. Based on the topic-entity relationship graph of the target research project, the historical core personnel under the same topic are recommended for the target research project.

5. The method according to claim 1, 2, or 3, characterized in that, The historical data includes structured data, contract data, achievement data, personnel data, unit data, semi-structured data, unstructured data, and image data. The temporal features in the structured data are subjected to Fourier transform to convert them into periodic vectors. The unstructured data is segmented, stop words are removed, lexical reconstruction is performed, and a dictionary specific to the scientific research field is constructed. Numerical data in the historical data is processed using Z-score standardization to obtain numerical features. Text data in the historical data is processed using BERT to obtain semantic features. Image data in the historical data is extracted using Faster R-CNN to obtain key image features.

6. The method according to claim 1, characterized in that, Based on the relationship between each target entity and historical entities in the target research project, the theme evolution path and association rules of the target research project, the hierarchical structure of the target research project is constructed and input into the visualization engine to generate a dynamic network diagram of the target research project; project progress warning is issued based on the dynamic network diagram: combined with the theme evolution path and association rules of the target research project, warnings are triggered for projects that deviate from the time sequence of output results of projects with the same theme.

7. The method according to claim 1, characterized in that, The text vector of the target research project is generated based on the project proposal or paper, and the image vector of the target research project is generated based on the image data of the target research project. The text vector and image vector are then input into a dual-stream attention module for fusion to obtain the multimodal features of the target research project. The pre-trained heterogeneous graph vectorization model is invoked to generate the semantic vector of each target entity based on the local heterogeneous graph of the target project and the multimodal features of the target scientific research project.

8. A research project context intelligent generation system based on graph neural networks, characterized in that, It includes a preprocessing module, a pre-trained model library generation module, a topic evolution path generation module, a topic similarity calculation module, an association rule generation module, and a dynamic context graph generation module; The preprocessing module is used to extract features from the historical data of scientific research projects to obtain a historical preprocessed dataset; wherein the feature data of each scientific research project is a training sample in the historical preprocessed dataset. The pre-trained model library generation module is used to train a heterogeneous graph vectorization model, a dynamic LDA topic model, and a weighted semantic association rule model based on historical preprocessed datasets. The topic evolution path generation module is used to construct a local heterogeneous graph of the target project based on the collected target scientific research project data, call a pre-trained heterogeneous graph vectorization model to generate a semantic vector of each target entity based on the local heterogeneous graph of the target project; then calculate the similarity between the semantic vector of each target entity and the semantic vector of each entity in the historical entity vector library, determine the association between each target entity and historical entities, and generate the topic evolution path of the target scientific research project. The topic similarity calculation module is used to call a pre-trained dynamic LDA topic model to obtain the topic distribution of the target research project based on the text vector of the target research project; and to match the topic distribution of the target research project with the historical topic evolution tree to calculate the topic similarity. The association rule generation module is used to call a pre-trained weighted semantic association rule model to generate association rules that are suitable for the target scientific research project based on the similarity between semantic vectors of entities and the similarity of the topic. The dynamic context diagram generation module is used to generate a dynamic context diagram of the target research project based on the relationship between each target entity and historical entities in the target research project, the theme evolution path of the target research project, and the association rules.

9. A server, characterized in that, It includes a memory and a processor, the memory storing a computer program configured to be executed by the processor, the computer program including instructions for performing the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 7.

Images