Patent classification method and device and storage medium

By building a patent classification model, using data cleaning and model training, the subjective problem of patent classification modeling in the existing technology is solved, patent stability and accuracy of victory and defeat prediction are improved, and the accuracy of patent classification results is improved .

CN118940087APending Publication Date: 2024-11-12HUAZHONG UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410959371.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-17
Publication Date
2024-11-12

AI Technical Summary

Technical Problem

In the prior art, patent stability prediction and patent victory or defeat prediction rely on the empirical analysis of professionals, it is highly subjective and lacks universality, and fails to effectively solve the problem of patent classification modeling.

Method used

By obtaining a large amount of patent data, performing data cleaning, encoding and preprocessing, building a patent classification model, using logistic regression, decision tree, gradient enhancement tree and other models for patent classification, eliminating irrelevant indicators, and improving the accuracy of classification results.

Benefits of technology

It has achieved an improvement in the accuracy of patent classification results, provided a patent classification model for specific needs, and improved patent stability and accuracy of victory and defeat prediction, which is objective and universal.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118940087A_ABST
    Figure CN118940087A_ABST
Patent Text Reader

Abstract

The invention discloses a patent classification method, which comprises the following steps of: extracting indexes and data of patents to be classified, and inputting the indexes and the data into a patent classification model to obtain a patent classification result; the patent classification model is constructed according to the following steps: obtaining indexes and classification results of a plurality of patents; performing data cleaning and normalization on the numerical data, and performing data coding on the classified data and the classified result; preprocessing the numerical index and the classification index; and constructing a data set by using the preprocessed indexes and data, and training the initial classification model by using the data set to obtain a patent classification model. According to the method, a large amount of patent data is acquired and processed to fit the patent classification model for specific requirements, so that the accuracy of a patent classification result can be effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of big data and artificial intelligence technology, and more specifically, to a patent classification method, device and storage medium. Background Art

[0002] Patents are both a carrier of technical information and a legal document. With the increasing number of patent infringement lawsuits, the prediction of patent stability and the prediction of patent success or failure have become the focus of increasing attention. In the existing technology, the prediction of patent stability and the prediction of patent success or failure mainly rely on the experience analysis of professionals, which is highly subjective and not universal.

[0003] Patent stability prediction and patent success or failure prediction can be regarded as a patent classification modeling problem. With the development of big data and artificial intelligence technology, it has become possible to classify patent documents using big data. Patent classification in the prior art is generally considered to refer to the disassembly of patent documents by artificial intelligence, the analysis of their technical fields and technical solutions, and the classification of them into the corresponding International Patent Classification (IPC) according to the technical fields and / or technical solutions. However, patent classification in specific classification situations such as patent stability and success or failure prediction has not yet been seen in the prior art. Summary of the invention

[0004] In response to at least one defect or improvement need of the prior art, the present invention provides a patent classification method, device and storage medium, which obtains and processes a large amount of patent data to fit a patent classification model for specific needs, thereby effectively improving the accuracy of patent classification results.

[0005] To achieve the above objectives, according to a first aspect of the present invention, a patent classification method is provided, the method comprising:

[0006] Extract the numerical index, classification index and corresponding numerical data and classification data of the patent to be classified, input them into the patent classification model, and obtain the patent classification result; characterized in that the patent classification model is constructed according to the following steps:

[0007] Obtaining numerical indicators and their corresponding numerical data, classification indicators and their corresponding classification data, and classification results of multiple patents;

[0008] Perform data cleaning and normalization on numerical data, and perform data encoding on categorical data and classification results;

[0009] Using the numerical data after data cleaning, the coded classification data and the coded classification results, the numerical indicators and the classification indicators are preprocessed to obtain the preprocessed numerical indicators and the preprocessed classification indicators;

[0010] Using the preprocessed numerical indicators and their corresponding numerical data, and the preprocessed classification indicators and their corresponding classification data to construct a data set, and using the data set to train an initial classification model to obtain a patent classification model;

[0011] Among them, the classification result is binary data, and the patent classification model is a binary classifier.

[0012] Furthermore, the above patent classification method also includes:

[0013] The classification data and classification results are encoded as 1 and 0 respectively according to the different data.

[0014] Furthermore, the above patent classification method also includes:

[0015] Conduct significance analysis on numerical indicators and classification indicators, and eliminate those that are not significant;

[0016] Calculate the variance of the numerical data corresponding to each numerical indicator, and eliminate the numerical indicators whose variance is less than the variance threshold;

[0017] Conduct correlation analysis on numerical indicators and eliminate those that are not relevant.

[0018] Furthermore, the above patent classification method also includes:

[0019] Assume that the null hypothesis is "there is no significant difference between the patent classification results and the corresponding numerical indicators", and the alternative hypothesis is "there is a significant difference between the patent classification results and the corresponding numerical indicators". Use the Mann-Whitney U test to calculate the p-value and eliminate the numerical indicators with a p-value greater than the first significance threshold.

[0020] Furthermore, the above patent classification method also includes:

[0021] Assume that the null hypothesis is "there is no significant difference between the patent classification results and the corresponding classification indicators", and the alternative hypothesis is "there is a significant difference between the patent classification results and the corresponding classification indicators". Fisher's exact probability method is used to calculate the p-value and eliminate the classification indicators with a p-value greater than the first significance threshold.

[0022] Furthermore, the above patent classification method also includes:

[0023] The point biserial correlation coefficients of the numerical indicators and the classification results and their corresponding significance p-values ​​are calculated, and the numerical indicators whose significance p-values ​​of the point biserial correlation coefficients are greater than the second significance threshold are eliminated.

[0024] Furthermore, the above patent classification method also includes:

[0025] The initial classification model is one of a logistic regression model, a decision tree model, a gradient boosting tree model and a support vector machine model.

[0026] Furthermore, the above patent classification method also includes:

[0027] When the initial classification model is a gradient boosting tree model, the gradient boosting classifier in the Scikit-learn library of Python is called, and the number of iterations, the learning rate, the maximum depth of the decision tree, and the minimum number of samples contained in each leaf node are adjusted, so that the gradient boosting tree model achieves the best performance in patent classification.

[0028] According to the second aspect of the present invention, there is also provided a patent classification device, which includes at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the computer program is executed by the processing unit, the processing unit performs the steps of any one of the above-mentioned methods.

[0029] According to a third aspect of the present invention, there is also provided a storage medium storing a computer program executable by an access authentication device, wherein when the computer program runs on the access authentication device, the access authentication device executes the steps of any one of the above methods.

[0030] In general, the above technical solutions conceived by the present invention can achieve the following beneficial effects compared with the prior art:

[0031] (1) The patent classification method provided by the present invention obtains a large amount of patent data and processes it to fit a patent classification model for specific needs, which can effectively improve the accuracy of patent classification results.

[0032] (2) The patent classification method provided by the present invention adopts different screening methods for common numerical indicators and classification indicators in patent analysis, accurately eliminates irrelevant indicators, and further improves the accuracy of patent classification results. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0034] Figure 1 A flowchart of a modeling method for a patent classification model provided by an embodiment of the present invention;

[0035] Figure 2A schematic diagram of the ROC curve of a patent classification model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0037] The terms "first", "second", "third", etc. in the specification and claims of the present invention and the above-mentioned drawings are used to distinguish different objects, rather than to describe a specific order. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not limited to the listed steps or units, but optionally includes steps or units that are not listed, or optionally includes other steps or units inherent to these processes, methods, products or devices.

[0038] Different from the traditional international patent classification based on the technical content of the patent, the present invention focuses on the legal classification of patents. For example, considering the stability of the patent in the invalidation procedure, it can be divided into "maintaining validity" and "completely invalid". Or considering whether the patent is likely to be supported in the litigation procedure of infringement disputes, it can be divided into "winning" and "losing" from the perspective of the patent owner.

[0039] A patent, especially one that has been substantively examined or has a patent evaluation report issued, will have many indicators available for analysis. Although the results of a single patent invalidation procedure or patent infringement dispute litigation will be affected by many human factors, the results of a large number of patents in the above procedures can be extracted through big data for analysis, and the corresponding classification model can be summarized and established. The classification model established is still objective and can predict the results of a single patent in the above procedures for reference by patent owners or other persons.

[0040] Therefore, as the first embodiment of the present invention, a patent classification method is provided, which specifically includes extracting various indicators and data of the patent to be classified, including numerical indicators, classification indicators, numerical data and classification data, and inputting them into a patent classification model to obtain the result of patent classification.

[0041] For how to obtain the patent classification model, see Figure 1 , this embodiment provides the following steps:

[0042] S1. Obtain numerical indicators and their corresponding numerical data, classification indicators and their corresponding classification data, and classification results of multiple patents.

[0043] In this embodiment, the various indicators of the patent are divided into numerical indicators and classified indicators. Numerical indicators may include indicators with specific numbers such as the number of first rights and authorization period, and classified indicators may include indicators with only "yes" and "no" such as whether there is an agency and whether the applicant is a natural person. This embodiment also includes collecting data corresponding to the above indicators.

[0044] The classification results of patents can be obtained through various databases. For example, if the classification target is whether the patent will be invalidated, the invalidation results of the patent and the above-mentioned indicators can be obtained through commercial patent databases such as Zhihuiya and Patentics. If the classification target is whether the patent will win in the infringement lawsuit, the results of the patent in the litigation process can be obtained through commercial judgment document databases such as Peking University Law Treasure and WestLaw. Through the analysis of the judgment documents and by associating the patent commercial database, the above-mentioned indicators and data of the corresponding patent can be obtained.

[0045] It should be noted that the classification results of patents in this embodiment only include two categories. For example, if the classification target is the invalidation result of the patent, the classification results can be classified into only two categories: "maintaining validity" and "completely invalid". For the two categories of "maintaining validity in part" and "maintaining validity after modification", they can be considered as "maintaining validity" or "completely invalid" as needed. At the same time, considering the difference between design patents and invention patents and utility model patents in patent texts, the "patent" referred to in this invention is limited to invention patents and utility model patents.

[0046] S2. Perform data cleaning and normalization on numerical data, and perform data encoding on categorical data and classification results.

[0047] In this embodiment, not all the original data extracted in the previous step are applicable to the subsequent modeling process of the classification model, and data cleaning and data encoding are required.

[0048] The first step is data cleaning for numerical data. Specifically, it is to detect numerical data, process missing values, remove outliers and duplicate records. The missing values ​​of numerical data are replaced by the median value of the numerical indicator.

[0049] Furthermore, the numerical data needs to be normalized. In this embodiment, the following normalization method can be used, namely:

[0050]

[0051] Data encoding is required for classification data and classification results. Since the classification data usually extracted are "yes" and "no", the classification results can also be considered as "yes" and "no", so the encoding process can be regarded as encoding one of the data as 1 and the other as 0.

[0052] For example, if the established classification model is a prediction model for patent stability, that is, predicting whether the result of the patent in the invalidation procedure is "maintained valid" or "completely invalid", then in the classification result, "maintained valid" can be encoded as 1, and "completely invalid" can be encoded as 0. For classification indicators and classification data, for example, when the classification indicator is "the applicant is a natural person", then in the classification data, "the applicant is a natural person" can be encoded as 1, and "the applicant is not a natural person" can be encoded as 0. Since the classification results and classification data in this embodiment are binary data, that is, there are only data similar to "yes" and "no", therefore, the encoding process only needs to encode one category as 1 and the other category as 0 to distinguish the two categories of data.

[0053] S3. Preprocess the numerical indicators and classification indicators using the numerical data after data cleaning, the encoded classification data, and the encoded classification results to obtain preprocessed numerical indicators and preprocessed classification indicators.

[0054] Considering that there are many types of numerical indicators and classification indicators extracted in the first step, incorporating too many indicators into the training of the classification model will not only seriously prolong the training time, but also lead to poor training results. Therefore, it is necessary to preprocess various indicators.

[0055] The pre-processing step may include the following steps.

[0056] S31. Perform significance analysis on numerical indicators and classification indicators, and eliminate those that are not significant.

[0057] The traditional significance analysis method requires setting the null hypothesis and the alternative hypothesis. Usually, the null hypothesis is that the two variables are unrelated, and the alternative hypothesis is the opposite of the null hypothesis. The p-value is calculated based on the collected data. If the p-value is less than the significance level (usually set to 0.05), the null hypothesis is rejected and the alternative hypothesis is supported. Otherwise, the null hypothesis is accepted. The p-value reflects the probability of the observed data when the null hypothesis is true. The smaller the p-value, the less likely it is to be caused by random chance. The above is a general method for significance analysis.

[0058] In this embodiment, the classification results of the patents can be regarded as dependent variables, and various indicators can be regarded as independent variables. Therefore, in the significance analysis of numerical indicators and classification indicators, the null hypothesis is "there is no significant difference between the patent classification results and the corresponding numerical indicators / classification indicators", and the alternative hypothesis is "there is a significant difference between the patent classification results and the corresponding numerical indicators / classification indicators".

[0059] Further, considering the difference between numerical indicators and classification indicators, in this embodiment, different test methods are required for numerical indicators and classification indicators. For numerical indicators, considering the large amount of data and the fact that each indicator does not obey the normal distribution, the Mann-Whitney U test is used for numerical indicators. For classification indicators, Fisher's exact probability method has a better effect, so Fisher's exact probability method is used for classification indicators in this embodiment. Indicators with a calculated p value greater than the first significance threshold can be eliminated, and the first significance threshold can be set to the common 0.05.

[0060] The above-mentioned significance analysis method is used for preprocessing to screen out the numerical indicators and classification indicators with significant differences after preprocessing.

[0061] S32. Calculate the variance of the numerical data corresponding to each numerical indicator, and eliminate the numerical indicators whose variance is less than the variance threshold.

[0062] Variance is a statistical indicator that measures the degree of data change. For an indicator, if its variance is small, it means that the sample value on this indicator does not change much; on the contrary, if the variance is large, it means that the sample value on this indicator changes a lot. Numerical indicators with little change in value contain insufficient information and are not suitable as data input for classification model training.

[0063] S33. Perform correlation analysis on the numerical indicators and eliminate the numerical indicators that are not relevant.

[0064] Correlation analysis can be performed by calculating the point biserial correlation coefficient. The point biserial correlation coefficient, also known as the Pearson correlation coefficient, is a statistical indicator that measures the strength and direction of the linear relationship between two variables. In this embodiment, the two variables are the classification result and the numerical indicator.

[0065] Calculate the point-by-point correlation coefficient and the corresponding significance p-value of the point-by-point correlation coefficient at the same time, and eliminate the indicators whose significance p-value of the point-by-point correlation coefficient is greater than the second significance threshold (usually set to 0.05), that is, the numerical indicators without correlation can be eliminated.

[0066] S4. Use the preprocessed numerical indicators and their corresponding numerical data, the preprocessed classification indicators and their corresponding classification data to construct a data set, and use the data set to train the initial classification model to obtain a patent classification model.

[0067] Among them, the initial classification model is one of a logistic regression model, a decision tree model, a gradient boosting tree model and a support vector machine model.

[0068] As a second embodiment of the present invention, a specific implementation method of the patent classification method of the first embodiment applied to the prediction of patent stability is also provided. Patent stability refers to the ability of a patent to remain valid when facing the challenge of invalidation. The stability factor is both different and related to the "technical" factor that previous studies have focused on. Specifically, the technical factor mainly focuses on the innovation and advancement of the patent technology, and pays attention to the characteristics and advantages of the technology itself. The more advanced and innovative the technology protected by the patent, the more stable the patent, but the opposite is not necessarily true. The patent stability factor is also related to multiple factors such as the quality of patent writing, the strictness of examination, the number of lawsuits, etc., covering the entire process of patent application, approval and protection. Considering that the results of patent stability prediction can be simplified to "maintain validity" and "all invalid", the patent stability prediction involved in this embodiment can be converted into a patent classification problem as mentioned in the first embodiment.

[0069] The first step is to obtain data. This embodiment obtains 17,705 patent cases that have experienced patent infringement lawsuits in China from January 1, 1999 to December 5, 2023, which were published by the Supreme People's Court on the "Judgment Documents Network". Then, the patent data that has been declared invalid is further searched through the WisdomBud patent database, and a total of 8,204 patent cases are obtained. The above cases are divided into two groups: those that are still valid and those that have been declared invalid. The initial data set is formed by extracting relevant data from the "Judgment Documents Network" and the WisdomBud patent database.

[0070] Regarding the selection of indicators, the following common indicators related to patent stability analysis are selected in this embodiment, as shown in Table 1.

[0071] Table 1 Selection of numerical and classification indicators in patent stability prediction

[0072]

[0073]

[0074] The second step is to clean and normalize the numerical data, and encode the classification data and classification results. The encoding method of the classification data can also refer to Table 1. In the classification results, "maintained valid" is encoded as 1, and "completely invalid" is encoded as 0. For patents that are partially valid and remain valid after modification, all are considered "maintained valid".

[0075] The third step is the screening of indicators. The first step is to conduct a significance analysis. The Mann-Whitney U test is used for numerical indicators. The sample group is patents with the result of "maintaining validity" and the control group is patents with the result of "all invalid". The significance analysis results are shown in Table 2.

[0076] Table 2 Significance analysis of numerical indicators

[0077]

[0078] Among them, *** indicates significance at the 0.01 confidence level, ** indicates significance at the 0.05 confidence level, and * indicates significance at the 0.1 confidence level.

[0079] The results show that the p-values ​​of 16 numerical indicators, including the total number of cited patents in the simple same family, the number of litigation judgments, the year from application to filing, the number of inventors, the number of patent applications in the simple same family, the number of first right words, the authorization cycle, the total number of cited patents, the total number of cited patents, the number of pages of literature, the number of claims, the total number of cited patents in the same family of Zhihuiya, the total number of cited patents in the same family of Zhihuiya, the number of lawsuits, the number of licenses, and the number of pledges, are significant at different levels and are important indicators that need to be introduced into the next step of analysis. However, the p-values ​​of the four numerical indicators, namely the number of citations within 3 years, the number of citations within 5 years, the number of original applicants, and the number of IPC classifications, are not significant, that is, there is no significant difference.

[0080] Fisher's exact probability method was used for the classification indicators, and the results of significance analysis are shown in Table 3.

[0081] Table 3 Significance analysis of classification indicators

[0082]

[0083] Table 3 shows that the p-value of the test result for judging whether the applicant is related to foreign countries or Hong Kong, Macao and Taiwan is less than 0.05, which is significant at the 5% confidence level. That is, in the two groups of patents that remain valid and those that are completely invalid, there is a significant difference in the proportion of applicants related to foreign countries or Hong Kong, Macao and Taiwan, which is an important indicator and needs to be introduced in the next analysis. The other five classification indicators, namely, the applicant is a natural person, the patentee is related to foreign countries or Hong Kong, Macao and Taiwan, the court is in the same region as the patentee, whether the plaintiff is the same as the applicant, and whether there is an agency at the time of application, do not have significant differences and are eliminated in the next analysis.

[0084] The next step is to calculate the variance of the numerical data corresponding to each numerical indicator, as shown in Table 4.

[0085] Table 4 Variance of each numerical index

[0086]

[0087]

[0088] The variance threshold selected in this embodiment, i.e., the variance threshold, is 10. As can be seen from Table 4, the variances of the four indicators, namely, the number of inventors, the total number of cited patents, the number of licenses, and the number of pledges, are less than 10, and should be removed in the next analysis. The remaining 12 indicators: the total number of cited patents of the simple same family, the number of litigation judgments, the year from application to filing, the number of patent applications of the simple same family, the number of first right words, the authorization cycle, the total number of cited patents, the number of document pages, the number of claims, the total number of cited patents of the same family of Wisdom Buds, the total number of cited patents of the same family of Wisdom Buds, and the number of lawsuits have variances greater than 10, and can be introduced into the next analysis.

[0089] The correlation between the numerical index and the classification results is further analyzed. Since the significance and correlation of the classification index are the same, the classification index does not need to calculate the correlation. The point-biserial correlation coefficient analysis method is used to obtain the correlation calculation results of the numerical index as shown in Table 5.

[0090] Table 5 Correlation analysis of numerical indicators

[0091]

[0092] As shown in Table 5, all indicators are positively correlated, and the p-value of the point-biserial correlation coefficient of each indicator is less than 0.05, which means that the point-biserial correlation coefficient of each indicator is significant at the confidence level of 0.05. Therefore, this embodiment selects 12 numerical indicators with obvious correlation with patent stability, including the total number of cited patents of the same family, the number of litigation judgments, the year from application to filing, the number of patent applications of the same family, the number of first rights, the authorization cycle, the total number of cited patents, the number of pages of documents, the number of claims, the total number of cited patents of the same family of Zhihuiya, the total number of cited patents of the same family of Zhihuiya, and the number of lawsuits, and the classification indicator of whether the applicant is involved in foreign countries or Hong Kong, Macao and Taiwan for modeling.

[0093] In this embodiment, the above 12 numerical indicators and 1 classification indicator, as well as the corresponding classification data and numerical data, are combined into a data set, and the data set is divided into a training set and a test set in a ratio of 4:1. Four initial classification models, namely, a logistic regression model, a decision tree model, a gradient boosting tree model and a support vector machine model, are trained respectively. The accuracy of each model is verified using the test set, and the results are shown in Table 6.

[0094] Table 6 Accuracy of different models

[0095]

[0096] It can be seen from Table 6 that the accuracy of the gradient boosting tree is the highest. Therefore, the initial classification model selected in this embodiment is the gradient boosting tree model.

[0097] Furthermore, this embodiment also includes calling the Gradient Boosting Classifier in the Scikit-learn library of Python, which includes multiple hyperparameters for adjusting the performance of the model. The following are some main hyperparameters.

[0098] n_estimators (number of iterations): represents the number of base decision trees and is the number of iterations in the gradient boosting tree.

[0099] learning_rate: Controls how much weights are adjusted for each base model. Smaller learning rates usually require more iterations.

[0100] max_depth (maximum depth of decision tree): limits the depth of decision tree to prevent overfitting.

[0101] min_samples_leaf (the minimum number of samples contained in each leaf node): used to prevent overfitting and limit the minimum number of samples for each leaf node.

[0102] By tuning these hyperparameters, an optimal parameter combination was found so that the model achieved the best performance in the classification task.

[0103] This embodiment also includes evaluating the performance of the gradient boosting tree model through a ROC curve (Receiver Operating Characteristic Curve, ROC), the curve is as follows Figure 2 As shown. By observing the ROC curve, we know that the AUC of the model is 0.86, which is close to 1, indicating that it has good performance. In order to have a more comprehensive understanding of the performance of the model, four key evaluation indicators of the gradient boosting tree model are further calculated:

[0104] 1) Accuracy is 77.51%: This represents the overall proportion of predictions that the model made correctly, which is an important measure of overall performance.

[0105] 2) Precision is 75.89%: Precision reflects the proportion of samples that are actually unstable when the model predicts them to be unstable, which is crucial for application scenarios that focus on reducing false positives.

[0106] 3) Recall is 72.98%: Recall describes the proportion of cases that the model successfully predicts as unstable out of all cases that are actually unstable, with a particular focus on reducing false negatives.

[0107] 4) F1 score is 73.94%: The F1 score is the harmonic mean of precision and recall, which takes into account the balanced performance of the classifier.

[0108] The excellent performance of these indicators shows the excellent performance of the model in this classification task. It is worth noting that the accuracy, precision, recall and F1 score are all over 70%, which shows that the model has achieved a good balance in predicting the stability and instability of patents, that is, it verifies that the modeling method described in the first embodiment has a relatively good effect.

[0109] As a third embodiment of the present invention, a patent classification device is also provided, which includes at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the computer program is executed by the processing unit, the processing unit performs the steps of any one of the above-mentioned methods.

[0110] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when executed by a processor. The computer-readable storage medium may include, but is not limited to, any type of disk, including a floppy disk, an optical disk, a DVD, a CD-ROM, a micro drive, and a magneto-optical disk, a ROM, a RAM, an EPROM, an EEPROM, a DRAM, a VRAM, a flash memory device, a magnetic card or an optical card, a nanosystem (including a molecular memory IC), or any type of medium or device suitable for storing instructions and / or data.

[0111] This invention was funded by the China 21st Century Agenda Management Center Commissioned Project: Research on Key Technology Equipment List and Intellectual Property Rights for Carbon Capture, Utilization and Storage (CCUS) (20170986) and the Ministry of Education Humanities and Social Sciences Fund Youth Fund Project (20YJC630201): Research on Patent Path Lock-in Effect and Breakthrough Countermeasures for Key Core Technologies.

[0112] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0113] It will be easily understood by those skilled in the art that the above description is only 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 in the protection scope of the present invention.

Claims

1. A patent classification method, comprising extracting numerical indicators, classification indicators and corresponding numerical data and classification data of patents to be classified, inputting them into a patent classification model, and obtaining patent classification results; characterized in that: The patent classification model is constructed according to the following steps: Obtaining numerical indicators and their corresponding numerical data, classification indicators and their corresponding classification data, and classification results of multiple patents; Perform data cleaning and normalization on numerical data, and perform data encoding on categorical data and classification results; Using the numerical data after data cleaning, the coded classification data and the coded classification results, the numerical indicators and the classification indicators are preprocessed to obtain the preprocessed numerical indicators and the preprocessed classification indicators; Using the preprocessed numerical indicators and their corresponding numerical data, and the preprocessed classification indicators and their corresponding classification data to construct a data set, and using the data set to train an initial classification model to obtain a patent classification model; Among them, the classification result is binary data, and the patent classification model is a binary classifier.

2. The patent classification method according to claim 1, characterized in that: Data encoding for classification data and classification results specifically includes: The classification data and classification results are encoded as 1 and 0 respectively according to the different data.

3. The patent classification method according to claim 1, characterized in that: Preprocessing specifically includes: Conduct significance analysis on numerical indicators and classification indicators, and eliminate those that are not significant; Calculate the variance of the numerical data corresponding to each numerical indicator, and eliminate the numerical indicators whose variance is less than the variance threshold; Conduct correlation analysis on numerical indicators and eliminate those that are not relevant.

4. The patent classification method according to claim 3, characterized in that: The significance analysis of numerical indicators includes: Assume that the null hypothesis is "there is no significant difference between the patent classification results and the corresponding numerical indicators", and the alternative hypothesis is "there is a significant difference between the patent classification results and the corresponding numerical indicators". Use the Mann-Whitney U test to calculate the p-value and eliminate the numerical indicators with a p-value greater than the first significance threshold.

5. The patent classification method according to claim 3, characterized in that: The significance analysis of classification indicators includes: Assume that the null hypothesis is "there is no significant difference between the patent classification results and the corresponding classification indicators", and the alternative hypothesis is "there is a significant difference between the patent classification results and the corresponding classification indicators". Fisher's exact probability method is used to calculate the p-value and eliminate the classification indicators with a p-value greater than the first significance threshold.

6. The patent classification method according to claim 3, characterized in that: The correlation analysis of numerical indicators includes: The point biserial correlation coefficients of the numerical indicators and the classification results and their corresponding significance p-values ​​are calculated, and the numerical indicators whose significance p-values ​​of the point biserial correlation coefficients are greater than the second significance threshold are eliminated.

7. The patent classification method according to claim 1, characterized in that: The initial classification model is one of a logistic regression model, a decision tree model, a gradient boosting tree model and a support vector machine model.

8. The patent classification method according to claim 7, characterized in that: When the initial classification model is a gradient boosting tree model, the gradient boosting classifier in the Scikit-learn library of Python is called, and the number of iterations, the learning rate, the maximum depth of the decision tree, and the minimum number of samples contained in each leaf node are adjusted, so that the gradient boosting tree model achieves the best performance in patent classification.

9. A patented classification device, characterized in that, The method comprises at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the computer program is executed by the processing unit, the processing unit executes the steps of the method according to any one of claims 1 to 8.

10. A storage medium, characterized in that: It stores a computer program executable by an access authentication device. When the computer program runs on the access authentication device, the access authentication device executes the steps of the method according to any one of claims 1 to 8.