Scientific and technological intelligence analysis-oriented dynamic weighted cross-domain graph pre-training method and system

Abstract

The invention relates to a scientific and technological intelligence analysis-oriented dynamic weighted cross-domain graph pre-training method and system, and the method comprises the following steps: S1, constructing a cross-domain unified graph: obtaining scientific and technological open source intelligence source data from a plurality of different fields, and constructing a plurality of source domain graph data, constructing a cross-domain unified graph through semantic unit mapping and semantic anchor point bridging; s2, weighted comparison pre-training: carrying out pre-training on the cross-domain unified graph by adopting comparison learning based on dynamic confidence coefficient weight; and S3, downstream task migration adaptation: adapting the pre-trained model to target domain graph data through a semantic-guided weighted fine tuning strategy to complete a downstream task. The system comprises a semantic unit mapping and cross-domain unified graph construction module, a dynamic weight calculation module, a weighted pre-training module and a task adaptation module. According to the method, the problems of cross-domain semantic gaps and noise interference can be effectively solved, and the generalization ability and cross-domain migration performance of the graph model in science and technology intelligence analysis are greatly improved.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and information security, specifically relating to a dynamic weighted cross-domain graph pre-training method and system for scientific and technological intelligence analysis. Background Technology

[0002] Graph neural networks, as powerful tools for processing relational data, have achieved remarkable results in fields such as social analytics and recommender systems. In recent years, inspired by the success of the "pre-training-fine-tuning" paradigm in natural language processing and computer vision, researchers have begun to explore applying graph pre-training techniques to a wider range of scenarios. Existing graph pre-training methods mainly follow two technical routes: one is single-domain pre-training, which learns general graph representations by designing self-supervised tasks (such as contrastive learning and feature reconstruction) within a specific domain (such as academic citation networks and e-commerce networks); the other is cross-domain pre-training, which attempts to integrate graph data from several sources and hopes that the model can learn more generalized knowledge.

[0003] However, when researchers turn their attention to the complex and high-value application area of ​​"full-domain open-source technology intelligence analysis," the aforementioned existing technical approaches reveal serious limitations. Open-source technology intelligence encompasses a variety of information sources, including structured academic papers and patents, semi-structured news and industry reports, and unstructured social media discussions. This multi-source heterogeneous data can be naturally constructed into graph structures (such as citation graphs, collaboration graphs, and event co-occurrence graphs), but the unified analysis of this multi-source heterogeneous data faces the following four core challenges: 1. "Intelligence silos" caused by semantic space fragmentation: Since graph data from different sources are in isolated semantic spaces, existing cross-domain pre-training methods usually use simple feature projection or add global virtual nodes to each graph. This pre-training scheme can only perform coarse alignment at the numerical level and cannot achieve deep semantic communication. The adverse consequences are that a model trained on paper data cannot understand the "intelligence silos" formed by news discussions on the same technical topic, and cannot support cross-domain correlation analysis and knowledge discovery. 2. Uneven data quality and noise interfere with model robustness: The data quality in open-source intelligence environments varies greatly, filled with a large amount of irrelevant information, duplicate content, and false signals. Existing pre-training methods, especially those that rely on fixed thresholds or random sampling to construct training samples, cannot distinguish the value and credibility of the above-mentioned data information. Therefore, these pre-training methods treat high-value intelligence and noise equally, which can easily lead to the model being contaminated by low-quality data, thereby affecting the accuracy and robustness of the judgment of real technology trends in the unified analysis process.

[0004] 3. Negative transfer risk restricts the effectiveness of knowledge reuse: In cross-domain transfer learning, when the source domain and target domain have significant differences in distribution, direct transfer not only fails to improve performance but also leads to a "negative transfer" effect. That is, the source domain knowledge interferes with the model's learning of the target domain. For example, in the context of scientific and technological intelligence, there is a natural distribution difference between the rigor of academic paper data and the randomness of social media data. Existing pre-training methods lack explicit modeling and adaptation mechanisms for inter-domain differences, which causes the performance of the pre-trained model to be unstable and often deteriorate when transferring from the academic paper domain to the news or social media domain, making it impossible to safely and reliably reuse cross-domain knowledge. 4. Existing static training models cannot adapt to the dynamic evolution of intelligence: Scientific and technological intelligence is a rapidly evolving dynamic system. New technological breakthroughs, research hotspots, and industry trends will continue to emerge. However, most existing pre-trained models adopt a "one-time" static training model. The model's capacity and learning strategy will remain fixed after training. This will make it difficult for the pre-trained model to quickly absorb and understand data from emerging fields, and it will be unable to respond agilely to the dynamic changes in the intelligence landscape, thus failing to meet the decision-making needs of real-time situational awareness.

[0005] As can be seen from the above, the technical shortcomings of existing graph pre-training techniques include: 1. Negative transfer effects caused by semantic space mismatch between graph data from different domains; 2. The impact of noisy samples on training stability in multi-source domain data; 3. Fixed threshold strategies struggle to adapt to dynamic changes in sample quality during training; 4. The technical challenge of effectively integrating graph structural information and semantic information.

[0006] Clearly, existing graph pre-training techniques still have significant shortcomings in terms of semantic interoperability, transfer robustness, noise resistance, and dynamic adaptability when dealing with open-source technology intelligence analysis across the entire domain.

[0007] Chinese invention patent CN119782822A discloses a method and system for pre-training cross-domain graph models based on learnable graph patches, belonging to the field of graph model training technology. The method involves acquiring a pre-trained graph dataset from multiple source domains; decomposing the node features of each graph data in the pre-trained dataset into a set of node tokens for each channel, with the node token set for each channel and its corresponding graph structure serving as the key graph patch for that channel; encoding and aggregating the key graph patches for each channel in the graph data as a graph data representation; and pre-training each module based on the graph data representation results using a feature masking restoration task and a graph neighbor context prediction task. The pre-trained module combination serves as a cross-domain graph model for extracting target domain graph data representations, enabling downstream tasks for the target domain graph data. This invention achieves the extraction of transferable information from graph data from different source domains without auxiliary information using the decomposed key graph patches, improving the transfer performance of graph models in multi-domain tasks. It is evident that this invention is a cross-domain pre-training method based on local structural reconstruction, which achieves knowledge transfer by decomposing node features into graph patches and learning the structural relationships between patches. However, this method may have the following limitations in certain scenarios: 1. The semantic alignment level is shallow: Because the alignment mechanism of this method relies on the numerical similarity of node features and local structural reconstruction, it is difficult to achieve deep cross-domain semantic communication and also difficult to solve the problem of information silos. 2. Static and rigid training process: This method does not distinguish between all data and has a fixed training strategy, so it is sensitive to data noise and has low training efficiency and stability. 3. Insufficient transfer robustness: When adapting to the target domain, the general pre-training-fine-tuning paradigm has a high risk of negative transfer. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a cross-domain graph model pre-training method and system based on semantic unit mapping and dynamic weighting. The graph model pre-training method achieves semantic space alignment of multi-source domain graph data by constructing a universal semantic dictionary, establishes cross-domain connection mechanisms using semantic anchors, and introduces dynamic confidence weights to achieve an adaptive training process. This invention not only effectively solves the technical problems of semantic differences, noise interference, and negative transfer in existing cross-domain graph pre-training, but also significantly improves the generalization performance and downstream task adaptability of the pre-trained model, fundamentally breaking down the semantic barriers between heterogeneous intelligence sources and achieving robust, efficient, and adaptive cross-domain knowledge fusion and reasoning graph pre-training.

[0009] This invention provides a dynamic weighted cross-domain graph pre-training method for scientific and technological intelligence analysis, comprising the following steps: S1: Cross-domain unified graph construction: Acquire open-source technology intelligence data from several different fields and construct several source domain graph data. Construct a cross-domain unified graph through semantic unit mapping and semantic anchor bridging. S2: Weighted contrastive pre-training: Pre-training is performed on the cross-domain unified graph using contrastive learning based on dynamic confidence weights; S3: Downstream Task Transfer Adaptation: Adapt the pre-trained model to the target domain graph data through a semantically guided weighted fine-tuning strategy to complete the downstream task.

[0010] Furthermore, S1 specifically includes the following steps: S11: Feature dimension unification: The node features of the several source domain graph data are projected to a unified dimension through a feature projector; S12: Discrete semantic encoding: The continuous features after unifying the dimensions are mapped to the discrete semantic space based on the general semantic dictionary by the discrete semantic encoder. Each node is mapped to K semantic units (K>1). The discrete semantic encoder realizes the transformation from continuous features to discrete semantic indexes. S13: Construction of a general semantic dictionary: This dictionary is jointly trained on all source domain data to serve as a semantic alignment benchmark. The dictionary training adopts joint optimization of reconstruction loss and dictionary quantization loss. S14: Cross-domain unified graph structure generation: Instantiate the semantic units in the general semantic dictionary as semantic anchors, retain the original connection relationships within the source domain graphs, and establish connections between each original node and the K semantic anchors mapped to each original node. Bridge graph data from different domains through the semantic anchors to form a cross-domain unified graph structure containing all the original nodes and semantic anchors.

[0011] Furthermore, S2 specifically includes the following steps: S21: Calculate the initial confidence weight of the node, the initial confidence weight value being based on the reconstruction quality and semantic consistency of the node features; S22: A learnable weight updater is used to dynamically adjust the node weights. The inputs of the weight updater include node embedding, neighbor aggregation information, and historical weights. S23: Implement a progressive relaxation strategy to schedule the weights in S22, dynamically adjust the confidence threshold parameter, and output the final dynamic weights of the current training step. S24: Perform data augmentation on the cross-domain unified graph to generate several augmented views. Use a graph neural network encoder to calculate the embedding representation of nodes in each augmented view. Calculate the weighted contrast loss based on the final dynamic weights to narrow the embedding distance of positive sample pairs and widen the embedding distance of negative sample pairs, thus completing model optimization.

[0012] Furthermore, S21 specifically includes the following steps: S211: Calculate the reconstruction confidence based on the reconstruction quality of the node features; S212: Calculate semantic confidence based on the semantic consistency between the node and the semantic anchor points associated with the node; S213: Integrate the reconstruction confidence and semantic confidence to form the initial confidence weight of the node.

[0013] Furthermore, the progressive relaxation strategy in S23 includes the following steps: S231: Preset initial confidence threshold and ,in ; S232: The initial confidence threshold is dynamically adjusted according to the training progress, and the formula for the dynamic adjustment threshold is: ; S233: The sample weights are cropped and scaled according to the adjusted threshold to achieve a training process from easy to difficult.

[0014] Furthermore, the semantically guided weighted fine-tuning strategy in S3 includes the following steps: S31: Connect the target domain graph data to the general semantic dictionary through the trained semantic units; S32: Initialize the downstream model using a pre-trained graph neural network encoder; S33: Calculate the sample weights based on the mapping results of the target domain nodes on the general semantic dictionary, and fine-tune them using a weighted loss function.

[0015] To address the aforementioned problems in the prior art, this invention also provides a dynamically weighted cross-domain graph pre-training system for science and technology intelligence analysis, used to implement the preferred method for cross-domain graph model pre-training for full-domain open-source science and technology intelligence analysis, comprising: The semantic unit mapping and cross-domain unified graph construction module is configured to perform the mapping of node features to a general semantic dictionary, the joint training of the general semantic dictionary, and the integration of semantic anchors to construct a cross-domain unified graph. The dynamic weight calculation module is configured as the initial confidence weight of the computing node and dynamically adjusts the weight through a learnable weight updater. The weighted pre-training module is configured to implement a progressive relaxation strategy to perform data augmentation and node embedding computation on the cross-domain unified graph and perform weighted contrastive learning. The task adaptation module is configured to align the target domain graph data with the general semantic dictionary, initialize the downstream model with a pre-trained encoder, and achieve model transfer through weighted fine-tuning.

[0016] Compared with the prior art, the beneficial effects of the present invention are: 1. To address the issues of semantic space fragmentation and intelligence silos, this invention constructs a universal semantic dictionary across intelligence sources, mapping heterogeneous nodes such as "technical concepts" in papers, "event entities" in news, and "innovative solutions" in patents to a unified semantic space. This fundamentally breaks down information silos, enabling deep integration and correlation reasoning of cross-domain and cross-modal intelligence, and laying a solid foundation for panoramic intelligence analysis. 2. To address the issues of uneven data quality and noise interference, this invention adaptively evaluates the quality and value of each intelligence sample by setting dynamic confidence weights and a progressive relaxation strategy. During training, it automatically focuses on high-confidence and high-value intelligence signals and effectively filters out low-quality and conflicting noise information, thereby greatly enhancing the robustness of the model in complex open-source environments and the accuracy of intelligence perception. 3. To address the problem of negative transfer risk restricting knowledge reuse, this invention provides a unified semantic coordinate benchmark for intelligence data from different sources by setting up a semantic dictionary and a semantic anchor mechanism. This enables the knowledge learned by the model in the source domain to be safely and smoothly transferred to the target domain through shared semantic units, significantly reducing the risk of negative transfer and improving the efficiency and stability of cross-domain intelligence knowledge reuse. 4. To address the problem that static models cannot adapt to dynamic evolution, this invention adopts a progressive relaxation strategy and an efficient downstream migration mechanism, enabling the model to simulate the learning process of intelligence analysts and continuously learn from new intelligence streams in an easy-to-difficult manner. This achieves rapid understanding and agile response to emerging technology fields and rapidly evolving intelligence landscapes, meeting the decision-making needs of real-time situational awareness. 5. This invention is particularly applicable to the unified representation, semantic association, and in-depth mining of multi-source, heterogeneous, and dynamic open-source intelligence data, thereby providing strong support for key decision-making tasks such as technology trend monitoring, competitive situation awareness, risk assessment and early warning, and innovation opportunity discovery. The open-source intelligence data includes, but is not limited to, academic papers, patent documents, news information, industry reports, social media information, and publicly available government documents.

[0017] The breakthroughs and innovations of this invention compared to existing technologies are shown in Table 1: Table 1: Comparison Dimensions Existing technology This invention Multi-domain semantic unification Using linear projection or adding global virtual nodes only achieves feature alignment at the numerical level and cannot establish deep semantic relationships. By constructing a general semantic dictionary based on vector quantization, nodes with similar semantics in different domains are mapped to a unified semantic space. Cross-domain semantic alignment and topological fusion are achieved through semantic anchors, thereby effectively solving the problem of semantic space fragmentation. Confidence weight calculation In contrastive learning or pre-training, sample weights are set to fixed values ​​or calculated based on simple heuristics. A dynamic confidence weighting mechanism is introduced. The initial confidence is calculated based on the reconstruction quality of node features and the semantic consistency between nodes and semantic anchors. The weights are dynamically adjusted through a learnable weight updater to achieve adaptive sample weighting. Sample scheduling strategy Setting fixed difficulty levels based on experience makes it impossible to dynamically adjust according to the actual learning status of the model, and it is difficult to adapt to the characteristics of different data distributions and model architectures. We propose an intelligent progressive relaxation strategy that dynamically adjusts the confidence threshold to schedule training samples from easy to difficult. This not only ensures a smooth and efficient training process but also ensures that the training process always stays on the optimal learning track. Dynamic training regulation Using a fixed threshold to select samples or an equal weight strategy makes it impossible to perceive changes in the training state, which can lead to oscillations or getting stuck in local optima during the training process. A weight updater is set up to analyze node embeddings, neighbor information, and historical weights in real time and form a closed-loop feedback to achieve fine-tuning of the model training process. Migration adaptation mechanism Using direct parameter transfer or simple fine-tuning strategies, there is a lack of in-depth analysis of the semantic distribution differences between the source and target domains. By analyzing the distribution characteristics of the source and target domains in the semantic dictionary and based on these analysis results, the system can automatically adjust the migration strategy to effectively adapt to the unique characteristics of the target domain and achieve stable and reliable knowledge transfer. ; Attached image description: Figure 1 This is a flowchart of a preferred embodiment of the dynamic weighted cross-domain graph pre-training method for science and technology intelligence analysis according to the present invention; Figure 2 This is a schematic diagram of semantic unit mapping in this invention; Figure 3 This is a schematic diagram of the construction of the cross-domain unified graph in this invention; Figure 4 This is a schematic diagram of the progressive relaxation weights driven by dynamic confidence in this invention. Figure 5 This is a schematic diagram of the unified graph feature and structure joint encoding in this invention; Figure 6 This is a schematic diagram of the downstream migration task in this invention; Figure 7 This is a comparison chart of the pre-training loss convergence curves of the present invention; Figure 8 This is a comparison chart of downstream task migration performance in this invention. Detailed Implementation

[0018] The preferred embodiments of the present invention will be described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0019] To address the aforementioned problems in the prior art, this invention provides a dynamic weighted cross-domain graph pre-training method for scientific and technological intelligence analysis, such as... Figure 1-6 As shown, it includes the following steps: S1: Cross-domain unified graph construction: Through semantic unit mapping technology, the node features of graph data from various source domains are unified into a discrete semantic space based on a general semantic dictionary. Then, a cross-domain unified graph is constructed using semantic units as anchors. The specific process includes the following steps: S11: Feature Dimension Unification: Singular Value Decomposition (SVD) or Learnable Multilayer Perceptron (MLP) is used as the feature projector to project the feature matrices of different source domain nodes (which may have different dimensions, such as Cora being 1433-dimensional and Citeseer being 3703-dimensional) to the same dimension (e.g., d=128), thus eliminating dimensional differences. S12: Discrete semantic encoding: Using a discrete semantic encoder (based on the vector quantization-variable autoencoder VQ-VAE architecture), the continuous features after unification of dimensions are converted into discrete indices pointing to a general semantic dictionary. Each node is mapped to K semantic units (K is an integer greater than 1, preferably K=3) to capture the multifaceted semantic characteristics of the node. S13: Construction of a general semantic dictionary: This dictionary is jointly trained on all source domain data, containing a fixed number (preferably |C|=50) of cross-domain shared semantic units (such as basic concepts like "entity-relationship" and "attribute-category") as a benchmark for semantic alignment. The dictionary training employs a joint optimization of "reconstruction loss + dictionary quantization loss": the reconstruction loss restores the original features of nodes through the decoder (calculated using mean squared error MSE), and the quantization loss ensures that node features are aligned with the embedding space of semantic units. S14: Generation of a unified cross-domain graph structure: Semantic units in the general semantic dictionary are instantiated as "semantic anchors" to construct a unified cross-domain graph containing "original nodes + semantic anchors". Specifically, the following steps are included: First, the connection relationships of the original nodes within each source domain graph are preserved (without destroying the domain topology); Second, undirected edges (with a weight of 1) are established between each original node and its K mapped semantic anchors; Finally, cross-domain connections are achieved through semantic anchors to ultimately form a unified graph structure (denoted as G_uni).

[0020] S2: Weighted contrastive pre-training to address noise interference and training rigidity: On the cross-domain unified graph G_uni constructed in S1, weighted contrastive learning is achieved through dynamic confidence weights and a progressive relaxation strategy. The specific process includes the following steps: Dynamic confidence weight calculation: Each node is assigned a weight that is adjusted in real time to regulate its contribution to the loss function. The weight consists of two parts: Reconstruction confidence: Calculated based on node features, the smaller the reconstruction loss (MSE), the higher the confidence. ; in: Original features; For reconstructing features; Semantic confidence: Calculated based on the semantic consistency between nodes and associated semantic anchors, using cosine similarity. ; in: Embedded for nodes; The embedding of the k-th associated semantic anchor point; Weights are updated dynamically: such as Figure 4 As shown, a learnable weight updater (small MLP) is used to take the current embedding of a node, neighbor aggregation information, and historical weights as inputs, and outputs the final dynamic weights, thereby achieving precise dynamic control of node weights: ; in: The learnable coefficient is initially set to 0.5. Progressive relaxation strategy: Implementing a training pace from easy to difficult, avoiding early noise interference, including the following steps: Initial threshold setting: Preset two initial confidence thresholds. (satisfy Preferred , The samples were divided into high-confidence ( ), medium confidence level ( ) and low confidence ( Three categories; Threshold dynamic adjustment: The threshold is linearly increased with each training epoch (denoted as t), as shown in the formula: ; in: For threshold increment, preferred Ensure that the threshold does not exceed 1 in the later stages of training; Sample weight pruning and scaling: High-confidence samples retain a weight of 1, while medium-confidence samples are pruned and scaled. Scaling: Set the weight of low-confidence samples to 0 (they will not participate in training for the time being). Weighted contrast loss optimization includes the following steps: Graph data augmentation: Two augmentation operations are performed on G_uni to generate multiple views: the first is node feature perturbation (adding Gaussian noise with a variance of 0.1); the second is edge structure perturbation (randomly deleting 5% of the edges). Node embedding computation: such as Figure 5 As shown, the encoder uses a graph convolutional network (GCN) or a graph attention network (GAT) to simultaneously fuse feature information and structural information of a cross-domain unified graph by calculating the embedding representation of nodes in each enhanced view. Weighted contrastive loss: The loss is calculated based on dynamic confidence weights, pushing positive sample pairs (different augmented views of the same node) closer together and negative sample pairs (augmented views of different nodes) further apart. The loss formula is as follows: ; in: For nodes Dynamic confidence weights; Embeddings for positive and negative samples, respectively; For cosine similarity, the complete formula is: ;

[0021] Representing vectors and dot product, They represent and norm, Temperature parameters (preferred) ); S3: Downstream Task Transfer Adaptation: Adapting the pre-trained model to the target domain graph data, and achieving accurate transfer through weighted fine-tuning. The specific process includes the following steps: Target domain semantic alignment: Using the S1-trained feature projector and discrete semantic encoder, the target domain node features are mapped to a general semantic dictionary, establishing a connection between the target domain nodes and semantic anchors; Model initialization: The downstream model is initialized using the graph neural network (GCN / GAT) encoder pre-trained with S2, reusing cross-domain semantic and structural knowledge; Weighted fine-tuning: Sample weights are calculated based on the mapping results of target domain nodes in the semantic dictionary (higher semantic consistency results in greater weights). The cross-entropy loss model is used for optimization, and the formula is as follows: ;

[0022] in: The number of nodes in the target domain; For target domain nodes The weights; This is a real label; For predicted labels.

[0023] To address the aforementioned technical problems, this invention also provides a dynamically weighted cross-domain graph pre-training system for science and technology intelligence analysis, used to implement the cross-domain graph model pre-training method for full-domain open-source science and technology intelligence analysis described in any of the preferred embodiments above, comprising: The semantic unit mapping and cross-domain unified graph construction module is configured to perform the mapping of node features to a general semantic dictionary, the joint training of the general semantic dictionary, and the integration of semantic anchors to construct a cross-domain unified graph. The dynamic weight calculation module is configured as the initial confidence weight of the computing node and dynamically adjusts the weight through a learnable weight updater. The weighted pre-training module is configured to implement a progressive relaxation strategy to perform data augmentation and node embedding computation on the cross-domain unified graph and perform weighted contrastive learning. The task adaptation module is configured to align the target domain graph data with the general semantic dictionary, initialize the downstream model with a pre-trained encoder, and achieve model transfer through weighted fine-tuning.

[0024] The following section presents experimental data and results, combined with... Figure 7 and Figure 8 The following are the beneficial effects of the preferred embodiments of the present invention: To fully verify the effectiveness of the proposed method, this section evaluates it through two sets of experiments: the first set of experiments assesses the independent contributions of dynamic confidence weights and progressive relaxation strategies during the pre-training phase, and the second set of experiments assesses the overall performance improvement of the progressive weighting strategy during downstream task transfer. All experiments are conducted on standard graph datasets (such as Cora and Citeseer) to ensure the reproducibility and comparability of the results.

[0025] Group 1 Experiments: Modular Validation of Progressive Relaxation Strategy in Pre-training Phase: This group of experiments aims to verify the role of dynamic confidence weights and progressive relaxation strategy as independent modules in the pre-training process.

[0026] For the weighted contrastive pre-training task, the following three strategies are compared: Strategy A (the present invention): Dynamic weights + progressive relaxation (as described in S2, including dynamic confidence weight calculation and progressive threshold adjustment).

[0027] Strategy B (Fixed Weight): All sample weights are fixed at 1, with no dynamic adjustment mechanism.

[0028] Strategy C (Static Threshold): Use a fixed confidence threshold. Sample screening was conducted without gradual adjustments.

[0029] This group of experiments performed 400 training cycles on a unified graph structure, recording the weighted contrastive loss value every 50 cycles. The results are as follows: Figure 7 As shown, the analysis indicates that: Convergence efficiency: Strategy A has the fastest convergence speed, improving by approximately 33.9% and 28.5% compared to strategies B and C, respectively, indicating that dynamic weights and incremental adjustments can effectively accelerate model optimization.

[0030] Stability and robustness: The standard deviation of the loss curve of strategy A is only 0.02, which is significantly lower than that of strategy B (0.15) and strategy C (0.09), proving that it can smooth the training process and suppress the interference of noisy samples.

[0031] Final performance: The final loss value of strategy A is 20.5% lower than that of strategy B, and its loss curve has the smallest mean squared error (MSE), which verifies the specific adaptability of the strategy to the semantic features of graph structures.

[0032] Our experiments demonstrate that progressive strategies, as pluggable modules, can independently improve the convergence efficiency and robustness of pre-training, providing crucial support for solving noise interference and training rigidity problems.

[0033] Group 2 Experiments: Overall Effect of Progressive Weighting in Downstream Task Migration: This group of experiments evaluates the impact of introducing a progressive weighting strategy on model performance during the downstream task migration adaptation phase.

[0034] Taking the academic citation network scenario as an example, we performed a node classification task on the Cora and Citeseer datasets, comparing the following two conditions: Conditions: The weighted fine-tuning strategy described in S3 is adopted, the sample weights are calculated based on the mapping results of the target domain nodes on the semantic dictionary, and progressive relaxation adjustment is applied.

[0035] Unconditional: The pre-trained model is used directly for fine-tuning, and all samples have equal weights (i.e., no progressive weighting).

[0036] This group of experiments used the same Graph Neural Network Encoder (GCN) and hyperparameter settings, with accuracy (ACC), normalized mutual information (NMI), adjusted Rand index (ARI), and F1 score as evaluation metrics. Five independent replicate experiments were conducted (random seeds 1234–5678). Results are as follows: Figure 8 As shown: Classification accuracy: On the Cora dataset, the conditional policy improves ACC by about 5.2% and ARI by over 14%; on the Citeseer dataset, the NMI improves by over 10% and the F1 score by about 6.8%.

[0037] Clustering Consistency: The conditional strategy significantly outperforms the unconditional strategy in both NMI and ARI metrics, indicating that it can enhance the model's ability to capture the semantic structure of nodes and improve clustering quality.

[0038] Statistical significance: All improvements passed the t-test (p-value < 0.05), demonstrating that the performance improvement was statistically significant.

[0039] Our experiments demonstrate that the progressive weighting strategy can effectively utilize semantic dictionary information and adaptively adjust sample weights in downstream migration, thereby improving the model's adaptability to target domain data and task performance.

[0040] The experimental results from the two sets above consistently demonstrate that the dynamic confidence weights and progressive relaxation strategies proposed in this invention can not only optimize the pre-training process as independent modules, but also significantly enhance the transfer effect of downstream tasks. This invention effectively solves the problems of cross-domain semantic gap, noise interference, and negative transfer through modular design and semantically driven weighting mechanism, providing reliable technical support for open-source science and technology intelligence analysis across the entire domain.

[0041] Based on the above preferred embodiments and experimental data and results, it can be seen that the breakthroughs and beneficial effects of the present invention compared with the prior art include: 1. Deep semantic integration: This invention achieves semantic alignment of cross-domain nodes through a general semantic dictionary and semantic anchors, fundamentally breaking down semantic barriers between heterogeneous intelligence sources; 2. Dynamic confidence weighting mechanism: This invention introduces reconstruction confidence + semantic confidence and combines it with a learnable weight updater to achieve adaptive sample weighting during the training process, effectively suppressing noise; 3. Progressive relaxation training strategy: This invention achieves training scheduling from easy to difficult by dynamically adjusting the confidence threshold, thereby improving convergence efficiency and model stability; 4. Weighted fine-tuning guided by semantics: In downstream tasks, this invention performs weighted fine-tuning based on the mapping results of the target domain nodes in the semantic dictionary, which significantly reduces the risk of negative migration and improves the accuracy of cross-domain adaptation.

[0042] In summary, this invention not only effectively solves the technical problems of semantic differences, noise interference, and negative transfer in existing cross-domain graph pre-training, but also significantly improves the generalization performance and downstream task adaptability of pre-trained models. It fundamentally breaks down the semantic barriers between heterogeneous intelligence sources and achieves robust, efficient, and adaptive cross-domain knowledge fusion and reasoning graph pre-training.

[0043] The above-described technical solutions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic weighted cross-domain graph pre-training method for scientific and technological intelligence analysis, characterized in that, Includes the following steps: S1: Cross-domain unified graph construction: Acquire open-source technology intelligence data from several different fields and construct several source domain graph data. Construct a cross-domain unified graph through semantic unit mapping and semantic anchor bridging. S2: Weighted contrastive pre-training: Pre-training is performed on the cross-domain unified graph using contrastive learning based on dynamic confidence weights; S3: Downstream Task Transfer Adaptation: Adapt the pre-trained model to the target domain graph data through a semantically guided weighted fine-tuning strategy to complete the downstream task.

2. The dynamic weighted cross-domain graph pre-training method for scientific and technological intelligence analysis as described in claim 1, characterized in that, Furthermore, S1 specifically includes the following steps: S11: Feature dimension unification: The node features of the several source domain graph data are projected to a unified dimension through a feature projector; S12: Discrete semantic encoding: The continuous features after unifying the dimensions are mapped to the discrete semantic space based on the general semantic dictionary by the discrete semantic encoder. Each node is mapped to K semantic units. The discrete semantic encoder realizes the transformation from continuous features to discrete semantic indexes. S13: Construction of a general semantic dictionary: This dictionary is jointly trained on all source domain data to serve as a semantic alignment benchmark. The dictionary training adopts joint optimization of reconstruction loss and dictionary quantization loss. S14: Cross-domain unified graph structure generation: Instantiate the semantic units in the general semantic dictionary as semantic anchors, retain the original connection relationships within the source domain graphs, and establish connections between each original node and the K semantic anchors mapped to each original node. Bridge graph data from different domains through the semantic anchors to form a cross-domain unified graph structure containing all the original nodes and semantic anchors.

3. The dynamic weighted cross-domain graph pre-training method for scientific and technological intelligence analysis as described in claim 1, characterized in that, S2 specifically includes the following steps: S21: Calculate the initial confidence weight of the node, the initial confidence weight value being based on the reconstruction quality and semantic consistency of the node features; S22: A learnable weight updater is used to dynamically adjust the node weights. The inputs of the weight updater include node embedding, neighbor aggregation information, and historical weights. S23: Implement a progressive relaxation strategy to schedule the weights in S22, dynamically adjust the confidence threshold parameter, and output the final dynamic weights of the current training step. S24: Perform data augmentation on the cross-domain unified graph to generate several augmented views. Use a graph neural network encoder to calculate the embedding representation of nodes in each augmented view. Calculate the weighted contrast loss based on the final dynamic weights to narrow the embedding distance of positive sample pairs and widen the embedding distance of negative sample pairs, thus completing model optimization.

4. The dynamic weighted cross-domain graph pre-training method for scientific and technological intelligence analysis as described in claim 3, characterized in that, S21 specifically includes the following steps: S211: Calculate the reconstruction confidence based on the reconstruction quality of the node features; S212: Calculate semantic confidence based on the semantic consistency between the node and the semantic anchor points associated with the node; S213: Integrate the reconstruction confidence and semantic confidence to form the initial confidence weight of the node.

5. The dynamic weighted cross-domain graph pre-training method for scientific and technological intelligence analysis as described in claim 3, characterized in that, The progressive relaxation strategy in S23 includes the following steps: S231: Preset initial confidence threshold and ,in ; S232: The initial confidence threshold is dynamically adjusted according to the training progress, and the formula for the dynamic adjustment threshold is: ； S233: The sample weights are cropped and scaled according to the adjusted threshold to achieve a training process from easy to difficult.

6. The dynamic weighted cross-domain graph pre-training method for scientific and technological intelligence analysis as described in claim 1, characterized in that, The semantically guided weighted fine-tuning strategy in S3 includes the following steps: S31: Connect the target domain graph data to the general semantic dictionary through the trained semantic units; S32: Initialize the downstream model using a pre-trained graph neural network encoder; S33: Calculate the sample weights based on the mapping results of the target domain nodes on the general semantic dictionary, and fine-tune them using a weighted loss function.

7. A dynamic weighted cross-domain graph pre-training system for scientific and technological intelligence analysis, used to implement the dynamic weighted cross-domain graph pre-training method for scientific and technological intelligence analysis as described in any one of claims 1-6, characterized in that, include: The semantic unit mapping and cross-domain unified graph construction module is configured to perform the mapping of node features to a general semantic dictionary, the joint training of the general semantic dictionary, and the integration of semantic anchors to construct a cross-domain unified graph. The dynamic weight calculation module is configured as the initial confidence weight of the computing node and dynamically adjusts the weight through a learnable weight updater. The weighted pre-training module is configured to implement a progressive relaxation strategy to perform data augmentation and node embedding computation on the cross-domain unified graph and perform weighted contrastive learning. The task adaptation module is configured to align the target domain graph data with the general semantic dictionary, initialize the downstream model with a pre-trained encoder, and achieve model transfer through weighted fine-tuning.

Images