Patent-to-technical standard conversion potential evaluation method and system based on feature interaction

Abstract

The invention relates to a method and a system for evaluating potential of conversion from a patent to a technical standard based on feature interaction. The method comprises the following steps: acquiring original data of a to-be-evaluated patent and corresponding technical standard data; screening features having significant influence on a prediction result from a technology association dimension, a patent maturity dimension, a standard demand dimension, a technology influence dimension and an application scene dimension; preprocessing the features according to the feature types; constructing a feature interaction item for capturing the influence of the multi-feature joint effect on the patent transformation potential, and adding the feature interaction item into the preprocessed features; constructing a transformation potential prediction model based on XGBoost, and obtaining a convergent transformation potential prediction model through data division based on stratified sampling, class imbalance processing and iterative training; and inputting the features with the feature interaction items into the convergent transformation potential prediction model to obtain a transformation potential evaluation label, and realizing patent-to-technical standard transformation potential evaluation of the to-be-evaluated patent.

Description

Technical Field

[0001] This invention relates to the field of patent data analysis technology, and in particular to a method and system for evaluating the potential of patents to be transformed into technical standards based on feature interaction. Background Technology

[0002] The coordinated transformation of technical standards and patents is key to the industrialization of patents and the improvement of industry standards. Especially in the power industry, patent standardization decisions directly affect the company's bargaining power and resource efficiency. It is necessary to complete three core tasks: predicting transformation potential, recommending standard types, and calculating the matching degree of technical directions. Moreover, the objectives of each task are significantly different.

[0003] Existing technologies use traditional machine learning methods to build models, but these methods suffer from numerous problems due to insufficient adaptability: (1) Feature sharing in multiple tasks leads to target conflicts. Existing technologies often use shared features combined with a single model to handle multiple tasks without customizing features for task objectives. For example, conversion potential prediction needs to focus on features such as technology relevance to prevent omissions, and standard type recommendation needs to focus on balanced features such as technology value level. Shared features are prone to weight conflicts, resulting in the accuracy of each task not meeting the standard.

[0004] (2) The algorithm selection is not closely related to the business. Linear models such as logistic regression cannot capture the nonlinear interaction of features, and the recall rate of conversion potential prediction is low, which is difficult to meet the requirements. Ensemble models such as random forest are sensitive to the redundancy of high-dimensional features. Lightweight CNNs are poorly adapted to structured features and cannot balance accuracy and recall.

[0005] (3) The feature engineering is not targeted enough, and core features such as secondary technical term similarity and standard gap matching degree are missing; the text features are not filtered with the power-specific dictionary and the numerical features are not normalized, resulting in low calculation accuracy; no feature interaction items are constructed, and key patterns such as high-tech matching combined with high gap filling cannot be identified.

[0006] In summary, existing technologies cannot meet the needs of the power industry for standardized, accurate identification, reasonable recommendation, and clear explanation of patents, and there is an urgent need for machine learning solutions with dedicated modeling capabilities. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of the existing technology by providing a method and system for evaluating the potential of patents to be converted into technical standards based on feature interaction. This invention aims to solve problems in existing methods for evaluating the potential of patents to be converted into technical standards, such as target conflicts caused by multi-task feature sharing, disconnect between algorithms and the business needs of the power industry, and lack of specificity in feature engineering.

[0008] The objective of this invention can be achieved through the following technical solutions: One aspect of the present invention provides a method for evaluating the potential for patent to be transformed into a technology standard based on feature interaction, comprising the following steps: Obtain the original data and corresponding technical standard data of the patent to be evaluated; We selected features that significantly impacted the prediction results from the dimensions of technology relevance, patent maturity, standard demand, technology influence, and application scenarios. Preprocess the features according to their type; Construct feature interaction terms that capture the combined effect of multiple features on patent commercialization potential, and add them to the preprocessed features; A conversion potential prediction model based on XGBoost is constructed. Through data partitioning based on stratified sampling, handling of class imbalance, and iterative training, a convergent conversion potential prediction model is obtained. By inputting features with interactive features into a convergent conversion potential prediction model, a conversion potential assessment label is obtained, enabling the assessment of the patent conversion potential of the patent to be evaluated into a technical standard.

[0009] As a preferred technical solution, the feature interaction items include: The first interaction term is obtained by multiplying the second-level technical term similarity with the standard gap matching degree, and is used to characterize the patent features of high technical matching and high gap filling. The second interaction term, obtained by multiplying the grant period by the number of citations, is used to characterize mature and high-impact patent features.

[0010] As a preferred technical solution, the features that significantly affect the prediction results include: The similarity of secondary technical terms in the technology association dimension is obtained based on the vector cosine similarity of secondary technical terms in patents and standards; The normalized value of the patent maturity dimension's grant years is calculated as (current year - grant year) / (maximum grant years - minimum grant years). The standard gap matching degree of the standard demand dimension is obtained based on the similarity of patented technologies with LDA topics in areas not covered by standards; The logarithmic value of citation counts for the technology influence dimension is calculated as log (1 + patent citation counts); The number of application scenarios covered in the application scenario dimension is based on the number of industry application scenarios mentioned in the patent specification.

[0011] As a preferred technical solution, the data partitioning, class imbalance handling, and iterative training based on stratified sampling include the following steps: The dataset is divided into training, validation and test sets by a stratified sampling strategy, keeping the ratio of positive to negative samples in each subset consistent with the original data. By using SMOTE oversampling to synthesize positive samples to balance the ratio of positive to negative samples and adjusting the prediction threshold, class imbalance can be addressed. The conversion potential prediction model is trained iteratively. Initial hyperparameters are set. After each preset round of iteration, the core evaluation index is calculated on the validation set. Training is terminated when the core index on the validation set fails to improve for a consecutive preset round.

[0012] As a preferred technical solution, the SMOTE oversampling process includes: Synthetic samples are generated from positive samples based on the K-nearest neighbor algorithm, averaging the ratio of positive to negative samples. When the model output probability is greater than or equal to the prediction threshold, it is determined to be a potential patent.

[0013] As a preferred technical solution, the preprocessing of features according to feature type includes the following steps: For numerical features, scale normalization is performed; for text features, domain dictionary filtering is performed; and for categorical features, encoding conversion is performed.

[0014] As a preferred technical solution, features that significantly affect the prediction results are selected by using the XGBoost feature importance gain value.

[0015] As a preferred technical solution, the original data includes patent technical text, authorization information, cited data, and application scenario description.

[0016] As a preferred technical solution, the technical standard data includes a standard technical terminology database and information on standard gap areas.

[0017] Another aspect of the present invention provides a patent-to-technology-standard transformation potential assessment system based on feature interaction, for implementing the aforementioned patent-to-technology-standard transformation potential assessment method, the system comprising: The salient feature filtering module is used to obtain the original data of the patent to be evaluated and the corresponding technical standard data, and to filter features that have a significant impact on the prediction results from the dimensions of technology relevance, patent maturity, standard demand, technology influence and application scenario. The feature interaction term modeling module is used to construct feature interaction terms that capture the combined effect of multiple features on the patent transformation potential and add them to the preprocessed features; The conversion potential prediction model building module is used to build a conversion potential prediction model based on XGBoost. Through data partitioning based on stratified sampling, handling of class imbalance, and iterative training, a converged conversion potential prediction model is obtained. The patent-to-technology-standard conversion potential assessment module is used to input the features with interactive features into a convergent conversion potential prediction model, obtain conversion potential assessment labels, and realize the patent-to-technology-standard conversion potential assessment of the patent to be assessed.

[0018] Compared with the prior art, the present invention has at least one of the following beneficial effects: (1) Achieving a multi-dimensional and accurate characterization of patent transformation potential: This invention selects features that significantly affect the prediction results from the dimensions of technology association, patent maturity, standard demand, technology influence and application scenarios. The features are preprocessed according to the feature type, and feature interaction terms that capture the combined effect of multiple features on patent transformation potential are constructed and added to the preprocessed features. This achieves a multi-dimensional and accurate characterization of patent transformation potential, improves the contribution of core features to model prediction, and enhances the sensitivity of high-potential patent identification by feature interaction terms, thus laying a feature foundation for evaluation accuracy.

[0019] (2) Improve the recall rate of the model: The present invention constructs a model based on the XGBoost algorithm. Through data partitioning based on stratified sampling, handling class imbalance, and iterative training, a convergent conversion potential prediction model is obtained. After each preset round of iteration, the core evaluation index is calculated on the validation set. Training is terminated when the core index on the validation set fails to improve for a consecutive preset round, thus avoiding overfitting and balancing evaluation accuracy and interpretability. (3) Improve positive sample recognition capability: In SMOTE oversampling, the present invention generates synthetic samples for positive samples based on the K-nearest neighbor algorithm, so that the ratio of positive and negative samples is averaged. When the model output probability is greater than or equal to the prediction threshold, it is judged as a potential patent, effectively balancing the sample distribution and improving the positive sample recall rate. Attached Figure Description

[0020] Figure 1 This is a flowchart of the patent-to-technology-standard conversion potential assessment method based on feature interaction in the embodiment. Figure 2 This is a schematic diagram of a patent-to-technology-standard conversion potential assessment system based on feature interaction, as shown in the embodiment. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0022] Example 1 To address the aforementioned problems in the existing technologies, this embodiment provides a method for evaluating the potential of patents to be converted into technical standards based on feature interaction. This method is designed for the scenario of evaluating the potential of patents to be converted into technical standards in the power industry. By constructing a feature system containing core features and exclusive interaction items, employing XGBoost, and combining SMOTE oversampling and hyperparameter optimization, it achieves high recall and high accuracy in evaluating the potential of patent conversion. This accurately identifies high-potential patents to avoid the risk of missed detection, while controlling the false detection rate of low-potential patents, thus providing reliable technical support for patent standardization decisions.

[0023] This method aims to predict the potential for commercialization, assessing the likelihood and value of synergistic commercialization of technical standards or patents in the future. This requires comprehensive consideration of multiple factors, such as the advancement of technical standards, the technological innovation of patents, the urgency of market demand, and the support of the policy environment. By analyzing and modeling these factors, it predicts which technical standards and patents have high commercialization potential, providing decision-making support for enterprises and relevant institutions to allocate resources rationally and prioritize the synergistic commercialization of high-potential technical standards and patents. The prediction process requires collecting and analyzing a large amount of historical data, including successfully commercialized technical standards and patent cases, development trends in related technological fields, and market dynamics, to establish an effective predictive model and accurately assess commercialization potential.

[0024] The task of predicting conversion potential is a binary task, namely, determining whether a patent is "potential" or "unpotential." Its core objective is risk avoidance, with a particular emphasis on prioritizing the prevention of the risk of overlooking high-potential patents. In the context of the power industry, if high-potential patents such as energy storage safety control technology are mistakenly judged as "unpotential," it may lead to the company losing its voice in setting industry standards. The resulting hidden losses far outweigh the waste of evaluation costs caused by mistakenly judging low-potential patents as "potential." Therefore, the core orientation of the entire evaluation system is set as "not overlooking any patent with conversion potential."

[0025] See Figure 1 The method includes the following steps: Step S1: Obtain the original data and corresponding technical standard data of the patent to be evaluated.

[0026] Step S2 involves selecting features that significantly impact the prediction results from the dimensions of technology relevance, patent maturity, standard demand, technology influence, and application scenarios.

[0027] This step revolves around the core judgment logic of "transformation potential," accurately capturing whether the patented technology meets standard requirements and has the conditions for implementation, providing solid feature support for the assessment of transformation potential. The system designs eight core features and uses XGBoost feature importance gain values ​​for selection. Its working logic leverages the quantitative evaluation of the contribution of each feature during the XGBoost model training process, prioritizing the retention of features that have a greater impact on the model's prediction results. This enhances the relevance and effectiveness of the feature system, ensuring that subsequent evaluations focus more on key elements.

[0028] Among them, the similarity of secondary technical terms in the technology association dimension, by calculating the vector cosine similarity of the secondary technical terms of the patent and the standard, can quantify the degree of matching between the patent technology and the core technology of the standard, directly reflecting the degree of fit between the two at the technical detail level, and is the basis for judging whether the patent can form an effective association with the standard.

[0029] The normalized value of the patent maturity dimension, after being converted through a specific calculation method, can reflect the stability of the patent technology. Generally speaking, the higher the maturity of the patent, the more thoroughly it has been tested by the market and time, and the lower the technical risks it faces when it is converted into a standard. This provides a time-dimensional reference for the feasibility of patent conversion.

[0030] The standard gap matching degree in the standard demand dimension can effectively assess the ability of a patent to fill the standard gap by measuring the similarity between the patent technology and the LDA topic in the direction not covered by the standard. The higher the gap matching degree, the greater the value of the patent in improving the standard system and meeting the technical needs of the industry that are not covered, and the more prominent the transformation potential.

[0031] The logarithmic value of the number of citations in the dimension of technological influence can reflect the industry recognition of the patented technology. The more citations, the higher the influence and acceptance of the patented technology in the industry, and the easier it is to gain widespread recognition and support in the industry when it is transformed into a standard.

[0032] The number of application scenarios covered by the application scenario dimension can reflect the universality of the technology by counting the number of industry application scenarios mentioned in the patent specification. The more scenarios covered, the wider the scope of the technology can be adapted, and the broader the fields that can be applied after being transformed into a standard, further enhancing its transformation value.

[0033] The key features selected comprehensively characterize the patent's transformation potential from multiple dimensions, including technological relevance, maturity, ability to supplement standards, industry influence, and application scope. This provides multi-dimensional and in-depth feature basis for the model to accurately assess the transformation possibility of new patents, helping to improve the scientificity and accuracy of transformation potential assessment.

[0034] See Table 1 for key features and their business significance.

[0035] Table 1 Key Features and Business Significance Step S3: Preprocess the features according to their type.

[0036] As a crucial link connecting raw data and model training, the core objective of this step is to improve the accuracy and stability of the model's judgment on patent conversion potential in binary classification tasks by eliminating data noise, unifying feature scale, and strengthening effective information through targeted processing. This lays a solid data foundation for the subsequent model to learn historical correlation patterns and output reliable prediction results.

[0037] For numerical features, the Min-Max normalization method is used to process "grant period" and "citation count" to map them to the [0,1] interval. This operation can effectively eliminate the influence of different features due to differences in units, avoid a certain feature from occupying an unreasonable dominant position in model training due to an excessively large numerical range, and ensure that each numerical feature has a balanced weight in the model, thereby more accurately reflecting its actual impact on patent commercialization potential.

[0038] In text feature processing, the focus is on refining the "secondary technical term similarity." Through "domain dictionary filtering," relying on a power industry-specific terminology database, words irrelevant to the technical correlation judgment (such as "general algorithm") are removed from the text. This approach focuses on core technical terms, reduces noise interference, and significantly improves the accuracy of similarity calculation between patent and standard secondary technical terms, making the text features more accurately reflect the degree of technical correlation between the two.

[0039] Based on category characteristics, one-hot encoding is implemented for "patent type" (invention / utility model). Invention patents are assigned values ​​of [1,0], and utility model patents are assigned values ​​of [0,1]. Invention patents typically have stronger technological innovation and higher potential for conversion into standards. This encoding method not only reflects the essential differences between the two types of patents but also transforms category information into a numerical form that the model can recognize. This facilitates the model in capturing the correlation between different patent types and conversion potential, thereby improving the accuracy of classification judgments.

[0040] Step S4: Construct feature interaction terms that capture the combined effect of multiple features on the patent transformation potential, and add them to the preprocessed features.

[0041] As a crucial step in deeply exploring the correlation patterns between features, the core objective of this step is to manually construct feature interaction terms to capture the combined effect of multiple features on patent commercialization potential, enhance the ability to identify key patterns, thereby providing the model with richer and more representative information and improving the accuracy of judging patent commercialization potential.

[0042] For example, two types of feature interaction items were constructed: The first interaction is "Secondary technical term similarity × Standard gap matching degree". This interaction highlights the patent characteristics of "high technical matching + high gap filling". For example, when the secondary technical term similarity is 0.9 and the standard gap matching degree is 0.8, the product of the two is 0.72. Such patents have significantly higher conversion potential because they are highly matched with the core technology of the standard and can effectively fill the gap in the standard. Through this interaction, the model can more keenly capture the high conversion potential signal brought about by this dual advantage.

[0043] Second, the interaction term is "grant term × number of citations". This interaction term aims to strengthen the characteristic signal of "mature and high-impact" patents. For example, when the grant term is 3 years and the number of citations is 50, the product of the two is 150. Such patents have both high technical stability (maturity) and industry recognition (high impact). This interaction term allows the model to more clearly identify the high conversion potential characteristics of patents with both high technical stability and high industry recognition, further improving the model's ability to capture key patterns.

[0044] Step S5: Construct a conversion potential prediction model based on XGBoost. Through data partitioning based on stratified sampling, handling of class imbalance, and iterative training, a converged conversion potential prediction model is obtained.

[0045] The core positioning of the conversion potential prediction model is to accurately identify patents with standard conversion potential, focusing on solving the business pain point of "missing high-potential patents". In the patent layout of the power industry, missing a core patent that meets future standard requirements may cause the company to lose its industry influence, while the cost of misjudging low-potential patents is only the cost of ineffective evaluation. Therefore, the model design prioritizes high recall while also considering accuracy, ultimately achieving the business value of accurate screening and minimal omissions.

[0046] This step involves a multi-algorithm comparative analysis based on business objectives to select models that not only meet core indicator requirements but also are highly adaptable to business scenarios, providing reliable algorithmic support for judging patent commercialization potential. Logistic regression, as a linear model, cannot capture the interaction relationships between features, resulting in a low recall rate that is difficult to meet the business requirement of zero tolerance for missed detections. Although random forest is an ensemble model, its ensemble effect is weaker than XGBoost, and it is more sensitive to redundancy in high-dimensional features. Lightweight CNNs are strong in text feature processing, but their adaptability to structured features is poor, and their recall rate fails to meet business standards. XGBoost, on the other hand, demonstrates excellent non-linear modeling capabilities, with the highest recall rate among various algorithms, while also supporting feature importance output and providing excellent interpretability.

[0047] The key focus was on "recall rate" and "business adaptability," and XGBoost was ultimately chosen. The core reason is that its recall rate can meet the business requirements for false negative rate, and its strong interpretability allows business personnel to directly trace the decision-making logic, which is highly consistent with the project's core objective of "prioritizing the prevention of false negative risks of high-potential patents."

[0048] Specifically, the model training process includes the following steps: (1) Data partitioning To ensure the scientific rigor of model training and the objectivity of evaluation, a stratified sampling strategy was adopted to divide the original dataset (i.e., patent data in the field of new energy equipment). The positive and negative sample ratios of each subset were strictly kept consistent with the original data to avoid interference from differences in data distribution on model training and performance evaluation.

[0049] The specific division ratio is as follows: the training set accounts for 70%, which serves as the core data for the model to learn historical correlation patterns and is used for parameter fitting and pattern capture; the validation set accounts for 20%, which is mainly used for hyperparameter optimization and early stopping judgment during the training process, and to monitor the performance changes of the model on unseen data; the test set accounts for 10%, which is independent of the training and optimization process and is used to finally evaluate the model's generalization ability and adaptability to actual business.

[0050] This partitioning method ensures sufficient training data and effectively avoids overfitting of the model to specific data by setting independent validation and test sets, thus providing a reliable data foundation for subsequent training and evaluation.

[0051] (2) Training process Model training follows a closed-loop process of "initial setup - iterative optimization - early stopping and convergence" to ensure efficient learning of feature patterns while avoiding overfitting risks. Initial hyperparameter settings reference XGBoost default values ​​and industry experience, such as learning_rate=0.1, max_depth=8, and subsample=0.7, providing a basic parameter framework for model training. During iterative optimization, after every 50 iterations, core metrics such as recall and F1-score are calculated on the validation set, and model parameter and performance changes are recorded, forming a dynamic monitoring curve of "iteration rounds - performance metrics".

[0052] The early stopping mechanism is set as follows: when the validation set recall rate does not improve for three consecutive iterations (fluctuation range ≤ 0.01), training is automatically terminated, and the current optimal parameters are used as the final model parameters. This mechanism can prevent the model from overfitting due to excessive iteration (such as memorizing noisy data in the training set), and also avoid underperformance caused by insufficient iteration, so as to achieve a balance between sufficient learning and generalization ability.

[0053] (3) Model hyperparameter optimization With the balance between recall and precision as the core optimization direction, the model maintains good overall performance while meeting recall requirements through fine-tuning of parameters.

[0054] The hyperparameters were optimized using a grid search + 5-fold cross-validation approach, focusing on key parameters affecting recall, such as max_depth and subsample. The final core hyperparameters each had a clear optimization objective and business significance (Table 2). `learning_rate` controls the iteration step size to avoid model oscillations and ensure convergence to the optimal solution; `max_depth` limits the tree depth to prevent overfitting to noisy data and balance the model's generalization ability; `subsample` is the sample sampling ratio to improve the model's adaptability to unseen data; `colsample_bytree` is the feature sampling ratio to reduce the impact of high-dimensional feature redundancy, allowing the model to focus on core features; `reg_lambda` suppresses overfitting through L2 regularization strength, ensuring model stability on new data; and `scale_pos_weight` is set based on class weights (positive samples / negative samples ≈ 1:2.3) to improve the recall of positive samples.

[0055] Table 2 Core Hyperparameters (4) Handling class imbalance To address the issue that only 30% of the patents in the training set are considered "potential" (while negative samples make up 70%), a dual strategy combining SMOTE oversampling and threshold adjustment is employed to balance the class distribution and ensure the model's ability to identify positive samples. SMOTE oversampling is a method to solve data class imbalance by synthesizing minority class samples to balance the ratio of positive to negative samples and prevent the model from being biased towards the majority class.

[0056] SMOTE oversampling uses synthetic sampling to generate synthetic samples based on K-nearest neighbors, adjusting the positive-to-negative sample ratio from 1:2.3 to 1:1. This effectively prevents the model from biasing towards negative sample predictions due to an excessively high proportion of negative samples. Simultaneously, the prediction threshold is adjusted, lowering the default threshold from 0.5 to 0.6. When the model output probability is ≥0.6, it is considered "potential." Validation tests show that this threshold improves recall while maintaining accuracy, satisfying the need to identify high-potential patents while ensuring overall accuracy.

[0057] Step S6 involves inputting the feature interaction terms into the convergent conversion potential prediction model to obtain a conversion potential evaluation label, thereby realizing the evaluation of the patent conversion potential of the patent to be evaluated into a technical standard.

[0058] The conversion potential prediction (binary classification task) is achieved using a convergent model. The core objective is to "not miss high-potential patents". Therefore, recall rate is used first to measure the coverage of positive samples. Accuracy rate is used to control the risk of misjudgment and avoid low-potential patents from consuming evaluation resources. The overall discrimination ability (ROC-AUC) is used to comprehensively evaluate the model's stability in distinguishing between "potential / non-potential" patents.

[0059] Example 2 Based on Example 1, this example provides a patent-to-technology-standard transformation potential assessment system based on feature interaction, used to implement the aforementioned patent-to-technology-standard transformation potential assessment method. The system includes: (1) The significant feature screening module is used to obtain the original data of the patent to be evaluated and the corresponding technical standard data, and to screen features that have a significant impact on the prediction results from the dimensions of technology association, patent maturity, standard demand, technology influence and application scenario. (2) Feature interaction term modeling module, used to construct feature interaction terms that capture the combined effect of multiple features on the patent transformation potential, and add them to the preprocessed features; (3) Conversion potential prediction model construction module, used to construct a conversion potential prediction model based on XGBoost. Through data partitioning based on stratified sampling, class imbalance processing and iterative training, a converged conversion potential prediction model is obtained. (4) The patent to technology standard conversion potential assessment module is used to input the feature with feature interaction items into the conversion potential prediction model to obtain the conversion potential assessment label, thereby realizing the patent to technology standard conversion potential assessment of the patent to be assessed.

[0060] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for evaluating the potential for patent transformation into technical standards based on feature interaction, characterized in that, Includes the following steps: Obtain the original data and corresponding technical standard data of the patent to be evaluated; We selected features that significantly impacted the prediction results from the dimensions of technology relevance, patent maturity, standard demand, technology influence, and application scenarios. Preprocess the features according to their type; Construct feature interaction terms that capture the combined effect of multiple features on patent commercialization potential, and add them to the preprocessed features; A conversion potential prediction model based on XGBoost is constructed. Through data partitioning based on stratified sampling, handling of class imbalance, and iterative training, a convergent conversion potential prediction model is obtained. By inputting features with interactive features into a convergent conversion potential prediction model, a conversion potential assessment label is obtained, enabling the assessment of the patent conversion potential of the patent to be evaluated into a technical standard.

2. The method for evaluating the potential for patent to be transformed into a technical standard based on feature interaction according to claim 1, characterized in that, The feature interaction items include: The first interaction term is obtained by multiplying the second-level technical term similarity with the standard gap matching degree, and is used to characterize the patent features of high technical matching and high gap filling. The second interaction term, obtained by multiplying the grant period by the number of citations, is used to characterize mature and high-impact patent features.

3. The method for evaluating the potential for patent transformation into technical standards based on feature interaction according to claim 1, characterized in that, The features that significantly affect the prediction results include: The similarity of secondary technical terms in the technology association dimension is obtained based on the vector cosine similarity of secondary technical terms in patents and standards; The normalized value of the patent maturity dimension's grant years is calculated as (current year - grant year) / (maximum grant years - minimum grant years). The standard gap matching degree of the standard demand dimension is obtained based on the similarity of patented technologies with LDA topics in areas not covered by standards; The logarithmic value of citation counts for the technology influence dimension is calculated as log (1 + patent citation counts); The number of application scenarios covered in the application scenario dimension is based on the number of industry application scenarios mentioned in the patent specification.

4. The method for evaluating the potential for patent transformation into technical standards based on feature interaction according to claim 1, characterized in that, The data partitioning, class imbalance handling, and iterative training based on stratified sampling include the following steps: The dataset is divided into training, validation and test sets by a stratified sampling strategy, keeping the ratio of positive to negative samples in each subset consistent with the original data. By using SMOTE oversampling to synthesize positive samples to balance the ratio of positive to negative samples and adjusting the prediction threshold, class imbalance can be addressed. The conversion potential prediction model is trained iteratively. Initial hyperparameters are set. After each preset round of iteration, the core evaluation index is calculated on the validation set. Training is terminated when the core index on the validation set fails to improve for a consecutive preset round.

5. The method for evaluating the potential for patent transformation into technical standards based on feature interaction according to claim 4, characterized in that, The SMOTE oversampling process includes: Synthetic samples are generated from positive samples based on the K-nearest neighbor algorithm, averaging the ratio of positive to negative samples. When the model output probability is greater than or equal to the prediction threshold, it is determined to be a potential patent.

6. The method for evaluating the potential for patent transformation into technical standards based on feature interaction according to claim 1, characterized in that, The preprocessing of features based on feature type includes the following steps: For numerical features, scale normalization is performed; for text features, domain dictionary filtering is performed; and for categorical features, encoding conversion is performed.

7. The method for evaluating the potential for patent to be transformed into a technical standard based on feature interaction according to claim 1, characterized in that, Features that significantly impact prediction results are selected using XGBoost feature importance gain values.

8. The method for evaluating the potential for patent to be transformed into a technical standard based on feature interaction according to claim 1, characterized in that, The original data includes patent technology text, authorization information, cited data, and application scenario descriptions.

9. The method for evaluating the potential for patent transformation into technical standards based on feature interaction according to claim 1, characterized in that, The technical standard data includes a standard technical glossary and information on standard gap areas.

10. A patent-to-technology-standard conversion potential assessment system based on feature interaction, characterized in that, The system for implementing the patent-to-technical-standard conversion potential assessment method as described in any one of claims 1-9 includes: The salient feature filtering module is used to obtain the original data of the patent to be evaluated and the corresponding technical standard data, and to filter features that have a significant impact on the prediction results from the dimensions of technology relevance, patent maturity, standard demand, technology influence and application scenario. The feature interaction term modeling module is used to construct feature interaction terms that capture the combined effect of multiple features on the patent transformation potential and add them to the preprocessed features; The conversion potential prediction model building module is used to build a conversion potential prediction model based on XGBoost. Through data partitioning based on stratified sampling, handling of class imbalance, and iterative training, a converged conversion potential prediction model is obtained. The patent-to-technology-standard conversion potential assessment module is used to input the features with interactive features into a convergent conversion potential prediction model, obtain conversion potential assessment labels, and realize the patent-to-technology-standard conversion potential assessment of the patent to be assessed.

Images