Edge innovation patent prediction method and system

Abstract

The invention relates to the technical field of technical opportunity recognition, and discloses an edge innovation patent prediction method and system, and the method comprises the steps: obtaining the information of all patents in the technical field of a target patent according to a patent edge innovation opportunity recognition task, and constructing a patent data set; dividing the patent data set according to a time sequence, performing marginal evaluation, and constructing a marginal patent data set and a central patent data set based on a time sequence; edge innovation patents are screened out according to the time sequence, the conversion duration is recorded, and an edge innovation patent data set is constructed; constructing a feature system of edge innovation opportunities, obtaining sample features of an edge innovation patent data set, and constructing a prediction model based on a neural network to perform supervised training about a conversion time task; and predicting that the target patent is converted into the edge innovation patent to obtain the conversion time. According to the development of the time sequence characteristic dynamic tracking technology of patent data, the technology with growth potential can be identified in an early stage.

Description

Technical Field

[0001] The present invention relates to the technical field of technology opportunity identification, and in particular to a marginal innovation patent prediction method and system. Background Art

[0002] With the rapid development of global science and technology, the application scenarios of innovative technologies are constantly expanding, and enterprises and research institutions have an increasing demand for the identification and prediction of cutting-edge technologies. Patent literature is an important record and source of technological innovation and can reflect the dynamics of technological development. At present, there are still great challenges in predicting edge innovative technologies. Edge technologies usually refer to technologies that have not yet been widely used or lack sufficient market attention. These technologies may have great application potential in the future due to their innovative or breakthrough characteristics. However, due to the relatively hidden development path, traditional technology forecasting methods, such as trend analysis based on historical data and expert review, are difficult to identify the potential of these technologies in a timely manner, resulting in a slow response to technological change trends, which in turn affects the formulation of innovation strategies and technology layout.

[0003] Technological opportunities are to discover the latest technological trends and infer the possible technological forms or technological development points in a certain technical field by mining the existing technological development trends and their interrelationships in a certain technical field. As the basis for innovation decision-making, technological opportunities are an important factor that must be considered in any technological innovation. Technological opportunity identification methods include qualitative and quantitative analysis. The rapid development of data mining and natural language processing technologies has made rapid progress in quantitative analysis and mining of technological opportunities. The quantitative analysis methods for technological opportunity identification mainly include: patent map method, morphological analysis method, outlier detection method, semantic TRIZ method, and link prediction-based methods. Among them, the patent map method is to perform dimensionality reduction visualization based on the vectorized representation of the keywords of the patent, and find blank areas in the patent map as potential technological opportunities; the core steps of the outlier detection method include the identification and evaluation of outlier patents, among which the evaluation step is usually based on qualitative methods such as expert experience and is highly subjective; the link prediction method uses network structure information to predict the possibility of connection between two unconnected nodes in the network, and find opportunities. Most of these methods are built based on single data or single relationships, and there are certain industrial limitations.

[0004] The patent application with publication number CN 114817567 A discloses a method for identifying technology opportunities, including: extracting classification numbers from patent data and constructing a text collection, obtaining the semantic vector and co-occurrence vector of each classification number through the Doc2vec model, constructing classification number co-occurrence networks for different time periods, dividing positive samples and negative samples, and obtaining a training sample set; counting all classification numbers that have appeared in all patents in the target field, and obtaining the target field node vector after vector weighted average; adding edges between the target field node and all nodes in the target field classification number set, and corresponding all nodes one by one to generate a sample set to be tested; inputting each sample to be tested in the sample set into the graph neural network model trained by the training sample set, and obtaining the probability of each sample to be tested generating an edge. The existing technology mainly relies on the classification number co-occurrence network, which cannot fully capture the multi-dimensional characteristics of the patent document such as the time dynamics and the inventor relationship network; at the same time, the existing technology only predicts the probability of connection between nodes through the graph neural network, which is relatively limited in the specific identification of marginal innovation opportunities.

[0005] The patent application with publication number CN 117634502 A discloses a method for identifying technology opportunities, including: collecting an initial literature data set corresponding to the technology field to be analyzed according to the technology field to be analyzed; preprocessing the initial literature data set to obtain a pre-analysis data set; screening the pre-analysis data set to determine the main technology path corresponding to the technology field to be analyzed, that is, the evolutionary development process of a technology theme; calculating the development maturity of the main technology path to determine the development stage of the main technology path; based on the determined development stage, determining the potential technology opportunities in the technology field to be analyzed. The existing technology mainly identifies technology opportunities by evaluating the development stage of the main technology path, which easily ignores the technology opportunities at the edge of the technology field and lacks a comprehensive evaluation of the market potential of the technology and marginal innovation technology opportunities. Summary of the invention

[0006] The purpose of this invention is to solve the deficiencies of the prior art in the identification and prediction of patent technology opportunities. Through a composite technology opportunity prediction method, the advantages of single prediction methods are complementary, and patent opportunities are analyzed dynamically and multi-dimensionally in combination with time changes. Potential future technology opportunities are identified, providing a scientific basis for technology strategic decision-making, which helps to plan future technology innovation directions in advance.

[0007] A marginal innovation patent prediction method, comprising:

[0008] Step 1: Based on the patent edge innovation opportunity identification task published by the user that includes the target patent, obtain the information of all patents in the technical field to which the target patent belongs and perform preprocessing to construct a patent dataset;

[0009] Step 2: Divide the patent dataset into chronological order, conduct marginal evaluation on the two dimensions of outlier and niche, and construct marginal patent dataset and central patent dataset based on time series;

[0010] Step 3: Based on the time-series-based marginal patent dataset and the central patent dataset, filter out marginal patents that have been converted into central patents in chronological order, record the conversion time, and construct a marginal innovation patent dataset;

[0011] Step 4: Construct a feature system for edge innovation opportunities. Based on the feature system, obtain the sample features of each edge innovation patent in the edge innovation patent dataset as model input, use the conversion time as a label, and perform supervised training on the conversion time task based on the prediction model constructed by the neural network;

[0012] Step 5: Based on the trained prediction model, predict the conversion of the target patent in the patent marginal innovation opportunity identification task into a marginal innovation patent to obtain the conversion time of the target patent.

[0013] Furthermore, in step 1, the information of all patents in the technical field to which the target patent belongs is obtained and preprocessed to construct a patent dataset, including:

[0014] The patent information includes international patent classification number, patent text, inventor co-occurrence information and publication date information, wherein the patent text includes patent title and abstract;

[0015] For the international patent classification number, retain the first 4 digits of data to obtain preprocessed classification number data;

[0016] For the patent text, the patent title and abstract are merged, symbols, numbers, stop words, and common words are removed, and morphological restoration is performed to obtain preprocessed patent text data;

[0017] For the inventor co-occurrence information, patents with only a single inventor are eliminated, and the inventor information of the remaining patents is extracted to obtain the pre-processed inventor information;

[0018] The patent data set is constructed after all patent preprocessing.

[0019] Furthermore, in step 2, the patent dataset is divided according to the time sequence, and marginal evaluations are performed on the two dimensions of outlier and niche, respectively, to construct a marginal patent dataset and a central patent dataset based on time sequence, including:

[0020] For the patent dataset, from the starting year to the ending year, the starting year is used as the benchmark and a new sub-dataset is formed for each additional year of data, thereby forming a series of patent datasets based on time series.

[0021] In the time-series-based patent data set, the patent text data is vectorized and converted into patent text vectors; the method used for vectorizing the patent text data includes the BERT algorithm;

[0022] The classification number data is vectorized and converted into a classification number vector; the method used for vectorizing the classification number data includes a term frequency-inverse document frequency algorithm.

[0023] The text vector and the classification number vector are fused and spliced, and the spliced ​​vector is subjected to outlier analysis by adopting the cross analysis method. The three outlier detection algorithms, namely local outlier factor, K nearest neighbor algorithm and isolation forest, are used for analysis. The intersection of the three analysis results is taken to obtain a time-series-based outlier edge patent data set as the outlier evaluation result.

[0024] In the patent data set based on time series, the inventors are taken as nodes, the cooperation relationship of the co-inventors is taken as edges, and the actual number of cooperations between the inventors is taken as the connection weight, so as to establish a network diagram of the cooperation relationship of the inventors;

[0025] Input the inventor cooperation relationship network graph into the graph convolutional network to obtain a reconstructed node feature matrix, and then use the graph autoencoder to perform link prediction on the reconstructed node feature matrix to obtain the link prediction scores of all edges in the inventor cooperation relationship network graph as the potential cooperation possibility between the inventors;

[0026] Taking inventors as nodes, establishing edges between nodes whose cooperation possibilities are higher than a set value, and using cooperation possibilities as connection weights, reconstructing the network diagram of inventors' potential cooperation relationships;

[0027] Based on the inventor collaboration network diagram and the inventor potential collaboration network diagram, the importance of each node in the two network diagrams is calculated respectively through the web page ranking algorithm. The nodes with low importance are selected and then the union is taken to obtain the information of inventors with strong niche and the patents of which they are the first author. The time-series-based niche marginal patent dataset is obtained as the niche evaluation result.

[0028] The intersection of the outlier evaluation result and the niche evaluation result is defined as a time-series-based marginal patent dataset; and the remaining patents that are not marginal patent data in the time-series-based patent dataset are defined as a time-series-based central patent dataset.

[0029] Furthermore, in step 3, based on the marginal patent dataset and the central patent dataset based on time series, the marginal patents that have been converted into central patents are screened out in chronological order, the conversion time is recorded, and the marginal innovation patent dataset is constructed, including:

[0030] For the marginal patent dataset and the central patent dataset based on time series, the types of patents are marked as marginal patents and central patents respectively; each patent is sorted from small to large according to the year, forming a time series trajectory containing the patent type mark for each patent;

[0031] The time series trajectory of each patent is analyzed using a sliding window method to check whether there is a state transition that indicates that the patent is a marginal patent in the publication year and becomes a central patent in the subsequent year;

[0032] Patents with status transitions are extracted, and the patent name, patent publication year, and transition year are recorded. All marginal innovation patents are screened out and their transformation time is obtained to construct a marginal innovation patent dataset.

[0033] Furthermore, in step 4, constructing a feature system of edge innovation opportunities includes:

[0034] According to the concept and characteristics of edge innovation patents, a feature system for evaluating edge innovation opportunities is designed based on the three dimensions of technological novelty, market attractiveness, and cross-diversity. Each dimension in the feature system has multiple quantifiable indicators.

[0035] The quantifiable indicators of technical novelty include the number of citations, the number of family citations, the number of cited scientific and technological documents, the number of claims, the number of independent claims, the number of dependent claims, the number of document pages, and patent indicators;

[0036] The quantifiable indicators of market attractiveness include the number of citations, the number of patent families, the number of countries in the patent family, and the number of emerging industry classifications;

[0037] The quantifiable indicators of cross-diversity include: the number of inventors, the number of applicants or patent holders, the number of IPC classifications, and the proportion of corporate types among applicants.

[0038] The present invention also provides a marginal innovation patent prediction system, comprising:

[0039] The first data set construction module is used to obtain information on all patents in the technical field to which the target patent belongs and perform preprocessing to construct a patent data set based on the patent marginal innovation opportunity identification task including the target patent published by the user;

[0040] The second data set construction module is used to divide the patent data set according to the time sequence, conduct marginal evaluation in two dimensions, namely, outlier and niche, and construct a marginal patent data set and a central patent data set based on time sequence;

[0041] The third data set construction module is used to filter out marginal patents that have been converted into central patents in chronological order based on the marginal patent data set and the central patent data set based on time series, record the conversion time, and construct a marginal innovation patent data set;

[0042] The prediction model training module is used to build a feature system for edge innovation opportunities. Based on the feature system, the sample features of each edge innovation patent in the edge innovation patent dataset are obtained as model input. The conversion time is used as a label, and the prediction model built based on the neural network is used for supervised training on the conversion time task.

[0043] The marginal innovation patent prediction module is used to predict the conversion of the target patent in the patent marginal innovation opportunity identification task into a marginal innovation patent based on the trained prediction model, and obtain the conversion time of the target patent.

[0044] The present invention identifies the hot spots of marginal innovation technology through a method based on time series analysis, solving the problem that the existing technology is difficult to capture early marginal innovation technology. Compared with the existing technology, the present invention evaluates the outlier and niche nature of marginal patents and screens marginal innovation patents; uses the time series characteristics of patent data to continuously track marginal innovation patents in historical data; analyzes the changes in the activity pattern of marginal innovation patents through marginal innovation characteristic indicators, realizes multi-dimensional dynamic tracking of the development and evolution of innovative technologies, and accurately identifies marginal patents with potential value at an early stage. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 The following are the execution steps of a marginal innovation patent prediction method in an embodiment of the present invention;

[0046] Figure 2 The specific process of evaluating and predicting marginal innovation patents in the embodiments of the present invention;

[0047] Figure 3 This is a marginal innovation patent prediction system in an embodiment of the present invention. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical solution and beneficial effects of the present invention clearer, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0049] like Figure 1 As shown, a marginal innovation patent prediction method includes:

[0050] S1. According to the patent edge innovation opportunity identification task published by the user containing the target patent, obtain the information of all patents in the technical field to which the target patent belongs and perform preprocessing to construct a patent dataset;

[0051] The specific process is as follows Figure 2 As shown in S101-S102:

[0052] From the Incopat patent database, obtain patent information S101 from 2004 to 2020 in the technical field to which the target patent belongs, including the International Patent Classification (IPC) number, patent text, inventor co-occurrence information and publication date information, among which the patent text includes the patent title and abstract;

[0053] The patent information S101 is preprocessed to construct a patent data set S102. The specific preprocessing method includes:

[0054] For the IPC number, the first 4 digits (IPC4) are retained to obtain the preprocessed classification number data; for the patent text, the patent title and abstract are merged, symbols, numbers, stop words, and common words are removed, and word form restoration is performed to obtain the preprocessed patent text data; for the inventor co-occurrence information, patents with only a single inventor are eliminated, and the inventor information of the remaining patents is extracted to obtain the preprocessed inventor information;

[0055] The data after all patent preprocessing are constructed into the patent dataset S102.

[0056] S2. Divide the patent dataset in chronological order, conduct marginal evaluation in two dimensions: outlier and niche, and construct marginal patent dataset and central patent dataset based on time series;

[0057] The specific process is as follows Figure 2 As shown in S201-S208:

[0058] For the patent dataset S102, from the start year to the end year, the start year is taken as the benchmark, and a new sub-dataset is formed for each additional year of data, to form a series of patent datasets S201 based on time series. The specific process is as follows:

[0059] (1) Extract the data of 2004 from the original data set to form the first patent data set;

[0060] (2) Extract the data of 2005 from the original data set and merge it with the first patent data set to form the second patent data set. That is, the second patent data set contains the data of 2004 and 2005.

[0061] (3) Extract the data of 2006 from the original data set and merge it with the second patent data set to form the third patent data set. That is, the third patent data set contains the data of 2004, 2005 and 2006.

[0062] (4) The data of each subsequent year will be merged with the data of all previous years in an annual increment to form a new data set until all data up to 2020 are included.

[0063] Based on the above process, a series of patent data sets S201 based on time series that are accumulated year by year are constructed, in which each sub-dataset is based on the previous sub-dataset and adds all the data of the next year.

[0064] In the time series-based patent data set S201, the patent text data is vectorized and converted into a patent text vector of a fixed length; the method used for vectorizing the patent text data is the BERT algorithm;

[0065] The classification number data is vectorized and converted into a classification number vector that can highlight the differences in patent technology. The method used to vectorize the classification number data is the Term Frequency-Inverse Document Frequency (TF-IDF) algorithm, which is specifically expressed as follows:

[0066]

[0067] IPC j =(tfidf 1,j ,tfidf 2,j ,...,tfidf x,j )

[0068] Among them, tfidf i,j represents the TF-IDF value of IPC4 classification number data i in patent j, tf i,j represents the frequency of IPC4 classification number data i appearing in patent j, N represents the total number of patents in the entire time-series-based patent dataset S201, df i,j represents the number of patents containing IPC4 classification number data i in the patent data set S201 based on the time series; δ is a smoothing parameter. In this embodiment, in order to ensure the validity of the function and prevent the denominator from being zero, it is set to δ=1; IPC j represents the classification code vector of patent j, and x is the total number of IPC4 classification code data; in this way, each patent can be represented by a vector, and the dimension of this vector is equal to the number of IPC4 classification code data.

[0069] The patent text vector and the classification number vector are fused and spliced, and the spliced ​​vector S202 is subjected to outlier analysis by adopting the cross analysis method. The three outlier detection algorithms, namely, Local Outlier Factor (LOF), K-Nearest Neighbors (KNN) and Isolation Forest (IF), are used for analysis. The intersection of the three analysis results is taken to obtain the time-series-based outlier edge patent data set S203 as the outlier evaluation result.

[0070] In the patent data set S201 based on time series, the inventors are taken as nodes, the cooperation relationship of the co-inventors is taken as edges, and the actual number of cooperations between the inventors is taken as the connection weight, so as to establish a network diagram of the cooperation relationship of the inventors S204;

[0071] The inventor cooperation network graph is input into the graph convolutional network to obtain the node feature vector reconstructed for each node, and then the link prediction of the reconstructed node feature vector is performed through the graph autoencoder (GAE) to obtain the link prediction scores of all edges in the inventor cooperation network graph as the potential cooperation possibility between the inventors, which is specifically expressed as:

[0072] score uv =GAE(h u ,h v )

[0073] Among them, h u and h v They represent the node feature vectors reconstructed by the graph convolutional network for node u and node v, respectively, and score uv represents the link prediction score of the edge between node u and node v; in this embodiment, the link prediction score score uv Greater than or equal to 0.6 indicates a high likelihood of cooperation;

[0074] Taking the inventors as nodes, establishing edges between nodes with high cooperation possibilities, using the cooperation possibilities as connection weights, and reconstructing the potential cooperation relationship network graph of the inventors S205;

[0075] Based on the inventor cooperation network diagram S204 and the inventor potential cooperation network diagram S205, the importance of each node in the two network diagrams is calculated respectively by the Page-Rank algorithm, and the nodes with low importance are selected and then the union is taken to obtain the information of inventors with strong niche and the patents of which they are the first authors, and the time-series-based niche marginal patent data set S206 is obtained as the niche evaluation result.

[0076] Among them, the Page-Rank algorithm is specifically expressed as:

[0077]

[0078] Where p represents the node to be calculated, PR(p) is the Page-Rank value of node u; d is the damping factor, which represents the probability of random walk to other nodes. In this embodiment, d=0.85; M is the total number of nodes in the network; B q It represents the set of all nodes q connected to node p, PR(q) is the Page-Rank value of node q, and L(q) represents the number of edges connected to node q. After sorting the Page-Rank values ​​of all nodes from high to low, the last 10% of nodes are regarded as nodes of low importance.

[0079] The intersection of the outlier evaluation result and the niche evaluation result is defined as a time-series-based marginal patent dataset S207;

[0080] The remaining patents that are not marginal patent data in the time-series-based patent dataset S201 are defined as a time-series-based central patent dataset S208.

[0081] S3. Based on the time-series-based marginal patent dataset and the central patent dataset, filter out marginal patents that have been converted into central patents in chronological order, record the conversion time, and construct a marginal innovation patent dataset;

[0082] For the marginal patent dataset and the central patent dataset based on time series, the types of patents are marked as marginal patents and central patents respectively; each patent is sorted from far to near in chronological order to form a time series trajectory containing patent type marks for each patent;

[0083] The time series trajectory of each patent is analyzed using a sliding window method to check whether there is a state transition that indicates that the patent is a marginal patent in the publication year and becomes a central patent in the subsequent year;

[0084] Patents with status transitions are extracted, and the patent name, patent publication year, and transition year are recorded. All marginal innovation patents are screened out and their transformation time is obtained to construct a marginal innovation patent dataset.

[0085] S4. Construct a feature system for edge innovation opportunities. Based on the feature system, obtain the sample features of each edge innovation patent in the edge innovation patent dataset as model input, use the conversion time as a label, and perform supervised training on the conversion time task based on the prediction model constructed by the neural network.

[0086] Specific as Figure 2 As shown in S401-S406:

[0087] According to the concept and characteristics of edge innovation patents, a feature system for evaluating edge innovation opportunities is designed based on the three dimensions of technological novelty S401, market attractiveness S402, and cross-diversity S403. Each dimension in the feature system has multiple quantifiable indicators, as shown in Table 1:

[0088] Table 1 Characteristic system for evaluating marginal innovation opportunities

[0089]

[0090]

[0091] Based on the feature system, the quantifiable indicators of all marginal innovation patents in the marginal innovation patent dataset S301 from the year of patent publication to each year in 2020 are obtained from the Incopat patent library as multi-perspective patent data time series vectors, which can consider the development trajectory of patents from multiple dimensions at different time nodes;

[0092] Combining the multi-view patent data time series vectors into a multi-view patent data time series matrix as the sample feature S405; performing normalization processing on the sample feature S405 to eliminate the scale difference between the features and use it as the input data set;

[0093] Construct a Transformer-based prediction model S406, divide the input data set into a training set and a test set in a ratio of 8:2, use the conversion time as a label, and perform supervised training on the conversion time task, which is specifically expressed as:

[0094] y=f Transformer ([X0,X1,...,X T ])

[0095] Among them, X T represents the multi-view patent data time series vector of each patent in the Tth year after publication, y represents the output conversion time, and f Transformer Represents the Transformer prediction model based on the multi-head attention mechanism.

[0096] S5. Based on the trained prediction model, the target patent in the patent marginal innovation opportunity identification task is predicted to be converted into a marginal innovation patent, and the conversion time of the target patent is obtained; among which, the disclosure time of the target patent is from 2021 to 2024.

[0097] like Figure 3 As shown, this embodiment provides a marginal innovation patent prediction system, including:

[0098] The first data set construction module 1 is used to obtain information on all patents in the technical field to which the target patent belongs and perform preprocessing to construct a patent data set based on the patent marginal innovation opportunity identification task including the target patent published by the user;

[0099] The second data set construction module 2 is used to divide the patent data set according to the time sequence, conduct marginal evaluation in two dimensions, namely, outlier and niche, and construct a marginal patent data set and a central patent data set based on the time sequence;

[0100] The third data set construction module 3 is used to filter out marginal patents that have been converted into central patents in chronological order based on the marginal patent data set and the central patent data set based on time series, record the conversion time, and construct a marginal innovation patent data set;

[0101] Prediction model training module 4 is used to construct a feature system of marginal innovation opportunities, obtain sample features of each marginal innovation patent in the marginal innovation patent data set based on the feature system as model input, use conversion time as a label, and perform supervised training on conversion time tasks based on the prediction model constructed by the neural network;

[0102] The marginal innovation patent prediction module 5 is used to predict the conversion of the target patent in the patent marginal innovation opportunity identification task into a marginal innovation patent based on the trained prediction model, and obtain the conversion time of the target patent.

[0103] The present invention identifies the hot spots of marginal innovation technology through a method based on time series analysis, solving the problem that the existing technology is difficult to capture early marginal innovation technology. Compared with the existing technology, the present invention evaluates the outlier and niche nature of marginal patents and screens marginal innovation patents; uses the time series characteristics of patent data to continuously track marginal innovation patents in historical data; analyzes the changes in the activity pattern of marginal innovation patents through marginal innovation characteristic indicators, realizes multi-dimensional dynamic tracking of the development and evolution of innovative technologies, and accurately identifies marginal patents with potential value at an early stage.

[0104] The preferred embodiments of the present invention are described above with reference to the accompanying drawings, but the scope of the rights of the embodiments of the present invention is not limited thereto. Any modification, equivalent substitution and improvement made by those skilled in the art without departing from the scope and essence of the embodiments of the present invention shall be within the scope of the rights of the embodiments of the present invention.

Claims

1. A method for predicting marginal innovation patents, characterized in that: include: Step 1: Based on the patent edge innovation opportunity identification task published by the user that includes the target patent, obtain the information of all patents in the technical field to which the target patent belongs and perform preprocessing to construct a patent dataset; Step 2: Divide the patent dataset into chronological order, conduct marginal evaluation on the two dimensions of outlier and niche, and construct marginal patent dataset and central patent dataset based on time series; Step 3: Based on the time-series-based marginal patent dataset and the central patent dataset, filter out marginal patents that have been converted into central patents in chronological order, record the conversion time, and construct a marginal innovation patent dataset; Step 4: Construct a feature system for edge innovation opportunities. Based on the feature system, obtain the sample features of each edge innovation patent in the edge innovation patent dataset as model input, use the conversion time as a label, and perform supervised training on the conversion time task based on the prediction model constructed by the neural network; Step 5: Based on the trained prediction model, predict the conversion of the target patent in the patent marginal innovation opportunity identification task into a marginal innovation patent to obtain the conversion time of the target patent.

2. A marginal innovation patent prediction method according to claim 1, characterized in that: In step 1, the information of all patents in the technical field to which the target patent belongs is obtained and preprocessed to construct a patent dataset, including: The patent information includes international patent classification number, patent text, inventor co-occurrence information and publication date information, wherein the patent text includes patent title and abstract; For the international patent classification number, retain the first 4 digits of data to obtain preprocessed classification number data; For the patent text, the patent title and abstract are merged, symbols, numbers, stop words, and common words are removed, and morphological restoration is performed to obtain preprocessed patent text data; For the inventor co-occurrence information, patents with only a single inventor are eliminated, and the inventor information of the remaining patents is extracted to obtain the pre-processed inventor information; The patent data set is constructed after all patent preprocessing.

3. A marginal innovation patent prediction method according to claim 1, characterized in that: In step 2, the patent dataset is divided according to the time sequence, and marginal evaluations are performed on the two dimensions of outlier and niche, respectively, to construct a marginal patent dataset and a central patent dataset based on time sequence, including: For the patent dataset, from the starting year to the ending year, the starting year is used as the benchmark, and a new sub-dataset is formed for each additional year of data, thereby forming a series of patent datasets based on time series; Conducting marginal evaluation on the patent dataset based on time series in two dimensions, namely, outlier and niche, and defining the intersection of the two evaluation results as a marginal patent dataset based on time series; The remaining patents that are not marginal patent data in the time-series-based patent dataset are defined as a time-series-based central patent dataset.

4. A marginal innovation patent prediction method according to claim 3, characterized in that: The marginality of outliers is evaluated on the time series-based patent dataset, including: In the patent data set based on time series, the patent text data is vectorized and converted into patent text vectors; the classification number data is vectorized and converted into classification number vectors; The text vector and the classification number vector are fused and spliced, and the spliced ​​vector is subjected to outlier analysis by adopting the cross analysis method. The three outlier detection algorithms, namely local outlier factor, K nearest neighbor algorithm and isolation forest, are used for analysis. The intersection of the three analysis results is taken to obtain a time-series-based outlier edge patent data set as the outlier evaluation result.

5. A marginal innovation patent prediction method according to claim 3, characterized in that: The method used for vectorizing patent text data includes the BERT algorithm; the method used for vectorizing classification number data includes the term frequency-inverse document frequency algorithm.

6. A marginal innovation patent prediction method according to claim 3, characterized in that: A niche marginal assessment of the time-series-based patent dataset is performed, including: In the patent data set based on time series, the inventors are taken as nodes, the cooperation relationship of the co-inventors is taken as edges, and the actual number of cooperations between the inventors is taken as the connection weight, so as to establish a network diagram of the cooperation relationship of the inventors; Input the inventor cooperation relationship network graph into the graph convolutional network to obtain the node feature vector reconstructed by each node, and then perform link prediction on the reconstructed node feature vector through the graph autoencoder to obtain the link prediction scores of all edges in the inventor cooperation relationship network graph as the potential cooperation possibility between the inventors; Taking inventors as nodes, establishing edges between nodes with high cooperation possibilities, using cooperation possibilities as connection weights, and reconstructing the network diagram of inventors' potential cooperation relationships; Based on the inventor collaboration network diagram and the inventor potential collaboration network diagram, the importance of each node in the two network diagrams is calculated respectively through the web page ranking algorithm. The nodes with low importance are selected and then the union is taken to obtain the information of inventors with strong niche and the patents of which they are the first author. The time-series-based niche marginal patent dataset is obtained as the niche evaluation result.

7. A marginal innovation patent prediction method according to claim 1, characterized in that: In step 3, based on the marginal patent dataset and the central patent dataset based on time series, the marginal patents that have been converted into central patents are screened out in chronological order, the conversion time is recorded, and the marginal innovation patent dataset is constructed, including: For the marginal patent dataset and the central patent dataset based on time series, the types of patents are marked as marginal patents and central patents respectively; each patent is sorted from far to near in chronological order to form a time series trajectory containing patent type marks for each patent; The time series trajectory of each patent is analyzed using a sliding window method to check whether there is a state transition that indicates that the patent is a marginal patent in the publication year and becomes a central patent in the subsequent year; Patents with status transitions are extracted, and the patent name, patent publication year, and transition year are recorded. All marginal innovation patents are screened out and their transformation time is obtained to construct a marginal innovation patent dataset.

8. A marginal innovation patent prediction method according to claim 1, characterized in that: In step 4, a characteristic system of edge innovation opportunities is constructed, including: According to the concept and characteristics of edge innovation patents, a feature system for evaluating edge innovation opportunities is designed based on the three dimensions of technological novelty, market attractiveness, and cross-diversity. Each dimension in the feature system has multiple quantifiable indicators. The quantifiable indicators of technical novelty include the number of citations, the number of family citations, the number of cited scientific and technological documents, the number of claims, the number of independent claims, the number of dependent claims, the number of document pages, and patent indicators; The quantifiable indicators of market attractiveness include the number of citations, the number of patent families, the number of countries in the patent family, and the number of emerging industry classifications; The quantifiable indicators of cross-diversity include: the number of inventors, the number of applicants or patent holders, the number of IPC classifications, and the proportion of corporate types among applicants.

9. A marginal innovation patent prediction system, characterized in that: include: The first data set construction module is used to obtain information on all patents in the technical field to which the target patent belongs and perform preprocessing to construct a patent data set based on the patent marginal innovation opportunity identification task including the target patent published by the user; The second data set construction module is used to divide the patent data set according to the time sequence, conduct marginal evaluation in two dimensions, namely, outlier and niche, and construct a marginal patent data set and a central patent data set based on time sequence; The third data set construction module is used to filter out marginal patents that have been converted into central patents in chronological order based on the marginal patent data set and the central patent data set based on time series, record the conversion time, and construct a marginal innovation patent data set; The prediction model training module is used to build a feature system for edge innovation opportunities. Based on the feature system, the sample features of each edge innovation patent in the edge innovation patent dataset are obtained as model input. The conversion time is used as a label, and the prediction model built based on the neural network is used for supervised training on the conversion time task. The marginal innovation patent prediction module is used to predict the conversion of the target patent in the patent marginal innovation opportunity identification task into a marginal innovation patent based on the trained prediction model, and obtain the conversion time of the target patent.

Images