Construction method, technical opportunity identification method and system of classification co-occurrence network

Abstract

The application discloses a construction method and a system of a classification co-occurrence network and a technical opportunity identification method, and belongs to the field of technical opportunity identification. Semantic information and co-occurrence information of a joint patent classification (CPC) are combined to form a CPC co-occurrence network, hidden connection modes between CPC nodes are mined through a graph neural network model, then CPC nodes that are likely to be connected with an artificial node representing a target field are predicted, and finally technical development points that are likely to appear in the future in the target field are identified, so that certain decision support is provided for the formulation of a technical development strategy. The application can fully combine the semantic information and the co-occurrence information of the CPC to mine potential correlations between technologies, rich node features provide a good basis for learning of the model, and the technical opportunity identification can be better performed.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of technical opportunity identification, and more particularly relates to a construction method of a classification code co-occurrence network, a technical opportunity identification method and system. BACKGROUND

[0002] A technical opportunity is discovered by mining the development trend and mutual relationship of existing technologies in a certain technical field, finding the latest technical trend, and inferring the possible technical form or technical development point in the field. As the basis for innovation decision, a technical opportunity is an important factor that must be considered for any technical innovation.

[0003] The main methods for identifying technical opportunities currently include five categories: abnormal value detection, patent map, link prediction, scientific and technological association method, and composite method. The core steps of the abnormal value detection method include identification and evaluation of outlier patents. The identification step mainly relies on quantitative data mining methods, while the evaluation step is mostly based on qualitative methods such as expert experience or TRIZ technical system evolution rules. Different knowledge accumulation of experts will lead to strong subjectivity of the analysis results. The patent map method is based on the vectorization of patent keywords, followed by dimensionality reduction visualization, and then finding the blank area in the patent map as a potential technical opportunity. However, it is still difficult to obtain a clear technical meaning from the "patent blank point" represented by a group of keywords around it. The link prediction method uses network structure information to predict the possibility of connection between two nodes in the network that have not occurred yet. The process of finding potential technical opportunities is to find the missing connections in the network. This method mainly has two directions: combination of link prediction method and co-word network based on text content, and combination of link prediction method and citation network based on patent classification number. The former takes the technical association represented by "keyword pairs" as the final prediction result, which may cause deviation in the understanding of the technical development trend by R&D personnel due to the ambiguous semantic relationship. The latter represents technical opportunities in the form of "patent classification number pairs", but it is not conducive to refining specific technical details due to the macroscopic meaning of patent classification numbers. The scientific and technological association method mainly finds the theme that is more active in scientific research but has not yet appeared in the technical field by mining the technical evolution path of papers and patent data and performing correlation analysis. The clustering method used only retains the main part of the data, which may lose valuable information in the sparse citation field. The composite method is a combination of multiple analysis methods for technical opportunity identification, which can improve the accuracy of identification by complementary advantages between methods. It is a direction that has not yet matured and still needs to be explored. SUMMARY

[0004] In view of the defects of the prior art and the improvement needs, the present application provides a construction method of a classification code co-occurrence network, a technical opportunity identification method and system, which aims to combine the semantic information and co-occurrence information of the Cooperative Patent Classification (CPC) to form a CPC co-occurrence network, then mine the hidden connection patterns between the CPC nodes through a graph neural network model, and then predict the CPC nodes that may be connected with the artificial nodes representing the target field, so as to finally identify the technical development points that may appear in the future in the target field and provide certain decision support for the formulation of technical development strategies. The present application can fully combine the semantic information and co-occurrence information of the CPC to mine the potential associations between technologies, and the rich node features provide a good foundation for the learning of the model, which can help better identify technical opportunities.

[0005] To achieve the above-mentioned purpose, according to the first aspect of the present application, a construction method of a classification code co-occurrence network is provided, which comprises the following steps:

[0006] S1. For each patent in the target field and the field to be analyzed, determine the classification code co-occurrence information of the patent; for all patents in the target field and the field to be identified, count all the classification codes that appear to form a classification code set;

[0007] S2. For each classification code in the classification code set, based on the text information of all patents in which the classification code appears, construct the text information of the classification code;

[0008] S3. Take the classification code as a document and the text information of the classification code as a word, train a first Doc2vec model to obtain the semantic vector of each classification code in the classification code set; take the patent as a document and the classification code as a word, train a second Doc2vec model to obtain the co-occurrence vector of each classification code in the classification code set;

[0009] S4. Combine the semantic vector and the co-occurrence vector of each classification code in the classification code set into a classification code vector corresponding to the classification code;

[0010] S5. Abstract each classification code in the classification code set as a network node, and take the corresponding classification code vector as the attribute of the network node; if the co-occurrence times between two network nodes exceed a set threshold, add an edge between the network nodes corresponding to the two classification codes, thereby realizing the construction of the classification code co-occurrence network.

[0011] To achieve the above-mentioned purpose, according to the second aspect of the present application, a technical opportunity identification method is provided, which comprises the following steps:

[0012] S1. Constructing a classification number co-occurrence network of a past time period, a classification number co-occurrence network of a current time period, and a classification number co-occurrence network containing the past time period and the current time period respectively by using the method of the first aspect;

[0013] S2. Taking a node pair corresponding to an edge that does not exist in the past time period but exists in the current time period as a positive sample, and taking a node pair corresponding to an edge that does not exist in the past time period and the current time period as a negative sample, to obtain a training sample set;

[0014] S3. Statistically counting all classification numbers appearing in all patents in the target field to form a target field classification number set;

[0015] S4. Weighted averaging the classification number vectors corresponding to the classification numbers in the target field classification number set to obtain a target field node vector, adding the target field node to the classification number co-occurrence network containing the past time period and the current time period, adding edges between the target field node and all nodes in the target field classification number set, and corresponding to all nodes outside the target field classification number set one by one to generate a test sample set;

[0016] S5. Inputting each test sample in the test sample set into the graph neural network model trained by the training sample set to obtain the probability of each test sample generating an edge.

[0017] Preferably, in step S4, the target field node is added to the classification number co-occurrence network containing the past time period and the current time period by the following way:

[0018] (1) Counting the number of occurrences of each classification number in the target field classification number set;

[0019] (2) Taking the number of occurrences of a specific classification number divided by the maximum value of the number of occurrences of the classification numbers as the weight of the classification number;

[0020] (3) Weighted averaging the classification number vectors corresponding to the classification numbers in the target field classification number set to obtain a target field node vector;

[0021] (4) Adding edges between the target field node and all nodes in the target field classification number set, and the weight of the edge is the weight of the corresponding classification number.

[0022] Beneficial effects: The application constructs a virtual node representing the target field by aggregating the classification number vectors contained in the target field, providing a simple and effective combination of technical elements. Compared with the method of clustering the classification numbers of the target field and then performing correlation measurement, this method can directly measure the correlation between the target field and the classification number, avoiding the problem of information loss. In addition, the introduction of weight highlights the important technical information composition in the target field, so that the final recognition result can focus on the core technical direction of the target field.

[0023] Preferably, the graph neural network model is trained by the following method:

[0024] 1) Construct a closed subgraph of node pairs:

[0025] The closed subgraph is obtained by searching k-order neighbor nodes outward from the center node that needs to predict the link, and traversing each positive sample node pair and negative sample node pair, saving all node indexes in the closed subgraph of each node pair;

[0026] 2) Distinguish the relative position of nodes:

[0027] Distinguish importance by node marking method, the formula is as follows:

[0028] d = d x +d y

[0029] Label(i) = 1 + min(d x ,d y ) + (d / 2) [(d / 2) + (d%2) -1]

[0030] In the formula, d x , d y are the distances of nodes in the closed subgraph to two center nodes x and y, Label(i) represents the label of node i, and the labels of center nodes x and y are defined as 1; if for node i, (d x ,d y ) = (1,1), then assign label Label(i) = 2; if for node i, (d x ,d y ) = (1,2) or (d x ,d y ) = (2,1), then assign label Label(i) = 3; and so on.

[0031] 3) Obtain the adjacency matrix and node information matrix of the closed subgraph:

[0032] The adjacency matrix contains the information of edges between nodes in the closed subgraph; the node information matrix contains the characteristics of all nodes in the closed subgraph, and one row in the matrix represents a node, and each row is composed of a semantic vector, a co-occurrence vector and a label; one training sample of the graph neural network model is the adjacency matrix and the node information matrix of the closed subgraph.

[0033] Preferably, the technical opportunity identification method further comprises:

[0034] Screening the classification numbers that may become technical opportunities, forming a primary technology list, the primary technology being the technical connotation of the classification number;

[0035] Searching for patents containing technical opportunity classification numbers in the target field to extract specific technical items, and forming a secondary technology list, the secondary technology being the technical items extracted from the patents corresponding to the classification numbers.

[0036] Beneficial effects: The technical opportunity identification method provided by the present application further screens the output to obtain a technology list with a secondary structure, providing more specific technical information.

[0037] Preferably, the secondary technology list is obtained by:

[0038] For each classification number in the primary technology list, all patents in the technical opportunity field corresponding to the classification number within 5 years are obtained, and a topological clustering algorithm is used to cluster the obtained patents into multiple clusters, each cluster corresponding to a sub-field of the technical opportunity;

[0039] The word frequency of the words extracted from the patent text in each cluster is counted respectively to obtain a high-frequency word set of each cluster;

[0040] According to the high-frequency word set, patents with high-frequency word occurrence frequency exceeding a threshold value are selected from each cluster;

[0041] For each patent selected from each cluster, the importance of the patent is calculated by singular value centrality;

[0042] For each cluster after high-frequency word screening, patents with importance exceeding a threshold value are selected from each cluster;

[0043] For the patents in each cluster after two layers of screening, specific technical items are extracted from the patent text to generate multiple secondary technologies subordinate to the classification number.

[0044] Preferably, the technical opportunity identification method further comprises:

[0045] Using all patents under the technical opportunity classification number and target field patents to construct a patent citation network;

[0046] All data is divided into patents containing only target field classification numbers, patents containing only technical opportunity classification numbers, and patents containing both target field classification numbers and technical opportunity classification numbers, and patents are clustered into multiple clusters using a topology clustering algorithm, and the clusters are defined as technical topics;

[0047] The aggregation trend changes of the three types of nodes over time are analyzed and added to the technology list as a description of the development trend of the primary technical opportunity.

[0048] Beneficial effects: The present application uses a time series-based patent citation network clustering technique to mine the relationship between technical opportunities and target fields, providing the trend of combining technical opportunities and target fields, rather than just considering CPC with relatively broad connotations as technical opportunities, providing more effective decision support for the development of technical development strategies.

[0049] To achieve the above-mentioned purpose, according to the third aspect of the present application, a technical opportunity identification system is provided, comprising: a computer readable storage medium and a processor;

[0050] The computer readable storage medium is used to store executable instructions;

[0051] The processor is used to read the executable instructions stored in the computer readable storage medium, and execute the technical opportunity identification method described in the second aspect.

[0052] To achieve the above-mentioned purpose, according to the fourth aspect of the present application, a computer readable storage medium is provided, and the computer readable storage medium stores a computer program, and the computer program is executed by a processor to realize the construction method of the classification number co-occurrence network according to the first aspect or the technical opportunity identification method according to the second aspect.

[0053] Overall, the above technical solutions conceived by the present application can achieve the following beneficial effects:

[0054] (1) The present application proposes a construction method of a classification number co-occurrence network, which uses machine learning methods to generate classification number node features containing network structure information and text information, providing a more abundant information source and a good foundation for model learning, which can improve the comprehensiveness of the technical opportunity identification results to a certain extent.

[0055] (2) The present application proposes a technical opportunity identification method, which uses a current advanced graph neural network model for technical opportunity identification. This model has good adaptability to different types of networks, avoids the relatively single prediction mode of heuristic link prediction methods based on specific assumptions, can automatically learn the diversified potential relationship patterns between nodes, avoids the constraints brought by traditional link prediction based on heuristic methods based on specific assumptions, and can provide more accurate and reliable identification results. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 is a technical opportunity identification flowchart based on a joint patent classification co-occurrence network;

[0057] Figure 2 is an architecture diagram of a Doc2vec model extracting a CPC vector;

[0058] Figure 3 is a patent citation network clustering analysis flowchart. DETAILED DESCRIPTION

[0059] In order to make the objectives, technical solutions and advantages of the present application clearer, further detailed description will be made to the present application in combination with the accompanying drawings and examples. It should be understood that the specific examples described herein are only used to explain the present application and do not limit the present application. In addition, the technical features involved in each embodiment of the present application described below can be combined with each other as long as they do not conflict with each other.

[0060] The present application proposes a construction method of a classification number co-occurrence network and a technical opportunity identification method, which is suitable for any type of patent classification number, such as IPC, EPC, CPC, etc.

[0061] The preferred technical opportunity identification method of the present embodiment fuses the semantic information of the joint patent classification (CPC) and the co-occurrence information. The main idea is to extract the CPC vector using the semantic information and the co-occurrence information of the joint patent classification, then construct the CPC co-occurrence network, then mine the hidden connection patterns between the CPC nodes through the graph neural network, then predict the technical opportunity CPC that may be connected with the artificial node representing the target field, finally identify the technical opportunity that may appear in the future in the target field, and extract the secondary technical list. As shown in Figure 1 The method comprises the following steps:

[0062] S1. Divide all patent data into a t1 time period (before 2015) and a t2 time period (2015-2020), then respectively perform preprocessing: on the one hand, extract the CPC contained in each patent to obtain CPC co-occurrence information. On the other hand, perform text segmentation and stop word removal processing on the title and abstract of the patent, then according to the CPC contained in the patent, allocate the patent text to obtain the semantic information corresponding to the CPC;

[0063] S2. Use the machine learning Doc2vec model to learn the semantic vector and co-occurrence vector of the CPC;

[0064] S3. Combine the semantic vectors of CPCs and co-occurrence vectors to construct a co-occurrence network. The nodes of the network represent CPCs, and the edges between nodes represent the co-occurrence of the two CPCs in patents. First, according to the earliest time of the appearance of the edges between nodes, they are divided into t1 time period and t2 time period, and the CPC co-occurrence networks of t1 time period and t2 time period are constructed respectively;

[0065] S4. Obtain training data from the constructed CPC co-occurrence network. Node pairs that have no edges in t1 time period but have edges in t2 time period are positive samples, and node pairs that have no edges in t1 and t2 time periods are negative samples. The number of positive and negative samples is the same, and they are divided into training set and test set in the ratio of 9:1;

[0066] S5. Select the target field of study, count the CPC vectors contained in the patents of the target field to obtain the target field vector set, take the mean to obtain the artificial node vector of the target field, add the vector to the co-occurrence network of t2 time period to generate a new prediction network, and extract the new artificial node and CPC node pairs outside the target field CPC set from the prediction network as prediction data;

[0067] S6. Preprocess the divided training data, then input it into the graph neural network model for iterative training, and use the trained model to predict the possibility of connection between the artificial node and the CPC node outside the target field;

[0068] S7. Combine expert knowledge to screen CPCs that may become technology opportunities, and refine specific technology items from the patents corresponding to the CPCs to form a technology list with a two-level structure. The first-level technology is the technical connotation of the CPC, and the second-level technology is the technology item extracted from the patents corresponding to the CPC;

[0069] S8. Use all patents under the technology opportunity CPC and target field patents to construct a patent citation network, divide all data into three time slices, and use a topology clustering algorithm to cluster patents into multiple clusters, respectively. Compare the distribution of patents in different clusters over time to obtain the combination trend of technology opportunities and target fields.

[0070] In step S1, two aspects of patent data preprocessing work are mainly performed:

[0071] 1) CPC co-occurrence information extraction: CPC of a patent can be decomposed into "department - major category - minor category - major group - minor group" five parts. Considering that using complete CPC will lead to too small technical granularity, the CPC of all patents is truncated to obtain 8-bit CPC, which corresponds to the information of "department - major category - minor category - major group" four parts. The data form used for CPC co-occurrence information extraction is: {Patent number: a list containing multiple 8-bit CPCs}. Since the window size of the subsequent Doc2vec model is k, filter out the patent data with less than k CPCs (such as k = 3).

[0072] 2) CPC semantic information extraction: Building CPC semantic information needs to use the title and abstract of the patent. The patent text preprocessing method is: a. Merge the title and abstract of the patent; b. Remove punctuation marks and stop words, and perform morphological reduction; c. Segment the continuous text, and store a patent as a list containing multiple words. After processing the text information of each patent, assign the patent text according to the CPC contained by the patent. A specific CPC text information is obtained by splicing the text information of multiple patents. The data form used for CPC semantic information extraction is: {CPC: CPC text information}.

[0073] In step S2, two Doc2vec models need to be trained respectively for the semantic information and co-occurrence information of CPC. This model can vectorize all documents and words contained in the documents in the corpus, and its application method is migrated to the present application. As shown in Figure 2 , when extracting semantic information, CPC is regarded as a document, and CPC text information is regarded as a word; when extracting co-occurrence information, a patent is regarded as a document, and CPC is regarded as a word. That is, CPC is trained as a document and a word vector in two different models. Extract the semantic vector and co-occurrence vector of CPC as the features of the nodes in the network. The main parameters of the model are the window size and the output vector dimension. Considering that the CPC text information will be relatively long, the window size k1 of the Doc2vec model for extracting semantic information can be relatively large (such as k1 = 20), and the window size k2 of the Doc2vec model for extracting co-occurrence information can be small (such as k2 = 3). The dimensions n of the document vector and the word vector output by the two models can be consistent (such as n = 50). CPC is trained as a document and a word vector in two different models, and the semantic vector and the co-occurrence vector trained are spliced to obtain the CPC node vector.

[0074] In step S3, a co-occurrence network is constructed based on the CPC co-occurrence relationship in the patent data. First, the co-occurrence number of each pair of CPCs in the patent is counted. If the co-occurrence number of two CPCs is greater than a threshold, it is considered that the two CPCs have occurred closely, and an edge is added between the two nodes in the network, and the weight of the edge is calculated according to the co-occurrence number. If the co-occurrence number of two CPCs is less than the threshold or no co-occurrence occurs, no edge is added between the two nodes, and the network construction rule is as follows:

[0075]

[0076] In the formula, CPC i and CPC j represent two different CPC nodes in the network, w ij represents the weight of the edge between CPC i and CPC j , v ij represents the co-occurrence number of CPC i and CPC j in all patent data, and threshold is the threshold value which can be set according to the experimental effect.

[0077] Since the present application is a future-oriented dynamic link prediction, it is necessary to construct two time period co-occurrence networks to capture the potential association between the "state of node pair in the t1 time period network" and "whether the node pair will produce a connection in the t2 time period network". That is, according to the earliest time of the appearance of the edge between the nodes, it is divided into the t1 time period (before 2015) and the t2 time period (2015-2020), and the CPC co-occurrence networks of the t1 time period and the t2 time period are constructed respectively.

[0078] In step S4, training data is obtained from the CPC co-occurrence network obtained in step S3, and is divided into a training set and a test set. The node pair that does not exist in the t1 time period but exists in the t2 time period is taken as a positive sample, and the node pair that does not exist in the t1 and t2 time periods is taken as a negative sample. The number of positive and negative samples is the same, and is divided into a training set and a test set in a ratio of 9:1. Finally, the CPC co-occurrence networks of the t1 and t2 time periods, the divided training set and test set are stored in a database.

[0079] In step S5, the CPC co-occurrence network for prediction is reconstructed. First, the target field to be studied is selected, and the data of the target field should be a subset of the entire data used in the foregoing steps. The CPCs contained in the patents of the target field are counted to obtain the target field CPC set. The vectors corresponding to the CPCs are averaged to obtain the artificial node vector (the artificial node can be regarded as a special CPC that integrates the technical connotation of the target field), which is then added to the CPC network of the t2 time period constructed in step S3, and edges are added between the artificial node and all CPC nodes in the target field CPC set to generate the CPC co-occurrence network for prediction. Subsequently, the artificial node is corresponded to all CPC nodes other than the target field CPC set one by one to generate prediction node pairs, and the indexes of the nodes are stored in a txt file, with one line representing one node pair and the nodes being separated by spaces. Finally, the CPC co-occurrence network for prediction and the prediction node pair file are stored in the database.

[0080] In step S6, the training data needs to be processed into the input form required by the graph neural network model, which includes the following three steps:

[0081] 1) Construct a closed subgraph of node pairs: The closed subgraph is obtained by searching k-order neighbor nodes outward from the center node of the link to be predicted (k is generally 1 or 2, and a too large k value can easily cause the subgraph to be too large, thereby greatly increasing the memory consumption in the calculation process). Each positive sample node pair and negative sample node pair is traversed, and all node indexes in the closed subgraph of each node pair are saved.

[0082] 2) Distinguish the relative positions of nodes: The distances of the nodes in the closed subgraph to the center nodes of the link to be predicted are different, and the importance of these nodes with different relative positions to link prediction is also different. A node labeling method is used to distinguish the importance, and the formula is as follows:

[0083] d x +d y

[0084] Label(i)=1+min(d x ,d y )+(d / 2)[(d / 2)+(d%2)-1]

[0085] In the formula, d x , d y are the distances (the number of edges between nodes) of the nodes in the closed subgraph to the two center nodes x and y. Label(i) represents the label of node i. The labels of the center nodes x and y are defined as 1; if for node i, d x ,dy If for node i, (d x ,d y ) = (1, 2) or (d x ,d y ) = (2, 1), assign label Label(i) = 3; and so on.

[0086] 3) Obtain the adjacency matrix and node information matrix of the closed subgraph: the adjacency matrix contains the information of the edges between the nodes in the closed subgraph; the node information matrix contains the characteristics of all nodes in the closed subgraph, and a row in the matrix represents a node, and each row is composed of the semantic vector, co-occurrence vector and label of the node. A training sample of a graph neural network model is the adjacency matrix and node information matrix of a closed subgraph.

[0087] Further, the prediction data is preprocessed according to the above three steps, the training data is input into the graph neural network model for iterative training, and then the prediction data is input into the trained model to predict the possibility of connection between the new artificial node and the CPC node outside the target field CPC set. The value range of the prediction result is [0, 1], which can be regarded as the probability of generating a link between the artificial node and the CPC node outside the target field CPC set. The prediction result is sorted in descending order and output to an Excel file, with the first column being CPC and the second column being the possibility of connection between CPC and the target field.

[0088] In step S7, a technology list is formulated according to the link prediction result of step S6. First, a preliminary screening is performed to delete the CPCs with a connection probability less than a threshold value m (the threshold value m can be set to 0.8) with the target field. Then, the CPCs with a greater possibility of being a technology opportunity of the target field are selected as a primary technology list.

[0089] Further, the secondary technology list is screened based on the following two aspects: first, obtain the patents of the technology opportunity field in the past 5 years, use the topological clustering algorithm to cluster the patents into multiple clusters, count the word frequency of the single words extracted from the patent texts in the clusters, obtain the high-frequency word set of each cluster, and screen the patents with more high-frequency words according to the high-frequency word set; second, evaluate the importance of the patents by singular value centrality (SVC), that is, obtain the Schatten matrix norm by summing the pth power of the first k largest singular values of the similarity matrix S:

[0090]

[0091] Further, the importance of the jth patent is measured by the following formula, that is, the change of the Schatten matrix norm before and after removing the jth patent:

[0092]

[0093] In the formula, S(-j) represents the similarity matrix after removing the j-th patent, and σ ii Let be the i-th singular value of S.

[0094] Patents with high frequency of keywords and significant importance are selected based on these two aspects. Specific technical items are then extracted from these patents as secondary technologies. In step S8, trend analysis is conducted to combine technological opportunities with the target field. The specific process is as follows: Figure 3 As shown, firstly, all patents under the Technology Opportunity CPC and patents in the target field are divided into three parts according to time slices, and patent citation networks are constructed for each part. In the patent citation network, nodes of three colors represent patents containing only the target field CPC, patents containing only the Technology Opportunity CPC, and patents containing both the target field CPC and the Technology Opportunity CPC, respectively. Then, a topological clustering algorithm is used to cluster the patents into multiple clusters. The clusters are defined as technology themes, and the aggregation trend of the three types of nodes over time is analyzed and added to the technology list as a description of the development trend of primary technology opportunities. Over time, technology opportunity patents may have already formed a relatively obvious connection and cluster with target field patents, or they may still not have much overlap.

[0095] By following the steps above, a list of technological opportunities in a particular field can be obtained. This list not only includes trends in the combination of technological opportunities with the field, but also provides more specific secondary technological points under the primary technological opportunities, thus providing more effective decision support for the formulation of technology development strategies.

[0096] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for constructing a classification number co-occurrence network, characterized in that, The construction method includes the following steps: S1. For each patent in the target field and the field to be analyzed, determine the co-occurrence information of the patent's classification number; for all patents in the target field and the field to be identified, count all the classification numbers that have appeared to form a classification number set; S2. For each classification number in the classification number set, construct the text information of that classification number based on all patent texts that have appeared under that classification number; S3. Treating the classification number as a document and the text information of the classification number as a word, train the first Doc2vec model to obtain the semantic vector of each classification number in the classification number set; treating the patent as a document and the classification number as a word, train the second Doc2vec model to obtain the co-occurrence vector of each classification number in the classification number set; S4. Combine the semantic vector and co-occurrence vector of each classification number in the classification number set into the classification number vector corresponding to that classification number; S5. Abstract each classification number in the classification number set into a network node, and use its corresponding classification number vector as the attribute of the network node. If the number of co-occurrences between two network nodes exceeds a set threshold, add an edge between the network nodes corresponding to the two classification numbers, thereby realizing the construction of the classification number co-occurrence network.

2. A method for identifying technological opportunities, characterized in that, The identification method includes the following steps: S1. Using the method described in claim 1, construct a co-occurrence network of classification numbers for past time periods, a co-occurrence network of classification numbers for the current time period, and a co-occurrence network of classification numbers that includes both past and current time periods; S2. Node pairs corresponding to edges that did not exist in the past time period but exist in the current time period are taken as positive samples, and node pairs corresponding to edges that did not exist in either the past time period or the current time period are taken as negative samples, thus obtaining the training sample set; S3. The set of classification codes for all patents appearing in the target domain constitutes the target domain classification code set; S4. Take the weighted average of the classification vectors corresponding to each classification number in the target domain classification number set to obtain the target domain node vector. Add the target domain nodes to the classification number co-occurrence network that includes past time periods and current time periods. Add edges between the target domain nodes and all nodes in the target domain classification number set. Assign each target domain node to all nodes outside the target domain classification number set to generate the test sample set. S5. Input each test sample in the test sample set into the graph neural network model trained by the training sample set to obtain the probability of each test sample generating an edge.

3. The method for identifying technological opportunities as described in claim 2, characterized in that, In step S4, the target domain node is added to the co-occurrence network of classification numbers containing past and current time periods in the following manner: (1) Count the number of times each classification number appears in the target domain classification number set; (2) The weight of a specific classification number is determined by dividing the number of times a particular classification number appears by the maximum number of times that classification number appears. (3) Take the weighted average of the classification number vectors corresponding to each classification number in the target domain classification number set to obtain the target domain node vector; (4) Add edges between the target domain node and all nodes in the target domain classification number set. The weight of the edge is the weight of the corresponding classification number.

4. The method for identifying technological opportunities as described in claim 2, characterized in that, The graph neural network model is trained in the following way: (1) Construct a closed subgraph of node pairs: The closed subgraph is obtained by searching outward from the central node of the link that needs to be predicted to the k-th order neighbor nodes. It traverses each positive sample node pair and negative sample node pair, and saves the index of all nodes in the closed subgraph of each node pair. (2) Distinguish the relative positions of nodes: Importance is distinguished using node labeling, as shown in the following formula: d=d x +d y Label(i)=1+min(d x ,d y )+(d / 2)[(d / 2)+(d%2)-1] In the formula, d x d y Let x and y be the distances from a node in the closed subgraph to the two center nodes x and y, respectively. Let Label(i) represent the label of node i, and define the labels of the center nodes x and y as 1. If for node i, (d x ,d y If (d) = (1,1), then assign label Label(i) = 2; if for node i (d) x ,d y ) = (1,2) or (d x ,d y If (i) = (2,1), then assign label Label(i) = 3; and so on. (3) Obtain the adjacency matrix and node information matrix of the closed subgraph: The adjacency matrix contains information about the edges between nodes in a closed subgraph; the node information matrix contains the features of all nodes in the closed subgraph. Each row in the matrix represents a node, and each row consists of the node's semantic vector, co-occurrence vector, and label. A training sample of a graph neural network model is the adjacency matrix and node information matrix of a closed subgraph.

5. The method for identifying technological opportunities as described in claim 2, characterized in that, The method for identifying technological opportunities also includes: The classification numbers that may become technological opportunities are screened to form a primary technology list, wherein the primary technology is the technical content of the classification number; The search target field contains patents with technology opportunity classification numbers. Specific technology items are extracted, and a secondary technology list is extracted. The secondary technologies are the technology items extracted from the patents containing the corresponding classification numbers.

6. The method for identifying technological opportunities as described in claim 5, characterized in that, The list of secondary technologies can be obtained through the following methods: For each category number in the primary technology list, obtain all patents within 5 years in the corresponding technology opportunity field. Use a topological clustering algorithm to cluster the obtained patents into multiple clusters, with each cluster corresponding to a subfield of the corresponding technology opportunity. The word frequencies of the words extracted from the patent text in each cluster are counted to obtain the high-frequency word set of each cluster; Based on the high-frequency word set, patents whose high-frequency words appear more than a threshold are selected from each cluster; For each patent selected from each cluster, the importance of the patent is calculated using singular value centrality. For each cluster after filtering high-frequency words, patents whose importance exceeds the threshold are selected from each cluster; For patents in each cluster after two layers of screening, specific technical items are extracted from the patent text to generate multiple secondary technologies belonging to that classification number.

7. The method for identifying technological opportunities as described in claim 5 or 6, characterized in that, The method for identifying technological opportunities also includes: Construct a patent citation network using all patents under the Technology Opportunity Classification and patents in the target field; All data are divided into patents containing only the target domain classification number, patents containing only the technology opportunity classification number, and patents containing both the target domain classification number and the technology opportunity classification number. The topological clustering algorithm is used to cluster the patents into multiple clusters, and the clusters are defined as technology topics. Analyze the aggregation trend changes of the three types of nodes over time and add them to the technology list as a description of the development trend of primary technology opportunities.

8. A technology opportunity identification system, characterized in that, include: Computer-readable storage media and processors; The computer-readable storage medium is used to store executable instructions; The processor is configured to read executable instructions stored in the computer-readable storage medium and execute the technical opportunity identification method according to any one of claims 2 to 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method for constructing a classification number co-occurrence network as described in claim 1 or the method for identifying technical opportunities as described in any one of claims 2 to 7.

Images