A method for recommending potential partners of researchers based on structural hole dynamic evolution

Abstract

This invention provides a method for recommending potential partners to researchers based on the dynamic evolution of structural holes, belonging to the fields of artificial intelligence and data mining. The method includes: collecting sample data to establish a dynamic cooperation network sequence; defining node homogeneity loss, temporal smoothing loss, structural hole closure loss, and structural hole disintermediation loss; extracting referral and disintermediation strategy information from the node embedding vectors using social strategy parameters, and calculating the single-mediation connection probability and single-mediation disintermediation probability; based on the assumption of independent influence of common neighbors, aggregating the total connection probability and total disintermediation probability, and constructing the structural hole closure loss and structural hole disintermediation loss through negative maximum likelihood estimation; finally, combining the four types of losses to obtain the future cooperation probability, and generating a potential partner recommendation list accordingly. This invention improves prediction accuracy and interpretability, effectively identifying persistent, potential, and disappearing partners, providing decision support for R&D team building and innovation management.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and data mining technology, specifically involving a method for recommending potential partners of developers based on the dynamic evolution of structural holes. Background Technology

[0002] In the fields of technological innovation and R&D management, research institutes, corporate R&D departments, and intellectual property analysis agencies have an increasingly prominent need for accurate recommendations of R&D partners. This is especially true in patent-intensive and collaboratively innovative industries such as high-end equipment, electronic information, new materials, new energy, and biomedicine, where reliable partner recommendations can effectively improve R&D efficiency and innovation quality. However, existing methods for identifying potential partners mainly fall into two categories: one is similarity calculation based on topological features such as common neighbors and resource allocation. These methods are intuitive but struggle to depict complex evolutionary patterns. The other is representation learning methods based on DeepWalk, graph neural networks, etc. While these methods excel in feature fusion, they lack sufficient modeling of the intrinsic mechanisms of partnership formation, resulting in low accuracy and weak interpretability of recommendation results. Consequently, they cannot provide stable and reliable partner decision support for innovation entities.

[0003] Existing technologies for predicting partners based on dynamic R&D cooperation networks typically construct the dynamic network according to time series and use time-series graph network models to learn and predict cooperation relationships. These methods mostly focus only on the overall temporal changes of the network and the smooth transition of node embeddings, failing to effectively model the intrinsic evolutionary mechanisms of cooperation formation. Furthermore, they lack specific handling of structural hole patterns commonly found in dynamic networks, resulting in insufficient ability to capture key topological evolutionary features. In addition, existing dynamic network methods generally use uniform parameters and operational rules to represent all node relationships, failing to distinguish and adapt to different cooperation formation paths, leading to a lack of rationality and discriminative power in the model's inferences about future cooperation relationships. Summary of the Invention

[0004] To address the aforementioned issues, this invention provides a method for recommending potential partners to developers based on the dynamic evolution of structural holes.

[0005] To achieve the above objectives, this invention provides a method for recommending potential R&D partners based on the dynamic evolution of structural holes, comprising: Sample information is collected from multiple patent documents in the target field. The inventors of the patent documents are used as network nodes, and the common invention patent relationship between inventors is defined as the connection edge between nodes.

[0006] Based on nodes and connecting edges, a dynamic cooperative network sequence is established according to a preset time step; each sample information is initialized as an original node embedding vector; according to the structural hole triples in the dynamic cooperative network sequence, the total parameters of the original social strategy are determined. The total parameters of the original social strategy include social strategy parameters that characterize the possibility of non-cooperative node pairs establishing connecting edges through common neighbor nodes, and disintermediation social strategy parameters that characterize the possibility of non-cooperative node pairs directly establishing connecting edges.

[0007] The original node embedding vectors and the total parameters of the original social strategy are iteratively updated using stochastic gradient descent. When the joint total loss function converges to its minimum, the optimal social strategy parameters and the optimal node embedding vectors for each sample are obtained. The joint total loss function is constructed from the node homogeneity space loss function, the node temporal smoothing loss function, the structural hole closure loss function, and the structural hole disintermediation loss function. The node homogeneity space loss function describes the distance difference between node pairs with and without connecting edges in the embedding vector space. The node temporal smoothing loss function describes the Euclidean distance between the embedding vectors of the same sample at adjacent time steps. The structural hole closure loss function describes the difference between the probability distribution of non-cooperative node pairs establishing connecting edges through common neighbor nodes and the actual situation. The structural hole disintermediation loss function describes the difference between the probability distribution of non-cooperative node pairs establishing connecting edges and the actual situation.

[0008] Using the optimal node embedding vector and the optimal social strategy parameters, calculate the probability of cooperation for each non-cooperative node pair in the future time step; sort them according to the cooperation probability and select the top K as the potential partner recommendation list.

[0009] Preferably, the node homogeneity space loss function The calculation formula is: ; in, For nodes and Strength of the relationship; For nodes and Embedded vector and The distance between; For any real number Established, and Boundary values; functions It combines the weights of each sample with a measure of sample difference, and is usually defined as ; The sequence number represents the node pairs that have a cooperative relationship. It is a set of node pairs that have cooperative relationships. These are the sequence numbers of node pairs with no cooperative relationship. It is a set of node pairs with no cooperative relationship.

[0010] Preferably, the node time smoothing loss function The calculation formula is: in, Denotes the Euclidean norm; and Representing nodes respectively At time step and time step The embedding vector.

[0011] Preferably, the structural hole closure loss function The construction process includes: Define the dynamic proximity vector for each triple. The Used to quantify common neighbors For nodes and The local effects of establishing connections; based on and social strategy parameters Calculation on a single common node Under the influence, and exist The probability of forming a connection at any given moment ; Based on the assumption of independent influence of all common neighbors and Calculate all triples and exist The total probability of establishing a connection at any given moment; Based on negative maximum likelihood estimation, a structural hole closure loss function is constructed using the dynamic cooperative network sequence and the total probability. .

[0012] Preferably, the structural hole mediation loss function The construction process includes: Define the dynamic proximity vector for each triple. The Used to quantify common neighbors For nodes and The local effects of establishing connections; based on And disintermediation social strategy parameters Calculation on a single common node Under the influence, and exist The probability of achieving a disintermediate connection at any given time ; Based on the assumption of independent influence of all common neighbors and Calculate each triplet and exist The total probability of de-intermediation at any given moment; Based on negative maximum likelihood estimation, a structural hole closure loss function is constructed using the dynamic cooperative network sequence and the total probability. .

[0013] Preferably, the dynamic proximity vector of the triplet The calculation formula is: ; in, , and Forming a triplet, in which, time and No connection, yes and The common neighbor, i.e., the intermediary; Indicates in time and Relationship weights between them; Indicates in time and Relationship weights between them; express In time The embedding vector; express In time The embedding vector; express In time The embedding vector.

[0014] Preferably, the cooperation probability The calculation formula is: ; in, For nodes The transpose of the embedding vector at time t For nodes The embedding vector at time t; Represents a node and In the future The probability of cooperation at any given moment. This represents the Sigmoid activation function. It is the transpose symbol; This indicates the node number of different samples.

[0015] Preferably, before establishing the dynamic cooperation network sequence according to the preset time step, the method further includes: calculating the total similarity of inventors with the same name in three dimensions: technical field, patent collaborator, and patentee. If the total similarity is greater than a preset threshold, they are determined to be the same inventor.

[0016] This invention also provides a developer potential partner recommendation system based on the dynamic evolution of structural holes, comprising: The data acquisition module is used to collect sample information from multiple patent documents in the target field, taking the inventors of the patent documents as network nodes, and defining the common invention patent relationship between inventors as the connection edge between nodes.

[0017] A construction module is used to build a dynamic cooperative network sequence based on nodes and connection edges at preset time steps; initialize each sample information as an original node embedding vector; and determine the total parameters of the original social strategy according to the structural hole triples in the dynamic cooperative network sequence. The total parameters of the original social strategy include social strategy parameters that characterize the possibility of non-cooperative node pairs establishing connection edges through common neighbor nodes, and disintermediation social strategy parameters that characterize the possibility of non-cooperative node pairs directly establishing connection edges.

[0018] The computation module iteratively updates the original node embedding vectors and the total parameters of the original social strategy using stochastic gradient descent. When the joint total loss function converges to its minimum, it obtains the optimal social strategy parameters and the optimal node embedding vectors for each sample. The joint total loss function is constructed from a node homogeneity space loss function, a node temporal smoothing loss function, a structural hole closure loss function, and a structural hole disintermediation loss function. The node homogeneity space loss function describes the distance difference between node pairs with and without connecting edges in the embedding vector space. The node temporal smoothing loss function describes the Euclidean distance between the embedding vectors of the same sample at adjacent time steps. The structural hole closure loss function describes the difference between the probability distribution of non-cooperative node pairs establishing connecting edges through common neighbor nodes and the actual situation. The structural hole disintermediation loss function describes the difference between the probability distribution of non-cooperative node pairs establishing connecting edges and the actual situation.

[0019] The prediction module is used to calculate the probability of cooperation for each non-cooperative node pair in the future time step using the optimal node embedding vector and optimal social strategy parameters; the nodes are sorted according to their cooperation probabilities, and the top K nodes are selected as a potential partner recommendation list.

[0020] The present invention also provides a computer-readable storage medium storing a computer program that, when loaded by a processor, is capable of executing any of the steps in the developer-partner prediction method.

[0021] This invention provides a method for recommending potential partners for developers based on the dynamic evolution of structural holes, which has the following advantages: By explicitly modeling the dual dynamic evolution mechanism of triple closure and disintermediation in structural hole theory, this invention overcomes the shortcomings of existing technologies that treat network evolution as a black box and lack mechanistic explanation. It not only ensures the stability and continuity of the embedded representation by utilizing node homogeneity and time smoothing constraints, but also quantifies the cooperative decision-making mode of heterogeneity of developers by introducing independent social strategy parameters. Thus, while significantly improving the accuracy of partner prediction, it achieves interpretability in complex evolution scenarios such as continuous cooperation, new potential cooperation, and disappearance of cooperative relationships, providing a decision-making basis with both high accuracy and strong logical support for R&D team optimization and innovation management. Attached Figure Description

[0022] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating a method for recommending potential partners to developers based on the dynamic evolution of structural holes, according to an embodiment of the present invention. Figure 2 This is a general framework diagram of the technical solution of an embodiment of the present invention. Detailed Implementation

[0024] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0025] like Figure 2 As shown, this invention constructs a dynamic developer collaboration network sequence by acquiring historical patent data, performing inventor name disambiguation processing, and then... ,in Indicates time step Network snapshots, For the set of developer nodes, For time step The set of cooperative relationship edges is determined. Node embedding vectors and social strategy parameters are initialized, and a joint total loss function is constructed, which includes node homogeneity loss, temporal smoothing loss, structural hole closure loss, and structural hole disintermediation loss. Based on the joint total loss function, stochastic gradient descent is used to iteratively optimize the model parameters, obtaining the embedding vectors and social strategy parameters of each developer node at each time step. Using the optimized node embedding vectors, the cooperation probability of any two unconnected nodes in future time steps is calculated. Candidate node pairs are sorted according to the cooperation probability, and already cooperated nodes are removed to generate a Top-K potential partner recommendation list.

[0026] Based on this, the present invention provides a method for recommending potential partners of developers based on the dynamic evolution of structural holes. This embodiment takes patent data in the field of civilian unmanned aerial vehicles as an example. Figure 1 As shown, it includes the following steps: S1. Data Collection and Dynamic Collaborative Network Construction Patent data for the civilian unmanned aerial vehicle (UAV) field from 2005 to 2024 were retrieved from the patent database, resulting in 83,989 records after family merging. Inventor name disambiguation was performed, calculating the similarity of inventors with the same name across three dimensions: IPC classification number, collaborator set, and patentee. If the similarity exceeded a threshold, they were merged into a single unique ID, as shown in Table 1.

[0027] Table 1. Patent Data Display in the Civil Unmanned Aerial Vehicle (UAV) Field After disambiguation, a dynamic cooperation network sequence divided by year is constructed based on the joint invention patent relationship. The dynamic cooperative network sequence includes the set of all sample nodes, the set of sample cooperative relationships, and the sample relationship strength weights. Non-cooperative node pairs represent node pairs that have not established cooperative relationships.

[0028] Definition 1. Dynamic Developer Collaboration Network. Let... For the set of developer nodes, The observation period length. The dynamic R&D network sequence is defined as... Snapshot ( ) indicates time The connection status between each vertex. Let be the set of cooperative relationships at time step t; The relation strength weights are determined by the mapping function. Confirmed. This definition allows relationships between nodes to be created, persist, or dissolve over time, aligning with dynamic behaviors in R&D networks such as project initiation, collaboration interruption, and interest migration.

[0029] Definition 2. Dynamic Network Embedding. Given a sequence of dynamic networks... Dynamic network embedding is designed for each time step Learning mapping functions ,in A positive integer representing the number of embedding dimensions. This function... The goal is to maintain and At time step The similarity between the network structure and their tendency to develop relationships with other nodes in the future. For simplicity, this embodiment of the invention defines embedding vectors. , .

[0030] Definition 3. Structural Void. Let... For time step A snapshot of the developer collaboration network. If triplet satisfy , and Then the triple is said to constitute a structural hole, in which Act as an intermediary, responsible for bridging knowledge gaps; and As a third-party actor, representing the separated groups.

[0031] Definition 4. Closed Evolution. For a structural hole triplet at time step t. If in Always satisfied , and If the triplet undergoes closed evolution, then the triplet is said to have undergone closed evolution.

[0032] Definition 5. Demediation Evolution. For structural hole triples at time step t. If in Always satisfied , and If so, the triplet is said to have undergone disintermediation evolution.

[0033] The developer collaboration network constructed according to Definition 1 exhibits significant dynamic growth characteristics. As shown in Table 2, the network grew from 41 nodes and 54 edges in 2005 to approximately 43,000 nodes and 210,000 edges in 2023, with the node size increasing by about 1,000 times in 20 years, reflecting the rapid expansion of the R&D scale in the field of civilian unmanned aerial vehicles.

[0034] Table 2. Number of Nodes and Edges per Year in the Cooperation Network for the Civil Unmanned Aerial Vehicle (UAV) Sector S2, Model Building S201, Node Space Embedding Modeling Regarding nodes, the model ensures that highly connected nodes are close to each other in the embedding space, effectively preserving the network's structural information. Formally, this embodiment of the invention uses two vertices... and Embedded and The distance between them is defined as:

[0035] ; At the current time step t, this embodiment of the invention divides all vertex pairs into two sets: the edge set... Non-edge set Based on the similarity assumption, if two nodes are connected, they tend to have closer embeddings in the latent representation space. Therefore, this invention derives a node homogeneity space loss function based on ranking loss, expressed as follows:

[0036] ; in, For nodes and Strength of the relationship; For nodes and Embedded vector and The distance between; For any real number Established, and Boundary values; functions It combines the weights of each sample with a measure of sample difference, and is usually defined as ; The sequence number represents the node pairs that have a cooperative relationship. It is a set of node pairs that have cooperative relationships. These are the sequence numbers of node pairs with no cooperative relationship. It is a set of node pairs with no cooperative relationship.

[0037] S202, Time Smoothing Modeling To ensure the dynamic evolution characteristics of the model, this embodiment of the invention also designs a temporal smoothness loss, which makes the node embeddings change smoothly over time, avoiding the problem of the network being completely reconstructed at every time step. This embodiment of the invention assumes that the network will evolve smoothly over time, rather than being completely reconstructed at every time step. Therefore, this embodiment of the invention defines temporal smoothness by minimizing the Euclidean distance between the embedding vectors of adjacent time steps. Formally, the node temporal smoothness loss function is:

[0038] in, Denotes the Euclidean norm; and Representing nodes respectively At time step and time step The embedding vector.

[0039] S203, Structural Void Closure Probability Modeling During the structural hole closure process, the model accurately predicts how two nodes establish direct connections through common neighbors, forming a tight innovation cluster, using a ternary closure proximity vector. This embodiment of the invention starts from a point in time. Open triplet Example begins: Developers and They were strangers, but both were Partners. At this time, the developers... A decision needs to be made regarding whether to proceed in the next time step. Introduction and To get to know each other and establish cooperation between them. This invention's embodiments assume... Will depend on him and and The decision is made based on the proximity (in the potential space), which is determined by a length of... The dynamic proximity vector of triples To quantify, among which, the structural hole closure loss function and the structural hole mediation loss function Same.

[0040] ; in, , and Forming a triplet, in which, time and No connection, yes and The common neighbor, i.e., the intermediary; Indicates in time and Relationship weights between them; Indicates in time and Relationship weights between them; express In time The embedding vector; express In time The embedding vector; express In time The embedding vector. Furthermore, this embodiment of the invention defines a social strategy parameter. This is A dimensional vector is used to extract policy information embedded in the latent vector of each node.

[0041] Based on the above definition, embodiments of the present invention will open triples. At the node Under the recommendation (or influence) of [name], The time interval evolves into a closed triple (i.e. and The probability that a connection will be formed between them is defined as: It is worth mentioning that, and The relationship may be facilitated by multiple co-developers. Therefore, the next objective of this embodiment of the invention is to jointly model how multiple open triples with common unconnected vertex pairs evolve. To this end, this embodiment of the invention defines a set For time step hour and The common neighbors of, and define vectors When the open triplet In When the time evolves into a closed triplet In other words, in Under the influence and They will become partners. Clearly, once... Close, all with and All related open triples will be closed. Based on this, by further assuming that each co-developer... and The effects of potential connections are independent of each other; embodiments of the present invention will use time steps. New connection The probability is defined as:

[0042] ; At the same time, if node and If not influenced by any co-developers, no edge will be created. In this embodiment of the invention, the maintenance probability is defined as:

[0043] ; The above open triplet Combining the two possible evolutionary trajectories described above, the embodiments of the present invention define a set. To indicate at time step The link was successfully created, and the collection... This represents links that were not created. Subsequently, in this embodiment of the invention, the loss function for the triplet closure process is defined as the negative maximum likelihood estimate of the data:

[0044] S204, Probabilistic Model for Removing Mediations from Structural Vulnerabilities In the process of disintermediation of structural holes, a disintermediation proximity vector is introduced to accurately predict how two nodes can bypass their common neighbors to establish a direct connection, thereby eliminating the advantage of mediators and enabling dynamic reconstruction of the developer's collaborative network. Considering the time step... A structural hole triplet ,in , but The triplet is in Decentralization occurs at any given time if and only if: 1) a third-party node and Create a new direct connection ;2) Intermediate edge ( and A break occurs. The probability of this joint event depends on the interaction of two types of signals: signals that facilitate connection and signals that distance the mediator. It is noteworthy that demediation and triple closure share the same initial topology (i.e., open triples), but represent drastically different evolutionary directions: the former reflects network cohesion and the elimination of information redundancy, while the latter reflects the decay of the mediator role and path optimization. To model these two complementary mechanisms in a unified manner, this embodiment of the invention adopts a design paradigm of shared state vectors and independent policy parameters. That is, defining another social policy parameter... (Parameters for disintermediation-based social strategies). One A dimensional vector is used to extract policy information embedded in the latent vector of each node.

[0045] Based on the above definition, embodiments of the present invention will open triples. In The probability of a time step evolving into disintermediation is defined as: same, and It may be connected by multiple common intermediaries. Let... Define an indicator variable for its common neighbor set. When the triplet When disintermediation happens Otherwise, it is 0. It is worth noting that disintermediation is exclusive, meaning that once... Established and all intermediate edges broken, the rest involve and The triplet will lose its structural hole property. Assuming the exclusion effects of each mediator are independent, define the time step... The probability of disintermediation occurring is:

[0046] Conversely, if disintermediation does not occur, the probability is: ; The above open triplet Combining the two possible evolutionary trajectories described above, the embodiments of the present invention define a set. To indicate at time step The node pairs where disintermediation successfully occurs, and the set This represents node pairs that did not meet the disintermediation conditions and were therefore not linked. Subsequently, in this embodiment of the invention, the structural hole disintermediation loss function of the triple disintermediation process is defined as the negative maximum likelihood estimate of the data:

[0047] ; In summary, given The joint total loss function for the global optimization problem with time steps is: ; S3, Model Training The model employs a joint optimization framework, integrating four objective functions: node homogeneity loss, temporal smoothness loss, ternary closure loss, and disintermediation loss. The importance of different objectives is balanced through hyperparameters. To address the high computational complexity of large-scale networks, this embodiment of the invention uses the following sampling technique: For the node homogeneity loss function, when a positive sample (edge) exists at time step t... In this embodiment of the invention, the first step is to... and Randomly select a vertex (Right now Then, another vertex is randomly selected from the other vertices. ,if Then the sample is determined to be a valid negative sample. In this embodiment of the invention, each edge... Repeated sampling process, and the training set is defined as .

[0048] For the loss function of a ternary closed process, for each The embodiments of the present invention first start from { Randomly select a vertex from the array, let's say it's} The next goal is to sample a vertex. ,in and Thus, an open triple is obtained in this embodiment of the invention. .like and exist Constantly If the open triple influences the formation of a connection, then it is a positive example; otherwise, it is a negative example. In embodiments of the present invention, this part can be defined as... The loss function for the mediator removal process can be similarly used to obtain the open triplet. ,like and exist Constantly Influences the formation of connections at the same time and If the connection is broken, the open triplet is a positive example; otherwise, it is a negative example. In embodiments of the present invention, this part can be defined as... Finally, in order to minimize the joint total loss function proposed in this embodiment of the invention, stochastic gradient descent is used to iteratively optimize the model parameters, including the embedding vectors of all nodes and the social policy parameters.

[0049] Probability of Cooperation The calculation formula is: ; in, For nodes The transpose of the embedding vector at time t For nodes The embedding vector at time t; Represents a node and In the future The probability of cooperation at any given moment. This represents the Sigmoid activation function. It is the transpose symbol. This indicates the node number of different samples.

[0050] S4, Recommended Potential Partners The dynamic network embedding model learns the embedding vectors of each node at different time steps. These vectors contain rich structural information and evolutionary patterns. As the network dynamically evolves, this embodiment of the invention focuses not only on the current network structure but also on predicting potential future cooperative relationships. If it is observed that at time... ,node and Although no connection has been established, their embedding vectors are relatively close in space, which may suggest a high probability of future cooperation. To transform the embedding vectors into actionable predictions, embodiments of the invention use inner products to map these high-dimensional vectors to... Annual cooperation probability. Finally, sort all nodes for the predicted future year according to their cooperation probability from high to low, remove nodes that have already cooperated in the previous time window, and generate... Recommended list, among which The value is determined based on the actual application requirements.

[0051] Experimental verification Training was performed using the PyTorch framework on an NVIDIA RTX 4090 GPU. The optimizer was Adam, the learning rate was set to 0.1, and the training epochs were 150. The node embedding dimension was set to 128. 15% of the edges were selected as the test set, and the rest were used for training. Undersampling was used to address the class imbalance problem in the dataset, ensuring a 1:1 ratio of positive to negative samples. Hyperparameters... and All are set to 1.

[0052] As shown in Table 3, the method of this invention was compared with nine baseline methods, which covered four technical routes: node similarity-based methods (CN, AA, PA), static representation learning-based methods (DeepWalk, Node2Vec), machine learning methods based on handcrafted features (SVM, Random Forest), and deep learning methods based on graph neural networks (GCN, TGAE).

[0053] Table 3. Results of Algorithm Comparison Experiments Experimental results show that the method of the present invention achieves the best performance in all four evaluation indicators, and has significantly improved all indicators compared with the suboptimal method (TGAE), thus verifying the effectiveness of the proposed method.

[0054] Each component was removed one by one through ablation experiments to verify its contribution to the model's performance. As shown in Table 4,

[0055] Two variant models were designed for the experiment: removal of closure (modeling the removal of the closure process) and removal of mediators (modeling the removal of the mediator process).

[0056] Table 4 Ablation Experiment Results The ablation experiment results show that the model performance decreased significantly after removing any structural hole evolution mechanism, indicating that both closure and disintermediation evolution mechanisms have irreplaceable contributions to the model performance.

[0057] As shown in Table 5, by adjusting and The settings were used to test the performance of all parameter combinations on the validation set.

[0058] Table 5 Parameter Analysis Results The parameter analysis results demonstrate that the model in this embodiment of the invention has good robustness within a reasonable parameter range, and provides empirical evidence for parameter selection in practical applications.

[0059] The recommendation results are analyzed using developer "Ren*feng" as an example. K=10 is set to generate a list of the top 10 predicted partners (with corresponding cooperation probability scores). Based on the developer's dynamic cooperation relationships, partners are divided into three categories: continuous partners (cooperating in two consecutive time periods), potential partners (newly emerging predicted partners), and disappeared partners (partners who previously cooperated but no longer appear).

[0060] The analysis results show that the seven continuing partners constitute the core stable layer of the collaboration network, with highly compatible research directions; the emergence of seven potential partners reflects the expansion of research directions or the initiation of new projects; and the loss of four disappearing partners may stem from project-driven one-off collaborations or personnel changes. This invention's model can accurately capture the above complex evolutionary patterns and provide a specific collaboration probability score for each recommended partner, offering quantifiable support for practical R&D management decisions.

[0061] Existing potential partner identification technologies lack explicit modeling of the mechanisms underlying social network formation. While existing deep learning methods can learn network representations in an end-to-end manner, their internal mechanisms are often unexplainable, making it difficult to reveal the social mechanisms behind cooperation formation. The ternary closure, as a fundamental unit of network evolution, is not adequately modeled in existing models, resulting in a less in-depth explanation of the mechanisms underlying cooperation formation.

[0062] Existing research largely focuses on the single evolutionary path of ternary closure (closing), neglecting the equally important evolutionary mechanism of disintermediation. However, researcher collaboration networks exhibit both a closed process where researchers establish connections through collaborators and a disintermediation process where they bypass intermediaries to collaborate directly for efficiency. These two mechanisms reflect different social dynamics and need to be captured simultaneously in modeling. Although some GNN methods support temporal modeling, they typically learn evolutionary patterns in a data-driven manner, lacking explicit characterization of theoretical constraints such as network evolution continuity and node policy heterogeneity, thus limiting the model's interpretability and generalization ability. Furthermore, there is insufficient characterization of node heterogeneity decision-making patterns. Different researchers have different social strategy preferences in collaboration decisions (e.g., whether they prefer to establish collaboration through intermediaries or directly seek partners), and existing methods lack mechanisms to learn and express this heterogeneity.

[0063] The purpose of this invention is to provide a developer partner recommendation method based on the dynamic evolution of structural holes. This method can explicitly integrate the two evolutionary mechanisms of structural hole theory, namely closure and disintermediation, into a dynamic network embedding framework. By modeling the evolutionary path of developers establishing cooperation (closure) under the introduction of common neighbors or directly cooperating without intermediaries (disintermediation), and introducing time smoothing constraints to ensure the continuity of evolution and the heterogeneous decision-making patterns of social strategy parameter learning nodes, it can achieve accurate characterization of the evolution law of developer cooperation network and high-precision, interpretable recommendation of future partners.

[0064] This invention explicitly integrates the closure and disintermediation evolutionary mechanisms of structural hole theory into a dynamic network embedding framework. This is the first time this invention has explicitly modeled both the closure mechanism (forming direct connections through mutual neighbor referrals) and the disintermediation mechanism (direct cooperation bypassing intermediaries) in a prediction model, achieving a complete characterization of the binary evolutionary mechanism of developer cooperation networks. A collaborative mechanism between the ternary closure proximity vector and social policy parameters is designed. By transforming the conceptual description of structural hole theory into a computable mathematical model and the abstract social theory into a quantifiable predictive indicator, a bridge is achieved from qualitative theory to quantitative prediction. A unified modeling paradigm of shared state vectors and independent policy parameters is adopted to model both mechanisms. Closure and disintermediation share the same ternary closure proximity vector (reflecting their common triplet initial topology), but use independent social policy parameters. and This approach characterizes two different evolutionary directions, ensuring both modeling consistency and capturing the heterogeneous features of the two mechanisms. Temporal smoothing constraints guarantee the continuity of network evolution. It avoids the abrupt changes that static models encounter when dealing with dynamic networks, allowing the embedding vectors to change smoothly over time, thus more accurately reflecting the actual characteristics of the progressive evolution of collaborative networks among developers.

[0065] Compared with the prior art, the present invention has the following significant advantages: 1. Significantly Improved Prediction Accuracy. Empirical validation on actual patent data demonstrates that the proposed model outperforms existing state-of-the-art methods in all four metrics: F1, AUC, AUPR, and ACC. Compared to the suboptimal dynamic graph neural network method, the F1 score is improved by approximately 3.26%, AUC by approximately 3.11%, AUPR by approximately 13.36%, and ACC by approximately 5.80%. This indicates that explicitly integrating the structural hole evolution mechanism into the model significantly improves the prediction accuracy of partner identification.

[0066] 2. The model possesses strong interpretability. Unlike end-to-end black-box deep learning methods, this invention provides a clear social mechanism explanation for the prediction results by explicitly modeling two social mechanisms: closure and disintermediation. Specifically, it can distinguish whether new collaborations are established through referrals from mutual neighbors (closed path) or by bypassing intermediaries and establishing them directly (disintermediation path), and it can also clearly distinguish between persistent partners, potential partners, and disappearing partners.

[0067] 3. The model exhibits good robustness. Through sensitivity analysis of key hyperparameters, the model of this invention demonstrates stability within a reasonable parameter range and good robustness to parameter variations, making it suitable for deployment and use in practical industrial applications.

[0068] 4. High practical application value. This invention's model can assist technology R&D managers in optimizing team structure, identifying key partners, and predicting the evolution trend of collaborative networks in real-world R&D team management scenarios, providing an effective decision support tool for building an innovation ecosystem and improving R&D efficiency.

[0069] Based on the same inventive concept, this invention also provides a developer potential partner recommendation system based on the dynamic evolution of structural holes, including: The data acquisition module is used to collect sample information from multiple patent documents in the target field, taking the inventors of the patent documents as network nodes, and defining the common invention patent relationship between inventors as the connection edge between nodes.

[0070] A construction module is used to build a dynamic cooperative network sequence based on nodes and connection edges at preset time steps; initialize each sample information as an original node embedding vector; and determine the total parameters of the original social strategy according to the structural hole triples in the dynamic cooperative network sequence. The total parameters of the original social strategy include social strategy parameters that characterize the possibility of non-cooperative node pairs establishing connection edges through common neighbor nodes, and disintermediation social strategy parameters that characterize the possibility of non-cooperative node pairs directly establishing connection edges.

[0071] The computation module iteratively updates the original node embedding vectors and the total parameters of the original social strategy using stochastic gradient descent. When the joint total loss function converges to its minimum, it obtains the optimal social strategy parameters and the optimal node embedding vectors for each sample. The joint total loss function is constructed from a node homogeneity space loss function, a node temporal smoothing loss function, a structural hole closure loss function, and a structural hole disintermediation loss function. The node homogeneity space loss function describes the distance difference between node pairs with and without connecting edges in the embedding vector space. The node temporal smoothing loss function describes the Euclidean distance between the embedding vectors of the same sample at adjacent time steps. The structural hole closure loss function describes the difference between the probability distribution of non-cooperative node pairs establishing connecting edges through common neighbor nodes and the actual situation. The structural hole disintermediation loss function describes the difference between the probability distribution of non-cooperative node pairs establishing connecting edges and the actual situation.

[0072] The prediction module is used to calculate the probability of cooperation for each non-cooperative node pair in the future time step using the optimal node embedding vector and optimal social strategy parameters; the nodes are sorted according to their cooperation probabilities, and the top K nodes are selected as a potential partner recommendation list.

[0073] The present invention also provides a computer-readable storage medium storing a computer program that can be used to execute the aforementioned method for recommending potential partners to developers based on the dynamic evolution of structural holes.

[0074] Specific limitations regarding the computational system for the developer potential partner recommendation method based on the dynamic evolution of structural holes can be found in the limitations of the developer potential partner recommendation method based on the dynamic evolution of structural holes mentioned above, and will not be repeated here. Each module in the aforementioned developer potential partner recommendation system based on the dynamic evolution of structural holes can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0075] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. Furthermore, the above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for recommending potential partners of researchers based on the dynamic evolution of structural holes, characterized in that, Includes the following steps: Sample information is collected from multiple patent documents in the target field. The inventors of the patent documents are used as network nodes, and the common invention patent relationship between inventors is defined as the connection edge between nodes. Based on nodes and connecting edges, a dynamic cooperative network sequence is established according to a preset time step; each sample information is initialized as an original node embedding vector. Based on the structural hole triples in the dynamic cooperative network sequence, the total parameters of the original social strategy are determined. The total parameters of the original social strategy include social strategy parameters that characterize the possibility of non-cooperative node pairs establishing connection edges through common neighbor nodes, and disintermediation social strategy parameters that characterize the possibility of non-cooperative node pairs establishing connection edges directly. The original node embedding vectors and the total parameters of the original social strategy are iteratively updated using stochastic gradient descent. When the joint total loss function converges to its minimum, the optimal social strategy parameters and the optimal node embedding vectors for each sample are obtained. The joint total loss function is constructed from the node homogeneity space loss function, the node temporal smoothing loss function, the structural hole closure loss function, and the structural hole disintermediation loss function. The node homogeneity space loss function describes the distance difference between node pairs with and without connecting edges in the embedding vector space. The node temporal smoothing loss function describes the Euclidean distance between the embedding vectors of the same sample at adjacent time steps. The structural hole closure loss function describes the difference between the probability distribution of non-cooperative node pairs establishing connecting edges through common neighbor nodes and the actual situation. The structural hole disintermediation loss function describes the difference between the probability distribution of non-cooperative node pairs establishing connecting edges and the actual situation. Using the optimal node embedding vector and the optimal social strategy parameters, calculate the probability of cooperation for each non-cooperative node pair in the future time step; sort them according to the cooperation probability and select the top K as the potential partner recommendation list.

2. The method for recommending potential partners of developers based on the dynamic evolution of structural holes according to claim 1, characterized in that, The node homogeneity space loss function The calculation formula is: ； in, For nodes and Strength of the relationship; For nodes and Embedded vector and The distance between; For any real number Established, and Boundary values; functions It combines the weights of each sample with a measure of sample difference, and is usually defined as ; The sequence number represents the node pairs that have a cooperative relationship. It is a set of node pairs that have a cooperative relationship. These are the sequence numbers of node pairs with no cooperative relationship. It is a set of node pairs with no cooperative relationship.

3. The method for recommending potential R&D partners based on the dynamic evolution of structural holes according to claim 1, characterized in that, The node time smoothing loss function The calculation formula is: in, Denotes the Euclidean norm; and Representing nodes respectively At time step and time step The embedding vector.

4. The method for recommending potential R&D partners based on the dynamic evolution of structural holes according to claim 1, characterized in that, The structural hole closure loss function The construction process includes: Define the dynamic proximity vector for each triple. The Used to quantify common neighbors For nodes and The local effects of establishing connections; based on and social strategy parameters Calculation on a single common node Under the influence, and exist The probability of forming a connection at any given moment ; Based on the assumption of independent influence of all common neighbors and Calculate all triples and exist The total probability of establishing a connection at any given moment; Based on negative maximum likelihood estimation, a structural hole closure loss function is constructed using the dynamic cooperative network sequence and the total probability. .

5. The method for recommending potential R&D partners based on the dynamic evolution of structural holes according to claim 1, characterized in that, The structural hole mediation loss function The construction process includes: Define the dynamic proximity vector for each triple. The Used to quantify common neighbors For nodes and The local effects of establishing connections; based on And disintermediation social strategy parameters Calculation on a single common node Under the influence, and exist The probability of achieving a disintermediate connection at any given time ; Based on the assumption of independent influence of all common neighbors and Calculate each triplet and exist The total probability of de-intermediation at any given moment; Based on negative maximum likelihood estimation, a structural hole closure loss function is constructed using the dynamic cooperative network sequence and the total probability. .

6. A method for recommending potential R&D partners based on the dynamic evolution of structural holes according to any one of claims 4 or 5, characterized in that, The dynamic proximity vector of the triplet The calculation formula is: ； in, , and Forming a triplet, in which, time and No connection, yes and The common neighbor, i.e., the intermediary; Indicates in time and Relationship weights between them; Indicates in time and Relationship weights between them; express In time The embedding vector; express In time The embedding vector; express In time The embedding vector.

7. The method for recommending potential R&D partners based on the dynamic evolution of structural holes according to claim 1, characterized in that, The probability of cooperation The calculation formula is: ； in, For nodes The transpose of the embedding vector at time t. For nodes The embedding vector at time t; Represents a node and In the future The probability of cooperation at any given moment. This represents the Sigmoid activation function. It is the transpose symbol; This indicates the node number of different samples.

8. The method for recommending potential R&D partners based on the dynamic evolution of structural holes according to claim 1, characterized in that, Before establishing the dynamic cooperation network sequence according to the preset time step, the method further includes: calculating the total similarity of inventors with the same name in three dimensions: technical field, patent collaborator, and patentee. If the total similarity is greater than the preset threshold, they are determined to be the same inventor.

9. A developer potential partner recommendation system based on the dynamic evolution of structural holes, characterized in that, include: The data acquisition module is used to collect sample information from multiple patent documents in the target field, taking the inventors of the patent documents as network nodes, and defining the common invention patent relationship between inventors as the connection edge between nodes. The module is used to build a dynamic cooperative network sequence based on nodes and connecting edges at preset time steps; it initializes each sample information as an original node embedding vector. Based on the structural hole triples in the dynamic cooperative network sequence, the total parameters of the original social strategy are determined. The total parameters of the original social strategy include social strategy parameters that characterize the possibility of non-cooperative node pairs establishing connection edges through common neighbor nodes, and disintermediation social strategy parameters that characterize the possibility of non-cooperative node pairs establishing connection edges directly. The computation module iteratively updates the original node embedding vectors and the total parameters of the original social strategy using stochastic gradient descent. When the joint total loss function converges to its minimum, it obtains the optimal social strategy parameters and the optimal node embedding vectors for each sample. The joint total loss function is constructed from the node homogeneity space loss function, the node temporal smoothing loss function, the structural hole closure loss function, and the structural hole disintermediation loss function. The node homogeneity space loss function describes the distance difference between node pairs with and without connecting edges in the embedding vector space. The node temporal smoothing loss function describes the Euclidean distance between the embedding vectors of the same sample at adjacent time steps. The structural hole closure loss function describes the difference between the probability distribution of non-cooperative node pairs establishing connecting edges through common neighbor nodes and the actual situation. The structural hole disintermediation loss function describes the difference between the probability distribution of non-cooperative node pairs establishing connecting edges and the actual situation. The prediction module is used to calculate the probability of cooperation for each non-cooperative node pair in the future time step using the optimal node embedding vector and the optimal social strategy parameters. Sort the potential partners according to their probability of cooperation and select the top K as the recommended list of potential partners.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is loaded by the processor, it is able to perform the steps of the method according to any one of claims 1 to 8.

Images