A method, system and device for identifying convertible patents
By constructing an improved recognition model based on the Stacking algorithm, using the JC index to screen target classifiers, and combining digital and textual features to identify convertible patents, the accuracy problem of patent conversion recognition was solved, and the activity of intellectual property conversion was improved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANDONG UNIV
- Filing Date
- 2023-07-27
- Publication Date
- 2026-04-24
AI Technical Summary
In the existing technology, patents cannot be directly converted into real productivity, and there is a lack of effective machine learning algorithms combined with natural language processing in identifying convertible patents, resulting in low activity in intellectual property conversion.
An improved recognition model based on the Stacking algorithm is constructed. The target classifier is selected by screening at least two base classifiers and using the JC index. Patent conversion recognition is performed using digital and text features, including the extraction of feature data such as the number of applicants and inventors, and the calculation of Jaccard distance and cosine similarity, forming the Stacking2 ensemble learning model.
It improved the accuracy of identifying convertible patents, enhanced the activity of intellectual property conversion, and ensured the accuracy and diversity of the model.
Smart Images
Figure CN116881809B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of big data analytics, specifically to a patent identification method, system, and device. Background Technology
[0002] Currently, most patents cannot be directly converted into real productivity. Faced with a massive number of "dormant" patents and constantly emerging new patent resources, how to better utilize patent information to identify convertible patents and enhance the activity of intellectual property transformation in my country is the key to shaping new advantages for my country's future development and achieving high-quality development.
[0003] With the advent of the era of big data and artificial intelligence, many scholars have used machine learning algorithms (AIForScience) to conduct a series of research and applications. At the same time, many scholars have used fusion models to conduct corresponding research and applications. However, in terms of identifying whether a patent can be converted, few scholars have used machine learning algorithms combined with natural language processing and other technologies to identify whether a patent can be converted. Moreover, most scholars have not explored much about how to select a base classifier. Summary of the Invention
[0004] In view of the deficiencies in the prior art, the present invention provides a method, system and device for identifying convertible patents.
[0005] Firstly, a method for identifying convertible patents, comprising the following steps:
[0006] 1. Construct a recognition model based on an improved Stacking algorithm;
[0007] 2. Obtain the data to be identified; the data to be identified includes convertible patent data;
[0008] 3. The identification method is used to identify the data to be identified, and an identification result is obtained; the identification result includes whether the patent can be converted.
[0009] Furthermore, the construction of the recognition model based on the improved Stacking algorithm specifically involves:
[0010] The dataset is classified using at least two base classifiers to obtain at least two classification results;
[0011] If the classification result does not reach the set threshold, the base classifier corresponding to the classification result is removed.
[0012] If the classification result reaches the set threshold, the target classifier is selected by the JC index;
[0013] The recognition model is constructed using the target classifier and then validated. This involves setting a threshold based on actual conditions, which helps to select base learners with high recognition rates, and then using the JC metric to select the target classifier to improve the Stacking ensemble learning model.
[0014] Furthermore, the selection of the target classifier using the JC metric specifically involves:
[0015] For pairwise base classifiers, calculate the JC index value;
[0016] The classifier with the smaller JC index value is selected as the target classifier. This is achieved by calculating the similarity between base classifiers, which compares the similarity between each pair of base classifiers. The higher the similarity, the worse the fusion effect. Therefore, it is necessary to ensure the difference between base classifiers, thus providing theoretical support for improving the Stacking ensemble learning model.
[0017] Furthermore, the calculation of the JC index value specifically involves:
[0018] Select patent sample data that has already been transformed, and extract the feature data of the patent sample data; the feature data includes numerical features and text features;
[0019] The numerical characteristics include: number of applicants, number of inventors, number of citations, number of citations by examiners, number of cited patents, number of cited non-patent documents, number of claims, number of words in the sovereign claim, number of pages in the specification, number of Innojoy family members, number of Innojoy layout countries, and remaining validity period.
[0020] The text features include: patent name, patent type, applicant type, PCT international application, and US-Japan-EU tripartite patent;
[0021] The feature data is input into two base classifiers respectively to obtain the first classification result and the second classification result;
[0022] The Jaccard distance is calculated using the first classification result, the second classification result, and the pdist function in the scipy library, and the Jaccard coefficient is calculated based on the Jaccard distance.
[0023] The cosine similarity is calculated using the first classification result, the second classification result, and the cosine function in the scipy library.
[0024] The Jaccard distance is multiplied by the cosine similarity to obtain the JC index value. This is to find the pair of classifiers with the lowest similarity and the other base classifiers with the lowest similarity to this pair, with the aim of improving the fusion model and refining the Stacking ensemble learning model.
[0025] Secondly, a system for identifying convertible patents, comprising the following steps:
[0026] 1. Model Building Module: Used to build a recognition model based on the Stacking algorithm, including the following units:
[0027] The first unit is used to obtain the dataset and at least two base classifiers;
[0028] The second unit is used to classify the dataset using at least two base classifiers to obtain at least two classification results;
[0029] The third unit is used to remove the base classifier corresponding to the classification result if the classification result does not reach the set threshold.
[0030] The fourth unit is used to select a target classifier by means of the JC index if the classification result reaches a set threshold.
[0031] The fifth unit is used to construct the recognition model using the target classifier and to verify the recognition model.
[0032] 2. Acquisition Module: Used to acquire data to be identified; the data to be identified includes convertible patent data;
[0033] 3. Identification module: used to identify the data to be identified using the identification method, and obtain identification results; the identification results include whether the patent will be converted.
[0034] Furthermore, the fourth unit is used to select a target classifier using the JC metric if the classification result reaches a set threshold, specifically for:
[0035] If the classification result reaches the set threshold, then the JC index value is calculated for each pairwise base classifier.
[0036] The classifier with the smaller JC index value is selected as the target classifier.
[0037] The calculation of the JC index value is specifically as follows:
[0038] Select the patent sample data that has been transformed, and extract the feature data of the patent sample data; input the feature data into two base classifiers respectively to obtain the first classification result and the second classification result;
[0039] The Jaccard distance is calculated using the first classification result, the second classification result, and the pdist function in the scipy library, and the Jaccard coefficient is calculated based on the Jaccard distance.
[0040] The cosine similarity is calculated using the first classification result, the second classification result, and the cosine function in the scipy library; the Jaccard distance is multiplied by the cosine similarity to obtain the JC index value.
[0041] By comparing the similarity between each pair of base classifiers, the pair of base classifiers with the lowest similarity is identified as a classifier pair. The similarity between other base classifiers and this pair of classifiers is calculated using hierarchical clustering. The three base classifiers with the lowest similarity are then selected using hierarchical clustering.
[0042] This invention provides a device including a processor, an input device, an output device, and a memory. The processor, input device, output device, and memory are interconnected. The memory is used to store a computer program, which includes program instructions. The processor is configured to call the program instructions to execute the method described in the above embodiments.
[0043] This invention provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, cause the processor to perform the method described in the above embodiments.
[0044] Beneficial effects of the present invention
[0045] Compared with base learners and Stacking ensemble learning models, this improved recognition model based on the Stacking algorithm adaptively selects the target classifier using the JC index. While ensuring accuracy, it increases the differentiation between primary learners, enabling the improved model to more accurately identify convertible patents and enhance the activity of intellectual property transformation in my country. Attached Figure Description
[0046] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.
[0047] Figure 1 A flowchart illustrating the method for identifying convertible patents provided in this embodiment of the invention;
[0048] Figure 2 This is a flowchart for constructing a recognition model based on the Stacking algorithm.
[0049] Figure 3 A flowchart of an identification system for convertible patents is provided for this invention;
[0050] Figure 4 This is a structural block diagram of a device according to the present invention. Detailed Implementation
[0051] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.
[0052] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.
[0053] Example 1
[0054] Please refer to Figure 1 This invention provides a method for identifying convertible patents, the steps of which include:
[0055] S1, please refer to Figure 2 We constructed an improved recognition model based on the Stacking algorithm.
[0056] (1) Obtain the dataset and at least three base classifiers.
[0057] A large amount of patent data that has been converted was obtained through retrieval. The converted patent data was input into an input device to form a dataset, which includes text data and numerical data. Then, the BERT model and TextRNN model were used to extract data features from the convertible patent dataset using a processor, as shown in the table below, totaling 35 data features.
[0058] Table 1 Features extracted from patent texts
[0059]
[0060]
[0061] Preferably, classification algorithms such as KNN, logistic regression, Naive Ye Bayes, support vector machines, decision trees, random forests, and XGBoost are selected. At least three base classifiers are constructed using the training set and the classification algorithm. The base classifiers are then validated and their parameters are tuned. Five-fold cross-validation is then used, which involves dividing the dataset into five equal parts, training on four parts, and testing on the remaining part. This process is repeated five times, and the average AUC value is calculated. Finally, the AUC value is determined after validation. A higher AUC value indicates a higher recognition rate for convertible patents.
[0062] (2) Filter the base classifier according to the set threshold.
[0063] As shown in the table below, the threshold is determined based on the AUC value after validation of the base classifiers. The threshold is set according to the actual situation to achieve the expected recognition effect. The threshold is set to 0.8, and the selected base classifiers that meet the threshold are: KNN, Logistic Regression, and Naive Bayes. Base classifiers that do not meet the threshold are removed. Setting the threshold according to the actual situation helps to select base learners with high recognition rates.
[0064] Table 2 AUC values for different models
[0065]
[0066] (3) The base classifiers are combined to form a Stacking ensemble learning model.
[0067] Optionally, KNN, logistic regression, and Naive Bayesian classifiers are used as primary learners, and logistic regression base classifiers are used as secondary learners to form a Stacking ensemble learning model.
[0068] Preferably, the Stacking ensemble learning model is implemented using the TabularPredictor module in the Autogluon library. The parameters of the trained Stacking ensemble learning model are then fused and subjected to five-fold cross-validation. Finally, the AUC value reaches 0.8423, which is higher than that of a typical base classifier. Therefore, this Stacking ensemble learning model is constructed using the idea of stacking base classifiers.
[0069] (4) Calculate the JC index
[0070] The processor calculates the similarity between primary learners, known as the JC metric. The JC metric consists of two parts: the Jaccard coefficient and the cosine similarity. These include:
[0071] ①Jaccard coefficient
[0072] Jaccard coefficients are used to compare the similarity and differences between finite sample sets. A higher Jaccard coefficient indicates higher sample similarity. The Jaccard coefficient is defined as follows:
[0073]
[0074] ② Cosine similarity
[0075] The cosine similarity between two vectors reflects their degree of similarity. A larger cosine value indicates higher similarity between the two vectors. Cosine similarity is defined as:
[0076]
[0077] ③ The product of Jaccard coefficient and cosine similarity
[0078] Both the Jaccard coefficient and cosine similarity represent a high degree of difference between two vectors when their values are small. Therefore, by multiplying the two indices, a new index for measuring the similarity between vectors is obtained—the JC index.
[0079] JC = J(A, B) * cosθ
[0080] ④ Calculate the JC index in the computer.
[0081] Preferably, after selecting the test set, these data features are input into the KNN base classifier and the logistic regression base classifier with adjusted parameters, respectively, to obtain the vector sets of prediction results, denoted as follows: and
[0082] Furthermore, the Jaccard distance is calculated using the pdist function in the scipy library, and then the Jaccard coefficient is calculated:
[0083]
[0084] Furthermore, the cosine function from the scipy library is used to calculate... and Cosine similarity between:
[0085]
[0086] Furthermore, multiplying the Jaccard coefficient by the cosine similarity yields a measure of... and The JC index value for similarity between the two base classifiers is: s(A, logistic regression) = min(s), which is the similarity between the two base classifiers, KNN and logistic regression.
[0087] Similarly, the similarity between any two base classifiers can be obtained using the above method, and the similarity values are recorded in the table below.
[0088] Table 3 shows the AUC values between any two machine learning models.
[0089]
[0090]
[0091] (5) Improve the Stacking ensemble learning model
[0092] As shown in Table 3, the two learners with the lowest similarity are the Naive Bayes single machine learning model and the KNN single machine learning model.
[0093] Furthermore, find the primary learners with the lowest similarity; take the pair of primary learners with the lowest similarity as a pair of classifiers, and use hierarchical clustering to calculate the similarity between other learners and this pair of classifiers;
[0094] Specifically, the samples are clustered according to a certain level, and multiple clusters are iterated multiple times. This is equivalent to setting up cluster A (KNN, Naive Leaf Bayes) and calculating the similarity between cluster A and classifiers such as logistic regression, support vector machine, random forest, and XGBoost.
[0095] This involves calculating the minimum similarity between the KNN model, the Naive Yebayst model, and classifiers such as Logistic Regression, Support Vector Machine, Random Forest, and XGBoost, respectively, as cluster A, and then comparing these clusters with each of these classifiers. Finding the classifier with the lowest similarity improves the performance of the fusion model.
[0096] s(A, Logistic Regression) = min(s(KNN, Logistic Regression), s(Naive Ye Bayes, Logistic Regression)) = 0.1786
[0097] s(A, Support Vector Machine) = min(s(KNN, Support Vector Machine), s(Naive Leaf Bayes, Support Vector Machine)) = 0.3666
[0098] s(A, Random Forest) = min(s(KNN, Random Forest), s(Naive Bayes, Random Forest)) = 01777
[0099] s(A, Xgboost) = min(s(KNN, Xgboost), s(Naive Bayes, Xgboost)) = 0.1727
[0100] According to the processor calculation, the similarity between XGBoost and cluster A is 0.1727, which is the lowest, indicating the best model fusion and better model generalization performance.
[0101] Furthermore, the primary learners adaptively selected by the JC metric were KNN, Naive Bayes, and XGBoost. They were used to form the Stacking2 ensemble learning model, and the AUC value of the model after cross-validation was 0.8531.
[0102] The Stacking2 ensemble learning model achieves a higher recognition performance for convertible patents than the Stacking ensemble learning model and the base classifier. Therefore, the Stacking2 ensemble learning model, which is an improvement on the Stacking ensemble learning model, is used as the implementation model for recognizing convertible patents.
[0103] S2, acquire the data to be identified.
[0104] The BERT and TextRNN models are used to extract digital features, text features, and bibliographic features of the patent documents to be identified. The data to be identified includes patent data that has been converted.
[0105] S3, Identify the data to be identified.
[0106] The identification method is used to identify the data to be identified, and an identification result is obtained; the identification result includes whether the patent can be commercialized.
[0107] Compared to the base learner and the Stacking ensemble learning model, this embodiment uses the JC metric to adaptively select the target classifier, improving the Stacking ensemble learning model and increasing the differentiation between primary learners while ensuring accuracy.
[0108] Example 2
[0109] Please refer to Figure 3 A system for identifying convertible patents, comprising the following steps:
[0110] Furthermore, the model building module is specifically used for:
[0111] The first unit acquires digital data features, textual data features, and bibliographic data features of patents that have undergone commercialization, including: patent name, number of applicants, applicant type, number of inventors, number of citations, number of citations by examiners, number of cited patents, number of cited non-patent documents, patent type, number of claims, number of words in the main claim, number of pages in the specification, remaining validity period, number of Innojoy family patents, number of Innojoy layout countries, PCT international applications, third-party patents (US, Japan, Europe), number of IPC parts and parts A / B / C / D / E / F / G / H;
[0112] The second unit constructs at least three base classifiers using the training set and classification algorithm, then verifies the base classifiers and tunes their parameters, then uses 5-fold cross-validation, and finally obtains the AUC value of the base classifiers after verification.
[0113] The third unit determines a threshold based on the AUC values verified by multiple base classifiers, setting a threshold to achieve the desired recognition effect according to actual conditions. The process involves setting a threshold, selecting base classifiers that meet the threshold, and discarding those that do not.
[0114] The fourth unit selects the patent sample data that has been transformed and extracts the feature data of the patent sample data; the feature data is then input into two base classifiers to obtain a first classification result and a second classification result.
[0115] The Jaccard distance is calculated using the first classification result, the second classification result, and the pdist function in the scipy library, and the Jaccard coefficient is calculated based on the Jaccard distance.
[0116] The cosine similarity is calculated using the first classification result, the second classification result, and the cosine function in the scipy library; the Jaccard distance is multiplied by the cosine similarity to obtain the JC index value.
[0117] By comparing the similarity between each pair of base classifiers, the pair of base classifiers with the least similarity is found as a pair of classifiers. The similarity between other base classifiers and this pair of classifiers is calculated using hierarchical clustering. The three base classifiers with the least similarity are selected by hierarchical clustering.
[0118] In the fifth unit, three base classifiers with the lowest similarity are selected to form the Stacking2 model. After cross-validation, the AUC value of the Stacking2 model is found to be higher than that of the base classifiers and the Stacking ensemble learning model.
[0119] Furthermore, the acquisition module is specifically used to: use a language representation model to extract digital features, text features, and bibliographic features of the patent document to be identified using a processor, wherein the data to be identified includes convertible patent data.
[0120] Furthermore, the identification module is specifically used to: identify the data to be identified using the identification method, and obtain an identification result; the identification result includes whether the patent will be converted.
[0121] The system provided in Embodiment 2 of the present invention is described in a brief manner. For any parts not mentioned in the embodiment section, please refer to the corresponding content in Embodiment 1 above.
[0122] As another preferred embodiment of the present invention, please refer to Figure 4 The device may include one or more processors 101, one or more input devices 102, one or more output devices 103, and a memory 104, wherein the processors 101, input devices 102, output devices 103, and memory 104 are interconnected via a bus 105. The memory 104 is used to store a computer program, which includes program instructions, and the processor 101 is configured to invoke the program instructions to execute the method steps described in Embodiments 1 and 2.
[0123] It should be understood that, in this embodiment of the invention, the processor 101 may be a central processing unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or any conventional processor.
[0124] Input device 102 may include a keyboard, etc., and output device 103 may include a display (LCD, etc.), a speaker, etc.
[0125] The memory 104 may include read-only memory and random access memory, and provides instructions and data to the processor 101. A portion of the memory 104 may also include non-volatile random access memory. For example, the memory 104 may also store device type information.
[0126] In specific implementations, the processor 101, input device 102, and output device 103 described in the embodiments of the present invention can execute the implementation methods described in the embodiments of the patentable identification method and system provided by the present invention, which will not be repeated here.
[0127] It should be noted that for a more detailed description of the device's workflow, please refer to the aforementioned method implementation section, which will not be repeated here.
[0128] In addition, corresponding to the aforementioned method embodiments and devices, this embodiment of the invention provides a computer-readable storage medium storing a computer program, the computer program including program instructions, which, when executed by a processor, implement: a method and system for identifying convertible patents.
[0129] The computer-readable storage medium can be an internal storage unit of the device as described in any of the foregoing embodiments, such as a system hard drive or memory. The computer-readable storage medium can also be an external storage device of the system, such as a plug-in hard drive, Smart Media Card (SMC), Secure Digital (SD) card, or Flash Card. Furthermore, the computer-readable storage medium can include both internal storage units and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the system. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.
[0130] The aforementioned storage media include: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks or optical disks, and other media that can store program code.
[0131] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.
Claims
1. A method for identifying convertible patents, characterized in that, include: Construct a recognition model based on an improved Stacking algorithm; Obtain the data to be identified; The data to be identified includes convertible patent data; The identification method is used to identify the data to be identified, and an identification result is obtained; the identification result includes whether the patent can be commercialized. Construct a recognition model based on an improved Stacking algorithm, specifically as follows: Obtain the dataset and at least two base classifiers; The dataset is classified using at least two base classifiers to obtain at least two classification results; If the classification result does not reach the set threshold, the base classifier corresponding to the classification result is removed. If the classification result reaches the set threshold, the target classifier is selected by the JC index; The recognition model is constructed using the target classifier, and the recognition model is then validated. The target classifier is selected using the JC metric, specifically as follows: For pairwise base classifiers, calculate the JC index value; The classifier with the smaller JC index value is selected as the target classifier. The JC index value is calculated as follows: Select patent sample data that has already been transformed, and extract the feature data of the patent sample data; the feature data includes numerical features and text features; The numerical characteristics include: number of applicants, number of inventors, number of citations, number of citations by examiners, number of cited patents, number of cited non-patent documents, number of claims, number of words in the sovereign claim, number of pages in the specification, number of Innojoy family members, number of Innojoy layout countries, and remaining validity period. The text features include: patent name, patent type, applicant type, PCT international application, and US-Japan-EU tripartite patent; The feature data is input into two base classifiers respectively to obtain the first classification result and the second classification result; The Jaccard distance is calculated using the first classification result, the second classification result, and the pdist function in the scipy library, and the Jaccard coefficient is calculated based on the Jaccard distance. The cosine similarity is calculated using the first classification result, the second classification result, and the cosine function in the scipy library. The Jaccard distance is multiplied by the cosine similarity to obtain the JC index value.
2. A patent-transferable identification system, characterized in that, include: The model building module is used to build recognition models based on the improved Stacking algorithm; The acquisition module is used to acquire the data to be identified. The data to be identified includes convertible patent data; The identification module is used to identify the data to be identified using the identification method, and obtain an identification result; the identification result includes whether the patent will be converted into a patent. The model building module specifically includes: The first unit is used to obtain the dataset and at least two base classifiers; The second unit is used to classify the dataset using at least two base classifiers to obtain at least two classification results; The third unit is used to remove the base classifier corresponding to the classification result if the classification result does not reach the set threshold. The fourth unit is used to select a target classifier by means of the JC index if the classification result reaches a set threshold. The fifth unit is used to construct the recognition model using the target classifier and to verify the recognition model; The fourth unit is specifically used for: If the classification result reaches the set threshold, then the JC index value is calculated for each pairwise base classifier. The classifier with the smaller JC index value is selected as the target classifier. The calculation of the JC index value is specifically as follows: Select patent sample data that has already undergone conversion, and extract the feature data of the patent sample data; the feature data includes numerical features and textual features; The feature data is input into two base classifiers respectively to obtain the first classification result and the second classification result; The Jaccard distance is calculated using the first classification result, the second classification result, and the pdist function in the scipy library, and the Jaccard coefficient is calculated based on the Jaccard distance. The cosine similarity is calculated using the first classification result, the second classification result, and the cosine function in the scipy library. The Jaccard distance is multiplied by the cosine similarity to obtain the JC index value.
3. A patent-convertible identification device, characterized in that, The system includes a processor, an input device, an output device, and a memory, which are interconnected. The memory is used to store a computer program, which includes program instructions. The processor is configured to invoke the program instructions to perform the steps of the method as described in claim 1.
4. A computer-readable storage medium storing a computer program, the computer program comprising program instructions, characterized in that, When the program instructions are executed by the processor, they implement the steps of the method as described in claim 1.