Patent value intelligent evaluation method based on big data

Abstract

The invention provides a patent value intelligent evaluation method based on big data, and the method comprises the following steps: collecting multi-source heterogeneous data from a patent database, market data, industry reports, science and technology news and social media data through an automatic data acquisition pipeline; cleaning, converting and integrating data by using a data cleaning and integrating technology; features related to patent values are automatically selected by using machine learning and deep learning technologies; through model training and deployment, the market value of patents is evaluated, and real-time prediction and dynamic updating are supported.

Description

Technical Field

[0001] The present invention relates to the field of patent evaluation, and in particular to an intelligent patent value evaluation method based on big data. Background Art

[0002] Universities and other research institutions lack assessment of patent quality and value before applying for patents, resulting in low-value patents occupying a large amount of resources, while high-value patents cannot be fully explored and valued; the implementers of the results fail to disclose the research results to the management department in a timely manner, and there is a lack of an authoritative patent transformation information service platform, resulting in the inability to accurately connect innovation teams and enterprises; universities pay less attention to the training and capacity building of technology transfer personnel, resulting in poor channels for the transformation of scientific and technological achievements and exacerbating the phenomenon of information asymmetry.

[0003] Due to the above reasons, the owners often do not understand the true value of their patents and the market supply and demand situation. Patents involve a wide range of fields, such as patents, trademarks, copyrights, etc. Each field has its own unique laws and evaluation standards, which has led to the current situation where patent value evaluation is too complicated. The existing patent evaluation system usually provides standardized evaluation, and users can only query based on a fixed set of evaluation standards, which lacks flexibility and cannot be personalized according to the needs of different companies. Summary of the invention

[0004] In view of the deficiencies in the prior art, the present invention provides an intelligent patent value evaluation method based on big data, which provides a flexible evaluation module and allows users to customize evaluation criteria and parameters. The evaluation weight can be dynamically adjusted according to the user's industry background, market demand or specific technical field. For example, some industries may pay more attention to the life cycle of technology, while other industries are more concerned with the structural advantages of the citation network. Through customized models and parameter adjustments, users can be provided with more personalized and refined evaluation services.

[0005] In order to achieve the above object, the present invention adopts the following technical solutions: A patent value intelligent evaluation method based on big data, the method comprising the following steps: Collect heterogeneous data from multiple sources, including patent databases, market data, industry reports, technology news, and social media data, through automated data acquisition pipelines; Use data cleaning and integration techniques to clean, transform and integrate data; Use machine learning and deep learning techniques to automatically select features related to patent value; Through model training and deployment, the market value of patents can be evaluated, and real-time prediction and dynamic updates can be supported.

[0006] Furthermore, the data cleaning and integration step includes the following processes: Through the ETL pipeline, multi-source data is processed to clean noise and remove duplicate data; The Z-score standardization technique is used to standardize the data to prevent different data ranges from affecting model training.

[0007] Furthermore, the machine learning automatic feature selection step includes the following process: The correlation between each feature and patent value was calculated using the filtering method, and irrelevant features were filtered using the Pearson correlation coefficient; Use wrapper methods to further filter features based on model performance; Optimize feature weights during model training through embedding methods.

[0008] Furthermore, the deep learning feature extraction step includes the following process: Use convolutional neural networks to extract high-level features from patent texts; Use graph neural networks to analyze patent citation networks and capture the influence of each patent in the network.

[0009] Furthermore, the model selection and optimization step includes: Based on the characteristics of patent data, random forest and XGBoost ensemble learning models are selected to evaluate the value of patents; Automatically tune model parameters through AutoML technology, select the best model and perform hyperparameter adjustment.

[0010] Furthermore, the model training and real-time prediction steps include: Through the distributed computing platform, the model can be trained in parallel to improve the efficiency of model training; Utilize the stream data processing engine to realize dynamic processing and predictive update of real-time data.

[0011] Furthermore, the patent value assessment model includes a regularization term, which is defined by the following formula:

[0012] in, is the complexity control parameter of the tree, T is the number of leaf nodes of the tree, is the L2 regularization parameter, is the weight of the jth leaf node.

[0013] Furthermore, the system dynamically adjusts the model hyperparameters based on the Bayesian optimization algorithm, and its Bayesian optimization formula is:

[0014] in, is the next hyperparameter value to be evaluated, is the expected performance under the current hyperparameter x, σ(x) is the uncertainty of the hyperparameter, and k is the coefficient for balancing exploration and utilization.

[0015] Furthermore, the model is optimized using the stochastic gradient descent (SGD) algorithm, and the update formula is:

[0016] in, are the parameters of the model, is the learning rate, The loss function parameter gradient.

[0017] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention can evaluate patent value based on richer dimensions and provide a more comprehensive and dynamic perspective by integrating multi-source data (including patent data, market dynamics, industry trends, technology news, etc.). For example, combined with industry news or social media discussions, it can discover the sudden rise or fall of certain technologies and improve the ability to predict the future development of technologies.

[0018] 2. Based on deep learning (such as convolutional neural networks, graph neural networks) and machine learning (such as random forests, XGBoost), this invention can more effectively mine the complex nonlinear relationships between patent features. This method can capture deep features in dimensions such as patent text and citation networks, and through automated feature selection technology, it can improve the accuracy and generalization ability of the model without a lot of manual intervention.

[0019] 3. The present invention analyzes patent citation networks through Graph Neural Networks (GNNs), which can dig deeper into the network influence of patents in the entire technology field. Patent citation networks can not only reflect direct citation relationships, but also show the influence of patents indirectly related to citations, and can more accurately measure the importance of patents, especially their potential contribution to the entire technology ecosystem.

[0020] 4. The present invention combines streaming data processing technology (such as Kafka, Flink) and real-time data analysis to dynamically update market and technology trend forecasts by continuously monitoring technology news, industry reports and social media. The future value forecast of the patent can be quickly adjusted according to the latest market changes, making the evaluation results more forward-looking and dynamic.

[0021] 5. The present invention uses AutoML technology. This method can automatically select the best machine learning model based on the data and dynamically adjust the parameters to ensure that the model is always in the best state. In addition, combined with technologies such as reinforcement learning, it can perform self-optimization based on user feedback (such as actual market performance, changes in the company's patent portfolio), ensuring that the evaluation results are more closely related to the actual market situation.

[0022] 6. The present invention provides a flexible evaluation module that allows users to customize evaluation criteria and parameters. The evaluation weights can be dynamically adjusted according to the user's industry background, market demand or specific technical field. For example, some industries may pay more attention to the life cycle of technology, while other industries may pay more attention to the structural advantages of the citation network. Through customized models and parameter adjustments, users can be provided with more personalized and refined evaluation services. DETAILED DESCRIPTION

[0023] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments described below are only examples, and those skilled in the art can think of other obvious variations. The basic principles of the present invention defined in the following description can be used for other embodiments, variations, improvements, equivalents, and other technical solutions that do not deviate from the spirit and scope of the present invention.

[0024] The present invention provides a patent value intelligent evaluation method based on big data, which comprises the following steps: Step 1: Collect heterogeneous data from multiple sources, including patent databases, market data, industry reports, technology news, and social media data, through an automated data acquisition pipeline; Step 2: Use data cleaning and integration techniques to clean, transform and integrate data; Step 3: Use machine learning and deep learning techniques to automatically select features related to patent value; Step 4: Evaluate the market value of patents through model training and deployment, and support real-time prediction and dynamic updates.

[0025] In step one of the application, multi-source heterogeneous data such as patent databases, market data, and industry reports are aggregated into a unified platform through an automated data acquisition pipeline, reducing the complexity of manual data processing.

[0026] Patent database: Real-time capture of patent claims, applicants, citation counts and other information through API.

[0027] Market data: Combine market research reports and corporate financial data to obtain information on technology-related market size, trends, and more.

[0028] Technology News and Social Media Data: Provide cutting-edge intelligence by capturing industry-related news, patent trends, and discussion content.

[0029] In step 2 of this application, because big data technology is good at processing large-scale, complex and differently structured data, ETL (Extract, Transform, Load) technology is used to build a data processing pipeline to clean, transform and integrate data to achieve automated data management. Hadoop, Spark and other big data processing frameworks are used to quickly process data and store it in NoSQL databases such as HBase and Cassandra for efficient access and query. Data cleaning and integration technology is used to clean, transform and integrate data.

[0030] Furthermore, data cleaning and integration are key steps to deal with noise, missing values, duplicate data and other issues in different data sources. Through the ETL (Extract, Transform, Load) pipeline, the raw data is processed to make it usable for machine learning models. The specific implementation process of data cleaning and integration is as follows:

[0031] 1. Data cleaning steps: 1.1 Missing value processing: Use mean imputation, median imputation or interpolation to handle missing values.

[0032] If the feature There are missing values, which are usually filled using the following formula:

[0033] in: : The value of feature j, missing values ​​will be filled with their mean; n: the number of samples; : The value of sample i on feature j.

[0034] Or use the k-nearest neighbor algorithm (KNN) to fill in based on similar samples.

[0035] 1.2 Duplicate value cleaning: Identify duplicate patent records and remove duplicates based on features such as patent number and application date.

[0036] 1.3 Noise data filtering: Identify and remove extreme values ​​through statistical distribution and outlier detection (such as IQR method).

[0037] IQR (Interquartile Range) Method: IQR=Q3-Q1, Among them, Q1: the first quartile, indicating the 25% quantile of the data; Q3: The third quartile, indicating the 75% quantile of the data; IQR: interquartile range, used to identify outliers; Outlier judgment conditions: If the data exceeds the range Q1-1.5×IQR, Q3+1.5×IQR, it is an outlier.

[0038] 2. Data integration steps: 2.1 Data source integration: Use big data frameworks such as Apache Spark or Hadoop to merge and cross-analyze data from different sources (patents, markets, industry reports, etc.) in a distributed environment.

[0039] 2.2 Data standardization: Standardize numerical data to prevent different data ranges from affecting model training results. Z-score standardization is usually used:

[0040]

[0041] Where Z is the normalized value.

[0042] X: original eigenvalue.

[0043] ‌μ: mean of the feature.

[0044] σ: standard deviation of the feature.

[0045] In step 3 of this application, machine learning and deep learning techniques are used to automatically select features related to patent value. Automatic feature selection using machine learning:

[0046] Big data technology can automatically discover important features in high-dimensional data. Based on feature engineering automation (AutoFeatureEngineering) and feature selection technology (such as Lasso regression and feature importance ranking of tree models), it can automatically screen out features that are highly relevant to the value of the patent and avoid excessive manual intervention. Typical features include:

[0047] Patent text features: Use NLP technologies, such as TF-IDF and BERT models, to automatically extract important terms in patent claims.

[0048] Citation relationship network analysis: Build a patent citation network through citation relationships, use graph neural networks (GNNs) for value assessment, and automatically identify key patents and the most cited patents.

[0049] Feature extraction based on deep learning: Deep learning technology can automatically extract complex features from large amounts of data through an end-to-end model. The following technologies can be used:

[0050] Convolutional Neural Networks (CNNs) are used to extract high-level features from patent texts and technical documents.

[0051] Graph Neural Networks (GNNs) can automatically discover influential patents from patent citation networks and quantify their market value.

[0052] It should be noted that automatic feature selection using machine learning aims to select the features most relevant to the patent value from a large number of candidate features, including the following methods: 1. Filter Methods Calculate the correlation between each feature and the target value through statistical methods and select the most significant feature. Use the Pearson correlation coefficient:

[0053]

[0054] Features are retained when the correlation coefficient between them and the target value is above a certain threshold.

[0055] Among them, r(X,Y): the correlation coefficient between feature X and target Y.

[0056] : The i-th sample value of feature X.

[0057] : The mean of feature X.

[0058] : The i-th sample value of target Y.

[0059] : The mean of target Y.

[0060] 2. WrapperMethods Select features based on model performance. For example, the Recursive Feature Elimination (RFE) algorithm:

[0061] Using a base model (such as linear regression or SVM), gradually remove features and evaluate the model performance after each removal.

[0062] Assuming the model is a linear model, the loss function of RFE is:

[0063] in, : Loss function, which measures the prediction error of the model.

[0064] : Parameters of the model.

[0065] m: sample size.

[0066] : The predicted value of the model (linear model is ).

[0067] : The true value of the sample.

[0068] 3.EmbeddedMethods The embedding method directly selects features through model training. A typical example is Lasso regression, which compresses the weights of irrelevant features by introducing L1 regularization.

[0069] The loss function of Lasso regression is:

[0070] in, : The weight of the j-th feature.

[0071] : Regularization parameter, controlling the regularization strength.

[0072] Other symbols are the same as above.

[0073] It should be noted that deep learning is particularly suitable for processing complex structured or unstructured data such as patent texts and citation networks by automatically extracting complex features. Feature extraction based on deep learning includes:

[0074] 1. Convolutional Neural Network (CNN) to extract text features Use CNN to process important information in patent texts, especially patent abstracts and claims. The convolutional layers of CNN can capture local features such as technical terms and key concepts.

[0075] Convolution operation:

[0076] in, : The pixel point (i, j) in the feature map after the convolution operation.

[0077] X: Input feature matrix (e.g., word embedding matrix for patent text).

[0078] K: Convolution kernel (also called filter), used to extract local features.

[0079] M, N: size of the convolution kernel.

[0080] 2. Graph Neural Network (GNN) Extracts Citation Network Features GNN can process patent citation networks and reveal the influence of patents in citation networks. Each node represents a patent and an edge represents a citation relationship. Through the message passing mechanism, GNN can propagate information in the patent network and capture the upstream and downstream influence of each patent.

[0081] Message passing formula:

[0082] in, : The feature vector of node v in the kth layer.

[0083] : The set of neighbor nodes of node v.

[0084] : The weight matrix of the kth layer.

[0085] : Bias term.

[0086] : Activation function, such as ReLU.

[0087] In step 4 of this application, the market value of the patent is evaluated through model training and deployment, and real-time prediction and dynamic update are supported. Model training, deployment and optimization are as follows:

[0088] 1. Model selection and optimization Using large-scale patent data, select appropriate machine learning algorithms for value assessment. The following model combinations can be used to improve prediction accuracy:

[0089] Ensemble learning models such as random forest and XGBoost are good at capturing the nonlinear relationship between patent features and market value.

[0090] AutoML technologies, such as Google's AutoML or H2O.ai, can automatically optimize and select the best model, reducing manual parameter tuning work.

[0091] It should be noted that the model selection is based on the characteristics of patent data (such as high-dimensional sparse data, nonlinear relationships, etc.), and the following models are selected: (1) Random Forest: Reduces the overfitting risk of a single decision tree by integrating multiple decision trees.

[0092] The prediction formula of random forest is:

[0093] in, : The predicted value of the model (the average of multiple decision trees).

[0094] T: number of decision trees in the random forest.

[0095] : The predicted value of the tth decision tree.

[0096] (2) XGBoost: A gradient boosting tree-based model that is good at handling sparse data and nonlinear problems.

[0097] Loss function (squared error with regularization):

[0098] Where, L: loss function.

[0099] : true value.

[0100] : predicted value.

[0101] : Regularization term, measuring model complexity.

[0102] k: the number of decision trees.

[0103]

[0104] in, : Regularization term for the k-th tree.

[0105] : Parameters that control the complexity of the tree, the penalty coefficient for the number of leaf nodes in the tree.

[0106] T: The number of leaf nodes in the tree.

[0107] : L2 regularization parameter controlling the weights.

[0108] : The weight of the j-th leaf node.

[0109] Furthermore, the best model can be automatically selected through AutoML tools (such as Google AutoML and H20.ai), and the parameters can be automatically adjusted through Bayesian optimization. Bayesian optimization estimates model performance based on Gaussian processes and dynamically adjusts hyperparameters.

[0110] Bayesian optimization formula:

[0111] in, : The next hyperparameter value to evaluate.

[0112] : The expected model performance under the current hyperparameters x.

[0113] σ(x): uncertainty in the hyperparameter x.

[0114] k: coefficient to balance exploration and exploitation.

[0115] 2. Model training and real-time prediction 2.1 Model Training The model training process can be accelerated by parallelizing the training model through a distributed computing framework (such as Apache Spark MLlib). On large-scale data sets, the gradient descent method can be used to optimize the model parameters.

[0116] Stochastic Gradient Descent (SGD) Optimization:

[0117] in, : Parameters of the model.

[0118] : Learning rate, which controls the step size of parameter updates.

[0119] : Loss function Parameters gradient.

[0120] 2.2 Real-time prediction Combined with a stream data processing engine (such as Apache Flink or Kafka Streams), the model can make predictions and updates based on real-time input data, supporting dynamic adjustment of patent value assessment. Real-time predictions are based on sliding windows or streaming learning algorithms, ensuring that this method can provide the latest assessment results at any time.

[0121] The model in the present invention supports real-time prediction and dynamic update, and updates the data in the system by regularly capturing the latest patent applications, industry reports, market trends and other information. Using streaming data processing technology (such as Kafka, Flink), real-time monitoring of data updates ensures that the model is always based on the latest data for prediction. Through user feedback (such as the actual evaluation results of experts), the weights and parameters of the model are dynamically adjusted. Using reinforcement learning technology, this method can continuously optimize the evaluation criteria of patent value based on feedback to achieve self-improvement of the model.

[0122] Those skilled in the art should understand that the embodiments of the present invention shown in the above description are only examples and do not limit the present invention. The purpose of the present invention has been fully and effectively achieved. The functions and structural principles of the present invention have been demonstrated and explained in the embodiments, and the embodiments of the present invention may be deformed or modified in any way without departing from the principles.

Claims

1. A patent value intelligent evaluation method based on big data, characterized by: The method comprises the following steps: Collect heterogeneous data from multiple sources, including patent databases, market data, industry reports, technology news, and social media data, through automated data acquisition pipelines; Use data cleaning and integration techniques to clean, transform and integrate data; Use machine learning and deep learning techniques to automatically select features related to patent value; Through model training and deployment, the market value of patents can be evaluated, and real-time prediction and dynamic updates can be supported.

2. According to the big data-based intelligent patent value assessment method of claim 1, it is characterized by: The data cleaning and integration steps include the following processes: Through the ETL pipeline, multi-source data is processed to clean noise and remove duplicate data; The Z-score standardization technique is used to standardize the data to prevent different data ranges from affecting model training.

3. According to the big data-based intelligent patent value assessment method of claim 1, it is characterized by: The machine learning automatic feature selection step includes the following process: The correlation between each feature and patent value was calculated using the filtering method, and irrelevant features were filtered using the Pearson correlation coefficient; Use wrapper methods to further filter features based on model performance; Optimize feature weights during model training through embedding methods.

4. According to the big data-based intelligent patent value assessment method of claim 1, it is characterized by: The deep learning feature extraction step includes the following processes: Use convolutional neural networks to extract high-level features from patent texts; Use graph neural networks to analyze patent citation networks and capture the influence of each patent in the network.

5. According to the big data-based intelligent patent value assessment method of claim 1, it is characterized by: The model selection and optimization steps include: Based on the characteristics of patent data, random forest and XGBoost ensemble learning models are selected to evaluate the value of patents; Automatically tune model parameters through AutoML technology, select the best model and perform hyperparameter adjustment.

6. The method for intelligent patent value assessment based on big data according to claim 1 is characterized by: The model training and real-time prediction steps include: Through the distributed computing platform, the model can be trained in parallel to improve the efficiency of model training; Utilize the stream data processing engine to realize dynamic processing and predictive update of real-time data.

7. The method for intelligent patent value assessment based on big data according to claim 1 is characterized by: The patent value assessment model includes a regularization term, which is defined by the following formula: ; in, is the complexity control parameter of the tree, T is the number of leaf nodes of the tree, is the L2 regularization parameter, is the weight of the jth leaf node.

8. The method for intelligent patent value assessment based on big data according to claim 1 is characterized by: The model is based on the Bayesian optimization algorithm to dynamically adjust the model hyperparameters. The Bayesian optimization formula is: ; in, is the next hyperparameter value to be evaluated, is the expected performance under the current hyperparameter x, σ(x) is the uncertainty of the hyperparameter, and k is the coefficient for balancing exploration and utilization.

9. The method for intelligent patent value assessment based on big data according to claim 1 is characterized by: The model is optimized using the stochastic gradient descent (SGD) algorithm, and the update formula is: ; in, are the parameters of the model, is the learning rate, The loss function parameter gradient.

Images