Link prediction method, system and device based on multi-layer heterogeneous network

Abstract

The application belongs to the field of link prediction, and specifically discloses a link prediction method, system and device based on a multi-layer heterogeneous network, comprising: constructing a multi-layer heterogeneous network; the nodes of the first, second and third sub-networks are enterprises, universities and research institutions respectively; based on the conditions and frequencies of cooperative patent applications among the enterprises, universities and research institutions, determining the actual cooperation edges and corresponding weights between the nodes within and between layers, and based on the similar conditions and frequencies of the corresponding technical fields of non-cooperative patent applications among the enterprises, universities and research institutions, determining the potential cooperation edges and corresponding weights between the nodes; constructing an adjacency matrix and a degree matrix through the actual cooperation edges, potential cooperation edges and corresponding weights; constructing a transition probability vector of each node, and performing link prediction based on the transition probability matrix of each node by using an RWR algorithm to determine the potential cooperation partners of the target entity. Through the application, high-accuracy industry-university-research link prediction is realized.

Description

Technical Field

[0001] This application belongs to the field of link prediction, and more specifically, relates to a link prediction method, system and device based on multi-layer heterogeneous networks. Background Technology

[0002] As global innovation competition intensifies, artificial intelligence (AI) has become a core engine driving "new-quality productivity." The Chinese government's work report has repeatedly emphasized the "AI+" action initiative, highlighting AI's strategic role in industrial upgrading. Industry-university-research (IUR) collaboration, as a key mechanism driving technological innovation and industrial development, is increasingly important. However, there is a significant disconnect in current IUR collaboration: universities and research institutions possess abundant research results, but enterprises frequently face shortages of key technologies, and the rate of technology transfer and industrialization remains low.

[0003] From a network science perspective, the IUR system consists of three heterogeneous entities: enterprises, universities, and research institutions. Each entity possesses different organizational logics and interaction characteristics. Traditional single-level network models homogenize the data from multiple entities, failing to effectively preserve the heterogeneous attributes of each entity and the differentiated interaction characteristics within and between layers, leading to severe information homogenization problems. Multi-level heterogeneous network models, through hierarchical representation, can effectively overcome this limitation.

[0004] Link prediction is a core task in network analysis, aiming to identify pairs of nodes that are not yet connected but have a high probability of being connected, based on network structure. In the IUR (Integrated Network Regulator) cooperation scenario, link prediction not only helps identify potential partners and reduce search costs, but also provides a scientific basis for governments to formulate innovative policies and optimize resource allocation. However, link prediction in IUR cooperation networks faces the following challenges: (1) Network heterogeneity: Enterprises, universities and research institutions have different connection patterns and cooperation preferences; (2) Supernode bias: There are a few supernodes with extremely high connectivity in the network (such as State Grid, Tsinghua University, etc.), which have a disproportionate impact during random walks; (3) Decoupling phenomenon: There is a significant gap between technology similarity networks and actual cooperation networks, and a large number of potential technology cooperation relationships have not yet been transformed into actual cooperation.

[0005] Existing IUR collaborative forecasting technologies mainly include the following categories: (1) Content similarity-based methods: These methods utilize TF-IDF to establish a content similarity matrix or employ topic models to extract sub-technical topics for partner recommendations. These methods rely on the quality of the text data and struggle to capture the topological information of the network structure.

[0006] (2) Collaborative filtering-based methods: These methods introduce behavioral attribute modeling and recommendation system approaches, considering organizational attributes and historical cooperation preferences for partner recommendations. However, these methods have limited effectiveness in recommending new nodes (cold start problem).

[0007] (3) Methods based on single-level network analysis: measuring the similarity of technical topics through patent applicant-classification number coupling analysis, or constructing a multi-index evaluation system to rank potential edges. These methods are limited to single-level networks and cannot handle the heterogeneity of cross-type subjects.

[0008] (4) Link prediction methods based on social networks: Social network analysis is used to discover the scale-free and small-world characteristics of cooperative networks. However, these methods mainly describe the structural characteristics of the network and have limited predictive capabilities.

[0009] (5) Standard Random Walk Restart (RWR) Algorithm: Starting from a specific source node, it measures the structural relevance with other nodes through random walks and estimates the probability of potential cooperative relationships. The standard RWR algorithm has a significant bias when dealing with networks with supernodes: the transfer weights of high-level nodes are too large, causing random walks to frequently visit these nodes, thus masking other combinations of entities with technical affinity.

[0010] In summary, existing technologies lack a link prediction method that can effectively handle multi-layered heterogeneous network structures, overcome supernode bias, and combine the two dimensions of technological similarity and actual cooperation. Summary of the Invention

[0011] To address the shortcomings of existing technologies, the purpose of this application is to provide a link prediction method, system, and device based on multi-layer heterogeneous networks. This aims to solve the problem that existing technologies lack a method that can effectively handle multi-layer heterogeneous network structures, overcome supernode bias, and combine the two dimensions of technical similarity and actual cooperation for IUR cooperative link prediction.

[0012] The first aspect of this application relates to a link prediction method based on multi-layer heterogeneous networks, comprising: Construct a multi-layered heterogeneous network; the multi-layered heterogeneous network includes: a first sub-network, a second sub-network, and a third sub-network; the nodes of the first sub-network are enterprises, the nodes of the second sub-network are universities, and the nodes of the third sub-network are research institutions, and other multi-layered network structures are also applicable; Based on the circumstances and frequency of collaborative patent applications among enterprises, universities, and research institutions, the actual collaborative edges and corresponding weights between corresponding nodes in the multi-layer heterogeneous network are determined. Based on the similarity and frequency of non-collaborative patent applications among enterprises, universities, and research institutions in the corresponding technical fields, the potential collaborative edges and corresponding weights between corresponding nodes in the multi-layer heterogeneous network are determined. The adjacency matrix and degree matrix of the multi-layer heterogeneous network are constructed by the actual cooperative edges, potential cooperative edges and corresponding weights between nodes in the multi-layer heterogeneous network; and the transition probability matrix of the multi-layer heterogeneous network is constructed based on the adjacency matrix and degree matrix; the transition probability vector of each node is determined by combining the transition probability matrix and the preset penalty coefficient, wherein the preset penalty coefficient is used to reduce the dominant effect of supernodes in the RWR algorithm. Based on the transition probability vectors of each node, the random walk restart (RWR) algorithm is used to predict the links of target nodes corresponding to target enterprises, universities, or research institutions, and to identify their potential partners.

[0013] In some implementations, the RWR algorithm is used to predict links to nodes corresponding to target enterprises, universities, or research institutions, specifically including: Initialize node probability distribution vector , ;in, express A real vector space of dimension n, where n represents the total number of nodes in the multi-layer heterogeneous network. , express The value of the i-th element in the array, where s is the sorting value of the target node among the n nodes in the multi-layer heterogeneous network; The node probability vector is iteratively updated by performing random walks on the nodes: ;in: Indicates the first The node probability distribution vector after round of iteration, let the th iteration be... t After one round of iterations, the path moves to node m; Indicates the first The node probability distribution vector after round of iteration, let the th iteration be... After one round of iterations, the journey reaches node n; P represents the probability of transitioning from node m to node n, which is the element value of node n in the transition probability vector of node m; The preset probability of returning to the target node after a random walk restart; When the Euclidean distance between the node probability vectors obtained in two adjacent iterations is less than a preset convergence threshold, the iterative update of the node probability vector is stopped, and the finally obtained node probability vector is taken as the node probability vector of the steady-state distribution. Obtain the similarity between each element in the steady-state distribution of the node probability vector and the s-th element corresponding to the target node, determine the candidate nodes corresponding to the elements with the highest similarity ranking, and regard the entities corresponding to the candidate nodes as potential partners of the entities corresponding to the target node.

[0014] In some implementations, the transition probability vector of each node is determined by combining the transition probability matrix and a preset penalty coefficient, including: Determine the nodes corresponding to each element on the diagonal of the transition probability matrix; The vector corresponding to each row of the transition probability matrix is ​​used as the transition probability vector of the node on the diagonal. When a node on the diagonal is a supernode, the transition probability vector is multiplied by a preset penalty coefficient; the preset penalty coefficient is greater than 0 and less than 1.

[0015] In some implementations, an adjacency matrix is ​​constructed based on the actual cooperative edges, potential cooperative edges, and corresponding weights between nodes in a multi-layered heterogeneous network, including: Construct the first-layer inner adjacency matrix based on the actual cooperative edges, potential cooperative edges, and corresponding weights between any two nodes in the first sub-network; Construct the second-layer inner adjacency matrix based on the actual cooperative edges, potential cooperative edges, and corresponding weights between any two nodes in the second sub-network; Construct the inner adjacency matrix of the third layer based on the actual cooperative edges, potential cooperative edges and corresponding weights between any two nodes in the third sub-network; Construct the first cross-layer adjacency matrix based on the actual cooperative edges, potential cooperative edges, and corresponding weights between any node in the first sub-network and any node in the second sub-network; Construct a second cross-layer adjacency matrix based on the actual cooperative edges, potential cooperative edges, and corresponding weights between any node in the first sub-network and any node in the third sub-network; Construct a third cross-layer adjacency matrix based on the actual cooperative edges, potential cooperative edges, and corresponding weights between any node in the second sub-network and any node in the third sub-network; The first layer inner adjacency matrix, the second layer inner adjacency matrix, the third layer inner adjacency matrix, the first cross-layer adjacency matrix, the second cross-layer adjacency matrix, and the third cross-layer adjacency matrix are combined into a diagonal matrix to obtain the adjacency matrix of the multi-layer heterogeneous network. Among them, the matrix elements corresponding to any two nodes in the first-level inner adjacency matrix, the second-level inner adjacency matrix, the third-level inner adjacency matrix, the first cross-level adjacency matrix, the second cross-level adjacency matrix, and the third cross-level adjacency matrix are obtained by weighted summation of the weights of the actual cooperative edges and the weights of the potential cooperative edges between the two nodes.

[0016] In some implementations, a degree matrix is ​​constructed based on the actual cooperative edges, potential cooperative edges, and corresponding weights among nodes in a multi-layer heterogeneous network, including: The degree matrix is ​​constructed based on the adjacency matrix of a multi-layer heterogeneous network; the degree matrix is ​​a diagonal matrix, and the elements on its main diagonal are the sum of all elements in the corresponding row of the adjacency matrix.

[0017] In some implementations, determining the actual cooperative edges and their corresponding weights, as well as the potential cooperative edges and their corresponding weights, between corresponding nodes in the multi-layer heterogeneous network includes: When any two enterprises, any two universities, any two research institutions, or any two entities among the three (enterprises, universities, and research institutions) have jointly applied for at least one patent, there is an actual cooperation edge between the corresponding two nodes. The weight of this actual cooperation edge is the frequency of the corresponding two entities jointly applying for patents. If any two enterprises, any two universities, any two research institutions, or any two entities among the three (enterprises, universities, and research institutions) have a common IPC classification number in their non-cooperative patent applications, then there is a potential cooperative edge between the two corresponding nodes. The weight of this potential cooperative edge is the co-occurrence frequency of the common IPC classification number in the non-cooperative patent applications of the two entities.

[0018] In some implementations, the transition probability matrix of a multilayer heterogeneous network is constructed based on the adjacency matrix and the degree matrix, including: The adjacency matrix is ​​then symmetrically normalized using the degree matrix to obtain the updated adjacency matrix. The updated adjacency matrix is ​​row normalized to obtain the transition probability matrix of the multilayer heterogeneous network.

[0019] In some implementations, the link prediction method further includes: determining the degree of each node based on the degree matrix; and determining the node as a supernode when the degree of a node is ranked before a preset ranking value.

[0020] Secondly, this application provides a link prediction system based on a multi-layer heterogeneous network, comprising: A network construction module is used to construct a multi-layer heterogeneous network. The multi-layer heterogeneous network includes a first sub-network, a second sub-network, and a third sub-network. The nodes of the first sub-network are enterprises, the nodes of the second sub-network are universities, and the nodes of the third sub-network are research institutions. Based on the situation and frequency of collaborative patent applications among enterprises, universities, and research institutions, the module determines the actual collaborative edges and corresponding weights between the corresponding nodes in the multi-layer heterogeneous network. Based on the similarity and frequency of non-collaborative patent applications in the technical fields corresponding to the patent applications among enterprises, universities, and research institutions, the module determines the potential collaborative edges and corresponding weights between the corresponding nodes in the multi-layer heterogeneous network. The vector acquisition module is used to construct the adjacency matrix and degree matrix of the multi-layer heterogeneous network through the actual cooperative edges, potential cooperative edges and corresponding weights between nodes in the multi-layer heterogeneous network; and to construct the transition probability matrix of the multi-layer heterogeneous network based on the adjacency matrix and degree matrix; and to determine the transition probability vector of each node by combining the transition probability matrix and the preset penalty coefficient; wherein, the preset penalty coefficient is used to reduce the dominant effect of supernodes in the RWR algorithm; The link prediction module is used to predict the industry-university-research links of target nodes corresponding to target enterprises, universities or research institutions based on the transition probability vectors of each node and using the random walk restart (RWR) algorithm to identify their potential partners.

[0021] Thirdly, this application provides an electronic device, including a memory and one or more processors; The memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions; The one or more processors invoke the computer instructions to cause the electronic device to perform the method as described in the first aspect or any of the embodiments of the first aspect above.

[0022] Fourthly, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to perform the methods described in the first aspect or any of the embodiments of the first aspect.

[0023] Fifthly, this application provides a computer program product that, when run on a processor, causes the processor to perform the methods described in the first aspect or any of the embodiments of the first aspect.

[0024] Overall, the technical solutions conceived in this application have the following beneficial effects compared with the prior art: This application provides a link prediction method, system, and device based on multi-layer heterogeneous networks. It proposes a multi-layer heterogeneous network modeling method for industry-academia-research collaboration. The multi-layer heterogeneous network model effectively preserves the heterogeneous attributes and differentiated interaction characteristics of three types of entities: enterprises, universities, and research institutions. Simultaneously, it constructs two dimensions: a "technology similarity network" and a "practical cooperation network." The technology similarity network establishes potential technological cooperation relationships based on IPC co-occurrence, while the practical cooperation network establishes actual cooperation relationships based on joint patent applications. Comparative analysis of the two networks improves the accuracy of link prediction for industry-academia-research collaboration. Furthermore, this application proposes a random walk restart algorithm with supernode penalty. Based on the standard RWR algorithm, the transition probability vector of supernodes (institutions with extremely high connectivity) in the network is multiplied by a penalty coefficient β=0.5, effectively mitigating the disproportionate influence of highly connected nodes in random walks and significantly reducing the dominance effect of highly connected nodes. This method achieves excellent prediction performance of AUC=0.8875 and F1-score=0.86 on the AI ​​patent dataset, significantly outperforming the standard RWR algorithm.

[0025] This application provides a link prediction method, system, and device based on multi-layer heterogeneous networks. It proposes a three-dimensional verification framework to empirically verify the link prediction results from three dimensions: technological synergy, geographical rationality, and policy consistency. This framework not only verifies the accuracy of the prediction algorithm but also reveals the institutional logic behind the predicted links, ensuring that the predicted high-potential links are feasible for transformation into actual cooperation. Enterprises can use this method to quickly identify potential partners with technological complementarity, reducing partner search and evaluation costs; governments can use the prediction results to formulate innovative strategic plans and promote targeted industry-academia-research cooperation; researchers can use this method to accelerate the connection between research results and market demands, promoting the industrialization of scientific and technological achievements. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating the link prediction method based on a multi-layer heterogeneous network provided in the embodiments of this application.

[0027] Figure 2 This is a research framework diagram for link prediction based on multi-layer heterogeneous networks provided in the embodiments of this application.

[0028] Figure 3 This is an architecture diagram of a multi-layer heterogeneous network provided in the embodiments of this application.

[0029] Figure 4 This is a schematic diagram of the adjacency matrix construction provided in an embodiment of this application.

[0030] Figure 5 This is an architecture diagram of a link prediction system based on a multi-layer heterogeneous network provided in an embodiment of this application.

[0031] Figure 6 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0032] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0033] In the description of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0034] Furthermore, throughout this specification, references to "an embodiment"; "an embodiment," "an example," or similar language indicate that a particular feature, structure, or characteristic described in connection with that embodiment is included in at least one embodiment of this application. Therefore, the appearance of the phrase "in one embodiment;" throughout this specification, and similar language, may, but not necessarily, refer to the same embodiment.

[0035] The embodiments of this application are described below with reference to the accompanying drawings.

[0036] Figure 1 This is a flowchart illustrating the link prediction method based on a multi-layer heterogeneous network provided in an embodiment of this application. Figure 1 As shown, it includes the following steps: Step S101: Construct a multi-layer heterogeneous network; the multi-layer heterogeneous network includes: a first sub-network, a second sub-network, and a third sub-network; the nodes of the first sub-network are enterprises, the nodes of the second sub-network are universities, and the nodes of the third sub-network are research institutions. Step S102: Based on the situation and frequency of collaborative patent applications among enterprises, universities and research institutions, determine the actual collaborative edges and corresponding weights between corresponding nodes in the multi-layer heterogeneous network; based on the similarity and frequency of non-collaborative patent applications in the technical fields corresponding to enterprises, universities and research institutions, determine the potential collaborative edges and corresponding weights between corresponding nodes in the multi-layer heterogeneous network. Step S103: Construct the adjacency matrix and degree matrix of the multi-layer heterogeneous network through the actual cooperative edges, potential cooperative edges and corresponding weights between nodes in the multi-layer heterogeneous network; and construct the transition probability matrix of the multi-layer heterogeneous network based on the adjacency matrix and degree matrix; combine the transition probability matrix and the preset penalty coefficient to determine the transition probability vector of each node, wherein the preset penalty coefficient is used to reduce the dominant effect of supernodes in the RWR algorithm. Step S104: Based on the transition probability vector of each node, the random walk restart (RWR) algorithm is used to predict the link of the target node corresponding to the target enterprise, university or research institution to determine its potential partners.

[0037] Specifically, the purpose of this application is to provide a link prediction method in multilayer heterogeneous networks based on the RWR algorithm with supernode penalty, see [link to relevant documentation]. Figure 2 As shown, it mainly includes the following aspects: (1) Network Modeling Level: Construct a multi-layered heterogeneous network comprising three types of entities: enterprises, universities, and research institutions. Build networks based on two dimensions: "technological similarity" and "actual cooperation." Through comparative analysis, reveal potential cooperation paths obscured by organizational barriers. See also Figure 3As shown, B1, B2, B3, etc. represent enterprise nodes, which are in the first sub-network layer; U1, U2, U3, etc. represent university nodes, which are in the second sub-network layer; and R1, R2, R3, etc. represent research institution nodes, which are in the third sub-network layer.

[0038] (2) Algorithm optimization level: This application proposes an RWR algorithm with supernode penalty. By applying a penalty coefficient to the transition probability of supernodes, it effectively alleviates the disproportionate influence of height nodes in random walks and improves the accuracy and fairness of link prediction.

[0039] (3) Prediction verification level: Establish a three-dimensional verification framework of "technology collaboration - regional layout - policy guidance" and verify the effectiveness of the prediction results through actual cases to ensure that the high-potential links predicted can be transformed into actual partnerships.

[0040] This application proposes a link prediction method in multi-layer heterogeneous networks based on the RWR algorithm with supernode penalty. The basic principle is as follows: First, a multi-layer heterogeneous network containing three types of entities—enterprises, universities, and research institutions—is constructed, establishing the network from two dimensions: technological similarity and actual cooperation. Then, a symmetric normalized transition matrix with supernode penalty is used, and the structural similarity between each node and the source node is calculated through a random walk restart algorithm. Finally, link prediction is performed based on a steady-state distribution, and the prediction results are verified using a three-dimensional framework.

[0041] In multi-layered heterogeneous networks, a few institutions (such as State Grid, JD.com, Tsinghua University, and the Chinese Academy of Sciences) have significantly higher connectivity than ordinary nodes, forming supernodes. In the standard RWR algorithm, the random walk's tendency to converge towards these supernodes is too strong, leading to a high concentration of prediction results in a few leading institutions, thus obscuring potential partners with genuine technological affinity but moderate connectivity. This application effectively weakens the disproportionate influence of supernodes by multiplying their transition probability vectors by a penalty coefficient β (default 0.5), resulting in more balanced and accurate prediction results.

[0042] In one specific embodiment, the link prediction method provided in this application includes the following steps: (1) Data collection and preprocessing: Collect valid invention patent data in the field of AI from patent databases (such as Qichacha platform), and perform data cleaning (remove design patents, exclude applicants outside the scope of IUR (that is, exclude applicants who are not enterprises, universities and research institutions), standardize entity names, and merge parent and subsidiary companies).

[0043] Taking the prediction of industry-academia-research collaboration in the AI ​​field as an example, valid invention patent data in the AI ​​field (such as 142,070 patents from 1985 to 2022) is collected from patent databases. Design patents are removed, applicants outside the IUR scope are excluded, entity names are standardized, and parent and subsidiary companies are merged to obtain a valid patent set. Node identification and classification: Institutional units are identified into three types of nodes: enterprises, universities, and research institutions. Aliases of the same entity are deduplicated, and parent and subsidiary companies are grouped under the same brand.

[0044] (2) Construction of multi-layer heterogeneous network: Construct a multi-layer heterogeneous network G containing three sub-networks: G=(V,E) Wherein, V represents the set of nodes, specifically including: enterprise nodes, university nodes, and research institution nodes; E Denotes the set of edges. E= : Including inner edges , , and cross-layer edge ; 2.1: Actual Cooperation Network: Actual cooperation relationships are established based on joint patent applications, with edge weights equal to the frequency of actual cooperation. For example, if two nodes have jointly applied for at least one patent, an actual cooperation edge is established, with the weight equal to the frequency of cooperation. 2.2: Technology Similarity Network: Potential technology cooperation relationships are established based on the co-occurrence frequency of IPCs, with edge weights equal to the co-occurrence frequency of shared IPC categories. For example, if the patents of two nodes belong to the same IPC category, a potential technology cooperation edge is established, with the weight equal to the shared IPC frequency.

[0045] Furthermore, all edges in the original network are randomly divided into a training edge set (80%) and a test edge set (20%). The training set is used for model training, and the test set is used to evaluate the recommendation accuracy.

[0046] (3) Construction of adjacency matrix A and degree matrix D: See Figure 4 As shown, a weighted adjacency matrix A and a degree matrix D are constructed for a multi-layer heterogeneous network. The three sub-networks (enterprise-enterprise, university-university, and research institution-research institution) are represented as adjacency matrices X, Y, and Z, respectively, and cross-layer connection matrices XY (enterprise-university), XZ (enterprise-research institution), and YZ (university-research institution) are constructed.

[0047] Specifically, the first-layer inner adjacency matrix is ​​constructed based on the actual cooperative edges, potential cooperative edges, and corresponding weights between any two nodes in the first sub-network. (See [link to documentation]). Figure 4 The red XX matrix; The second-layer inner adjacency matrix is ​​constructed based on the actual cooperative edges, potential cooperative edges, and corresponding weights between any two nodes in the second sub-network. See [link / reference]. Figure 4 The medium-gray YY matrix; the third-layer inner adjacency matrix is ​​constructed based on the actual cooperative edges, potential cooperative edges, and corresponding weights between any two nodes in the third sub-network. See [link / reference]. Figure 4 The blue ZZ matrix.

[0048] Construct the first cross-layer adjacency matrix based on the actual cooperative edges, potential cooperative edges, and corresponding weights between any node in the first sub-network and any node in the second sub-network; see [link to relevant documentation]. Figure 4 The XY matrix in the diagram. The second cross-layer adjacency matrix is ​​constructed based on the actual cooperative edges, potential cooperative edges, and corresponding weights between any node in the first sub-network and any node in the third sub-network; see [link to documentation]. Figure 4 The XZ matrix in [the original text is missing]. The third cross-layer adjacency matrix is ​​constructed based on the actual cooperative edges, potential cooperative edges, and corresponding weights between any node in the second sub-network and any node in the third sub-network; see [the original text is missing]. Figure 4 The YZ matrix in the diagram.

[0049] The first-layer inner adjacency matrix, the second-layer inner adjacency matrix, the third-layer inner adjacency matrix, the first cross-layer adjacency matrix, the second cross-layer adjacency matrix, and the third cross-layer adjacency matrix are combined into a diagonal matrix to obtain the adjacency matrix A of the multi-layer heterogeneous network, as shown below. Figure 4 As shown; where, Figure 4 In the first column vertically and the first column horizontally, E1~E5,... represent different enterprise nodes, U1~U4,... represent different university nodes, and R1~R3,... represent different research institution nodes; the elements at the intersection of the horizontal and vertical columns are the matrix element values ​​of the corresponding nodes, for example: X 12 This represents the matrix elements corresponding to the first and second enterprise nodes, XY. 23 This represents the matrix elements corresponding to the second enterprise node and the third university node, XZ 41 YZ represents the matrix element corresponding to the fourth enterprise node and the first research institution node. 13 This represents the matrix elements corresponding to the first university node and the third research institution node.

[0050] It is understandable that the element values ​​corresponding to the elements X11, X22, ..., Z11, Z22, ... on the diagonal of the above adjacency matrix A are 0.

[0051] The matrix elements corresponding to any two nodes in the first-level inner adjacency matrix, the second-level inner adjacency matrix, the third-level inner adjacency matrix, the first cross-level adjacency matrix, the second cross-level adjacency matrix, and the third cross-level adjacency matrix are obtained by weighted summation of the weights of the actual cooperative edges and the weights of the potential cooperative edges between the two nodes.

[0052] In one example, if the weight of the actual cooperative edge between two nodes is q1 and the weight of the potential cooperative edge is q2, then the matrix element value of the two nodes is β1*q1 + β2*q2, where β1 and β2 represent the weighting ratio of the two weights, which can be determined empirically. When there is no actual cooperative edge between the two nodes, q1 is 0, and when there is no potential cooperative edge, q2 is 0.

[0053] In some implementations, a degree matrix is ​​constructed based on the actual cooperative edges, potential cooperative edges, and corresponding weights among nodes in a multi-layer heterogeneous network, including: The degree matrix is ​​constructed based on the adjacency matrix of a multi-layer heterogeneous network; the degree matrix is ​​a diagonal matrix, and the elements on its main diagonal are the sum of all elements in the corresponding row of the adjacency matrix. That is, the degree matrix D has elements only on its diagonal, and the element value on the diagonal of each row is the sum of all elements in that row.

[0054] (4) Construction of the transition matrix: The construction of the transition probability matrix adopts a two-step normalization: Step 1, symmetric normalization: A'=( ) Where D is the degree matrix. Symmetric normalization can weaken the influence of height-in-degree nodes and height-out-degree nodes.

[0055] The second step is row normalization: P = RowNormalize(A') Ensure that the sum of the transition probabilities for each row equals 1.

[0056] (5) Supernode Penalty Module: Supernode penalty is the core technological innovation of this application. In the standard RWR algorithm, row normalization of the transition probability matrix dilutes the outgoing edge transition probability of height nodes, but the effect of convergence through incoming edges is still strong. The supernode penalty mechanism directly reduces the probability of walking from any node to the supernode by multiplying the transition probability vector corresponding to the supernode by a penalty coefficient β (default 0.5), thereby effectively mitigating supernode bias.

[0057] Supernode identification criteria: Nodes with a connectivity degree significantly higher than the network average degree, such as the top 1% of height nodes. The penalty coefficient β ranges from (0,1], with β=1 indicating no penalty (standard RWR), and smaller β indicating stronger penalty. Experiments have verified that β=0.5 achieves the best prediction performance on the AI ​​patent dataset.

[0058] (6) RWR calculation module: Starting from the source node s, initialize the probability vector r0 (only the dimension of s is 1), and iteratively calculate until convergence: Where α∈(0,1) is the restart probability (default 0.3), and the convergence threshold is... Convergence condition: .

[0059] Specifically, given a network graph The total number of nodes is The RWR algorithm targets the source node. Starting with the source node, the structural correlation between other nodes in the network and the source node is calculated by simulating the mechanism of continuously returning to the source node during a random walk. Specifically, this includes: 6.1 Initialize the node probability vector First, construct the initial probability vector: in, express A 3D real vector space. Vectors Source node only The corresponding position is set to 1, and all other node positions are set to 0, that is: This initialization method means that the initial state of the random walk is entirely located at the source node s, that is, the source node has an initial probability of 1, and the initial probabilities of the other nodes are all 0.

[0060] 6.2 Iterative Update of Probability Distribution In the During this iteration, the node access probability vector is updated according to the following rules: in: Indicates the first The node probability distribution vector after round of iteration, let the th iteration be... t After one round of iterations, the path moves to node m; Indicates the first The node probability distribution vector after round of iteration, let the th iteration be... After one round of iterations, the journey reaches node n; P represents the probability of transitioning from node m to node n, which is the element value of node n in the transition probability vector of node m; This is the transpose of the transition matrix, used to satisfy probability propagation calculations in column vector form; To determine the probability of restarting, this application sets: , indicating that in each iteration: with probability Random walk propagation along the network topology; with probability Return to the initial source node They started wandering again.

[0061] Therefore, RWR can not only use network topology information to propagate node relationships, but also maintain attention to the source node through a periodic restart mechanism, avoiding excessive deviation from the initial node during random walks.

[0062] 6.3 Iterative Convergence Judgment To determine whether the random walk process has reached a steady state, the Euclidean distance between two adjacent iterations is used as the convergence criterion: in: express Norm, used to measure the overall difference between two probability distribution vectors; The convergence threshold is set as follows in this method: When the change in the probability vector obtained from two consecutive iterations is less than the threshold, the random walk process is considered to have reached a stable state, and the iteration stops.

[0063] 6.4 Steady-state similarity calculation and link prediction After multiple iterations, a stable probability distribution was finally obtained: The steady-state vector is the first One element: This indicates the nodes visited during a random walk. The stable probability. Because the walk process always starts from the source node. As a basis for restarting, this probability value can reflect the node's condition. With source node The degree of correlation in the network topology.

[0064] In the link prediction task, the node probability values ​​in the steady-state distribution vector are used as the node association score: The higher the score, the better the node. With source node The stronger the structural similarity and potential connection probability between nodes, the higher the ranking of candidate nodes, thus enabling the prediction of unknown links. The value of each node in the steady-state distribution r represents the structural similarity between that node and the source node s, serving as a recommendation score for link prediction.

[0065] (7) Link prediction and ranking module: Based on the value of each node in the steady-state distribution as the structural similarity (recommendation score) with the source node, the candidate edges are ranked and the Top-K results are selected.

[0066] For example, based on the recommendation scores from highest to lowest, select nodes with the highest number of nodes (such as the top 50 enterprise nodes, the top 30 university nodes, and the top 20 research institution nodes) as the starting point for recommendations, execute the RWR algorithm, and select Top-K=50 candidate target nodes. Sort all candidate edges by score and retain the global Top-N=50 results.

[0067] (8) Three-dimensional verification module: The effectiveness of the prediction results is empirically verified through a three-dimensional framework of "technology collaboration - regional layout - policy guidance". For example, in the dimension of technological synergy: This involves analyzing and predicting the overlap, complementarity, and collaborative innovation potential of the collaborating parties. For instance, the University of Electronic Science and Technology of China and China Three Gorges University both have deep expertise in power electronics and smart grids, achieving technological synergy through jointly applying for National Natural Science Foundation of China projects and holding innovation forums. In the dimension of regional layout: This involves analyzing the impact of geographical proximity on cooperation. Local proximity promotes resource flow (e.g., the AI ​​alliance between Shandong University and Haier Group), while cross-regional cooperation breaks through spatial limitations (e.g., the space breeding cooperation between the China Academy of Launch Vehicle Technology and Hainan University). In the dimension of policy guidance: This involves analyzing whether the cooperation aligns with national strategic policies. For example, the "dual carbon" goal guided the State Grid Corporation of China and Zhejiang University to jointly release the "Industrial Carbon Efficiency Code 1.0"; the "new infrastructure" and digital economy policies promoted the cooperation between Founder Group and the Shandong Artificial Intelligence Research Institute to develop a blockchain-based digital copyright protection system.

[0068] (9) Model performance evaluation. The link prediction performance was evaluated using AUC and F1-score. AUC measures the overall ranking ability, and F1-score measures the harmonic mean of precision and recall.

[0069] As verified by the embodiments, the method provided in this application achieves a prediction performance of 0.8875 area under the curve (AUC) and 0.86 F1 score on the AI ​​patent dataset.

[0070] In summary, the advantages of this application compared to existing technologies are as follows: 1. Heterogeneity preservation: The multi-layer heterogeneous network model effectively preserves the heterogeneous attributes and differentiated interaction characteristics of the three types of entities: enterprises, universities, and research institutions; 2. Supernode bias elimination: The supernode penalty mechanism multiplies the transition probability of supernodes by a penalty coefficient β=0.5, significantly reducing the dominance effect of high-level nodes; 3. Two-dimensional comparison: Through comparative analysis of technology similarity networks and actual cooperation networks, the degree of decoupling between industry, academia, and research cooperation is quantitatively revealed; 4. High prediction accuracy: A prediction performance of AUC=0.8875 and F1-score=0.86 is achieved on the AI ​​patent dataset; 5. Strong interpretability: The prediction results are empirically verified through a three-dimensional verification framework (technology-region-policy), which has clear practical guiding significance.

[0071] Figure 5 This is a diagram of the link prediction system architecture based on a multi-layer heterogeneous network provided in the embodiments of this application, such as... Figure 5 As shown, it includes: Network construction module 510 is used to construct a multi-layer heterogeneous network; the multi-layer heterogeneous network includes: a first sub-network, a second sub-network, and a third sub-network; the nodes of the first sub-network are enterprises, the nodes of the second sub-network are universities, and the nodes of the third sub-network are research institutions; and based on the situation and frequency of collaborative patent applications among enterprises, universities, and research institutions, the actual collaborative edges and corresponding weights between corresponding nodes in the multi-layer heterogeneous network are determined; based on the similarity and frequency of non-collaborative patent applications in the technical fields corresponding to enterprises, universities, and research institutions, the potential collaborative edges and corresponding weights between corresponding nodes in the multi-layer heterogeneous network are determined. The vector acquisition module 520 is used to construct the adjacency matrix and degree matrix of the multi-layer heterogeneous network through the actual cooperative edges, potential cooperative edges and corresponding weights between nodes in the multi-layer heterogeneous network; and to construct the transition probability matrix of the multi-layer heterogeneous network based on the adjacency matrix and degree matrix; and to determine the transition probability vector of each node by combining the transition probability matrix and the preset penalty coefficient; wherein, the preset penalty coefficient is used to reduce the dominant effect of supernodes in the RWR algorithm. The link prediction module 530 is used to predict the industry-university-research links of target nodes corresponding to target enterprises, universities or research institutions based on the transition probability vectors of each node using the random walk restart RWR algorithm, and to identify their potential partners.

[0072] It should be understood that the above system is used to execute the methods in the above embodiments. The corresponding program modules in the system are similar in implementation principle and technical effect to those described in the above methods. The working process of the system can be referred to the corresponding process in the above methods, and will not be repeated here.

[0073] Based on the methods in the above embodiments, this application provides an electronic device, such as... Figure 6 As shown, the electronic device may include a processor 610, a communication interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communication interface 620, and the memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions in the memory to execute the methods in the above embodiments.

[0074] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0075] Based on the methods in the above embodiments, this application provides a computer-readable storage medium storing a computer program that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0076] Based on the methods in the above embodiments, this application provides a computer program product that, when run on a processor, causes the processor to execute the methods in the above embodiments.

[0077] It is understood that the processor in the embodiments of this application can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. A general-purpose processor can be a microprocessor or any conventional processor.

[0078] The method steps in this application embodiment can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in random access memory (RAM), flash memory, read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), registers, hard disks, portable hard disks, CD-ROMs, or any other form of storage medium known in the art. An exemplary storage medium is coupled to the processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and the storage medium can reside in an ASIC.

[0079] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially as a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted through the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid-state disk (SSD)).

[0080] It is understood that the various numerical designations used in the embodiments of this application are merely for the convenience of description and are not intended to limit the scope of the embodiments of this application.

[0081] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A link prediction method based on multi-layer heterogeneous networks, characterized in that, include: Constructing multi-layered heterogeneous networks; The multi-layered heterogeneous network includes: a first sub-network, a second sub-network, and a third sub-network; the nodes of the first sub-network are enterprises, the nodes of the second sub-network are universities, and the nodes of the third sub-network are research institutions. Based on the circumstances and frequency of collaborative patent applications among enterprises, universities, and research institutions, the actual collaborative edges and corresponding weights between corresponding nodes in the multi-layer heterogeneous network are determined. Based on the similarity and frequency of non-collaborative patent applications among enterprises, universities, and research institutions in the corresponding technical fields, the potential collaborative edges and corresponding weights between corresponding nodes in the multi-layer heterogeneous network are determined. The adjacency matrix and degree matrix of the multi-layer heterogeneous network are constructed by the actual cooperative edges, potential cooperative edges and corresponding weights between nodes in the multi-layer heterogeneous network; and the transition probability matrix of the multi-layer heterogeneous network is constructed based on the adjacency matrix and degree matrix; the transition probability vector of each node is determined by combining the transition probability matrix and the preset penalty coefficient, wherein the preset penalty coefficient is used to reduce the dominant effect of supernodes in the RWR algorithm. Based on the transition probability vectors of each node, the random walk restart (RWR) algorithm is used to predict the links of target nodes corresponding to target enterprises, universities, or research institutions, and to identify their potential partners.

2. The link prediction method according to claim 1, characterized in that, The RWR algorithm is used to predict links to nodes corresponding to target enterprises, universities, or research institutions, specifically including: Initialize node probability distribution vector , ;in, express A real vector space of dimension n, where n represents the total number of nodes in the multi-layer heterogeneous network. , express The value of the i-th element in the array, where s is the sorting value of the target node among the n nodes in the multi-layer heterogeneous network; The node probability vector is iteratively updated by performing random walks on the nodes: ;in: Indicates the first The node probability distribution vector after round of iteration, let the th iteration be... t After one round of iterations, the path moves to node m; Indicates the first The node probability distribution vector after round of iteration, let the th iteration be... After one round of iterations, the journey reaches node n; P represents the probability of transitioning from node m to node n, which is the element value of node n in the transition probability vector of node m; The preset probability of returning to the target node after a random walk restart; When the Euclidean distance between the node probability vectors obtained in two adjacent iterations is less than a preset convergence threshold, the iterative update of the node probability vector is stopped, and the finally obtained node probability vector is taken as the node probability vector of the steady-state distribution. Obtain the similarity between each element in the steady-state distribution of the node probability vector and the s-th element corresponding to the target node, determine the candidate nodes corresponding to the elements with the highest similarity ranking, and regard the entities corresponding to the candidate nodes as potential partners of the entities corresponding to the target node.

3. The link prediction method according to claim 1, characterized in that, The transition probability vector of each node is determined by combining the transition probability matrix and the preset penalty coefficient, including: Determine the nodes corresponding to each element on the diagonal of the transition probability matrix; The vector corresponding to each row of the transition probability matrix is ​​used as the transition probability vector of the node on the diagonal. When a node on the diagonal is a supernode, the transition probability vector is multiplied by a preset penalty coefficient; the preset penalty coefficient is greater than 0 and less than 1.

4. The link prediction method according to claim 1, characterized in that, The adjacency matrix is ​​constructed based on the actual cooperative edges, potential cooperative edges, and corresponding weights between nodes in a multi-layer heterogeneous network, including: Construct the first-layer inner adjacency matrix based on the actual cooperative edges, potential cooperative edges, and corresponding weights between any two nodes in the first sub-network; Construct the second-layer inner adjacency matrix based on the actual cooperative edges, potential cooperative edges, and corresponding weights between any two nodes in the second sub-network; Construct the inner adjacency matrix of the third layer based on the actual cooperative edges, potential cooperative edges and corresponding weights between any two nodes in the third sub-network; Construct the first cross-layer adjacency matrix based on the actual cooperative edges, potential cooperative edges, and corresponding weights between any node in the first sub-network and any node in the second sub-network; Construct a second cross-layer adjacency matrix based on the actual cooperative edges, potential cooperative edges, and corresponding weights between any node in the first sub-network and any node in the third sub-network; Construct a third cross-layer adjacency matrix based on the actual cooperative edges, potential cooperative edges, and corresponding weights between any node in the second sub-network and any node in the third sub-network; The first layer inner adjacency matrix, the second layer inner adjacency matrix, the third layer inner adjacency matrix, the first cross-layer adjacency matrix, the second cross-layer adjacency matrix, and the third cross-layer adjacency matrix are combined into a diagonal matrix to obtain the adjacency matrix of the multi-layer heterogeneous network. Among them, the matrix elements corresponding to any two nodes in the first-level inner adjacency matrix, the second-level inner adjacency matrix, the third-level inner adjacency matrix, the first cross-level adjacency matrix, the second cross-level adjacency matrix, and the third cross-level adjacency matrix are obtained by weighted summation of the weights of the actual cooperative edges and the weights of the potential cooperative edges between the two nodes.

5. The link prediction method according to claim 1, characterized in that, Based on the actual cooperative edges, potential cooperative edges, and corresponding weights among nodes in a multi-layer heterogeneous network, a degree matrix is ​​constructed, including: The degree matrix is ​​constructed based on the adjacency matrix of a multi-layer heterogeneous network; the degree matrix is ​​a diagonal matrix, and the elements on its main diagonal are the sum of all elements in the corresponding row of the adjacency matrix.

6. The link prediction method according to claim 1, characterized in that, Determining the actual cooperative edges and corresponding weights between relevant nodes in the multi-layer heterogeneous network, as well as the potential cooperative edges and corresponding weights, includes: When any two enterprises, any two universities, any two research institutions, or any two entities among the three (enterprises, universities, and research institutions) have jointly applied for at least one patent, there is an actual cooperation edge between the corresponding two nodes. The weight of this actual cooperation edge is the frequency of the corresponding two entities jointly applying for patents. If any two enterprises, any two universities, any two research institutions, or any two entities among the three (enterprises, universities, and research institutions) have a common IPC classification number in their non-cooperative patent applications, then there is a potential cooperative edge between the two corresponding nodes. The weight of this potential cooperative edge is the co-occurrence frequency of the common IPC classification number in the non-cooperative patent applications of the two entities.

7. The link prediction method according to claim 1, characterized in that, The transition probability matrix of a multilayer heterogeneous network is constructed based on the adjacency matrix and the degree matrix, including: The adjacency matrix is ​​then symmetrically normalized using the degree matrix to obtain the updated adjacency matrix. The updated adjacency matrix is ​​row normalized to obtain the transition probability matrix of the multilayer heterogeneous network.

8. The link prediction method according to claim 1, characterized in that, Also includes: The degree of each node is determined based on the degree matrix. When a node's degree is ranked before a preset ranking value, that node is determined to be a supernode.

9. A link prediction system based on a multi-layer heterogeneous network, characterized in that, include: The network building module is used to construct multi-layered heterogeneous networks. The multi-layered heterogeneous network includes: a first sub-network, a second sub-network, and a third sub-network; the nodes of the first sub-network are enterprises, the nodes of the second sub-network are universities, and the nodes of the third sub-network are research institutions; and based on the situation and frequency of collaborative patent applications among enterprises, universities, and research institutions, the actual collaborative edges and corresponding weights between corresponding nodes in the multi-layered heterogeneous network are determined; and based on the similarity and frequency of non-collaborative patent applications in the technical fields corresponding to enterprises, universities, and research institutions, the potential collaborative edges and corresponding weights between corresponding nodes in the multi-layered heterogeneous network are determined. The vector acquisition module is used to construct the adjacency matrix and degree matrix of the multi-layer heterogeneous network through the actual cooperative edges, potential cooperative edges and corresponding weights between nodes in the multi-layer heterogeneous network; and to construct the transition probability matrix of the multi-layer heterogeneous network based on the adjacency matrix and degree matrix; and to determine the transition probability vector of each node by combining the transition probability matrix and the preset penalty coefficient; wherein, the preset penalty coefficient is used to reduce the dominant effect of supernodes in the RWR algorithm; The link prediction module is used to predict the industry-university-research links of target nodes corresponding to target enterprises, universities or research institutions based on the transition probability vectors of each node and using the random walk restart (RWR) algorithm to identify their potential partners.

10. An electronic device, characterized in that, Includes memory and one or more processors; The memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions; The one or more processors invoke the computer instructions to cause the electronic device to perform the method as described in any one of claims 1-8.

Images