A method and system for identifying key network nodes in technical network analysis

Abstract

This invention relates to the field of complex network analysis technology, and provides a method and system for identifying key network nodes in technical network analysis. The method includes: identifying a target patent dataset to be analyzed; constructing a patent topology model using target patent data in the target patent dataset as nodes and patent citation relationships in the target patent data as edges; obtaining the weight of each edge in the patent topology model based on the target patent dataset to obtain a reference weight set; optimizing the patent topology model based on the reference weight set to obtain a target topology model; training a pre-constructed graph convolutional network model to obtain an optimized graph convolutional network model; and using the optimized graph convolutional network model to identify preset key network nodes in the target topology model, thereby achieving the identification of key network nodes. This invention can improve the accuracy of identifying key network nodes.

Description

Technical Field

[0001] This invention relates to the field of complex network analysis technology, and in particular to a method and system for identifying key network nodes in technical network analysis. Background Technology

[0002] As a complex system composed of technological entities and their key relationships, the identification of key nodes in a technology network is of great significance for technological innovation planning and supply chain risk prevention and control.

[0003] Traditional methods rely on structural centrality metrics (such as degree centrality, betweenness centrality, K-shell, and PageRank) to rank nodes. However, these methods often take a single perspective, ignoring the heterogeneity and dynamism of technical networks, thus limiting the accuracy of identifying key nodes. Therefore, improving the accuracy of identifying key nodes in a network has become an urgent problem to be solved. Summary of the Invention

[0004] This invention provides a method for identifying key network nodes in technical network analysis and a computer-readable storage medium, the main purpose of which is to improve the accuracy of identifying key network nodes.

[0005] To achieve the above objectives, the present invention provides a method for identifying key network nodes in technical network analysis, comprising:

[0006] The target patent dataset to be analyzed has been identified. The target patent dataset includes multiple target patent data, and the target patent data includes the citation relationship between patents, the initial patent field set, and the patent text.

[0007] A patent topology model is constructed using the target patent data in the target patent dataset as nodes and the patent citation relationships between the target patent data in the target patent dataset as edges. The patent topology model includes multiple topology nodes, and each topology node corresponds one-to-one with the target patent data.

[0008] The weight of each edge in the patent topology model is obtained based on the target patent dataset to obtain a reference weight set. The patent topology model is then optimized based on the reference weight set to obtain the target topology model.

[0009] The pre-built graph convolutional network model is trained to obtain an optimized graph convolutional network model;

[0010] By utilizing an optimized graph convolutional network model, the key nodes of the target network are identified in the target topology model, thereby achieving the identification of key nodes of the target network.

[0011] Optionally, obtaining the weight of each edge in the patent topology model based on the target patent dataset to obtain a reference weight set includes:

[0012] Extract the topology nodes sequentially from the patent topology model, and perform the following operations on the extracted topology nodes:

[0013] An initial tree structure is constructed with the topology node as the root node and each patent initial field in the patent initial field set corresponding to the topology node as the first child node.

[0014] The initial claim text is extracted from the patent text corresponding to the topology node. The initial claim text is matched and replaced to obtain the target claim text. The target claim text is parsed using a pre-built language model to obtain the initial text set. The reference corpus set is identified based on the patent initial domain set. The initial text set is used to search and classify in the reference corpus set to obtain the parsed text set. The reference corpus in the reference corpus set corresponds one-to-one with the patent initial domain in the patent initial domain set.

[0015] For each parsed text group in the parsed text group set, perform the following operation:

[0016] The initial tree structure is retrieved and identified using parsed text groups to obtain the target tree structure, and a reference weight set is obtained based on the target tree structure.

[0017] Optionally, obtaining the reference weight set based on the target tree structure includes:

[0018] Identify the technical field of each parsed text group in the parsed text group set to obtain one or more search fields. Perform the following operation on each of the one or more search fields:

[0019] The number of initial texts in the initial text set corresponding to the target tree structure is counted to obtain the statistical quantity. The parsed text groups in the parsed text group set are classified and counted using the retrieval domain to obtain the retrieval quantity set.

[0020] For each search result in the search result set, perform the following operation:

[0021] Calculate the ratio of the number of searches to the number of statistics to obtain the domain percentage value. Compare the domain percentage value with the preset domain retention threshold. If the domain retention threshold is confirmed to be greater than or equal to the domain retention threshold, then use the search domain corresponding to the domain percentage value to mark the domain percentage value to obtain the marked percentage value.

[0022] The identifier percentage values ​​are summarized to obtain an identifier percentage value set, and a reference weight set is obtained based on the identifier percentage value set.

[0023] Optionally, obtaining the reference weight set based on the set of identifier proportion values ​​includes:

[0024] The initial abstract text is extracted from the patent text corresponding to the topology node, and the initial abstract text is optimized using a pre-built abstract optimization model to obtain the target abstract vector.

[0025] Using the extracted topological nodes, neighboring nodes are identified in the patent topology model. The neighboring abstract vectors are obtained using the patent texts corresponding to the neighboring nodes. The cosine distance between the target abstract vector and the neighboring abstract vectors is calculated to obtain the first matching degree.

[0026] Obtain the reference proportion value set of neighboring nodes and one or more reference domains, and perform a logical intersection operation on one or more reference domains and one or more search domains to obtain one or more overlapping domains;

[0027] Using one or more overlapping regions, a first percentage value set and a second percentage value set are identified in the reference percentage value set and the identifier percentage value set, respectively.

[0028] The first restored value set is calculated using the pre-constructed scaling formula and the first set of percentage values, and the second restored value set is obtained using the second set of percentage values.

[0029] The scaling relationship is as follows:

[0030]

[0031] in, Indicates the first set of restored values. The first restored value. This indicates the first percentage value set. The first percentage value. This indicates the first percentage value set. The first percentage value. This indicates that the first percentage value is in the total. The first percentage value;

[0032] The first restored value set and the second restored value set are matched using one or more overlapping neighborhoods to obtain one or more matching nodes. Each matching node corresponds to one of the overlapping neighborhoods and includes a first restored value and a second restored value.

[0033] A second matching degree is calculated based on the one or more matching nodes. Reference weights are calculated using the first matching degree and the second matching degree. The reference weights are then summarized to obtain a reference weight set.

[0034] Optionally, calculating the second matching degree based on the one or more matching nodes includes:

[0035] One or more matching nodes are shown below:

[0036]

[0037] in, This indicates the first matching node in a set of one or more matching nodes. This indicates the second matching node in a group of one or more matching nodes. Indicates that one or more matching nodes share a common feature. Matching nodes, They respectively represent the first matching node, the second matching node, and the third matching node in one or more matching nodes. The first restored value corresponding to each matching node. They respectively represent the first matching node, the second matching node, and the third matching node in one or more matching nodes. The second restored value corresponding to each matching node;

[0038] Calculate the Euclidean distance of each matching node in one or more matching nodes and sum them up to obtain the second matching degree.

[0039] Optionally, training the pre-built graph convolutional network model to obtain an optimized graph convolutional network model includes:

[0040] For each topological node in the target topology model, indexes are extracted to obtain node indices, as shown below:

[0041]

[0042] in, Indicates node metrics, Indicates degree centrality. Represents the eigenvector centrality. Indicates the degree of importance of the reference;

[0043] The network model of the target topology model is confirmed based on node indicators;

[0044] The network model of the target topology model is shown below:

[0045]

[0046] in, The network model representing the target topology model, This represents the set of topological nodes in the target topology model. This represents the set of edges in the target topology model. Represents the reference weight set;

[0047] Construct the network resilience assessment formula, as shown below:

[0048]

[0049] in, Indicates network resilience. This represents the number of nodes in the largest weakly connected subgraph in the network model. Indicates the first The maximum number of nodes in a weakly connected subgraph after removing topological nodes. Indicates the first The maximum number of nodes in a weakly connected subgraph after removing topological nodes;

[0050] An optimized graph convolutional network model is obtained based on the network model and the network damage assessment relation.

[0051] Optionally, the step of obtaining the optimized graph convolutional network model based on the network model and the network resilience assessment formula includes:

[0052] The graph structure is extracted from the network model, as shown below:

[0053]

[0054] Where S represents the graph structure;

[0055] Using a pre-built strategy function, a node is selected for removal from multiple topological nodes corresponding to the graph structure. The strategy function is shown below:

[0056]

[0057] in, This indicates the removal of the node. Indicates the first The next node to be removed. This represents the policy function. Indicates the first The graph structure after removing the node;

[0058] By defining the time reward function for the graph convolutional network model, the target convolutional network model is obtained. The time reward function is as follows:

[0059]

[0060] in, This represents the time-reward function. Indicates the first After removing the topology nodes, the network resilience rate of the network model is calculated using the network resilience assessment formula. Indicates the first Network resilience of the network model after removing topology nodes;

[0061] An optimized graph convolutional network model is obtained by using a policy function and a target convolutional network model.

[0062] Optionally, obtaining the optimized graph convolutional network model using the policy function and the target convolutional network model includes:

[0063] The iteration probability ratio is obtained based on the policy function, where the iteration probability ratio is as follows:

[0064]

[0065] in, Indicates that at time step is The ratio of iteration probabilities at time, This represents the parameters of the current policy network. The parameters represent the old policy network. This indicates that the topology node to be removed is selected;

[0066] Construct the advantage function, and then use the advantage function and the iterative probability ratio to construct the policy pruning loss function. The advantage function is shown below:

[0067]

[0068] in, Indicates time step as The dominant function value at that time, This indicates the preset termination time step. Indicates the discount factor. Represents the GAE coefficient. This indicates that the value network corresponds to the network model at time step [missing information]. The following value estimate, This indicates that the value network corresponds to the network model at time step [missing information]. The following is a value estimate;

[0069] The target convolutional network model is defined using the dominance function and the policy pruning loss function to obtain the optimized convolutional network model. The optimized convolutional network model is then trained to obtain the optimized graph convolutional network model.

[0070] Optionally, the strategy prunes the loss function as follows:

[0071]

[0072] in, This represents the policy pruning loss of the PPO algorithm. Indicates time step Expectations This represents the cutting factor.

[0073] To achieve the above objectives, the present invention also provides a network critical node identification system for technical network analysis, comprising:

[0074] The data collection module is used to identify the target patent dataset to be analyzed. The target patent dataset includes multiple target patent data, and the target patent data includes the citation relationship between patents, the initial field set of patents, and the patent text.

[0075] The data processing module is used to construct a patent topology model with the target patent data in the target patent dataset as nodes and the patent citation relationship between the target patent data in the target patent dataset as edges. The patent topology model includes multiple topology nodes, and each topology node corresponds one-to-one with the target patent data.

[0076] The graph model construction module is used to obtain the weight of each edge in the patent topology model based on the target patent dataset, obtain a reference weight set, and optimize the patent topology model based on the reference weight set to obtain the target topology model.

[0077] The network training and feature recognition module is used to train a pre-built graph convolutional network model to obtain an optimized graph convolutional network model.

[0078] By utilizing an optimized graph convolutional network model, the key nodes of the target network are identified in the target topology model, thereby achieving the identification of key nodes of the target network.

[0079] To address the above problems, the present invention also provides an electronic device, the electronic device comprising:

[0080] A memory that stores at least one instruction; and a processor that executes the instructions stored in the memory to implement the network critical node identification method for technical network analysis described above.

[0081] To address the aforementioned problems, the present invention also provides a computer-readable storage medium storing at least one instruction, which is executed by a processor in an electronic device to implement the network critical node identification method for technical network analysis described above.

[0082] To address the problems described in the background art, this invention identifies a target patent dataset to be analyzed. This dataset includes multiple target patent data points, each containing inter-patent citation relationships, an initial patent domain set, and patent text. A patent topology model is constructed using the target patent data points as nodes and the inter-patent citation relationships as edges. This model includes multiple topology nodes, each corresponding one-to-one with a target patent data point. The weight of each edge in the patent topology model is obtained based on the target patent dataset, resulting in a reference weight set. The patent topology model is then optimized based on this reference weight set to obtain the target topology model. This invention considers the dynamic nature of the technology network. Therefore, in this embodiment, the target patent dataset to be analyzed is first identified, and this dataset can be manually set. Furthermore, this invention integrates a graph construction method combining structural topology and semantic similarity. Based on patent citation relationships, a summary embedding vector and the patent's technical field are integrated as edge weights, forming a weighted network with both semantic and technical heterogeneous features. Furthermore, the weights of edges between nodes in different topologies are increased, making the weighted network more accurate, and consequently improving the accuracy of training graph convolutional network models using the weighted network. An optimized graph convolutional network model is obtained by training a pre-constructed graph convolutional network model. Using this optimized model, preset key nodes of the target network are identified in the target topology model. This invention constructs a deep reinforcement learning framework with GCN as the policy network and PPO as the optimizer. It combines centrality features and edge weights as inputs, introduces a custom graph environment module to simulate the node removal process, and uses the change in the proportion of the most connected subgraph to measure structural resilience as a reward signal to guide the policy model in learning the optimal node selection path, achieving efficient identification of key nodes and policy generalization. Therefore, this invention can improve the accuracy of identifying key network nodes. Attached Figure Description

[0083] Figure 1 This is a flowchart illustrating a method for identifying key network nodes in technical network analysis according to an embodiment of the present invention.

[0084] Figure 2 A functional block diagram of a network critical node identification system for technical network analysis provided in an embodiment of the present invention;

[0085] Figure 3 This is a schematic diagram of the structure of an electronic device that implements the network key node identification method for technical network analysis, according to an embodiment of the present invention.

[0086] Explanation of reference numerals in the attached figures:

[0087] 10. Electronic device; 11. Processor; 12. Memory; 13. Bus.

[0088] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0089] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0090] This application provides a method for identifying key network nodes in technical network analysis. The executing entity of this method includes, but is not limited to, at least one electronic device configured to execute the method provided in this application, such as a server or a terminal. In other words, the method for identifying key network nodes in technical network analysis can be executed by software or hardware installed on a terminal device or a server device, where the software may be a blockchain platform. The server includes, but is not limited to, a single server, a server cluster, a cloud server, or a cloud server cluster.

[0091] Reference Figure 1 The diagram shown is a flowchart illustrating a method for identifying key network nodes in technical network analysis according to an embodiment of the present invention. In this embodiment, the method for identifying key network nodes in technical network analysis includes:

[0092] S1. Identify the target patent dataset to be analyzed. The target patent dataset includes multiple target patent data, and the target patent data includes the citation relationship between patents, the initial patent field set, and the patent text.

[0093] It should be explained that the target patent dataset refers to the patent dataset to be analyzed, and the target patent dataset can be an existing publicly available dataset or manually set. Target patent data refers to the patent data to be analyzed, and the patent data includes, but is not limited to: inter-patent citation relationships, initial patent domain sets, and patent text. Inter-patent citation relationships refer to the citation relationships between different patents. The initial patent domain set refers to the set of domains involved in the target patent data. Optionally, the initial patent domain set can be obtained by searching the classification number of the patent documents; other techniques can achieve the same effect, and will not be elaborated further here. Patent text refers to the text corresponding to the patent application documents.

[0094] For example, in order to identify key patent data in a patent dataset in the field of image processing, so that researchers can reconstruct the critical path of technological evolution and devote more time and energy to research and innovation in key technologies, it is easy to understand that the critical path here is consistent with the content referred to by the key nodes of the target network in this embodiment. Both are used to describe target patent data that has a profound impact in the patent dataset. For the process of confirming the key nodes of the target network, please refer to the following embodiments.

[0095] S2. Construct a patent topology model using the target patent data in the target patent dataset as nodes and the patent citation relationships between the target patent data in the target patent dataset as edges. The patent topology model includes multiple topology nodes, and each topology node corresponds one-to-one with the target patent data.

[0096] Furthermore, the patent topology model is a topology diagram used to describe different target patent datasets. In the topology diagram, the target patent data is used as nodes and the inter-patent citation relationship of the target patent data is used as edges. That is, when two target patent data have an inter-patent citation relationship, the two target patent data are connected by a line. For example, if the target patent dataset contains first patent data, second patent data, third patent data, fourth patent data, and fifth patent data, and if any two of the target patent data in the first, second, and third patent data have a patent citation relationship, and the fourth patent data cites the fifth patent data, then a triangular initial patent topology model can be constructed using the first, second, and third patent data. In this triangular initial patent topology model, each vertex of the triangle is the target patent data, and the sides of the triangle are bidirectional. A line segment type initial patent topology model can be constructed using the fourth and fifth patent data. In this line segment type initial patent topology model, the two endpoints of the line segment are the fourth and fifth patent data, respectively, and the sides corresponding to the fourth and fifth patent data are unidirectional. Here, the initial patent topology model containing the most topological nodes among the two initial patent topology models is taken as the patent topology model.

[0097] S3. Obtain the weight of each edge in the patent topology model based on the target patent dataset to obtain a reference weight set. Optimize the patent topology model based on the reference weight set to obtain the target topology model.

[0098] Furthermore, the step of obtaining the weight of each edge in the patent topology model based on the target patent dataset to obtain a reference weight set includes:

[0099] Extract the topology nodes sequentially from the patent topology model, and perform the following operations on the extracted topology nodes:

[0100] An initial tree structure is constructed with the topology node as the root node and each patent initial field in the patent initial field set corresponding to the topology node as the first child node.

[0101] The initial claim text is extracted from the patent text corresponding to the topology node. The initial claim text is matched and replaced to obtain the target claim text. The target claim text is parsed using a pre-built language model to obtain the initial text set. The reference corpus set is identified based on the patent initial domain set. The initial text set is used to search and classify in the reference corpus set to obtain the parsed text set. The reference corpus in the reference corpus set corresponds one-to-one with the patent initial domain in the patent initial domain set.

[0102] For each parsed text group in the parsed text group set, perform the following operation:

[0103] The initial tree structure is retrieved and identified using parsed text groups to obtain the target tree structure, and a reference weight set is obtained based on the target tree structure.

[0104] It should be explained that the initial tree structure refers to a tree structure constructed with the topological node as the root node and each patent initial field in the patent initial field set as the first child node. The relationship between the first child nodes is a sibling relationship, and the technology used to construct the tree structure is prior art, which will not be elaborated further here. The initial claim text refers to the text in the patent text used to describe the claims, generally the claim sheet. Matching and replacing the initial claim text refers to retrieving the data flow and actions involved in the initial claim text within the patent text and using the search results to replace the initial claim text. It is easy to understand that after matching and replacing, the initial claim text will be replaced with complete claim text; that is, statements with citation relationships will be replaced by their explanatory content. For example, if sentences A and B in the independent claim are cited by dependent claim 2 and dependent claim 3 respectively, then sentences A and B will be replaced with the content of dependent claim 2 and dependent claim 3 respectively.

[0105] It should be understood that the language model is a model or technique used to remove unnecessary words from the target claim text. Optionally, a machine learning model and a corpus from the patent field can be used together as the language model. Other techniques can achieve the same effect, which will not be elaborated here. The initial text set refers to the text set after stop words are removed, which contains only nouns and actions. The reference corpus in the reference corpus set refers to the corpus corresponding to each patent initial field in the patent initial field set. The reference corpus can be an existing publicly available corpus or can be set by prior knowledge. Importantly, the reference corpus includes corpora corresponding to multiple different technical fields. For example, if the patent initial field is machine learning, the reference corpus may include: grayscale, binarization, image erosion, etc., corresponding to the field of image recognition technology, and text classification, sentiment recognition, etc., corresponding to the field of natural language processing technology.

[0106] It needs to be explained that retrieving and classifying the initial text set in the reference corpus to obtain the parsed text set refers to retrieving the initial text from the initial text set in the reference corpus, classifying the retrieved initial text, and obtaining parsed text sets that correspond one-to-one with the initial patent field. Using the parsed text sets to perform a retrieval and labeling operation on the initial tree structure refers to retrieving the corresponding first child node in the initial tree structure using the patent field corresponding to the parsed text set, and then labeling the retrieved first child node with the parsed text set. Generally, labeling the first child node yields a parsed text set that represents the target patent data within the patent field. Optionally, the labeling of the first child node can be achieved through text; other techniques can achieve the same effect, which will not be elaborated upon here.

[0107] It should be understood that obtaining the reference weight set based on the target tree structure includes:

[0108] Identify the technical field of each parsed text group in the parsed text group set to obtain one or more search fields. Perform the following operation on each of the one or more search fields:

[0109] The number of initial texts in the initial text set corresponding to the target tree structure is counted to obtain the statistical quantity. The parsed text groups in the parsed text group set are classified and counted using the retrieval domain to obtain the retrieval quantity set.

[0110] For each search result in the search result set, perform the following operation:

[0111] Calculate the ratio of the number of searches to the number of statistics to obtain the domain percentage value. Compare the domain percentage value with the preset domain retention threshold. If the domain retention threshold is confirmed to be greater than or equal to the domain retention threshold, then use the search domain corresponding to the domain percentage value to mark the domain percentage value to obtain the marked percentage value.

[0112] The identifier percentage values ​​are summarized to obtain an identifier percentage value set, and a reference weight set is obtained based on the identifier percentage value set.

[0113] It should be explained that the retrieval domain refers to the technical field corresponding to the parsed text group in the reference corpus, such as machine learning, image recognition, etc. It is easy to understand that since different technical fields are pre-defined in the reference corpus, and there may be overlap in the technical fields involved in different initial patent fields, the parsed text groups are categorized and statistically analyzed in this embodiment of the invention to determine the proportion of different target patent data in the technical fields. Here, the technical field refers to the specific technical points involved. For example, the initial patent field set includes machine learning and image data generation and processing. The technical field corresponding to machine learning may include image feature extraction, and the technical field corresponding to image data generation and processing may also include image feature extraction. Therefore, under the machine learning node, image feature extraction is included; under the image data generation and processing node, image feature extraction is also included. The parsed text groups whose technical field is image feature extraction are then summarized and cleaned to remove duplicate technical features. The number of parsed texts included in the cleaned parsed text group set is counted to obtain the retrieval count.

[0114] It should be understood that since the technical fields included in the target patent data may be numerous, and not all technical fields affect the novelty of the patent, this embodiment of the invention uses a preset field retention threshold to retain features that can characterize its technical field. Optionally, textual identifiers can be used to identify the proportion of fields based on the search fields. Other techniques can achieve the same effect, and will not be elaborated further here.

[0115] It should be explained that obtaining the reference weight set based on the set of identifier proportion values ​​includes:

[0116] The initial abstract text is extracted from the patent text corresponding to the topology node, and the initial abstract text is optimized using a pre-built abstract optimization model to obtain the target abstract vector.

[0117] Using the extracted topological nodes, neighboring nodes are identified in the patent topology model. The neighboring abstract vectors are obtained using the patent texts corresponding to the neighboring nodes. The cosine distance between the target abstract vector and the neighboring abstract vectors is calculated to obtain the first matching degree.

[0118] Obtain the reference proportion value set of neighboring nodes and one or more reference domains, and perform a logical intersection operation on one or more reference domains and one or more search domains to obtain one or more overlapping domains;

[0119] Using one or more overlapping regions, a first percentage value set and a second percentage value set are identified in the reference percentage value set and the identifier percentage value set, respectively.

[0120] The first restored value set is calculated using the pre-constructed scaling formula and the first set of percentage values, and the second restored value set is obtained using the second set of percentage values.

[0121] The scaling relationship is as follows:

[0122]

[0123] in, Indicates the first set of restored values. The first restored value. This indicates the first percentage value set. The first percentage value. This indicates the first percentage value set. The first percentage value. This indicates that the first percentage value is in the total. The first percentage value;

[0124] The first restored value set and the second restored value set are matched using one or more overlapping neighborhoods to obtain one or more matching nodes. Each matching node corresponds to one of the overlapping neighborhoods and includes a first restored value and a second restored value.

[0125] A second matching degree is calculated based on the one or more matching nodes. Reference weights are calculated using the first matching degree and the second matching degree. The reference weights are then summarized to obtain a reference weight set.

[0126] Further, the initial abstract text is the text in the patent text used to represent the specification abstract. The abstract optimization model refers to the ability to extract a high-dimensional semantic embedding vector representing the patent abstract from the initial abstract text; optionally, a pre-trained Sentence-BERT model is used as the abstract optimization model. The target abstract vector refers to the high-dimensional semantic embedding vector representing the initial abstract text. The neighboring node refers to a topological node that has an edge with the extracted topological node, and this edge is directional. The method for obtaining the neighboring abstract vector is the same as the method for obtaining the target abstract vector, and will not be repeated here. Calculating the cosine distance between the target abstract vector and the neighboring abstract vector refers to calculating the cosine similarity between the target abstract vector and the neighboring abstract vector, and the technique for calculating cosine similarity is existing technology, and will not be repeated here. The methods for obtaining the reference proportion value set and one or more reference domains are the same as the methods for obtaining the identifier proportion value set and one or more retrieval domains, and will not be repeated here. Performing a logical intersection operation on one or more reference domains and one or more retrieval domains refers to taking the intersection of one or more reference domains and one or more retrieval domains. It's easy to understand that because there are reference relationships between topological nodes and neighboring nodes, one or more reference domains and one or more search domains must intersect, and the resulting intersection is called the overlapping domain. The first proportion value set and the second proportion value set refer to the reference proportion value and the identifier proportion value corresponding to each overlapping domain in the reference proportion value set and the identifier proportion value set, respectively.

[0127] It should be explained that calculating the second matching degree based on the one or more matching nodes includes:

[0128] One or more matching nodes are shown below:

[0129]

[0130] in, This indicates the first matching node in a set of one or more matching nodes. This indicates the second matching node in a group of one or more matching nodes. Indicates that one or more matching nodes share a common feature. Matching nodes, They respectively represent the first matching node, the second matching node, and the third matching node in one or more matching nodes. The first restored value corresponding to each matching node. They respectively represent the first matching node, the second matching node, and the third matching node in one or more matching nodes. The second restored value corresponding to each matching node;

[0131] Calculate the Euclidean distance of each matching node in one or more matching nodes and sum them up to obtain the second matching degree.

[0132] Understandably, the first and second proportion value sets are subsets of the reference and identifier proportion value sets, respectively. Therefore, the purpose of calculating the first restored value set using the proportion scaling formula is to restore the similarity of patents with citation relationships in different technical fields. Matching the first and second restored value sets using one or more overlapping domains to obtain one or more matching nodes means matching the first restored value in the first restored value set and the second restored value in the second restored value set according to the overlapping domains, obtaining values ​​used to characterize the same technical field. Since there is a citation relationship between topological nodes and neighboring nodes, it can be considered that the number of texts involved between topological nodes and neighboring nodes in the same technical field is of reference significance. Calculating the reference weight using the first and second matching degrees means calculating the weighted sum of the reciprocals of the first and second matching degrees to obtain the reference weight. It is easy to understand that a larger first matching degree indicates a higher similarity between the two patent documents, while a smaller second matching degree indicates a higher similarity in the technical fields involved in the two patent documents. Therefore, in this embodiment of the invention, the reciprocal of the second matching degree is used to calculate the reference weight. Calculating and summing the Euclidean distances of each matching node in one or more matching nodes refers to calculating the Euclidean distances between the first and second restored values ​​of the matching nodes, and then summing them to obtain the second matching degree. Optimizing the patent topology model based on the aforementioned reference weight set to obtain the target topology model involves marking the edges in the patent topology model using their reference weights in the reference weight set. Optionally, textual marking can be used; other techniques can achieve the same effect, and will not be elaborated further here.

[0133] S4. Train the pre-built graph convolutional network model to obtain an optimized graph convolutional network model.

[0134] It should be explained that training the pre-built graph convolutional network model to obtain an optimized graph convolutional network model includes:

[0135] For each topological node in the target topology model, indexes are extracted to obtain node indices, as shown below:

[0136]

[0137] in, Indicates node metrics, Indicates degree centrality. Represents the eigenvector centrality. Indicates the degree of importance of the reference;

[0138] The network model of the target topology model is confirmed based on node indicators;

[0139] The network model of the target topology model is shown below:

[0140]

[0141] in, The network model representing the target topology model, This represents the set of topological nodes in the target topology model. This represents the set of edges in the target topology model. Represents the reference weight set;

[0142] Construct the network resilience assessment formula, as shown below:

[0143]

[0144] in, Indicates network resilience. This represents the number of nodes in the largest weakly connected subgraph in the network model. Indicates the first The maximum number of nodes in a weakly connected subgraph after removing topological nodes. Indicates the first The maximum number of nodes in a weakly connected subgraph after removing topological nodes;

[0145] An optimized graph convolutional network model is obtained based on the network model and the network damage assessment relation.

[0146] Furthermore, degree centrality is the most direct measure of node centrality in network analysis. Eigenvector centrality measures the importance or influence of nodes in a network, with reference importance referring to the value obtained using PageRank to evaluate node importance. The maximum weakly connected subgraph is the weakly connected subgraph containing the most nodes in the topology.

[0147] It is easy to understand that the main purpose of this invention is to identify important nodes in the target topology model. Therefore, the importance of the removed topology nodes in the target topology model can be identified through the network resilience assessment formula. It is easy to understand that the more important the removed network nodes are in the target topology model, the higher the calculated network resilience rate will be.

[0148] It should be explained that the method of obtaining the optimized graph convolutional network model based on the network model and the network resilience assessment formula includes:

[0149] The graph structure is extracted from the network model, as shown below:

[0150]

[0151] Where S represents the graph structure;

[0152] Using a pre-built strategy function, a node is selected for removal from multiple topological nodes corresponding to the graph structure. The strategy function is shown below:

[0153]

[0154] in, This indicates the removal of the node. Indicates the first The next node to be removed. This represents the policy function. Indicates the first The graph structure after removing the node;

[0155] By defining the time reward function for the graph convolutional network model, the target convolutional network model is obtained. The time reward function is as follows:

[0156]

[0157] in, This represents the time-reward function. Indicates the first After removing the topology nodes, the network resilience rate of the network model is calculated using the network resilience assessment formula. Indicates the first Network resilience of the network model after removing topology nodes;

[0158] An optimized graph convolutional network model is obtained by utilizing the policy function and the target convolutional network model.

[0159] It should be understood that obtaining the optimized graph convolutional network model using the policy function and the target convolutional network model includes:

[0160] The iteration probability ratio is obtained based on the policy function, where the iteration probability ratio is as follows:

[0161]

[0162] in, Indicates that at time step is The ratio of iteration probabilities at time, This represents the parameters of the current policy network. The parameters represent the old policy network. This indicates that the topology node to be removed is selected;

[0163] Construct the advantage function, and then use the advantage function and the iterative probability ratio to construct the policy pruning loss function. The advantage function is shown below:

[0164]

[0165] in, Indicates time step as The dominant function value at that time, This indicates the preset termination time step. Indicates the discount factor. Represents the GAE coefficient. This indicates that the value network corresponds to the network model at time step [missing information]. The following value estimate, This indicates that the value network corresponds to the network model at time step [missing information]. The following is a value estimate;

[0166] The target convolutional network model is defined using the dominance function and the policy pruning loss function to obtain the optimized convolutional network model. The optimized convolutional network model is then trained to obtain the optimized graph convolutional network model.

[0167] Furthermore, training the target convolutional network model refers to the iterative training of the parameters in the target convolutional network model. The parameters of the old policy network refer to the parameters before removing the topology nodes selected by the policy function, while the parameters of the current policy network refer to the parameters obtained after removing the topology nodes selected by the policy function and optimizing the parameters in the target convolutional network model. The value network refers to the Critic network, and the value estimation refers to the expected long-term cumulative reward that the agent can obtain by continuing to act according to the current policy sequence from the current network state, i.e., the graph structure at the specific topology node removal stage, calculated through the value network during reinforcement learning. The GAE coefficient refers to a hyperparameter with a value range of [0,1].

[0168] It should be understood that the pruning loss function of the strategy is as follows:

[0169]

[0170] in, This represents the policy pruning loss of the PPO algorithm. Indicates time step Expectations This represents the cutting factor.

[0171] It should be explained that PPO stands for Proximal Policy Optimization. The policy pruning loss is a loss function designed based on the probability ratio of iterations. Its purpose is to constrain the "step size" of policy updates, ensuring that the new policy always closely follows the old policy and preventing the agent's behavior from becoming uncontrollable due to excessively large single update increments. The pruning coefficient is a hyperparameter in the Proximal Policy Optimization algorithm. "Clip" refers to the clipping operation during policy updates, and its core function is to limit the fluctuation range of the probability ratio between the new and old policies.

[0172] Furthermore, the step of training the optimized convolutional network model to obtain an optimized graph convolutional network model includes:

[0173] A reference topology graph is obtained, and the optimized convolutional network model is trained based on the reference topology graph to obtain the training pruning loss sequence, the Euclidean distance sequence, and the reference network resilience sequence. Curve fitting is performed on the training pruning loss sequence, the Euclidean distance sequence, and the reference network resilience sequence to obtain the training pruning loss curve, the Euclidean distance curve, and the network resilience curve. Using a preset extraction step size, the pruning loss slope sequence is extracted from the training pruning loss curve. Using a preset pruning loss slope threshold and the pruning loss slope sequence, the target pruning loss curve is identified in the training pruning loss curve. The pruning loss slope corresponding to the target pruning loss curve is less than or equal to the pruning loss slope threshold.

[0174] Using the target clipping loss curve, the target Euclidean distance curve and the target damage resistance curve are identified from the parametric Euclidean distance curve and the network damage resistance curve, respectively.

[0175] Determine whether the parameter Euclidean distances corresponding to the target Euclidean distance curves are all less than or equal to the preset parameter Euclidean thresholds, and determine whether the reference network resilience rates corresponding to the target resilience curves are all greater than or equal to the preset reference network resilience thresholds.

[0176] If the parameter Euclidean distances corresponding to the target Euclidean distance curves are all less than or equal to the parameter Euclidean thresholds, and the reference network resilience rates corresponding to the target resilience curves are all greater than or equal to the reference network resilience thresholds, then the trained optimized convolutional network model is used as the optimized graph convolutional network model.

[0177] It should be explained that the method for obtaining the reference topology graph in the reference topology graph set is the same as the method for obtaining the target topology model, and will not be repeated here. The training pruning loss sequence refers to the sequence of policy pruning losses calculated using the policy pruning loss function when training the optimized convolutional network model. The parameter Euclidean distance sequence refers to the sequence ordered by the L2 norms of the parameters of the current policy network and the old policy network when training the optimized convolutional network model. The reference network resilience sequence refers to the sequence of network resilience calculated using the network resilience evaluation relation when training the optimized convolutional network model. It is easy to understand that in this embodiment of the invention, it is explicitly stated that when identifying key nodes in the reference topology graph, it is necessary to ensure that the network resilience is greater than or equal to the reference network resilience threshold. Therefore, the optimized convolutional network model may identify more than one node in the reference topology graph set. Furthermore, due to the setting of a time reward function, the optimized convolutional network model will identify the topology node with the greatest impact on the graph structure in the smallest possible time step. Here, the time step refers to the number of topology nodes identified. Optionally, a polynomial fitting method can be used to perform curve fitting on the training pruning loss sequence, the parametric Euclidean distance sequence, and the reference network resilience sequence, respectively, to obtain the training pruning loss curve, the parametric Euclidean distance curve, and the network resilience curve.

[0178] It should be understood that the extraction step size refers to the preset step size used to identify the pruning loss slope sequence in the training pruning loss curve. Optionally, 1 is taken as the extraction step size, where 1 refers to the reference topology graph in the reference topology graph set used to train the optimized convolutional network model.

[0179] It is understandable that when training an optimized convolutional network model using a reference topology graph, a training pruning loss and a reference network resilience can be obtained. Furthermore, a parametric Euclidean distance can be obtained between every two reference topology graphs. Therefore, there is a correspondence between the target pruning loss curve, the parametric Euclidean distance curve, and the network resilience curve. Using the target pruning loss curve, identifying the target Euclidean distance curve and the target resilience curve in the parametric Euclidean distance curve and the network resilience curve respectively means taking the parametric Euclidean distance curve and the network resilience curve in the corresponding relationship as the target Euclidean distance curve and the target resilience curve.

[0180] Furthermore, when the parameter Euclidean distances corresponding to the target Euclidean distance curves are all less than or equal to the parameter Euclidean thresholds, and the reference network resilience rates corresponding to the target resilience curves are all greater than or equal to the reference network resilience thresholds and their preconditions, it indicates that there is no room for optimization in the selection strategy of key topology node sets when training the optimized convolutional network model. It is applicable to different topology graph structures, and the network parameters tend to be stable. Therefore, it can be determined that the optimized convolutional network model has been trained, and the optimized graph convolutional network model is obtained.

[0181] S5. Using an optimized graph convolutional network model, identify the preset key nodes of the target network in the target topology model to achieve the identification of key nodes of the target network.

[0182] Furthermore, the target network critical nodes refer to multiple topology nodes identified in the target topology model using an optimized graph convolutional network model, wherein the ratio of the number of the multiple topology nodes to the total number of all topology nodes is less than or equal to a preset value, and after removing the identified multiple topology nodes, the network resilience of the target topology model is greater than or equal to the reference network resilience threshold.

[0183] To address the problems described in the background art, this invention identifies a target patent dataset to be analyzed. This dataset includes multiple target patent data points, each containing inter-patent citation relationships, an initial patent domain set, and patent text. A patent topology model is constructed using the target patent data points as nodes and the inter-patent citation relationships as edges. This model includes multiple topology nodes, each corresponding one-to-one with a target patent data point. The weight of each edge in the patent topology model is obtained based on the target patent dataset, resulting in a reference weight set. The patent topology model is then optimized based on this reference weight set to obtain the target topology model. This invention considers the dynamic nature of the technology network. Therefore, in this embodiment, the target patent dataset to be analyzed is first identified, and this dataset can be manually set. Furthermore, this invention integrates a graph construction method combining structural topology and semantic similarity. Based on patent citation relationships, a summary embedding vector and the patent's technical field are integrated as edge weights, forming a weighted network with both semantic and technical heterogeneous features. Furthermore, the weights of edges between nodes in different topologies are increased, making the weighted network more accurate, and consequently improving the accuracy of training graph convolutional network models using the weighted network. An optimized graph convolutional network model is obtained by training a pre-constructed graph convolutional network model. Using this optimized model, preset key nodes of the target network are identified in the target topology model. This invention constructs a deep reinforcement learning framework with GCN as the policy network and PPO as the optimizer. It combines centrality features and edge weights as inputs, introduces a custom graph environment module to simulate the node removal process, and uses the change in the proportion of the most connected subgraph to measure structural resilience as a reward signal to guide the policy model in learning the optimal node selection path, achieving efficient identification of key nodes and policy generalization. Therefore, this invention can improve the accuracy of identifying key network nodes.

[0184] like Figure 2 The diagram shown is a functional block diagram of a network critical node identification system for technical network analysis provided in an embodiment of the present invention.

[0185] The network critical node identification system 100 for technical network analysis described in this invention can be installed in an electronic device. Depending on the functions implemented, the network critical node identification system 100 for technical network analysis may include a data collection module 101, a data processing module 102, a graph model construction module 103, and a network training and feature recognition module 104. The module described in this invention can also be referred to as a unit, which refers to a series of computer program segments that can be executed by the processor of an electronic device and can perform a fixed function, and which are stored in the memory of the electronic device.

[0186] The data collection module 101 is used to identify the target patent dataset to be analyzed, wherein the target patent dataset includes multiple target patent data, and the target patent data includes patent citation relationships, patent initial field sets and patent text;

[0187] The data processing module 102 is used to construct a patent topology model with the target patent data in the target patent dataset as nodes and the patent citation relationship between the target patent data in the target patent dataset as edges. The patent topology model includes multiple topology nodes, and the topology nodes correspond one-to-one with the target patent data.

[0188] The graph model construction module 103 is used to obtain the weight of each edge in the patent topology model based on the target patent dataset, obtain a reference weight set, and optimize the patent topology model based on the reference weight set to obtain the target topology model.

[0189] The network training and feature recognition module 104 is used to train the pre-constructed graph convolutional network model to obtain an optimized graph convolutional network model.

[0190] By utilizing an optimized graph convolutional network model, the key nodes of the target network are identified in the target topology model, thereby achieving the identification of key nodes of the target network.

[0191] In detail, the modules in the network critical node identification system 100 for technical network analysis described in this embodiment of the invention employ the same methods as described above. Figure 1 The method used for identifying key network nodes in technical network analysis is the same as the method described in the previous section and can produce the same technical effect, so it will not be repeated here.

[0192] like Figure 3 The diagram shown is a structural schematic of an electronic device that implements a method for identifying key network nodes for technical network analysis, according to an embodiment of the present invention.

[0193] The electronic device 1 may include a processor 10, a memory 11 and a bus 12, and may also include a computer program stored in the memory 11 and executable on the processor 10, such as a network critical node identification method program for technical network analysis.

[0194] The memory 11 includes at least one type of readable storage medium, such as flash memory, portable hard drive, multimedia card, card-type memory (e.g., SD or DX memory), magnetic memory, magnetic disk, optical disk, etc. In some embodiments, the memory 11 can be an internal storage unit of the electronic device 1, such as a portable hard drive. In other embodiments, the memory 11 can be an external storage device of the electronic device 1, such as a plug-in portable hard drive, smart media card (SMC), secure digital card (SD), flash card, etc., equipped on the electronic device 1. Furthermore, the memory 11 includes both internal storage units and external storage devices of the electronic device 1. The memory 11 can be used not only to store application software and various types of data installed on the electronic device 1, such as code for a network key node identification method program used for technical network analysis, but also to temporarily store data that has been output or will be output.

[0195] In some embodiments, the processor 10 may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits with the same or different functions, including combinations of one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and various control chips. The processor 10 is the control unit of the electronic device, connecting various components of the entire electronic device via various interfaces and lines. It executes programs or modules stored in the memory 11 (e.g., a network key node identification method program for technical network analysis) and calls data stored in the memory 11 to perform various functions of the electronic device 1 and process data.

[0196] The bus 12 can be a peripheral component interconnect (PCI) bus or an extended industry standard architecture (EISA) bus, etc. The bus 12 can be divided into an address bus, a data bus, a control bus, etc. The bus 12 is configured to realize the connection and communication between the memory 11 and at least one processor 10, etc.

[0197] Figure 3 Only electronic devices with components are shown; those skilled in the art will understand that... Figure 3The structure shown does not constitute a limitation on the electronic device 1, and may include fewer or more components than shown, or combine certain components, or have different component arrangements.

[0198] For example, although not shown, the electronic device 1 may also include a power supply (such as a battery) to power the various components. Preferably, the power supply can be logically connected to the at least one processor 10 through a power management system, thereby enabling functions such as charging management, discharging management, and power consumption management through the power management system. The power supply may also include one or more DC or AC power supplies, recharging systems, power fault detection circuits, power converters or inverters, power status indicators, and other arbitrary components. The electronic device 1 may also include various sensors, Bluetooth modules, Wi-Fi modules, etc., which will not be described in detail here.

[0199] Furthermore, the electronic device 1 may also include a network interface. Optionally, the network interface may include a wired interface and / or a wireless interface (such as a Wi-Fi interface, a Bluetooth interface, etc.), which is typically used to establish communication connections between the electronic device 1 and other electronic devices.

[0200] Optionally, the electronic device 1 may further include a user interface, which may be a display, an input unit (such as a keyboard), and optionally, a standard wired interface or a wireless interface. Optionally, in some embodiments, the display may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen, etc. The display may also be appropriately referred to as a screen or display unit, used to display information processed in the electronic device 1 and to display a visual user interface.

[0201] The network key node identification method program for technical network analysis stored in the memory 11 of the electronic device 1 is a combination of multiple instructions, which, when run in the processor 10, can achieve the following:

[0202] The target patent dataset to be analyzed has been identified. The target patent dataset includes multiple target patent data, and the target patent data includes the citation relationship between patents, the initial patent field set, and the patent text.

[0203] A patent topology model is constructed using the target patent data in the target patent dataset as nodes and the patent citation relationships between the target patent data in the target patent dataset as edges. The patent topology model includes multiple topology nodes, and each topology node corresponds one-to-one with the target patent data.

[0204] The weight of each edge in the patent topology model is obtained based on the target patent dataset to obtain a reference weight set. The patent topology model is then optimized based on the reference weight set to obtain the target topology model.

[0205] The pre-built graph convolutional network model is trained to obtain an optimized graph convolutional network model;

[0206] By utilizing an optimized graph convolutional network model, the key nodes of the target network are identified in the target topology model, thereby achieving the identification of key nodes of the target network.

[0207] Specifically, the processor 10's implementation method for the above instructions can be found in [reference needed]. Figures 1 to 3 The descriptions of the relevant steps in the corresponding embodiments are not repeated here.

[0208] Furthermore, if the modules / units integrated in the electronic device 1 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. The computer-readable storage medium can be volatile or non-volatile. For example, the computer-readable medium may include: any entity or system capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, or a read-only memory (ROM).

[0209] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor of an electronic device, can perform the following:

[0210] The target patent dataset to be analyzed has been identified. The target patent dataset includes multiple target patent data, and the target patent data includes the citation relationship between patents, the initial patent field set, and the patent text.

[0211] A patent topology model is constructed using the target patent data in the target patent dataset as nodes and the patent citation relationships between the target patent data in the target patent dataset as edges. The patent topology model includes multiple topology nodes, and each topology node corresponds one-to-one with the target patent data.

[0212] The weight of each edge in the patent topology model is obtained based on the target patent dataset to obtain a reference weight set. The patent topology model is then optimized based on the reference weight set to obtain the target topology model.

[0213] The pre-built graph convolutional network model is trained to obtain an optimized graph convolutional network model;

[0214] By utilizing an optimized graph convolutional network model, the key nodes of the target network are identified in the target topology model, thereby achieving the identification of key nodes of the target network.

[0215] In the embodiments provided by this invention, it should be understood that the disclosed devices, systems, and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative, and actual implementations may have other classification methods.

[0216] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0217] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0218] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0219] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for identifying key network nodes in technical network analysis, characterized in that, The method includes: The target patent dataset to be analyzed has been identified. The target patent dataset includes multiple target patent data, and the target patent data includes the citation relationship between patents, the initial patent field set, and the patent text. A patent topology model is constructed using the target patent data in the target patent dataset as nodes and the patent citation relationships between the target patent data in the target patent dataset as edges. The patent topology model includes multiple topology nodes, and each topology node corresponds one-to-one with the target patent data. The weight of each edge in the patent topology model is obtained based on the target patent dataset to obtain a reference weight set. The patent topology model is then optimized based on the reference weight set to obtain the target topology model. The pre-built graph convolutional network model is trained to obtain an optimized graph convolutional network model; The step of training a pre-constructed graph convolutional network model to obtain an optimized graph convolutional network model includes: For each topological node in the target topology model, indexes are extracted to obtain node indices, as shown below: in, Indicates node metrics, Indicates degree centrality. Represents the eigenvector centrality. Indicates the degree of importance of the reference; The network model of the target topology model is confirmed based on node indicators; The network model of the target topology model is shown below: in, The network model representing the target topology model, This represents the set of topological nodes in the target topology model. This represents the set of edges in the target topology model. Represents the reference weight set; Construct the network resilience assessment formula, as shown below: in, Indicates network resilience. This represents the number of nodes in the largest weakly connected subgraph in the network model. Indicates the first The maximum number of nodes in a weakly connected subgraph after removing topological nodes. Indicates the first The maximum number of nodes in a weakly connected subgraph after removing topological nodes; Optimized graph convolutional network model is obtained based on network model and network resilience assessment relation; The method for obtaining the optimized graph convolutional network model based on the network model and the network resilience assessment formula includes: The graph structure is extracted from the network model, as shown below: Where S represents the graph structure; Using a pre-built strategy function, a node is selected for removal from multiple topological nodes corresponding to the graph structure. The strategy function is shown below: in, This indicates the removal of the node. Indicates the first The next node to be removed. This represents the policy function. Indicates the first The graph structure after removing the removed nodes; By defining the time reward function for the graph convolutional network model, the target convolutional network model is obtained. The time reward function is as follows: in, This represents the time-reward function. Indicates the first After removing the topology nodes, the network resilience rate of the network model is calculated using the network resilience assessment formula. Indicates the first Network resilience of the network model after removing topology nodes; An optimized graph convolutional network model is obtained by utilizing the policy function and the target convolutional network model. By utilizing an optimized graph convolutional network model, the key nodes of the target network are identified in the target topology model, thereby achieving the identification of key nodes of the target network.

2. The method for identifying key network nodes for technical network analysis as described in claim 1, characterized in that, The step of obtaining the weight of each edge in the patent topology model based on the target patent dataset to obtain a reference weight set includes: Extract the topology nodes sequentially from the patent topology model, and perform the following operations on the extracted topology nodes: An initial tree structure is constructed with the topology node as the root node and each patent initial field in the patent initial field set corresponding to the topology node as the first child node. The initial claim text is extracted from the patent text corresponding to the topology node. The initial claim text is matched and replaced to obtain the target claim text. The target claim text is parsed using a pre-built language model to obtain the initial text set. The reference corpus set is identified based on the patent initial domain set. The initial text set is used to search and classify in the reference corpus set to obtain the parsed text set. The reference corpus in the reference corpus set corresponds one-to-one with the patent initial domain in the patent initial domain set. For each parsed text group in the parsed text group set, perform the following operation: The initial tree structure is retrieved and identified using parsed text groups to obtain the target tree structure, and a reference weight set is obtained based on the target tree structure.

3. The method for identifying key network nodes for technical network analysis as described in claim 2, characterized in that, The process of obtaining the reference weight set based on the target tree structure includes: Identify the technical field of each parsed text group in the parsed text group set to obtain one or more search fields. Perform the following operation on each of the one or more search fields: The number of initial texts in the initial text set corresponding to the target tree structure is counted to obtain the statistical quantity. The parsed text groups in the parsed text group set are classified and counted using the retrieval domain to obtain the retrieval quantity set. For each search result in the search result set, perform the following operation: Calculate the ratio of the number of searches to the number of statistics to obtain the domain percentage value. Compare the domain percentage value with the preset domain retention threshold. If the domain retention threshold is confirmed to be greater than or equal to the domain retention threshold, then use the search domain corresponding to the domain percentage value to mark the domain percentage value to obtain the marked percentage value. The identifier percentage values ​​are summarized to obtain an identifier percentage value set, and a reference weight set is obtained based on the identifier percentage value set.

4. The method for identifying key network nodes for technical network analysis as described in claim 3, characterized in that, The step of obtaining the reference weight set based on the identifier proportion value set includes: The initial abstract text is extracted from the patent text corresponding to the topology node, and the initial abstract text is optimized using a pre-built abstract optimization model to obtain the target abstract vector. Using the extracted topological nodes, neighboring nodes are identified in the patent topology model. The neighboring abstract vectors are obtained using the patent texts corresponding to the neighboring nodes. The cosine distance between the target abstract vector and the neighboring abstract vectors is calculated to obtain the first matching degree. Obtain the reference proportion value set of neighboring nodes and one or more reference domains, and perform a logical intersection operation on one or more reference domains and one or more search domains to obtain one or more overlapping domains; Using one or more overlapping regions, a first percentage value set and a second percentage value set are identified in the reference percentage value set and the identifier percentage value set, respectively. The first restored value set is calculated using the pre-constructed scaling formula and the first set of percentage values, and the second restored value set is obtained using the second set of percentage values. The scaling relationship is as follows: in, Indicates the first set of restored values. The first restored value. This indicates the first percentage value set. The first percentage value. This indicates the first percentage value set. The first percentage value. This indicates that the first percentage value set has a total of The first percentage value; The first restored value set and the second restored value set are matched using one or more overlapping neighborhoods to obtain one or more matching nodes. Each matching node corresponds to one of the overlapping neighborhoods and includes a first restored value and a second restored value. A second matching degree is calculated based on the one or more matching nodes. Reference weights are calculated using the first matching degree and the second matching degree. The reference weights are then summarized to obtain a reference weight set.

5. The method for identifying key network nodes for technical network analysis as described in claim 4, characterized in that, The calculation of the second matching degree based on the one or more matching nodes includes: One or more matching nodes are shown below: in, This indicates the first matching node in a set of one or more matching nodes. This indicates the second matching node in a group of one or more matching nodes. Indicates that one or more matching nodes share a common feature. Matching nodes, They respectively represent the first matching node, the second matching node, and the third matching node in one or more matching nodes. The first restored value corresponding to each matching node. They respectively represent the first matching node, the second matching node, and the third matching node in one or more matching nodes. The second restored value corresponding to each matching node; Calculate the Euclidean distance of each matching node in one or more matching nodes and sum them up to obtain the second matching degree.

6. The method for identifying key network nodes for technical network analysis as described in claim 5, characterized in that, The process of obtaining an optimized graph convolutional network model using a policy function and a target convolutional network model includes: The iteration probability ratio is obtained based on the policy function, where the iteration probability ratio is as follows: in, Indicates that at time step is The ratio of iteration probabilities at time, This represents the parameters of the current policy network. The parameters represent the old policy network. This indicates that the topology node to be removed is selected; Construct the advantage function, and then use the advantage function and the iterative probability ratio to construct the policy pruning loss function. The advantage function is shown below: in, Indicates the time step as The dominant function value at that time, This indicates the preset termination time step. Indicates the discount factor. Represents the GAE coefficient. This indicates that the value network corresponds to the network model at time step [missing information]. The following value estimate, This indicates that the value network corresponds to the network model at time step [missing information]. The following is a value estimate; The target convolutional network model is defined using the dominance function and the policy pruning loss function to obtain the optimized convolutional network model. The optimized convolutional network model is then trained to obtain the optimized graph convolutional network model.

7. The method for identifying key network nodes for technical network analysis as described in claim 6, characterized in that, The pruning loss function for the strategy is as follows: in, This represents the policy pruning loss of the PPO algorithm. Indicates time step Expectations This represents the cutting factor.

8. A network critical node identification system for technical network analysis, characterized in that, The system includes: The data collection module is used to identify the target patent dataset to be analyzed. The target patent dataset includes multiple target patent data, and the target patent data includes the citation relationship between patents, the initial field set of patents, and the patent text. The data processing module is used to construct a patent topology model with the target patent data in the target patent dataset as nodes and the patent citation relationship between the target patent data in the target patent dataset as edges. The patent topology model includes multiple topology nodes, and each topology node corresponds one-to-one with the target patent data. The graph model construction module is used to obtain the weight of each edge in the patent topology model based on the target patent dataset, obtain a reference weight set, and optimize the patent topology model based on the reference weight set to obtain the target topology model. The network training and feature recognition module is used to train a pre-built graph convolutional network model to obtain an optimized graph convolutional network model. The step of training a pre-constructed graph convolutional network model to obtain an optimized graph convolutional network model includes: For each topological node in the target topology model, indexes are extracted to obtain node indices, as shown below: in, Indicates node metrics, Indicates degree centrality. Represents the eigenvector centrality. Indicates the degree of importance of the reference; The network model of the target topology model is confirmed based on node indicators; The network model of the target topology model is shown below: in, The network model representing the target topology model, This represents the set of topological nodes in the target topology model. This represents the set of edges in the target topology model. Represents the reference weight set; Construct the network resilience assessment formula, as shown below: in, Indicates network resilience. This represents the number of nodes in the largest weakly connected subgraph in the network model. Indicates the first The maximum number of nodes in a weakly connected subgraph after removing topological nodes. Indicates the first The maximum number of nodes in a weakly connected subgraph after removing topological nodes; Optimized graph convolutional network model is obtained based on network model and network resilience assessment relation; The method for obtaining the optimized graph convolutional network model based on the network model and the network resilience assessment formula includes: The graph structure is extracted from the network model, as shown below: Where S represents the graph structure; Using a pre-built strategy function, a node is selected for removal from multiple topological nodes corresponding to the graph structure. The strategy function is shown below: in, This indicates the removal of the node. Indicates the first The next node to be removed. This represents the policy function. Indicates the first The graph structure after removing the removed nodes; By defining the time reward function for the graph convolutional network model, the target convolutional network model is obtained. The time reward function is as follows: in, This represents the time-reward function. Indicates the first After removing the topology nodes, the network resilience rate of the network model is calculated using the network resilience assessment formula. Indicates the first Network resilience of the network model after removing topology nodes; An optimized graph convolutional network model is obtained by utilizing the policy function and the target convolutional network model. By utilizing an optimized graph convolutional network model, the key nodes of the target network are identified in the target topology model, thereby achieving the identification of key nodes of the target network.

Images