A regional green innovation cooperation network link prediction method fusing network topology and content features
By integrating node topology, path topology, and patent content features into a comprehensive index model, the problem of insufficient prediction accuracy in regional green innovation cooperation networks has been solved, enabling more accurate prediction of cooperation relationships and policy support.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- UNIV OF SCI & TECH BEIJING
- Filing Date
- 2026-04-23
- Publication Date
- 2026-08-04
AI Technical Summary
Existing link prediction models lack a unified prediction framework in regional green innovation cooperation networks and fail to effectively integrate network topology characteristics and patent content characteristics, resulting in insufficient prediction accuracy.
By integrating node topology features, path topology features, and patent content features, and employing the entropy weight method and the gray wolf optimization algorithm, a method for predicting links in regional green innovation cooperation networks is constructed, including a comprehensive index fusion model of node similarity, path similarity, and content similarity.
It significantly improves the accuracy of link prediction, enabling more accurate prediction of future cooperation relationships and providing recommendations for potential partners and suggestions for optimizing regional innovation policies.
Smart Images
Figure CN122513282A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of complex network analysis and data mining, specifically relating to a method and system for predicting links in a regional green innovation cooperation network that integrates network topology features and patent content features. Background Technology
[0002] In recent years, to promote the innovation-driving effect of developed regions on surrounding areas, my country has proposed a regional collaborative innovation development strategy, promoting regional innovation cooperation through measures such as resource allocation, industrial restructuring, and tax incentives. These collaborative development regions include the Beijing-Tianjin-Hebei region, the Yangtze River Delta region, and the Guangdong-Hong Kong-Macao Greater Bay Area. Against this policy backdrop, many scholars have studied regional innovation cooperation networks to assess the effects and trends of regional collaborative innovation. Meanwhile, green technology innovation networks have received considerable attention in the process of regional collaborative innovation due to their importance to regional ecology, environmental protection, and sustainable development. Therefore, studying regional green innovation cooperation networks, analyzing their characteristics, and revealing their evolutionary trends through link prediction has significant practical implications. On the one hand, by predicting the future development trends of green technology innovation cooperation within a region, network structure problems can be identified early, the driving effect of collaborative innovation can be assessed, and green technology collaborative innovation policies can be optimized. On the other hand, link prediction can help patent inventors identify potential partners and promote the achievement of green technology collaborative innovation.
[0003] Currently, while there is a wealth of research on link prediction based on cooperative networks, there is still a lack of research specifically on link prediction for regional green innovation cooperative networks. Constructing regional green innovation cooperative networks and performing link prediction on these networks can reveal their evolutionary trends and characteristics, providing support for scientific green collaborative innovation decision-making. Current link prediction models possess a very rich system of indicators, each with its own characteristics and applicable prediction scenarios. For example, prediction based on node similarity is a relatively mature method, which assumes that two nodes with common neighbors are most likely to cooperate. However, there are many indicators for measuring node similarity, and most studies tend to select the most accurate indicators for prediction without considering the combined effect of the indicators. Similarly, many studies have achieved good results using path similarity for cooperation prediction, but they do not consider the heterogeneity of paths. Furthermore, existing prediction models mainly focus on the topological characteristics of cooperative networks, failing to fully utilize patent text data related to innovation cooperation. At the same time, due to the increasing number of indicators in prediction models and the different focuses of each model, a unified prediction framework, model, and method are lacking.
[0004] Against this backdrop, the biggest challenge we face is how to select appropriate forecasting indicators based on the characteristics of regional innovation cooperation networks, and how to integrate multiple types of indicators to continuously improve forecasting accuracy. Summary of the Invention
[0005] The technical problem to be solved by the present invention is to provide a method and system for predicting links in a regional green innovation cooperation network that integrates network topology features and content features, thereby improving prediction accuracy by integrating node topology features, path topology features and patent content features.
[0006] The technical solution adopted in this invention is:
[0007] A method for predicting links in a regional green innovation cooperation network that integrates network topology and content characteristics includes the following steps:
[0008] Step S100: Network Construction
[0009] Based on green patent application data, an undirected weighted regional green innovation cooperation network is constructed, with patent application organizations as nodes and cooperative application relationships as links.
[0010] Step S200: Multidimensional Feature Extraction and Fusion
[0011] S201: Node similarity feature extraction. Calculate multiple classic node similarity indices between any two unconnected nodes in the network, and use the entropy weight method to couple the indices to obtain the comprehensive node similarity index NS.
[0012] S202: Path similarity feature extraction, based on second-order and third-order paths between nodes, while considering the heterogeneity of intermediate connecting edges and intermediate nodes, to construct a comprehensive path similarity index PS;
[0013] S203: Content similarity feature extraction. Based on the patent abstract text, the LDA topic model is used to extract the distribution of the patentee's technical topics and calculate the vector similarity between nodes. Then, the entropy weight method is used to couple multiple vector similarity indicators to obtain the comprehensive content similarity index CS.
[0014] S204: Multi-dimensional index fusion. With link prediction accuracy as the optimization goal, a weighted fusion model of three indicators, NS, PS, and CS, is constructed, and an optimization algorithm is used to solve for the optimal weight of each indicator to obtain the fused prediction index NPC.
[0015] Step S300: Model Training and Prediction
[0016] The network edges are divided into a training set, a test set, and a set of non-existent edges. The fusion model is trained using the training set, and the prediction accuracy is evaluated using the test set. Finally, the model is applied to predict links between unconnected node pairs to generate a predicted cooperative network.
[0017] Furthermore, in S201, the classic node similarity indicators include one or more of CN, Salton, Jaccard, Sorenson, HPI, HDI, LHN-I, PA, AA, and RA; the entropy weight method determines the weight of each indicator by calculating its information entropy, and then calculates the weighted sum to obtain the comprehensive node similarity index NS.
[0018] Furthermore, in S202, the construction of the comprehensive path similarity index PS includes:
[0019] Local path metrics calculate the number of second-order and third-order paths between nodes.
[0020] The heterogeneity index for intermediate links is defined by the product of the degrees of intermediate nodes in the path, with the reciprocal of the weight representing the impact on the probability of cooperation; and
[0021] The heterogeneity index of intermediate nodes is constructed based on the proportion of the connected paths between intermediate nodes and the two end nodes to their total connected paths.
[0022] The above indicators are weighted and combined to obtain the comprehensive path similarity index PS, where the weights of the third-order paths are determined through sensitivity analysis.
[0023] Furthermore, in S203, the extraction of content similarity features includes:
[0024] The patent abstract text is segmented and stop word removed to construct Doc2Bow vectors. The LDA topic model is used to extract topics from the text vectors, and the optimal number of topics K is determined using the CV consistency score method to obtain the topic distribution for each patent. Patent topics are assigned to corresponding patentee nodes, constructing a topic set vector for each node. Vector similarity between node pairs is calculated using one or more of the following metrics: Euclidean distance, Manhattan distance, Jaccard similarity, cosine similarity, Pearson correlation coefficient, and Tanimoto coefficient. Entropy weighting is used to couple multiple vector similarity metrics to obtain a comprehensive content similarity index, CS.
[0025] Furthermore, in S204, the optimization objective is to maximize the AUC value, and the optimization algorithm adopts the Grey Wolf Optimization Algorithm to solve for the optimal weight combination of the three indices NS, PS, and CS.
[0026] A regional green innovation cooperation network link prediction system that integrates network topology and content characteristics includes:
[0027] The data acquisition module is used to obtain green patent application data from patent databases or documents, and extract patentee information, cooperative application relationships and patent abstract text.
[0028] The network construction module is used to build and store an undirected weighted regional green innovation cooperation network based on the information extracted by the data acquisition module;
[0029] The feature extraction module is used to perform the node similarity feature extraction, path similarity feature extraction, and content similarity feature extraction as described in claim 1;
[0030] The model fusion and optimization module is used to build a weighted fusion model of NS, PS, and CS, and has a built-in Grey Wolf optimization algorithm to automatically solve for the optimal weight combination with the goal of maximizing the AUC value.
[0031] The link prediction module is used to load the optimal fusion model, calculate and sort the NPC values of all unconnected node pairs in the network, and output a list of potential cooperative relationships and a predicted network topology; and
[0032] The results analysis module includes visualization and statistical analysis tools to generate predictive network diagrams and perform statistical analysis on the type and geographical distribution of innovation entities, outputting analysis reports.
[0033] The positive effects of this invention are:
[0034] This invention improves the accuracy of node feature prediction by coupling multiple node similarity indices using the entropy weight method. It introduces heterogeneity of intermediate links and intermediate nodes into path similarity, thereby improving the predictive ability of traditional path indices. It also introduces the LDA topic model to extract patent content features, integrates textual semantic information into link prediction, and achieves optimal weighted fusion of multiple indices through the Grey Wolf optimization algorithm, significantly improving the overall performance of the prediction model. Attached Figure Description
[0035] Figure 1 This is a framework diagram of the method of the present invention;
[0036] Figure 2 This is a schematic diagram illustrating the principle of constructing the path similarity index for this invention.
[0037] Figure 3 The AUC values of the PS model under different α values in this invention;
[0038] Figure 4 This is a graph showing the consistency of the subject matter of this invention.
[0039] Figure 5 This is a network diagram for predicting green innovation cooperation in the Beijing-Tianjin-Hebei region, as presented in this invention.
[0040] Figure 6 This is a map showing the distribution and geographical distribution of the inventors' subject types in the prediction network of this invention. Detailed Implementation
[0041] Example 1
[0042] This embodiment provides a method for predicting links in a regional green innovation cooperation network that integrates network topology and content characteristics. The overall process is as follows: Figure 1 As shown, the specific steps include:
[0043] Step S100: Network Construction
[0044] This step builds a regional green innovation cooperation network based on green patent application data.
[0045] Data Acquisition: Data on green invention patent grants or applications for the target region (e.g., Beijing-Tianjin-Hebei region, Yangtze River Delta region, Guangdong-Hong Kong-Macao Greater Bay Area) within a specific time period (e.g., 2020-2023) is retrieved from official data sources such as the State Intellectual Property Office. The search strategy includes: determining the IPC classification number of green patents based on the International Patent Classification (IPC) Green Inventory published by the World Intellectual Property Organization; limiting the patent applicant's address or business registration address to the target region; and filtering out patents with more than one inventor.
[0046] Node definition: The network nodes are patent application organizations, removing individual inventors and making enterprises, universities, research institutes and other institutions the core nodes.
[0047] Connection definition: If two applicant organizations jointly file one or more green patent applications, an undirected connection is established between them. If multiple collaborations exist, the connection can be assigned corresponding weights.
[0048] Network Generation: Based on the above rules, an undirected weighted innovative cooperation network is constructed, consisting of a set of nodes and a set of edges. This network conforms to the small-world and scale-free characteristics of complex networks, making it suitable for using link prediction models based on complex networks.
[0049] Step S200: Multidimensional Feature Extraction and Fusion
[0050] S201: Extract node similarity features and construct a comprehensive node similarity index NS.
[0051] For any pair of non-directly connected nodes x and y in the network, calculate multiple classic node similarity indices between them. This embodiment preferably uses the 10 indices listed in Table 1. Specifically, in the formula... This represents the degree of node x. This represents the neighboring nodes of node x.
[0052]
[0053] To integrate the advantages of each indicator, the entropy weight method is used for coupling. The entropy weight method is used for coupling by performing maximum-minimum standardization on the indicator values and calculating the information entropy of each indicator. The smaller the information entropy, the higher the weight. The calculation of information entropy is shown in formulas (1) and (2).
[0054] (1)
[0055] (2)
[0056] in, This represents the overall ratio of all indicators. This represents the information entropy value of a certain indicator. is the vector matrix after max-min standardization of each indicator, where m is the number of samples.
[0057] Since entropy and the contribution of the indicator are inversely proportional, the entropy needs to be transformed to a positive relationship. Here, we use... The entropy value is uniformly transformed to represent the weight value of each indicator, so that... Proportional to the contribution, a joint index representing node similarity is finally obtained. As shown in formula (3).
[0058] (3)
[0059] Where x and y represent two nodes that are not connected, and J represents the number of indicators to be jointly predicted, selected based on the prediction accuracy of the indicators. The higher the indicator value, the greater the likelihood that the two nodes will cooperate based on the network topology.
[0060] Step S202: Extract path similarity features and construct a comprehensive path similarity index PS.
[0061] Besides node features, path features in cooperative networks can also predict cooperative relationships. Traditional path similarity metrics, including the global path model Katz and the local path model LP, both assume that the more indirect paths between two nodes, the greater the likelihood of a connection. However, even when two nodes have the same number of paths, other heterogeneity factors can still lead to different connection probabilities. Therefore, to improve prediction accuracy, this invention considers both the number of paths and path heterogeneity. To balance prediction accuracy and computational complexity, this invention only considers local path metrics and path heterogeneity metrics for second- and third-order paths, and improves existing path heterogeneity metrics by incorporating both intermediate edge heterogeneity and intermediate node heterogeneity into the prediction model. The construction principle is as follows: Figure 2 As shown.
[0062] For local path metrics, the number of second-order paths and the number of third-order paths between node x and node y are calculated based on the reachability matrix, denoted as follows: , The more paths there are, the greater the likelihood of cooperation.
[0063] The path heterogeneity index includes the heterogeneity of intermediate edges and the heterogeneity of intermediate nodes.
[0064] Intermediate link heterogeneity: The larger the degree value of the intermediate nodes in a path, the more dispersed the opportunities for cooperation between nodes, and the less likely a link will be formed. Therefore, the link weight is defined by the product of the degrees of the intermediate nodes, and the reciprocal of the link weight represents the impact on the probability of cooperation. For two nodes x and y in a network that are reachable, let M represent the number of intermediate nodes between x and y, and the path representing their reachability is ( Then the heterogeneity index of the middle connecting edge) The calculation is shown in formula (4):
[0065] (4)
[0066] in, They represent The degree value of the nodes, M is the number of intermediate nodes between x and y.
[0067] Intermediate node heterogeneity: Intermediate nodes refer to all nodes on a reachable path between two nodes, excluding the two end nodes. These nodes exhibit heterogeneity due to varying connectivity, and connectivity also influences the cooperative relationships between end nodes. For a network path, the larger the proportion of the connected paths between intermediate and end nodes in a node's total connected paths, the more likely the nodes at the ends of the path are to cooperate. Therefore, using | | represents the number of connected paths actually formed by all the surrounding neighbors of Z,| | represents the number of all possible connected paths that can be formed by the neighbors around Z, and scores the heterogeneity of the intermediate nodes between nodes x and y. This is denoted as formula (5):
[0068] (5)
[0069] Considering the rapid decrease in prediction accuracy with increasing path length, and the time complexity of the computation process, this paper only considers second-order and third-order paths, i.e. Construct a new path similarity index It can be expressed by formula (6):
[0070] (6)
[0071] Where α is the weight value of the third-order path index. In practical applications, the optimal value is obtained through sensitivity analysis. In practice, the value of α is determined by performing sensitivity analysis on the validation set. For example, in one embodiment of the invention, the model prediction accuracy (AUC) is highest when α = 0.03 (e.g., ...). Figure 3 (As shown).
[0072] Step S203: Extract content similarity features and construct a comprehensive content similarity index CS.
[0073] This step utilizes patent text data to mine the similarity of technical content between nodes.
[0074] We collect the abstract texts of all patents and perform preprocessing such as word segmentation and stop word removal. We then use the Doc2Bow model to vectorize the patent abstracts, obtaining lexical vector data for each abstract, with each lexical corresponding to a unique ID index.
[0075] After vectorizing the abstract text, an LDA topic model is applied to train the text vectors for topic extraction. Since consistency metrics typically outperform perplexity metrics, the CV consistency score method is used to determine the optimal number of topics K. The value of K should ensure that the topics have optimal interpretability and discriminability. For example, in one embodiment of the invention, topic consistency curves (such as...) are used to determine the optimal number of topics K. Figure 4 As shown, K=6 is determined to be the optimal value. After obtaining the optimal number of topics K, the text vectors are trained one by one using the LDA model. During the training process, the model calculates the topic distribution probability of each patent text according to formula (7):
[0076] (7)
[0077] in, This represents the probability of word w in document d. This represents the probability of word w in a specific topic t. This represents the probability of document d under a specific topic t.
[0078] After obtaining the topic of each patent, the topic is assigned to each inventor of that patent, indicating that the inventor's research direction includes that topic. After assigning values to each patent, a topic set vector is formed for each inventor, representing the complete set of their research directions. Finally, based on the topic set vector, the similarity of patent content among the inventors is calculated.
[0079] To improve prediction performance, similar to node similarity, the algorithm uses entropy weighting to jointly predict six classic vector similarity indices. These six classic vector similarity indices are shown in Table 2.
[0080]
[0081] Content similarity index based on multi-vector similarity metrics is denoted as... As shown in formula (8):
[0082] (8)
[0083] in, The weight values are obtained based on the entropy weight method. , Let I be the i-th vector similarity index, and let I be the number of vector similarity indices. The index is selected based on the prediction accuracy of each individual index. The higher the index value, the closer the technical research directions of the two patentee nodes are, and the easier it is for them to cooperate.
[0084] Step S204: Multi-dimensional index fusion to construct the fused prediction index NPC.
[0085] By integrating the above-mentioned node similarity, path similarity, and content similarity indicators, a new link prediction indicator is constructed, denoted as . The calculation is shown in formula (9):
[0086] (9)
[0087] in, These are the weights of the three categories of indicators. Since the weights have a significant impact on the prediction results, an optimization model with prediction accuracy as the objective function is constructed and solved to obtain the optimal weight values. The model is shown in formula (10):
[0088]
[0089] (10)
[0090] in, Let denot NPC as the prediction accuracy, and NS, PS, and CS be the similarity score matrices generated for the corresponding indicators. Since the model is non-differentiable, a metaheuristic algorithm is used for solving it. This embodiment preferably employs the Grey Wolf Optimizer (GWO) algorithm, which aims to maximize the AUC value through iterative searching, ultimately outputting the optimal weight combination that achieves the highest prediction accuracy.
[0091] Step S300: Model training and prediction.
[0092] All edges in the original network are divided into a training set (e.g., 90%) and a test set (e.g., 10%). Simultaneously, samples of similar size to the test set are randomly sampled from the set of non-existent edges to form a negative sample set.
[0093] The training set is used to calculate the various indicators in step S2, and the GWO algorithm is used to determine the optimal weights. Then, the trained model is used to calculate the NPC values of all node pairs in the test set and the negative sample set. AUC (Area Under the Curve) is used as the evaluation index to measure the prediction accuracy of the model. The AUC is calculated as follows: each time, an edge is randomly selected from the test set, and then an edge is randomly selected from the set of non-existent edges. If the similarity score of the edge selected from the test set is greater than the score of the edge selected from the negative sample set, it is recorded as "add 1 point". If the two values are equal, it is recorded as "add 0.5 points". After independent comparison n times, the arithmetic mean is taken as the AUC value. The calculation formula is shown in formula (11).
[0094] (11)
[0095] in Indicates that in n experiments there are The score for T2 is higher. Indicates that there is The two are equal. In this embodiment, n=672400 is set to ensure that the absolute error of AUC will not exceed one-thousandth at a 90% confidence level, and can remain at a relatively stable level.
[0096] The trained model is applied to all unconnected node pairs, their NPC values are calculated and sorted from high to low, and the top-ranked node pairs (e.g., the top 0.2%) are selected as candidate pairs most likely to cooperate in the future, thus constructing a predictive cooperation network. This network can be used for subsequent evolutionary trend analysis and partner recommendation.
[0097] Example 2
[0098] This embodiment provides a regional green innovation cooperation network link prediction system that integrates network topology and content features, including a data acquisition module, a network construction module, a feature extraction module, a model fusion and optimization module, a link prediction module, and a result analysis module.
[0099] The data acquisition module is equipped with a data interface for obtaining green patent application data from patent databases or documents, and extracting patentee information, collaborative application relationships, and patent abstract text.
[0100] The network building module is used to construct and store an undirected weighted regional green innovation cooperation network based on the information extracted by the data acquisition module.
[0101] The feature extraction module includes: a node feature extraction unit, used to calculate the similarity index of multiple nodes and perform entropy weight coupling to generate the NS index matrix; a path feature extraction unit, used to calculate the number of local paths, heterogeneity of intermediate edges and intermediate nodes, and generate the PS index matrix; and a content feature extraction unit, used to perform LDA topic modeling on the patent abstract, assign topics to inventors and aggregate them to the patentee nodes, calculate the content vector similarity between nodes, and perform entropy weight coupling to generate the CS index matrix.
[0102] The model fusion and optimization module is used to build a weighted fusion model of NS, PS, and CS, and has a built-in Grey Wolf optimization algorithm to automatically solve for the optimal weight combination with the goal of maximizing the AUC value.
[0103] The link prediction module is used to load the optimal fusion model, calculate and sort the NPC values of all unconnected node pairs in the network, and output a list of potential cooperative relationships and a predicted network topology.
[0104] The results analysis module includes visualization and statistical analysis tools for generating predictive network diagrams (such as...). Figure 5 As shown), and the distribution of innovation entities by type and geographical distribution (e.g. Figure 6 Statistical analysis is performed on (as shown in the figure), and an analysis report is output.
[0105] Example 3
[0106] This embodiment describes the model verification and application based on the Beijing-Tianjin-Hebei regional green innovation cooperation network. Using actual data from the Beijing-Tianjin-Hebei regional green innovation cooperation network, the prediction accuracy of the method described in Embodiment 1 is verified, and compared with various existing methods to demonstrate the beneficial effects of the present invention.
[0107] 1. Experimental setup
[0108] Data on green invention patent authorizations in the Beijing-Tianjin-Hebei region from 2020 to 2023 were obtained from the State Intellectual Property Office. The search criteria included green patent classification number (based on the WIPO Green List), inventor address (Beijing, Tianjin, Hebei), application date, and a number of inventors greater than one, excluding individual inventors.
[0109] Following step S100, the network was constructed, resulting in 1533 nodes and 7667 edges. Network feature analysis indicates that the network conforms to the small-world and scale-free characteristics of complex networks.
[0110] AUC was used as the metric for predictive accuracy. To reduce randomness, network edges were randomly divided into a training set (90%) and a test set (10%), and negative samples of the same size as the test set were selected from the set of non-existent edges. This random division was performed 100 times, and the average AUC was taken as the final result. In each AUC calculation, 672,400 independent comparisons were performed to ensure that the error did not exceed one-thousandth at a 90% confidence level.
[0111] 2. Comparative Example 1: Link Prediction Based on Node Similarity
[0112] The AUC values of the 10 classic node similarity indices listed in Table 1 were calculated respectively. The results are shown in Table 3.
[0113]
[0114] It can be seen that the prediction accuracy of the 10 indicators is not significantly different, indicating that the prediction of each indicator in this network has its merits.
[0115] According to step S201, this invention uses the entropy weight method to combine all 10 indicators (i.e., J=10 in Formula 3) to obtain the integrated NS indicator and its AUC value. Table 3 shows that the accuracy of the NS indicator (0.9330) is significantly higher than that of the highest-performing single indicator, the RA indicator (0.9171), demonstrating the effectiveness of the entropy weight method in indicator combination.
[0116] 3. Comparative Example 2: Link Prediction Based on Path Similarity
[0117] The path similarity index PS was calculated according to step S202. Sensitivity analysis was performed on α to determine a reasonable value for the third-order path weight α. Figure 3 The AUC accuracy of the PS model under different α values is presented. It can be observed that as the α value increases, the accuracy of the PS index first increases and then decreases, reaching its highest value when the α value is 0.03, at which point the prediction accuracy AUC of the PS model of this network is 0.9245.
[0118] 4. Comparative Example 3: Link Prediction Based on Content Similarity
[0119] After calculating node similarity and path similarity, content similarity between inventor nodes is calculated. First, patent abstracts from the training set are extracted, and after word segmentation and stop word removal, a dictionary and corresponding corpus are built, assigning a unique number to each word. Then, the CV consistency method is used to draw... Figure 4 The topic consistency plot is shown. Sensitivity analysis of the number of topics K reveals that the CV consistency is highest when the number of topics K=6, and there is a significant inflection point. Therefore, K=6 is chosen as the optimal number of topics.
[0120] After obtaining the number of topics, an LDA topic model is constructed. The topic probability distribution of each abstract is then extracted from the model, and the topic with the highest probability is taken as the final topic of the patent. The topic of each patent data is assigned to all inventors of that patent. By merging all topics of the inventors, the set of all topic vectors associated with each inventor is obtained. The six distance calculation methods listed in Table 2 are used to calculate the content similarity of the topic set vectors of node pairs. The AUC values corresponding to each distance are shown in Table 4. The results show that the accuracy of the Jaccard, Euclidean, and Manhattan indices is significantly higher than that of the Cosine, Pearson, and Tanimoto indices. Therefore, this invention selects these three indices and uses the entropy weight method for coupling (i.e., I=3 in Formula 8). The coupled index is denoted as CS, and its average AUC value is 0.7874, which is better than a single vector similarity index.
[0121]
[0122] 5. An embodiment of the present invention provides an NPC model prediction that integrates multi-dimensional features.
[0123] Following step S204, the three comprehensive indices NS, PS, and CS are weighted and fused into an NPC index. Since this model is a non-differentiable nonlinear model, an approximate optimization based on the metaheuristic Grey Wolf Optimizer (GWO) algorithm is used to obtain... , , At this point, the optimal accuracy The optimal weighting is the highest among all pairwise models. This indicates that node similarity has the greatest impact on prediction accuracy, while content similarity has the least. The comparison results with each pairwise model and pairwise combination model are shown in Table 5.
[0124]
[0125] The data in the first three rows of Table 5 show that when the NS, PS, and CS models are used individually for predicting innovation collaboration, the NS model, which emphasizes node similarity, has the highest prediction accuracy (0.9330), while the PS model, which emphasizes path similarity, has slightly lower accuracy. The CS model, which emphasizes content similarity, has a prediction accuracy of only 0.7874, indicating that network topology information is significantly more effective than content similarity in prediction alone.
[0126] The data in rows 4-6 of Table 5 represent the AUC accuracy of the pairwise optimized combinations of the NS, PS, and CS models (the optimal weights for each pairwise combination are also calculated using the GWO algorithm). Both the NS+PS and NS+CS combinations achieved higher accuracy than the NS single model, indicating that introducing path similarity (PS) and content similarity (CS) to the node similarity-dominated model can further improve prediction accuracy. However, the overall accuracy of the PS+CS combination model without the NS indicator is low, indirectly demonstrating the dominant role of node similarity. Furthermore, both PS and CS are auxiliary prediction indicators in the indicator combination, and their combined accuracy remains low.
[0127] The last row of Table 5 shows that when the combined NS, PS, and CS models are used to predict innovative cooperation, the prediction accuracy reaches 0.9634, significantly higher than other baseline models. Here, the node similarity index (NS) still has the highest weight, followed by the PS index, and the CS index has the lowest weight. This further confirms that the combined model, which integrates topological and content information and incorporates the three types of indicators, achieves the best prediction results and can more accurately depict future cooperative relationships. Although the content similarity index has a very small weight, it further increases the prediction accuracy from 0.9504 to 0.9634, thus playing an important supplementary predictive role. This also confirms the necessity of introducing content features into topological features for prediction.
[0128] 6. Application Analysis Based on Prediction Results
[0129] To further demonstrate the application value of the method of the present invention, this embodiment conducts an in-depth analysis of the prediction network generated in step S300.
[0130] 6.1 Recommendations for Potential Partners
[0131] The trained NPC model is used to predict all unconnected node pairs in the network, calculate their NPC values, and sort them in descending order. The top ten potential cooperative node pairs are shown in Table 6. These results can be directly output to the user interface to provide accurate partner recommendation services for green technology innovation entities (enterprises, universities, and research institutions).
[0132]
[0133] 6.2 Predictive Network Structure Analysis
[0134] To reveal the evolutionary trend of regional green innovation cooperation from a macro perspective, this embodiment selects the top 0.2% of node pairs in the NPC score (a total of 219 edges involving 232 nodes) to construct a prediction network for green innovation cooperation in the Beijing-Tianjin-Hebei region, such as... Figure 5As shown. Statistical analysis of this network yields the following conclusions, which verify the rationality and effectiveness of the prediction results obtained by the method of this invention:
[0135] like Figure 6 As shown, enterprises account for 73.58% of the nodes in the predicted network, indicating that enterprises will remain the core force for green technology innovation for some time to come. Meanwhile, although university nodes are few in number (approximately 11.32%), they have high degree values in the network and are prone to forming structural holes, proving that the method of this invention successfully captures the bridging role of universities in the collaborative network. Research institutions account for approximately 15.10%.
[0136] Mapping the nodes in the prediction network to their geographic coordinates yields a geographic distribution network. The analysis results show ( Figure 6 The number of cooperative connections within the Beijing region accounts for more than half of all predicted connections (approximately 53.42%), indicating that Beijing's innovation agglomeration effect remains significant. Cooperation within the Hebei region (11.95%) is higher than cooperation within the Tianjin region (8.52%), showing that Hebei has certain endogenous innovation potential. Among cross-regional cooperation, cooperation between Beijing and Hebei (12.18%) is higher than cooperation between Beijing and Tianjin (9.06%), while cooperation between Tianjin and Hebei is the weakest (only 2.94%).
[0137] This analysis quantitatively reveals the imbalance in cross-regional cooperation in green collaborative innovation in the Beijing-Tianjin-Hebei region, proving that the method of this invention can provide data support for identifying shortcomings in regional innovation policies and optimizing resource allocation.
[0138] Example 4
[0139] This embodiment provides a computer-readable storage medium on which a computer program (or instruction set) is stored. When executed by a processor, the program can implement the regional green innovation cooperation network link prediction method that integrates network topology and content characteristics as described in any step of Embodiment 1. The storage medium can be any medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), random access memory (RAM), magnetic disk, or optical disk.
[0140] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for predicting links in a regional green innovation cooperation network that integrates network topology and content characteristics, characterized in that... It includes the following steps: Step S100: Network Construction Based on green patent application data, an undirected weighted regional green innovation cooperation network is constructed, with patent application organizations as nodes and cooperative application relationships as links. Step S200: Multidimensional Feature Extraction and Fusion S201: Node similarity feature extraction. Calculate multiple classic node similarity indices between any two unconnected nodes in the network, and use the entropy weight method to couple the indices to obtain the comprehensive node similarity index NS. S202: Path similarity feature extraction, based on second-order and third-order paths between nodes, while considering the heterogeneity of intermediate connecting edges and intermediate nodes, to construct a comprehensive path similarity index PS; S203: Content similarity feature extraction. Based on the patent abstract text, the LDA topic model is used to extract the distribution of the patentee's technical topics and calculate the vector similarity between nodes. Then, the entropy weight method is used to couple multiple vector similarity indicators to obtain the comprehensive content similarity index CS. S204: Multi-dimensional index fusion. With link prediction accuracy as the optimization goal, a weighted fusion model of three indicators, NS, PS, and CS, is constructed, and an optimization algorithm is used to solve for the optimal weight of each indicator to obtain the fused prediction index NPC. Step S300: Model Training and Prediction The network edges are divided into a training set, a test set, and a set of non-existent edges. The fusion model is trained using the training set, and the prediction accuracy is evaluated using the test set. Finally, the model is applied to predict links between unconnected node pairs to generate a predicted cooperative network.
2. The method for predicting regional green innovation cooperation network links by integrating network topology and content characteristics according to claim 1, characterized in that... In S201, classic node similarity indices include one or more of CN, Salton, Jaccard, Sorenson, HPI, HDI, LHN-I, PA, AA, and RA; the entropy weight method determines the weight of each index by calculating its information entropy, and then calculates the weighted sum to obtain the comprehensive node similarity index NS.
3. The method for predicting regional green innovation cooperation network links by integrating network topology and content characteristics according to claim 1, characterized in that... In S202, the construction of the comprehensive path similarity index PS includes: Local path metrics calculate the number of second-order and third-order paths between nodes. The heterogeneity index for intermediate links is defined by the product of the degrees of intermediate nodes in the path, with the reciprocal of the weight representing the impact on the probability of cooperation; and The heterogeneity index of intermediate nodes is constructed based on the proportion of the connected paths between intermediate nodes and the two end nodes to their total connected paths. The above indicators are weighted and combined to obtain the comprehensive path similarity index PS, where the weights of the third-order paths are determined through sensitivity analysis.
4. The method for predicting regional green innovation cooperation network links that integrates network topology and content characteristics according to claim 1, characterized in that... In S203, the extraction of content similarity features includes: The patent abstract text is segmented and stop word removed to construct Doc2Bow vectors. The LDA topic model is used to extract topics from the text vectors, and the optimal number of topics K is determined using the CV consistency score method to obtain the topic distribution for each patent. Patent topics are assigned to corresponding patentee nodes, constructing a topic set vector for each node. Vector similarity between node pairs is calculated using one or more of the following metrics: Euclidean distance, Manhattan distance, Jaccard similarity, cosine similarity, Pearson correlation coefficient, and Tanimoto coefficient. Entropy weighting is used to couple multiple vector similarity metrics to obtain a comprehensive content similarity index, CS.
5. The method for predicting regional green innovation cooperation network links by integrating network topology and content characteristics according to claim 1, characterized in that... In S204, the optimization objective is to maximize the AUC value, and the optimization algorithm adopts the Grey Wolf Optimization Algorithm to solve for the optimal weight combination of the three indices NS, PS, and CS.
6. A regional green innovation cooperation network link prediction system that integrates network topology and content characteristics, characterized in that... It includes: The data acquisition module is used to obtain green patent application data from patent databases or documents, and extract patentee information, cooperative application relationships and patent abstract text. The network construction module is used to build and store an undirected weighted regional green innovation cooperation network based on the information extracted by the data acquisition module; The feature extraction module is used to perform the node similarity feature extraction, path similarity feature extraction, and content similarity feature extraction as described in claim 1; The model fusion and optimization module is used to build a weighted fusion model of NS, PS, and CS, and has a built-in Grey Wolf optimization algorithm to automatically solve for the optimal weight combination with the goal of maximizing the AUC value. The link prediction module is used to load the optimal fusion model, calculate and sort the NPC values of all unconnected node pairs in the network, and output a list of potential cooperative relationships and a predicted network topology. as well as The results analysis module includes visualization and statistical analysis tools to generate predictive network diagrams and perform statistical analysis on the type and geographical distribution of innovation entities, outputting analysis reports.