Method for analyzing dynamic evolution of industry-university-research network based on life cycle and topological structure

Abstract

This invention discloses a method for dynamic evolution analysis of industry-academia-research networks based on lifecycle and topology, belonging to the field of industry-academia-research network evolution analysis technology. Based on various entities and cooperative relationships within industry-academia-research collaborations, this invention combines time decay to calculate dynamic edge weights, constructs a dynamic heterogeneous knowledge graph, and periodically updates it incrementally to obtain a time-series snapshot sequence. Based on the time-series snapshot sequence, multi-scale topological representation is performed. An enhanced benefit-driven evolutionary game model based on the PPO algorithm is constructed, and based on topological features, the cooperative and default strategies of industry-academia-research entities are made, and the cooperative relationships are dynamically adjusted. A Transformer-VAE coupled model is constructed, and a comprehensive evolution index is obtained through VAE adaptive fusion. The lifecycle stages are divided and stage transition points are identified through time-series fusion Transformer. An evolutionary causal graph is constructed based on the DoWhy causal inference library, and multi-scenario evolution simulation is completed by combining counterfactual reasoning. This invention realizes dynamic evolution judgment and intelligent control of the entire process of industry-academia-research networks, and has profound application value.

Description

Technical Field

[0001] This invention relates to the field of industry-academia-research network evolution analysis technology, specifically to a method for dynamic evolution analysis of industry-academia-research networks based on lifecycle and topology. Background Technology

[0002] Industry-university-research collaborative innovation has become an important vehicle for technological innovation, industrial transformation and upgrading, and tackling key core technologies. A vast number of heterogeneous and dynamically changing cooperative relationships have formed among various enterprises, universities, research institutes, and technological elements. Currently, analytical techniques for industry-university-research collaborative networks have developed rapidly. Research paradigms have gradually shifted from traditional static statistical analysis to complex network modeling, time-series evolution analysis, and intelligent data mining. Related technologies are widely used in practical scenarios such as science and technology policy formulation, innovation resource allocation, cooperation risk assessment, and industrial evolution trend prediction, possessing significant theoretical and engineering application value. For example, Liu Guowei (2014). Research on the Evolution of Industry-University-Research Collaborative Innovation Networks (Doctoral-dissertation, Harbin Engineering University). He points out that "the networked relationship between industry, academia, and research promotes resource renewal and aggregation, and is the driving force behind the evolution of innovation networks."

[0003] Given the industry's application needs and technological evolution trends, how to accurately construct a time-series dynamic industry-academia-research network, achieve in-depth mining of multi-scale topological features, simulate the interest game behavior of innovation entities, scientifically divide the network life cycle evolution stages, and further complete multi-scenario simulation and interpretable decision generation through causal inference has become a key technical challenge that urgently needs to be solved in this field.

[0004] Chinese patent (publication number CN112364141A) discloses a document association mining method based on graph neural networks. By constructing a document association network and using graph neural networks to extract topological features, the method can realize the mining of the relationship between innovative subjects. However, this method is only applicable to static network modeling and lacks a regular incremental update mechanism, making it difficult to adapt to the needs of dynamic evolution analysis of industry-university-research networks.

[0005] Chinese patent (publication number CN117112784A) discloses a method for document identification and technology path evolution. This method constructs a network through document association and combines community segmentation and patent indicators to achieve preliminary segmentation of the technology life cycle. However, this method is based on document text features and traditional statistical indicators to segment stages, without integrating deep features of network topology, lacking the ability to couple and model with temporal deep learning, and lacking the functions of subject interest game simulation and causal counterfactual reasoning.

[0006] Chinese patent (publication number CN117010119A) discloses a network evolution method and system based on Markov chains. This method simulates the temporal evolution of network structure through Markov chain state transitions. However, this method is only applicable to the evolution of simple homogeneous networks and cannot simulate the cooperation and default game behavior of multiple types of entities in industry, academia and research, nor does it support multi-scenario simulation decision output.

[0007] In summary, existing technologies and academic research cannot fully characterize the intrinsic mechanisms of the entire process of industry-academia-research networks, from structural evolution and subject competition to stage transitions and simulation decision-making. There is an urgent need to propose a new integrated, refined, and interpretable method for dynamic evolution analysis of industry-academia-research networks. Summary of the Invention

[0008] To address the aforementioned technical issues, this application discloses a method for dynamic evolution analysis of industry-academia-research networks based on lifecycle and topology, specifically including:

[0009] The system acquires entities and relationships from industry-academia-research data, constructs a time-series industry-academia-research collaboration network in the form of a dynamic heterogeneous knowledge graph, and performs incremental updates regularly to obtain a continuous time-series snapshot sequence of the knowledge graph.

[0010] Based on the time-series snapshot sequence of the knowledge graph, a multi-scale topological representation is performed by fusing topological indicators and deep features to obtain a multi-scale topological representation result, which includes at least a comprehensive topological feature vector.

[0011] An enhanced benefit-driven evolutionary game model is constructed. The benefit-driven evolutionary game model uses a dynamic heterogeneous knowledge graph as the network environment, is based on PPO deep reinforcement learning, takes a comprehensive topological feature vector as input, and outputs the probability distribution of cooperation and default of the graph network nodes.

[0012] A temporal Transformer-VAE coupled lifecycle stage partitioning model is constructed. The partitioning model is based on multi-scale topological representation results and the running results of evolutionary game model, outputting the probability distribution of lifecycle stages corresponding to each time step, and detecting stage transition points in real time.

[0013] Based on the stage transition points, the DoWhy causal inference library is used to construct the causal graph of the industry-academia-research network evolution. Counterfactual reasoning is performed on the current stage and the upcoming stage transition points to obtain the results of multi-scenario evolution simulation and interpretable decision generation.

[0014] Preferably, the time-series industry-academia-research collaboration network specifically includes: entity types including enterprises, universities, research institutes, and technologies; and relationship types including collaboration applications, collaborative research and development, technology transfer, talent mobility, joint project undertaking, joint publication of papers, upstream and downstream linkages, and policy support.

[0015] Calculate dynamic edge weights based on time decay factor and relation type;

[0016] By using entity types as nodes and relationship types as edges, and combining dynamic edge weights, a collaborative network for industry, academia, and research can be constructed.

[0017] By statistically analyzing regularly updated industry-academia-research collaboration network data, a time-series snapshot sequence of the knowledge graph is obtained.

[0018] Preferably, the multi-scale topological representation specifically includes topological indices, node depth features, comprehensive topological features, network topology type, and community evolution entropy. The topological indices are obtained by calculating the basic topological indices for each time-series snapshot, including degree distribution power law exponent, clustering coefficient, average path length, intermediate central potential, and effective size of structural holes.

[0019] The node depth features are extracted from each time-series snapshot using the dynamic GraphSAGE algorithm. The model input is node attributes and adjacency matrix, and the output is node depth feature vector.

[0020] The comprehensive topological feature vector is obtained by fusing basic topological indices and node-level deep topological features through an attention mechanism.

[0021] Based on the global comprehensive topological feature vector, network topology types are identified, and the dynamic Louvain algorithm is used to discover communities and calculate community evolution entropy.

[0022] Preferably, the enhanced benefit-driven evolutionary game model is as follows: using a dynamic heterogeneous knowledge graph as the network operating environment, each node in the network is defined to have two strategy choices, namely a cooperative strategy and a default strategy. Based on its own state, the state of its neighbors and the global network state, the node uses the Proximal Policy Optimization (PPO) algorithm to simulate the decision-making process of bounded rational nodes and selects the corresponding strategy to maximize its own benefit.

[0023] Preferably, the near-end strategy optimization PPO algorithm specifically includes the PPO subject decision-making and dynamic reconnection process with partners;

[0024] Among them, the PPO subject decision takes the node's own comprehensive topological feature vector, the comprehensive topological feature vector of neighboring nodes and the global network state features as input, and outputs the probability distribution of choosing cooperation or default strategy through a multi-layer neural network to dynamically learn and optimize the strategy.

[0025] The dynamic reconnection mechanism targets low-revenue and low-reputation neighbors. Nodes disconnect with a fixed probability and use an attention-weighted preference connection method. The selection probability is calculated by combining the similarity of the comprehensive topological feature vectors between nodes, the historical average revenue and reputation value of candidate partners, and then a new partner is selected.

[0026] Preferably, the reward function of the near-end strategy optimization PPO algorithm is specifically: taking into account factors such as cooperation benefits, other party cooperation benefits, breach of contract penalties, opportunity benefits and reputation rewards, and dynamically adjusting the benefit distribution ratio in combination with the betweenness centrality coefficient in the comprehensive topological feature vector of the node.

[0027] Preferably, the Transformer-VAE coupled lifecycle stage division model is as follows:

[0028] Based on the results of multi-scale topological characterization, traditional structural entropy and deep feature entropy are calculated; the Gini coefficient of interest distribution, network default risk index, and average subject satisfaction are extracted from the results of evolutionary game model operation.

[0029] The structural orderliness and interest coordination are calculated separately to obtain a two-dimensional time series index sequence.

[0030] The two-dimensional index is encoded and compressed by variational autoencoder (VAE), the latent space weights of the two types of indexes are learned, and the one-dimensional comprehensive evolution index is generated by adaptive fusion.

[0031] The integrated evolutionary index sequence and the integrated topological feature vector sequence are input into the temporal fusion Transformer. The multi-head self-attention mechanism is used to capture long-term temporal dependencies and output the probability distribution of the network belonging to the five life cycle stages at the current moment.

[0032] Based on a preset judgment threshold, the system continuously monitors changes in stage probability, captures stage transition points in real time, and saves the stage category, stage probability, and stage transition time for each time step to obtain a complete life cycle evolution time series result.

[0033] Preferably, the variational autoencoder (VAE) specifically involves: processing each time node... Corresponding structural order Coordination with interests The input vector is concatenated to obtain a two-dimensional input vector. A multi-layer fully connected neural network encoder maps the two-dimensional input vector to a low-dimensional latent space, outputting the mean and variance vectors of the latent variables. A reparameterization technique is then used for random sampling to obtain the latent space vector. The latent space vector is then decoded. Map back to the original input dimension and output the reconstructed vector. ,in These are the reconstructed values ​​for structural orderliness and interest coordination, respectively.

[0034] Preferably, the temporal fusion Transformer specifically includes an input layer, a temporal embedding layer, a multi-head self-attention layer, a temporal fusion layer, a temporal decoder, and an output layer;

[0035] The temporal embedding layer maps the preprocessed dual temporal sequences into high-dimensional temporal embedding vectors, and introduces positional encoding to obtain dual temporal embedding sequences.

[0036] The dual temporal embedding sequences are concatenated and input into a multi-head self-attention layer. Multiple attention heads are used for parallel computation, and the attention-weighted temporal feature sequence is output.

[0037] The temporal fusion layer uses residual connections and layer normalization operations to fuse and optimize the attention-weighted temporal features to obtain a temporal fusion feature sequence.

[0038] The temporal decoder uses a multi-layer fully connected neural network and a Softmax activation function to calculate the probability distribution of the five lifecycle stages corresponding to each time step.

[0039] Preferably, the counterfactual reasoning for the current stage and the upcoming stage transition point specifically involves: using the graph structure construction module of the DoWhy causal inference library, taking the stage transition point as the time node, using the selected core variables as causal graph nodes, and combining the evolution mechanism of the industry-academia-research network to define the direct causal path and indirect causal path between variables, and generating a standardized industry-academia-research network evolution causal graph.

[0040] By using the backdoor adjustment method in the DoWhy library, the interference of confounding variables on causal effects is blocked. Based on the propensity score matching (PSM) algorithm, the propensity score of each sample is calculated.

[0041] By matching samples with and without intervention, the average causal effect ATE of the cause variable on the outcome variable is estimated, and the causal effect estimation results are obtained.

[0042] Focusing on the current life cycle stage and the upcoming stage transition point, we set up counterfactual scenarios, and based on the causal graph and causal effect estimation results, we simulated the time nodes of stage transition, network topology evolution trends, changes in node benefit distribution and cooperation stability under different counterfactual scenarios, and output counterfactual inference results.

[0043] Based on the counterfactual reasoning results and the incremental update characteristics of dynamic heterogeneous knowledge graphs, three core scenarios are set up: baseline scenario, optimization scenario, and risk scenario. With a monthly time step, the entire process of topological evolution, node strategy selection, and life cycle stage transition of the industry-academia-research network under each scenario is simulated to obtain multi-scenario evolution simulation results.

[0044] Compared with the prior art, the technical solution of this application has the following technical effects:

[0045] This invention constructs a dynamic heterogeneous knowledge graph and updates it incrementally on a monthly cycle, generating a continuous knowledge graph time-series snapshot sequence. It breaks through the limitations of existing technologies that can only construct static networks and require full reconstruction and updating. It can accurately adapt to the characteristics of the dynamic evolution of industry-university-research cooperation relationships over time, and greatly improve the efficiency and temporal continuity of network construction and data updates.

[0046] This invention realizes multi-scale topological representation of basic topological indicators and dynamic GraphSAGE deep features, and combines the dynamic Louvain algorithm to complete community discovery and community evolution entropy calculation, which can characterize network topological characteristics from multiple levels of macro structure, node micro features and community evolution.

[0047] This invention constructs an enhanced interest-driven evolutionary game model based on PPO deep reinforcement learning, integrates multi-dimensional payoff factors to construct a reward function, and simultaneously realizes the decision-making of principal cooperation default strategy and dynamic reconnection of partners, which is more in line with the behavioral decision-making mechanism of real industry-academia-research entities.

[0048] This invention designs a Transformer-VAE coupled lifecycle stage partitioning model. By adaptively fusing two time-series indicators, namely structural orderliness and interest coordination, the VAE then uses a time-series fusion Transformer to capture long-term time-series dependencies, automatically identifying lifecycle stages and stage transition points. It can accurately output the probability distribution of each time step belonging to different lifecycle stages, realizing intelligent identification of lifecycle stages and automatic real-time capture of stage transition points.

[0049] This invention introduces the DoWhy causal inference library to construct an evolutionary causal graph, and combines propensity score matching and counterfactual reasoning to conduct multi-scenario evolutionary simulations, outputting interpretable decision-making suggestions. It provides interpretable and implementable intelligent support for industry-academia-research network policy formulation, collaborative regulation, and risk warning.

[0050] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0051] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0053] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0054] Figure 1 This is a flowchart illustrating the overall process of analyzing the dynamic evolution of industry-academia-research networks based on lifecycle and topology.

[0055] Figure 2 This is an architecture diagram of a dynamic evolution analysis method for industry-academia-research networks based on lifecycle and topology.

[0056] Figure 3 This is a diagram illustrating the architecture of the enhanced interest-driven evolutionary game model in this application.

[0057] Figure 4 An architecture diagram of the lifecycle phase division model of the Transformer-VAE coupling in this application;

[0058] Figure 5 This is a schematic diagram of the evolution of the industry-academia-research network in the first eight years of this application embodiment;

[0059] Figure 6 As can be seen from the comparison chart of the evolution of industry-university-research network data and actual measurements in the last 5 years of this application embodiment;

[0060] Figure 7 This is a comparison chart of multi-scenario evolution simulation data results in the embodiments of this application;

[0061] Figure 8 This is a comparison chart of multi-scenario evolution simulation data results under different adjustment intensities in the embodiments of this application;

[0062] Figure 9 This is a comparison chart of the evolution results of various methods in the embodiments of this application;

[0063] Figure 10 This is a comparison chart of the evolution accuracy data of each method in the embodiments of this application over the last 5 years. Detailed Implementation

[0064] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0065] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0066] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0067] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0068] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0069] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0070] Example 1 mainly describes a method for dynamic evolution analysis of industry-academia-research networks based on lifecycle and topology, such as... Figure 1 , Figure 2 As shown, it specifically includes:

[0071] The system acquires entities and relationships from industry-academia-research data, constructs a time-series industry-academia-research collaboration network in the form of a dynamic heterogeneous knowledge graph, and performs incremental updates regularly to obtain a continuous time-series snapshot sequence of the knowledge graph.

[0072] Based on the time-series snapshot sequence of the knowledge graph, a multi-scale topological representation is performed by fusing topological indicators and deep features to obtain a multi-scale topological representation result, which includes at least a comprehensive topological feature vector.

[0073] An enhanced benefit-driven evolutionary game model is constructed. The benefit-driven evolutionary game model uses a dynamic heterogeneous knowledge graph as the network environment, is based on PPO deep reinforcement learning, takes a comprehensive topological feature vector as input, and outputs the probability distribution of cooperation and default of the graph network nodes.

[0074] A temporal Transformer-VAE coupled lifecycle stage partitioning model is constructed. The partitioning model is based on multi-scale topological representation results and the running results of evolutionary game model, outputting the probability distribution of lifecycle stages corresponding to each time step, and detecting stage transition points in real time.

[0075] Based on the stage transition points, the DoWhy causal inference library is used to construct the causal graph of the industry-academia-research network evolution. Counterfactual reasoning is performed on the current stage and the upcoming stage transition points to obtain the results of multi-scenario evolution simulation and interpretable decision generation.

[0076] Furthermore, the time-series industry-academia-research collaboration network specifically includes entities such as enterprises, universities, research institutes, and technology companies.

[0077] Relationship types include collaborative applications, collaborative research and development, technology transfer, talent mobility, joint project undertaking, joint publication of papers, upstream and downstream linkages, and policy support;

[0078] Data sources include patents, academic papers, research projects, technology transactions, news reports, and government announcements;

[0079] The dynamic edge weight is calculated based on the time decay factor and relation type, using the following formula:

[0080]

[0081] in, For the first Annual Entity With entity Dynamic edge weights between them The weights for relation types determined by the entropy weight method. For time window, The time decay coefficient, For the first Annual Entity With entity Between The number of class relations, This represents the total number of relations.

[0082] By using entity types as nodes and relationship types as edges, and combining dynamic edge weights, a collaborative network for industry, academia, and research can be constructed.

[0083] By statistically analyzing regularly updated industry-academia-research collaboration network data, a time-series snapshot sequence of the knowledge graph is obtained.

[0084] Furthermore, the regular incremental updates are specifically as follows:

[0085] Collect multi-source heterogeneous data related to industry, academia, and research on a monthly basis, including patents, academic papers, research projects, technology transactions, news reports, and government announcements.

[0086] The newly collected data is cleaned, deduplicated, and standardized to extract the entities (enterprises, universities, research institutes, and technologies) and the relationships between entities (eight types of relationships, including collaborative applications and collaborative research and development).

[0087] Based on new and historical data, the dynamic edge weights between entities are calculated to reflect the changes in the intensity of cooperation at different time points.

[0088] The pre-processed new entities, relations and the calculated dynamic edge weights are incrementally updated to the original heterogeneous knowledge graph to avoid full reconstruction and improve update efficiency.

[0089] After each monthly update, a knowledge graph snapshot of the current time node is generated, resulting in a continuous sequence of knowledge graph time-series snapshots, which provides a foundation for subsequent steps such as multi-scale topological representation and evolutionary game theory.

[0090] Furthermore, multi-scale topological representation specifically includes topological indices, node depth features, comprehensive topological features, network topology types, and community evolution entropy. Among them, the topological indices are obtained by calculating the basic topological indices of each time-series snapshot, including the degree distribution power law exponent, clustering coefficient, average path length, intermediate central potential, and effective size of structural holes.

[0091] The node depth features are extracted from each time-series snapshot using the dynamic GraphSAGE algorithm. The model input is node attributes and adjacency matrix, and the output is node depth feature vector.

[0092] The comprehensive topological feature vector is obtained by fusing basic topological indices and node-level deep topological features through an attention mechanism, as shown in the formula:

[0093]

[0094] in, For nodes The comprehensive topological feature vector, For nodes The basic topological index vector, For nodes The deep topological feature vector, , For learnable weight matrix, It is the sigmoid activation function;

[0095] Based on the global comprehensive topological feature vector, network topology types (scale-free / small-world / random / regular) are identified, and the dynamic Louvain algorithm is used to discover communities and calculate community evolution entropy.

[0096] Furthermore, the dynamic GraphSAGE algorithm is as follows: for each target node, a fixed number of neighboring nodes are randomly sampled from its adjacent nodes (the sampling number is set to 20) to construct the local neighborhood structure of the target node;

[0097] The mean aggregation strategy is adopted to aggregate the target node's own attribute features with the sampled neighbor node features to capture the local topological association information of the node. The formula is as follows:

[0098]

[0099] in, For the target node In the Feature vectors after layer aggregation For the first Learnable weight matrix for layer aggregation For the target node In the Feature vectors after layer aggregation For the set of adjacent nodes, To calculate the average, It is the ReLU activation function;

[0100] By using linear transformations and activation functions, nonlinear mapping is applied to the aggregated features to enhance their expressive power and output the preliminary deep features of the nodes.

[0101] For each knowledge graph time-series snapshot, repeat the above steps, update the node depth features based on the adjacency relationship and node attributes of the current snapshot, and adapt to the dynamic changes in network topology.

[0102] The iterative node depth features are normalized to 128 dimensions to obtain node-level deep topological feature vectors, which are used for subsequent fusion of comprehensive topological feature vectors.

[0103] Furthermore, the dynamic Louvain algorithm is used to discover communities and calculate community evolution entropy. Specifically, for the current knowledge graph time-series snapshot, each node is regarded as an independent community. At this time, the number of communities is the same as the number of nodes, and the initial modularity is calculated.

[0104] Iterate through each node and try to move the node to the community of its adjacent node. Calculate the change in modularity after the move. If the change is positive, keep the move operation; otherwise, cancel the move. Continue until the modularity no longer increases after all nodes have been moved.

[0105] Treat each community obtained in the previous step as a new super node, construct a new community adjacency matrix (the weight of the edge between communities is the sum of the weights of the edges between nodes within the original community), and repeat the local optimization process until the modularity reaches its maximum value, thus obtaining the optimal community partitioning result for the current snapshot. The formula is:

[0106]

[0107] in, For modularity, This represents the sum of the weights of all edges in the current knowledge graph snapshot. For nodes With nodes Edge weights between them They are nodes With nodes The weighting degree, For indicator functions, if node With nodes Belonging to the same club =1, otherwise 0 , They are nodes With nodes The club number to which it belongs;

[0108] For continuous knowledge graph time-series snapshots, compare the community division results of two adjacent snapshots, identify four evolutionary behaviors of communities: merging, splitting, disappearing, and emerging, and record the changes in the node composition of each community;

[0109] Based on the community division results of each snapshot and the community evolution behavior of adjacent snapshots, the community evolution entropy is calculated to quantify the orderliness and evolutionary dynamics of the community structure. The formula is as follows:

[0110]

[0111] in, Time node The entropy of community evolution, Time node At that time, the total number of online communities Time node At that time, the first The proportion of nodes in each community to the total number of nodes in the network.

[0112] Furthermore, such as Figure 3 The diagram shows the architecture of the enhanced interest-driven evolutionary game model. Specifically, the enhanced interest-driven evolutionary game model uses a dynamic heterogeneous knowledge graph as the network operating environment. Each node in the network is defined to have two strategy choices: cooperation strategy and default strategy. Based on its own state, the state of its neighbors, and the global network state, the node uses the Proximal Policy Optimization (PPO) algorithm to simulate the decision-making process of bounded rational nodes and chooses the corresponding strategy to maximize its own benefits.

[0113] Furthermore, the near-end strategy optimizes the PPO algorithm, specifically including the PPO subject decision-making and the dynamic reconnection process with partners;

[0114] Among them, the PPO subject decision takes the node's own comprehensive topological feature vector, the comprehensive topological feature vector of neighboring nodes and the global network state features as input, and outputs the probability distribution of choosing cooperation or default strategy through a multi-layer neural network to dynamically learn and optimize the strategy.

[0115] Dynamic partner reconnection targets low-reward and low-reputation neighbors. Nodes disconnect with a fixed probability and employ an attention-weighted preference-based connection method. The selection probability is calculated by combining the similarity of the comprehensive topological feature vectors between nodes, the historical average revenue of candidate partners, and their reputation values. The selection probability formula is as follows:

[0116]

[0117] in, Time node At that time, node Select node The probability of becoming a new partner For feature similarity weights, , , They are nodes , , The comprehensive topological feature vector, , Time node At that time, node , Historical average returns This is the historical return preference coefficient. , Time node At that time, node , Reputation value Reputation preference coefficient This is the set of all nodes in the current knowledge graph snapshot.

[0118] Furthermore, the enhanced benefit-driven evolutionary game model includes a dynamic partner reconnection mechanism, specifically:

[0119] Each node in the network disconnects from its low-reward, low-reputation neighbor nodes with a fixed probability of 0.1.

[0120] After a node disconnects from its original connection, the selection of a new partner adopts an attention-weighted preference connection method. The selection probability comprehensively considers three factors: the similarity of the comprehensive topological feature vectors between nodes, the historical average revenue of the candidate partner, and the reputation value of the candidate partner. By weighting the above factors, the selection probability of the new connection is determined, ensuring that the selection of the new partner has both structural adaptability and revenue stability.

[0121] Furthermore, the reward function of the near-end strategy optimization PPO algorithm is as follows: It comprehensively considers factors such as cooperation benefits, counterparty cooperation benefits, breach of contract penalties, opportunity benefits, and reputational rewards, and dynamically adjusts the benefit distribution ratio based on the betweenness centrality coefficient in the node's comprehensive topological feature vector. The formula is:

[0122]

[0123] in, Time node At that time, node The dynamic reward value, For the distribution coefficient of benefits, Time node At that time, the total cooperative revenue of all nodes in the network, , Time node At that time, node , The betweenness centrality coefficient, For breach of contract compensation, For the default indication function (when a node selects a default strategy) =1, or 0 when choosing a cooperation strategy). For the sake of the unit's opportunity and interests, Reputation reward weight, Time node At that time, node Reputation value.

[0124] Furthermore, the reward function comprehensively considers the cooperative benefits, the other party's cooperative benefits, the penalty for breach of contract, the opportunity benefits, and the reputational rewards; among them, the cooperative benefits and the other party's cooperative benefits are dynamically adjusted based on the total cooperative benefits at the current moment, the preset benefit distribution coefficient, and the betweenness centrality coefficient of the node;

[0125] The penalty for breach of contract is linked to the betweenness centrality coefficient of the node and the breaching behavior, and the penalty for breach of contract will only occur when the node chooses the breaching strategy; the opportunity benefit can only be obtained when the node chooses the breaching strategy; the reputation reward is determined based on the node's historical reputation value and the preset reputation reward weight.

[0126] The betweenness centrality coefficient of a node is calculated by the ratio of the node's own betweenness to the largest betweenness in the network; the reputation value of a node is calculated by the ratio of the number of times the node has cooperated in history to the total number of times it has played games in history.

[0127] Furthermore, the enhanced interest-driven evolutionary game model is built on a hierarchical structure based on the PPO policy network, which is divided into a state input layer, a feature mapping layer, a policy hiding layer, and a policy output layer.

[0128] The state input layer accesses node topology features, neighborhood association features, and global network state features to uniformly align data dimensions.

[0129] The feature mapping layer is set to a fully connected structure with 128 neurons, and the ReLU activation function is used to complete the feature transformation.

[0130] The policy hidden layer is set up with two fully connected layers. The first layer has 64 neurons and the second layer has 32 neurons, both of which use the ReLU activation function.

[0131] The strategy output layer has a 2-dimensional output dimension, is configured with a Softmax activation function, and outputs the selection probabilities corresponding to the cooperation strategy and the default strategy.

[0132] Furthermore, such as Figure 4 The diagram shows the architecture of the Transformer-VAE coupled lifecycle phase division model. The specific details of the Transformer-VAE coupled lifecycle phase division model are as follows:

[0133] Based on the results of multi-scale topological characterization, traditional structural entropy and deep feature entropy are calculated; the Gini coefficient of interest distribution, network default risk index, and average subject satisfaction are extracted from the results of evolutionary game model operation.

[0134] The structural orderliness and interest coordination are calculated separately to obtain a two-dimensional time-series index sequence, which characterizes the evolutionary law of the industry-academia-research network structure and the state of interest game. The formula is as follows:

[0135]

[0136]

[0137] in, , These are respectively structural orderliness and interest coordination. , , These are, respectively, the traditional structural entropy, the deep feature entropy, and the network maximum entropy; The Gini coefficient is used to allocate benefits. This is the network average default risk index obtained from evolutionary game theory statistics. The average satisfaction level of the main participants in the cooperation;

[0138] The two-dimensional index is encoded and compressed by variational autoencoder (VAE), the latent space weights of the two types of indexes are learned, and the one-dimensional comprehensive evolution index is generated by adaptive fusion.

[0139] The comprehensive evolution index sequence and comprehensive topological feature vector sequence are input into the temporal fusion Transformer. The multi-head self-attention mechanism is used to capture long-term temporal dependencies, explore the slow evolution trend and mutation characteristics of the network, and output the probability distribution of the network belonging to the five life cycle stages at the current time.

[0140] Based on a preset judgment threshold, the system continuously monitors changes in stage probability, captures stage transition points in real time, and saves the stage category, stage probability, and stage transition time for each time step to obtain a complete life cycle evolution time series result.

[0141] Furthermore, the traditional structural entropy is calculated using basic topological indices and community evolution entropy, and the depth feature entropy is calculated using the distribution of comprehensive topological feature vectors. The formula is as follows:

[0142]

[0143]

[0144] in, Time node Traditional structural entropy under the following conditions , These are the weights of the basic topological index entropy and the community evolution entropy, respectively. Time node The entropy of the basic topological index (which is calculated by comprehensively using the entropy weight method from five basic topological indexes: degree distribution power law exponent, clustering coefficient, average path length, central potential, and effective size of structural holes). Time node The entropy of community evolution; For deep feature entropy, Time node At that time, the total number of nodes in the knowledge graph network. Time node At that time, node The comprehensive topological feature vector of the first Normalized probabilities of eigenvalues. This represents the dimension of the integrated topological feature vector.

[0145] Furthermore, the variational autoencoder (VAE) specifically involves: dividing each time node... Corresponding structural order Coordination with interests The input vector is concatenated to obtain a two-dimensional input vector. A multi-layer fully connected neural network encoder maps the two-dimensional input vector to a low-dimensional latent space, outputting the mean and variance vectors of the latent variables. Based on the mean and variance of the latent variables, a reparameterization technique is used for random sampling to obtain the latent space vector. The latent space vector is then decoded. Map back to the original input dimension and output the reconstructed vector. ,in These are the reconstructed values ​​for structural orderliness and interest coordination, respectively.

[0146] Furthermore, the training process of the variational autoencoder (VAE) is as follows: A loss function is constructed that includes reconstruction loss and KL divergence, as shown in the formula:

[0147]

[0148]

[0149]

[0150] in, Let be the loss function of the variational autoencoder (VAE). To reconstruct the loss, Let KL divergence be the KL divergence. Time node Two-dimensional input vector, Time node The reconstructed vector below, This represents the total number of time-series snapshots. Time node The mean vector of the lower latent variables. The trace of the matrix, Time node The variance vector of the hidden variables. Let be the determinant of the matrix;

[0151] The parameters of the encoder and decoder are iteratively optimized through the backpropagation algorithm, while the fusion weight of structural orderliness and interest coordination is learned.

[0152] Once the model training converges (the loss function stabilizes), the weights learned in the latent space are extracted, and the structural orderliness is determined. Coordination with interests The weighted fusion yields a one-dimensional comprehensive evolution index, which is then output to the subsequent Transformer model for temporal feature mining.

[0153] Furthermore, the variational autoencoder is a symmetric encoding and decoding hierarchical structure, which includes an input layer, an encoding layer, a latent variable layer, a decoding layer, and a reconstructed output layer;

[0154] The input layer has a 2-dimensional dimension.

[0155] The encoding layer consists of two fully connected hidden layers. The first hidden layer has 64 neurons, and the second hidden layer has 32 neurons. Both layers are configured with the ReLU activation function. The encoding layer ends with two output branches: the mean branch and the variance branch. Both branches have an output dimension of 16, and the overall hidden space has a dimension of 16.

[0156] The hidden variable layer has 16 dimensions;

[0157] The decoding layer consists of two fully connected hidden layers. The first hidden layer has 32 neurons, and the second hidden layer has 64 neurons. Both layers use the ReLU activation function.

[0158] The reconstructed output layer is 2-dimensional and uses the Sigmoid activation function.

[0159] Furthermore, the temporal embedding layer of the temporal fusion Transformer maps the preprocessed dual temporal sequences into high-dimensional temporal embedding vectors, while introducing positional encoding to obtain the temporal embedding sequence. and ;

[0160] The dual temporal embedding sequences are concatenated and input into a multi-head self-attention layer. Multiple attention heads are used for parallel computation, and the attention-weighted temporal feature sequence is output.

[0161] The temporal fusion layer uses residual connections and layer normalization operations to fuse and optimize the attention-weighted temporal features to obtain a temporal fusion feature sequence.

[0162] The temporal decoder uses a multi-layer fully connected neural network and a Softmax activation function to calculate the probability distribution of the five lifecycle stages corresponding to each time step.

[0163] Furthermore, the input layer of the temporal fusion Transformer is fed with two types of temporal feature sequences to uniformly complete the dimension regularization;

[0164] The temporal embedding layer outputs a 128-dimensional dimension and has a built-in positional encoding structure, with the positional encoding dimension being the same as the embedding dimension.

[0165] The multi-head self-attention layer has 8 attention heads and 16-dimensional single-head features. Residual connections and layer normalization structures are set within the layer.

[0166] The temporal fusion layer performs multi-source temporal feature concatenation and fusion while maintaining the feature output dimension unchanged. The temporal decoder has a built-in two-layer fully connected network, with 256 neurons in the first layer and 128 neurons in the second layer, both using the ReLU activation function;

[0167] The output layer has a 5-dimensional output and is equipped with the Softmax activation function.

[0168] Furthermore, the time series preprocessing involves: processing the composite evolution index time series... (dimension is) (and the time series of integrated topological feature vectors) (dimension is) , Normalization is performed to eliminate dimensional differences and ensure the consistency of the input sequence.

[0169] Furthermore, for incremental updates of knowledge graph time-series snapshots, the input time-series sequence is updated in real time. The time-series fusion Transformer adapts to the dynamic characteristics of network evolution through an online fine-tuning mechanism, ensuring the timeliness and accuracy of stage predictions.

[0170] Furthermore, counterfactual reasoning is conducted for the current stage and the upcoming stage transition point. Specifically, using the graph structure construction module of the DoWhy causal inference library, the core variables selected are used as nodes in the causal graph, with the stage transition point as the time node. Combined with the evolution mechanism of the industry-academia-research network (topology evolution law, node interest game behavior, life cycle stage transition logic), the direct and indirect causal paths between variables are defined to generate a standardized causal graph of the industry-academia-research network evolution.

[0171] By using the backdoor adjustment method from the DoWhy library, the interference of confounding variables on causal effects is blocked. Based on the propensity score matching (PSM) algorithm, the propensity score for each sample is calculated using the following formula:

[0172]

[0173] in To favor scoring, For the set of covariates, This is a learnable parameter vector for a propensity score model (logistic regression);

[0174] Samples with and without intervention (causal variable value of 1) are matched to estimate the average causal effect ATE of the causal variables on the outcome variables, quantifying the influence of each causal variable on stage transitions and network evolution states. The formula is as follows:

[0175]

[0176] in, for, Let be the expected function. Intervention variables When intervention is applied, the values ​​of the outcome variables are... Intervention variables The values ​​of the outcome variables when no intervention is applied;

[0177] Focusing on the current life cycle stage and the upcoming stage transition point, multiple counterfactual scenarios are set up. Based on the constructed causal graph and causal effect estimation results, the time nodes of stage transition, network topology evolution trends, changes in node benefit distribution and cooperation stability under different counterfactual scenarios are simulated, and counterfactual inference results are output.

[0178] Based on the counterfactual reasoning results and the incremental update characteristics of dynamic heterogeneous knowledge graphs, three core scenarios are set up: baseline scenario, optimization scenario, and risk scenario. With a monthly time step, the entire process of topological evolution, node strategy selection, and life cycle stage transition of the industry-academia-research network under each scenario is simulated to obtain multi-scenario evolution simulation results.

[0179] This embodiment details a method for dynamic evolution analysis of industry-academia-research networks based on lifecycle and topology. Based on various entities and collaborative relationships within the industry-academia-research network, dynamic edge weights are calculated using time decay, a dynamic heterogeneous knowledge graph is constructed and periodically incrementally updated to obtain a time-series snapshot sequence. Based on this time-series snapshot sequence, multi-scale topological representation is performed. An enhanced benefit-driven evolutionary game model based on the PPO algorithm is built, and based on topological features, collaborative and default strategy decisions and dynamic adjustments to collaborative relationships among industry-academia-research entities are made. A Transformer-VAE coupled model is constructed, and a comprehensive evolutionary index is obtained through adaptive VAE fusion. The lifecycle stages are divided and stage transition points are identified via a time-series fusion Transformer. An evolutionary causal graph is built based on the DoWhy causal inference library, and multi-scenario evolution simulation is completed using counterfactual reasoning.

[0180] Example 2, based on Example 1, details the process of using the historical evolution data of the industry-university-research cooperation network in the equipment manufacturing industry as an empirical research object, and conducting life cycle evolution analysis of the industry-university-research topology using this method, as follows:

[0181] Data related to the industry, including joint invention patents from the State Intellectual Property Office, collaborative research projects, technology transfer, and talent mobility, were collected, covering three core innovation entities: enterprises, universities, and research institutes. Thirteen continuous time-series knowledge graph snapshots were constructed, with the first eight years' data used to build networks, learn topological features, train game strategies, and learn life cycle patterns. The last five years' data were used for model evolution prediction, stage transition identification, counterfactual reasoning, and multi-scenario simulation verification.

[0182] The network comprises 217 entities, including 152 equipment manufacturing companies, 41 science and engineering universities, and 24 professional research institutes. It has 436 effective cooperative edges, encompassing various relationships such as joint patent applications, collaborative project construction, technology transfer, and talent mobility. The network spans 13 annual snapshots, consistent with existing research timeframes, ensuring comparability. The average path length of the network's basic topology is 2.109, the clustering coefficient is 0.378, and the degree distribution conforms to scale-free characteristics.

[0183] Based on data from the first 8 years, a time decay factor was introduced to weight various types of cooperation relationships, and annual incremental updates and continuous time-series snapshots were performed to obtain the industry-academia-research network data for the first 8 years, as shown in Table 1.

[0184] Table 1. Industry-academia-research network data for the first 8 years

[0185] Year 1 17 12 3 2 34 Year 2 33 23 6 4 67 Year 3 50 35 9 6 101 Year 4 67 47 13 7 134 5th year 83 59 16 8 168 Year 6 100 70 19 11 201 7th year 117 82 22 13 235 8th year 134 94 25 15 268

[0186] According to Table 1 and as follows Figure 5The diagram shows the evolution of the industry-academia-research network over the first eight years. The network exhibits a pattern of increasing nodes and continuously strengthening cooperative relationships. Initially, it was dominated by scattered single-point cooperation, but later it gradually formed a core-periphery multi-core network structure, which is highly consistent with the scale-free, small-world evolution law proposed in existing literature.

[0187] Evolutionary testing was conducted based on data from the following five years. Multi-scale topological feature extraction was performed on time-series snapshots of each year, and basic topological indices were calculated, including degree distribution, clustering coefficient, average path length, central tendency, and effective size of structural holes. The dynamic GraphSAGE algorithm was used to extract node depth features. The basic indices and depth features were fused through an attention mechanism to obtain a comprehensive topological feature vector. The dynamic Louvain algorithm was used to divide communities and calculate community evolution entropy. The number of communities in each year was found to be 8-15, and the community evolution entropy was between 0.21 and 0.87.

[0188] An enhanced benefit-driven evolutionary game model is constructed based on PPO deep reinforcement learning. Node strategies include cooperation and default. The reward function integrates cooperation benefits, default penalties, opportunity benefits, and reputation rewards, and dynamically adjusts the benefit distribution ratio based on node betweenness centrality. Simultaneously, low-benefit and low-reputation neighbors are disconnected and reconnected, and new partners are selected by weighting feature similarity, historical benefits, and reputation value. The network's average annual cooperation rate is 75.3%, the average default risk index is 0.186, and the average Gini coefficient for benefit distribution is 0.317.

[0189] The structural orderliness is calculated based on topological representation results, and the interest coordination degree is calculated based on game theory output, forming a dual-time-series index sequence. Feature compression and adaptive weight fusion are performed using a VAE to obtain a one-dimensional comprehensive evolution index. This comprehensive index is input into a time-series fusion Transformer, which uses multi-head self-attention to capture long-term time-series dependencies, outputting the probability distributions of the three stages of formation, development, and maturity, and automatically identifying stage transition points. The comprehensive evolution accuracy is 93.42%. The evolution results for each year are compared with the measured results of historical data, resulting in the industry-academia-research network data for the last five years, as shown in Tables 2 and 3 below.

[0190] Table 2. Evolution data of the industry-academia-research network in the last 5 years.

[0191] 9th year 151 106 28 17 300 Year 10 165 116 31 18 338 11th year 181 127 36 18 366 12th year 194 136 36 22 395 13th year 208 145 39 24 418

[0192] Table 3 Historical measured data of industry-university-research network in the last 5 years.

[0193] 9th year 150 105 28 17 302 Year 10 167 117 32 18 335 11th year 184 129 35 20 369 12th year 200 140 38 22 402 13th year 217 152 41 24 436

[0194] According to Table 2, Table 3 and Figure 6The comparison chart of the evolution of industry-academia-research network data over the past 5 years shows that, with the increase of the evolution years, the number of network nodes and the number of cooperative edges have maintained a continuous growth trend. The predicted data and the historical measured data are highly consistent in the overall growth trend. The evolution results of the number of various nodes and topology structure indicate that the method of the present invention can accurately characterize the scale-free topological characteristics, main structure composition and dynamic evolution of cooperative relationships of industry-academia-research networks, and can provide reliable quantitative support for judging the long-term evolution trend of the network, optimizing resource allocation and managing the stability of cooperation.

[0195] Based on the model validated by the above tests, the transition points in years 7 and 11 were used as key nodes. A causal graph of the industry-academia-research network evolution was constructed using the DoWhy library. Confounding factors were eliminated through backdoor adjustments, and propensity score matching (PSM) was used to estimate the average causal effect. Three scenarios were set up: a baseline scenario, a policy optimization scenario, and a cooperation risk scenario. Multi-scenario evolution simulations were conducted with an annual step size, resulting in the multi-scenario evolution simulation data results shown in Table 4 below.

[0196] Table 4. Results of Multi-Scenario Evolution Simulation Data

[0197] Baseline Scenario 217 436 0.80 0.157 0.798 Policy optimization scenario 229 464 0.88 0.116 0.885 Cooperation Risk Scenario 197 396 0.68 0.281 0.617

[0198] Table 4 shows the policy optimization scenario, which refers to the ideal regulatory scenario in the process of industry-university-research cooperation, where the government increases policy support, strengthens guidance on interest distribution, increases penalties for breach of contract, and enhances cooperation incentives and subsidies, thereby improving the willingness of the main actors to cooperate, curbing opportunistic behavior, and promoting the orderly development of the network. The cooperation risk scenario refers to the risk evolution scenario in the industry-university-research network, where policy support weakens, interest distribution becomes unbalanced, the cost of breach of contract is low, and opportunistic behavior increases, leading to a decline in the willingness of the main actors to cooperate, unstable cooperative relationships, and a loose network structure. The intensity of both the policy optimization scenario and the cooperation risk scenario in Table 4 is 15%. This intensity indicates that under the policy optimization scenario, policy support, cooperation incentives, and breach of contract constraints are 15% stronger than in the baseline scenario; under the cooperation risk scenario, policy support and cooperation constraints are 15% weaker than in the baseline scenario.

[0199] According to Table 4 and Figure 7 As shown in the comparison chart of the multi-scenario evolution simulation data, under the policy optimization scenario, the overall network scale is larger, the connections are denser, the structure is more compact, the hub role of the core nodes is more prominent, the structural orderliness and cooperation stability are significantly improved, and the default risk is significantly reduced. This indicates that strengthening policy incentives, optimizing the distribution of benefits, and improving default constraints can effectively promote the network to evolve in a more efficient, orderly, and robust direction.

[0200] Under the cooperative risk scenario, the number of network nodes decreases, the number of cooperative edges drops significantly, the topology becomes more sparse and loose, the structural orderliness and cooperative stability are significantly reduced, and the risk of default increases sharply. This indicates that insufficient policy support, imbalance in the distribution of benefits, and weakened constraint mechanisms can significantly inhibit cooperative behavior, exacerbate opportunistic risks, and lead to a significant deterioration in the overall network evolution efficiency.

[0201] The evolution data under different regulation intensities were statistically analyzed, and the multi-scenario evolution simulation data results under different regulation intensities are shown in Tables 5 and 6 below:

[0202] Table 5. Multi-scenario evolution simulation data results under different policy optimization and adjustment intensities.

[0203] Adjusting intensity Total node size Effective number of cooperative edges Structural order Default Risk Index 0% 217 436 0.80 0.16 5% 222 449 0.83 0.14 10% 226 458 0.86 0.13 15% 229 463 0.88 0.12 20% 232 471 0.90 0.10 25% 234 477 0.92 0.09 30% 236 483 0.93 0.09 35% 237 486 0.94 0.08 40% 238 491 0.95 0.08

[0204] Table 6. Multi-scenario evolution simulation data results under different cooperation risk adjustment intensities.

[0205] Adjusting intensity Total node size Effective number of cooperative edges Structural order Default Risk Index 0% 217 436 0.80 0.16 5% 210 420 0.77 0.19 10% 203 407 0.74 0.22 15% 197 397 0.68 0.28 20% 189 378 0.62 0.34 25% 181 359 0.57 0.41 30% 174 343 0.52 0.47 35% 166 326 0.47 0.54 40% 159 308 0.43 0.62

[0206] According to Table 5, Table 6 and Figure 8 The comparison chart of multi-scenario evolution simulation data under different regulation intensities shows that, under the policy optimization scenario, as the regulation intensity gradually increases, the network node scale and the number of cooperative connections continue to expand, the structural orderliness steadily increases, the default risk continuously decreases, and the overall regulatory effect shows the characteristic of diminishing marginal benefits. The driving effect of high-intensity intervention on network quality and efficiency improvement gradually slows down. Under the cooperation risk scenario, the greater the impact intensity, the more obvious the network size contraction, the more severe the weakening of cooperative relationships, the continuous decline in structural regularity, the significant increase in default behavior of subjects, and the progressively worsening network operation stability.

[0207] This embodiment details an empirical study conducted using this method based on historical data of industry-academia-research collaboration in the equipment manufacturing industry. The method divides the time-series datasets before and after the model training and performance verification, respectively. Multiple algorithms are used to extract topological features, model evolutionary game theory, identify stages, and predict trends. The derived evolutionary data is compared with actual historical data, verifying that the overall evolutionary accuracy of the method reaches 93.42%. Causal inference and propensity score matching analysis are performed through identified stage transition nodes. Three simulation scenarios—benchmark, policy optimization, and cooperation risk—are set up, and simulation experiments are conducted with a uniform intensity of 15% to clarify the impact of different control modes on network structure and operational status. Furthermore, the control intensity range is broadened, and the fluctuation trends of various network indicators under intensity changes from 0% to 40% are quantitatively analyzed. The results conclude that policy incentives have diminishing marginal returns, and risk shocks will continuously weaken the stability of the cooperation system. This demonstrates the practical application capabilities of this method in characterizing network evolution, predicting development trends, and simulating control effects.

[0208] Example 3, based on Example 1 or 2, describes in detail the historical evolution data of the industry-academia-research network based on Example 2, and conducts implementation tests using several existing typical methods, including SNA (Social Network Analysis), BA (Scale-Free Evolution), ER (Random Network Evolution), GA (Genetic Algorithm Evolution), GNN (Graph Neural Network Evolution), and LPA (Label Propagation Evolution).

[0209] SNA (Social Network Analysis) collects real-world data on interactions, business collaborations, and information sharing among industry-academia-research entities to construct an initial network topology model. It calculates multiple topological characteristic indicators such as node degree centrality, betweenness centrality, and clustering coefficient, and dynamically adjusts the connection links between nodes based on the frequency of cooperation and the tightness of association among entities. The network structure is updated progressively according to a time series, quantitatively analyzing changes in group association patterns and predicting the overall network's evolutionary trend over time.

[0210] BA (Scale-Free Evolution) determines the initial set of basic nodes for the network and sets the node addition rate and connection adaptation rules. It continuously adds new nodes to the network, following a priority connection mechanism that allows new nodes to establish cooperative relationships with core nodes that have a larger number of connections. By continuously expanding the overall network size and simultaneously updating the node hierarchy and edge connection distribution, iteratively generating a dynamic scale-free evolutionary network with power-law distribution characteristics.

[0211] The ER (Random Evolution) method defines the range of the total number of nodes in the network and uniformly sets the probability of generating connections between any two nodes. It iterates through all node combinations, determines whether cooperative connections are established between nodes according to the preset probabilities, and generates internal network edges in batches. This connection generation process is repeated multiple times to continuously enrich the network connection structure, simulating the growth and evolution of the network structure under random cooperative conditions.

[0212] GA (Genetic Algorithm Evolutionary Approach) transforms the arrangement and combination of network nodes into coded individuals, forming an initial evolutionary population. Based on network structure fit evaluation criteria, fitness is assessed for each individual in the population, and selection, crossover, and mutation core operations are executed sequentially. Poorly performing combinations are eliminated, while high-performing network configurations are retained. The population is iteratively updated, and the topology layout is repeatedly optimized to obtain the optimal evolutionary network structure.

[0213] GNN (Graph Neural Network Evolution) extracts the attribute data of nodes themselves and the features associated with neighboring nodes to build a graph learning computation framework. It aggregates feature information from surrounding nodes layer by layer to uncover hidden association patterns and evolutionary logic within the topological structure. Relying on feature propagation and weight update mechanisms, it changes node state values ​​and connection weights in real time, simulating network morphology at different time points step by step, thus completing the temporal simulation calculation of network evolution.

[0214] LPA (Label Propagation Evolution) assigns initial feature labels to each node in the network and constructs information transmission paths based on the connections between nodes. Labels are synchronously diffused according to the similarity of labels between adjacent nodes, and the node classification is adjusted synchronously based on local community association characteristics. Multiple iterations update the label status of all nodes, reconstructing node cooperative connections based on changes in label distribution, and gradually promoting the synchronous evolution of community division and the overall network structure.

[0215] Each method was trained using historical data from the first eight years of the industry-academia-research network, and its evolution was tested based on data from the last five years. The evolution data of each method was statistically analyzed, and the evolution results of each method are shown in Table 7 below:

[0216] Table 7 Evolution results of each method

[0217] Method type Total node size Effective number of cooperative edges Structural order Default Risk Index Stability of cooperation SNA 196 385 0.65 0.25 0.64 BA 241 492 0.91 0.09 0.92 ER 185 361 0.58 0.32 0.53 GA 204 403 0.70 0.21 0.69 GNN 233 478 0.86 0.13 0.87 LPA 191 374 0.62 0.28 0.58 This method 208 418 0.74 0.18 0.77 Historical measured data 217 436 0.80 0.16 0.80

[0218] According to Table 7 and Figure 9 As shown in the comparison chart of the evolution results of each method, the network structure generated by the method in this paper is highly consistent with the historical measured network. The other comparative methods all have obvious structural sparsity, connection anomalies or scale deviations, which can effectively verify that the evolution accuracy of the method in this paper is better.

[0219] The evolution accuracy data of each method over the last 5 years in the industry-academia-research network were statistically analyzed, and the following table (Table 8) shows the evolution accuracy data of each method over the last 5 years:

[0220] Table 8 shows the evolution accuracy data for each method over the following 5 years.

[0221] years SNA BA ER GA GNN LPA This article's method 9th year 88.65% 90.12% 85.43% 92.37% 94.81% 87.58% 94.60% Year 10 84.21% 87.56% 81.89% 90.15% 93.26% 83.94% 94.30% 11th year 80.17% 85.32% 78.54% 88.62% 92.05% 78.61 93.68% 12th year 78.34% 83.15% 75.27% 87.14% 90.83% 75.45% 93.10% 13th year 76.73% 80.72% 71.36% 85.79% 89.60% 72.76% 91.42%

[0222] According to Table 8 and Figure 10 The comparison chart of the evolution accuracy data of each method over the last 5 years shows that the accuracy of all algorithms gradually decreases with the increase of the extrapolation time. The initial accuracy of this method is close to that of GNN, the curve remains high throughout the process, the decline is gradual, and the decay stability is better than the other comparison algorithms. It has a significant advantage in the accuracy of long-term evolution prediction.

[0223] This embodiment details the selection of six classic network evolution algorithms for industry-academia-research network simulation testing, elucidates the modeling and deduction process of each algorithm, uses historical data from the first eight years to train the model, and completes the evolution calculation using data from the last five years. The test process compares and statistically analyzes the network size, structural characteristics, and other indicators of various algorithms with that of this method. During the process, the evolution results of this method closely match the measured data, the accuracy decay trend is gradual, and the long-term deduction stability and prediction accuracy have significant advantages over other algorithms.

[0224] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. A method for dynamic evolution analysis of industry-academia-research networks based on lifecycle and topology, characterized in that, include: The system acquires entities and relationships from industry-academia-research data, constructs a time-series industry-academia-research collaboration network in the form of a dynamic heterogeneous knowledge graph, and performs incremental updates regularly to obtain a continuous time-series snapshot sequence of the knowledge graph. Based on the time-series snapshot sequence of the knowledge graph, a multi-scale topological representation is performed by fusing topological indicators and deep features to obtain a multi-scale topological representation result, which includes at least a comprehensive topological feature vector. An enhanced benefit-driven evolutionary game model is constructed. The benefit-driven evolutionary game model uses a dynamic heterogeneous knowledge graph as the network environment, is based on PPO deep reinforcement learning, takes a comprehensive topological feature vector as input, and outputs the probability distribution of cooperation and default of the graph network nodes. A temporal Transformer-VAE coupled lifecycle stage partitioning model is constructed. The partitioning model is based on multi-scale topological representation results and the running results of evolutionary game model, outputting the probability distribution of lifecycle stages corresponding to each time step, and detecting stage transition points in real time. Based on the stage transition points, the DoWhy causal inference library is used to construct the causal graph of the industry-academia-research network evolution. Counterfactual reasoning is performed on the current stage and the upcoming stage transition points to obtain the results of multi-scenario evolution simulation and interpretable decision generation.

2. The method for dynamic evolution analysis of industry-academia-research networks based on lifecycle and topology structure as described in claim 1, characterized in that, The aforementioned time-series industry-academia-research collaboration network specifically includes: entity types such as enterprises, universities, research institutes, and technologies; and relationship types such as collaboration applications, collaborative research and development, technology transfer, talent mobility, joint project undertaking, joint publication of papers, upstream and downstream linkages, and policy support. Calculate dynamic edge weights based on time decay factor and relation type; By using entity types as nodes and relationship types as edges, and combining dynamic edge weights, a collaborative network for industry, academia, and research can be constructed. By statistically analyzing regularly updated industry-academia-research collaboration network data, a time-series snapshot sequence of the knowledge graph is obtained.

3. The method for dynamic evolution analysis of industry-academia-research networks based on lifecycle and topology structure as described in claim 1, characterized in that, The multi-scale topological representation specifically includes topological indices, node depth features, comprehensive topological features, network topology types, and community evolution entropy. The topological indices are obtained by calculating the basic topological indices for each time-series snapshot, including the degree distribution power law exponent, clustering coefficient, average path length, intermediate central potential, and effective size of structural holes. The node depth features are extracted from each time-series snapshot using the dynamic GraphSAGE algorithm. The model input is node attributes and adjacency matrix, and the output is node depth feature vector. The comprehensive topological feature vector is obtained by fusing basic topological indices and node-level deep topological features through an attention mechanism. Based on the global comprehensive topological feature vector, network topology types are identified, and the dynamic Louvain algorithm is used to discover communities and calculate community evolution entropy.

4. The method for dynamic evolution analysis of industry-academia-research networks based on lifecycle and topology structure as described in claim 1, characterized in that, The enhanced benefit-driven evolutionary game model is as follows: using a dynamic heterogeneous knowledge graph as the network operating environment, each node in the network is defined to have two strategy choices, namely, a cooperative strategy and a default strategy. Based on its own state, the state of its neighbors and the global network state, the node uses the Proximal Policy Optimization (PPO) algorithm to simulate the decision-making process of bounded rational nodes and selects the corresponding strategy to maximize its own benefit.

5. The method for dynamic evolution analysis of industry-academia-research networks based on lifecycle and topology structure according to claim 4, characterized in that, The near-end strategy optimization PPO algorithm specifically includes the PPO subject decision-making and dynamic reconnection process with partners; Among them, the PPO subject decision takes the node's own comprehensive topological feature vector, the comprehensive topological feature vector of neighboring nodes and the global network state features as input, and outputs the probability distribution of choosing cooperation or default strategy through a multi-layer neural network to dynamically learn and optimize the strategy. The dynamic reconnection mechanism targets low-revenue and low-reputation neighbors. Nodes disconnect with a fixed probability and use an attention-weighted preference connection method. The selection probability is calculated by combining the similarity of the comprehensive topological feature vectors between nodes, the historical average revenue and reputation value of candidate partners, and then a new partner is selected.

6. The method for dynamic evolution analysis of industry-academia-research networks based on lifecycle and topology structure according to claim 5, characterized in that, The reward function of the near-end strategy optimization PPO algorithm is specifically as follows: it comprehensively considers factors such as cooperation benefits, other party cooperation benefits, breach of contract penalties, opportunity benefits, and reputation rewards, and dynamically adjusts the benefit distribution ratio in combination with the betweenness centrality coefficient in the comprehensive topological feature vector of the node.

7. The method for dynamic evolution analysis of industry-academia-research networks based on lifecycle and topology structure according to claim 1, characterized in that, The Transformer-VAE coupled lifecycle stage division model is as follows: Based on the results of multi-scale topological representation, traditional structural entropy and deep feature entropy are calculated. Extract the Gini coefficient for benefit distribution, the network default risk index, and the average satisfaction of the subjects from the results of the evolutionary game model; The structural orderliness and interest coordination are calculated separately to obtain a two-dimensional time series index sequence. The two-dimensional index is encoded and compressed by variational autoencoder (VAE), the latent space weights of the two types of indexes are learned, and the comprehensive evolution index is generated by adaptive fusion. The integrated evolutionary index sequence and the integrated topological feature vector sequence are input into the temporal fusion Transformer. The multi-head self-attention mechanism is used to capture long-term temporal dependencies and output the probability distribution of the network belonging to the five life cycle stages at the current moment. Based on a preset judgment threshold, the system continuously monitors changes in stage probability, captures stage transition points in real time, and saves the stage category, stage probability, and stage transition time for each time step to obtain a complete life cycle evolution time series result.

8. The method for dynamic evolution analysis of industry-academia-research networks based on lifecycle and topology structure according to claim 7, characterized in that, The variational autoencoder (VAE) specifically involves: processing each time node... Corresponding structural order Coordination with interests The input vector is concatenated to obtain a two-dimensional input vector. A multi-layer fully connected neural network encoder maps the two-dimensional input vector to a low-dimensional latent space, outputting the mean and variance vectors of the latent variables. A reparameterization technique is then used for random sampling to obtain the latent space vector. The latent space vector is then decoded. Map back to the original input dimension and output the reconstructed vector. ,in These are the reconstructed values ​​for structural orderliness and interest coordination, respectively.

9. The method for dynamic evolution analysis of industry-academia-research networks based on lifecycle and topology as described in claim 8, characterized in that, The temporal fusion Transformer specifically includes an input layer, a temporal embedding layer, a multi-head self-attention layer, a temporal fusion layer, a temporal decoder, and an output layer; The temporal embedding layer maps the preprocessed dual temporal sequences into high-dimensional temporal embedding vectors, and introduces positional encoding to obtain dual temporal embedding sequences. The dual temporal embedding sequences are concatenated and input into a multi-head self-attention layer. Multiple attention heads are used for parallel computation, and the attention-weighted temporal feature sequence is output. The temporal fusion layer uses residual connections and layer normalization operations to fuse and optimize the attention-weighted temporal features to obtain a temporal fusion feature sequence. The temporal decoder uses a multi-layer fully connected neural network and a Softmax activation function to calculate the probability distribution of the five lifecycle stages corresponding to each time step.

10. The method for dynamic evolution analysis of industry-academia-research networks based on lifecycle and topology structure according to claim 1, characterized in that, The counterfactual reasoning for the current stage and the upcoming stage transition point is specifically as follows: using the graph structure construction module of the DoWhy causal inference library, with the stage transition point as the time node, the selected core variables are used as nodes of the causal graph. Combining the evolution mechanism of the industry-academia-research network, the direct causal path and indirect causal path between variables are defined to generate a standardized causal graph of the industry-academia-research network evolution. By using the backdoor adjustment method in the DoWhy library, the interference of confounding variables on causal effects is blocked. Based on the propensity score matching (PSM) algorithm, the propensity score of each sample is calculated. By matching samples with and without intervention, the average causal effect ATE of the cause variable on the outcome variable is estimated, and the causal effect estimation results are obtained. Focusing on the current life cycle stage and the upcoming stage transition point, we set up counterfactual scenarios, and based on the causal graph and causal effect estimation results, we simulated the time nodes of stage transition, network topology evolution trends, changes in node benefit distribution and cooperation stability under different counterfactual scenarios, and output counterfactual inference results. Based on the counterfactual reasoning results and the incremental update characteristics of dynamic heterogeneous knowledge graphs, three core scenarios are set up: baseline scenario, optimization scenario, and risk scenario. With a monthly time step, the entire process of topological evolution, node strategy selection, and life cycle stage transition of the industry-academia-research network under each scenario is simulated to obtain multi-scenario evolution simulation results.

Images