Patent technology direction matching method and system based on multi-classification and regression combination

Abstract

The invention relates to a patent technology direction matching method and system based on multi-classification and regression combination. The method comprises the following steps: constructing a core feature system of patent technology direction matching; performing feature preprocessing and feature interaction engineering to obtain preprocessed features and feature interaction items; an XGBoost multi-classification and regression combined model is constructed, model hyper-parameters are optimized through grid search and multi-fold cross validation, and abnormal samples in the training process are processed by adopting a weighted training strategy, newly added abnormal mark features and direction cross features; and obtaining a target patent subjected to preorder transformation potential prediction and standard type recommendation screening and a corresponding standard type, inputting the trained XGBoost multi-classification and regression combination model, and outputting probability distribution of all potential technical directions under the standard type and a matching degree of the target patent and each technical direction standard. And determining a recommendation technical direction of the target patent based on a preset threshold value.

Description

Technical Field

[0001] This invention relates to the field of patent data analysis technology, and in particular to a patent technology direction matching method and system based on a combination of multi-classification and regression. Background Technology

[0002] In the coordinated development of technological innovation and industry standardization, patents, as the core carriers of technological innovation achievements, are crucial for the standardized and large-scale implementation of technologies in technology-intensive industries such as power, through their synergistic transformation with technical standards. While current patent standardization decision-making has incorporated machine learning technology, forming a preliminary system encompassing two-category patent transformation potential prediction and multi-category standard type recommendation, existing technical solutions still have significant shortcomings in the critical link of matching the technology direction of patent standardization. These shortcomings make it difficult to meet the core needs of the power industry for precise technology positioning, differentiation quantification, and traceability. Specific problems are as follows: (1) After completing the assessment of conversion potential and the recommendation of standard type, only the standard level of patent adaptation can be determined, but the subdivided technical direction under the standard type cannot be further clarified. The technical direction is vague and lacks precise positioning.

[0003] (2) Existing technology direction judgment models based on support vector machines and basic random forests are mostly black box models. They cannot know the probability distribution of each technology direction, nor can they trace the reasons for the formation of the matching results, and their interpretability is insufficient.

[0004] In summary, existing technologies suffer from problems such as ambiguous attribution, poor interpretability, and insufficient algorithm-industry compatibility in the patent technology direction matching process. There is a lack of a patent technology direction matching method and system to solve or partially solve the aforementioned problems. Summary of the Invention

[0005] The purpose of this invention is to overcome the defects of the existing technology by providing a patent technology direction matching method and system based on a combination of multi-classification and regression, so as to solve or partially solve the problems of ambiguity in patent technology direction attribution, lack of quantification of matching degree, poor interpretability, and insufficient industry adaptability.

[0006] The objective of this invention can be achieved through the following technical solutions: One aspect of the present invention provides a patent technology direction matching method based on a combination of multi-classification and regression, comprising: Construct a core feature system for matching patent technology directions; The core feature system is subjected to feature preprocessing and feature interaction engineering to obtain preprocessed features and feature interaction items; An XGBoost multi-class and regression combined model was constructed. The XGBoost multi-class and regression combined model was trained based on a pre-acquired patent-standard type-technology direction matching training set, preprocessed features and feature interaction terms. The model hyperparameters were optimized by grid search and multi-fold cross-validation. A weighted training strategy, new abnormal label features and direction cross features were added to the abnormal samples during the training process. Obtain the target patent and its corresponding standard type after prior conversion potential prediction and standard type recommendation screening. Input the trained XGBoost multi-classification and regression combined model, output the probability distribution of all potential technical directions under the standard type and the matching degree between the target patent and the standard of each technical direction, and determine the recommended technical direction of the target patent based on the preset threshold.

[0007] As a preferred technical solution, the core feature system includes: technical word vector distance in the semantic similarity dimension, quantitative indicator deviation rate in the parameter fit dimension, LDA topic overlap in the topic consistency dimension, functional goal matching rate in the efficacy achievement dimension, IPC classification intersection ratio in the technical field overlap dimension, professional terminology mapping rate in the terminology system compatibility dimension, and technology life cycle matching in the development trend fit dimension.

[0008] As a preferred technical solution, the LDA topic overlap is the intersection-union ratio of the patent and standard LDA topic distributions.

[0009] As a preferred technical solution, the functional target matching rate is the number of standard functional targets that the patent efficacy meets / the total number of standard functional targets.

[0010] As a preferred technical solution, the feature preprocessing includes the following steps: The deviation rate of the quantitative index for numerical features is processed by piecewise normalization. Specifically, when the deviation rate is ≤0, it is mapped to the interval [0.8,1]; when 0 < deviation rate ≤10%, it is mapped to the interval [0.5,0.8); and when the deviation rate >10%, it is mapped to the interval [0,0.5]. The distance between technical word vectors of text features is semantically expanded and optimized using a technical vocabulary list and thesaurus for different technical directions in the power industry. The Box-Cox transformation is used to eliminate the distribution skewness of the functional target matching rate and the intersection ratio of IPC classification for proportional features.

[0011] As a preferred technical solution, the feature interaction items include: Technical term vector distance of a technical direction × deviation rate of the quantitative indicator of that technical direction; LDA topic overlap of technical directions × functional goal matching rate of that technical direction; Technical term mapping rate × distance between technical word vectors in that technical field.

[0012] As a preferred technical solution, the hyperparameters include learning rate, maximum tree depth, sample sampling ratio, feature sampling ratio, minimum split loss, and negative / positive sample ratio.

[0013] As a preferred technical solution, the process of obtaining the training set includes: A stratified sampling strategy is adopted to divide the pre-obtained patent-standard type-technology direction matching sample set into a training set, a validation set, and a test set.

[0014] As a preferred technical solution, during the partitioning process, the technical directions and matching degree intervals under the standard type are layered to ensure that the technical direction distribution and matching degree distribution of each subset are consistent.

[0015] Another aspect of the present invention provides a patent technology direction matching system based on a combination of multi-classification and regression, for implementing the aforementioned patent technology direction matching method based on a combination of multi-classification and regression, the system comprising: The feature preprocessing and interaction module is used to construct a core feature system for matching patent technology directions, perform feature preprocessing and feature interaction engineering, and obtain preprocessed features and feature interaction items. The model building module is used to build an XGBoost multi-class and regression combined model. It trains the XGBoost multi-class and regression combined model based on a pre-acquired patent-standard type-technology direction matching training set, preprocessed features and feature interaction terms, and optimizes the model hyperparameters through grid search and multi-fold cross-validation. It also uses a weighted training strategy, adds new anomaly label features and direction cross features to handle abnormal samples during the training process. The target patent recommendation technology direction recommendation module is used to obtain the target patent and its corresponding standard type after the previous transformation potential prediction and standard type recommendation screening. The module inputs the trained XGBoost multi-classification and regression combined model, outputs the probability distribution of all potential technology directions under the standard type and the matching degree between the target patent and the standard of each technology direction, and determines the recommended technology direction of the target patent based on the preset threshold.

[0016] Compared with the prior art, the present invention has at least one of the following beneficial effects: (1) Achieving traceable patent technology direction matching with definite direction attribution: This invention constructs a system covering core features such as semantic similarity and parameter fit, comprehensively captures the association between patents and technology directions, fully characterizes multi-dimensional synergistic influence by designing feature interaction terms, constructs an XGBoost multi-classification and regression combined model, outputs the probability distribution and matching degree of each technology direction, and calculates the specific contribution value of each feature of a single sample to the result, so as to achieve direction attribution that can be determined, probability distribution that can be referenced, matching degree that can be quantified, and the cause of difference that can be traced. For example, the matching degree and probability of a certain patent in the direction of energy storage safety can be clearly identified by the contribution value of parameter deviation rate and the contribution value of technical word similarity, and the matching result corresponds to the actual technical difference.

[0017] (2) High accuracy in identifying niche technology directions: In view of the problem that the sample size of niche technology directions is small, the model bias leads to a high misjudgment rate, and it cannot adapt to the subdivided characteristics of power industry technology directions, this invention adopts a weighted training strategy for abnormal samples in the training process, adds abnormal label features and direction cross features to reduce the prediction error of abnormal samples and the misjudgment rate of niche technology directions, and improves the adaptability of the model to the subdivided technology directions of the industry. Attached Figure Description

[0018] Figure 1 The flowchart below shows the patent technology direction matching method based on a combination of multi-classification and regression in the embodiment. Figure 2 This is a schematic diagram of the patent technology direction matching system based on a combination of multi-classification and regression in the embodiment. Detailed Implementation

[0019] 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.

[0020] Example 1 To address the problems of the aforementioned existing technologies, this embodiment provides a patent technology direction matching method based on a combination of multi-classification and regression. The method follows a progressive logic of "assessment foundation - algorithm reserves - construction path - specific model," deepening layer by layer to form a complete technical system. Customized machine learning models are constructed for the three core tasks of conversion potential prediction, standard type recommendation, and technology direction matching degree calculation. Conversion potential prediction and standard type recommendation utilize existing models.

[0021] This method avoids target conflicts and accuracy dilution caused by sharing features across multiple tasks by accurately matching task characteristics with model capabilities. The calculation of technical direction matching requires refined quantification of the probability distribution and semantic similarity of each technical direction. Each model achieves maximum accuracy for a single task through dedicated feature engineering, algorithm selection, and parameter optimization. Finally, through a collaborative mechanism of "feature sharing - process concatenation - result verification," a technical system is formed that enables independent modeling without interference and highly reliable collaborative output, providing more accurate and interpretable support for patent standardization decisions.

[0022] The technology direction matching model constructed in this embodiment is the "technology locator" of the three-model system. Its core positioning is to output the probability distribution of all potential technology directions under the standard type and the matching degree between the patent and the standard of each direction for the patent and the corresponding standard type that have been screened and recommended by the first two levels of models. By setting a preset threshold, the technology direction with the highest probability is used as the recommended direction, realizing the dual-dimensional output of "probability-matching degree". It focuses on solving the business pain points of "ambiguity of technology direction attribution and unclear matching degree".

[0023] In the technology assessment of the power industry, merely judging a patent as "potential" and "suitable for industry standards" without clarifying its technological direction under those standards and its specific technological differences from those standards (such as parameter deviations and semantic discrepancies) may lead to a lack of precise improvement directions for subsequent transformation efforts. Therefore, the model design takes "dual-dimensional precision interpretation of probability and matching degree" as its core objective, ensuring that the output results can not only clearly identify the technological direction but also reflect the overall matching level with the standards of each direction and trace specific technological differences. Ultimately, it achieves the business value of "determinable direction attribution, referable probability distribution, quantifiable matching degree, and traceable reasons for differences."

[0024] This method calculates the matching degree of technical directions, primarily targeting all potential technical directions under the standard recommendation type. It outputs the probability distribution of a patent belonging to each technical direction, quantifies the matching degree between the patent and the standard in each direction, and uses a preset threshold to select the technical direction with the highest probability as the recommended direction. This achieves a two-dimensional output of "probability-matching degree," clearly defining the patent's technical direction affiliation and detailed matching degree. Calculating the matching degree of technical directions requires in-depth analysis of the technical content of the technical standards and patents, extracting key information such as technical keywords and features from different technical directions, converting them into vector representations, and then using similarity calculation algorithms, such as cosine similarity and Euclidean distance, combined with multi-classification algorithms to calculate the probability distribution and matching degree. For example, for a technical standard about smart grids and a patent about smart grid monitoring, in potential technical directions such as "smart dispatching" and "smart monitoring," by extracting their technical keywords, calculating the cosine similarity between vectors and the probability of their respective directions, their matching degree and affiliation in each technical direction can be evaluated. During the calculation process, factors such as technological development trends and the relevance of application fields can also be considered to further improve the accuracy of the matching degree calculation.

[0025] This task falls under the category of hybrid tasks, aiming to output the probability distribution of each technical direction and the continuous matching degree value in the 0-1 interval, thereby quantifying the strength of the technical association and the probability of attribution between patents and standards in each direction. Its core objective is a precise interpretation using both probability and matching degree. Specifically, the probability distribution clarifies the technical direction attribution of a patent under the recommended standard type, while the matching degree value clearly and accurately reflects the degree of matching between the patent and the standards in each direction in terms of technical details. In particular, different probability distributions and quantification values ​​correspond to specific differences in technical details. For example, "a probability of 0.8 and a matching degree of 0.85 for a certain technical direction" might mean that the patent is highly likely to belong to that technical direction and that the technical terms are perfectly matched, while "a probability of 0.3 and a matching degree of 0.65 for a certain technical direction" might mean that the patent is less likely to belong to that direction and that there are some discrepancies in the technical terms.

[0026] See Figure 1 The method includes the following steps: Step S1: Construct a core feature system for matching patent technology directions.

[0027] This step revolves around the core judgment logic of "consistency of technical direction and probability of attribution," accurately capturing the correlation strength of various technical directions under patent and standard types at the semantic, parameter, and topic levels, providing multi-dimensional feature support for matching quantification and probability distribution calculation. The system designs seven core features, which are selected through the feature importance of the XGBoost multi-classification and regression combined model. Its working logic is to leverage the quantitative evaluation of the model's contribution to feature prediction in mixed tasks, retaining the features most closely related to matching degree and probability distribution, ensuring that the probability changes output by the model are highly correlated with direction attribution, and the numerical changes in matching degree are highly correlated with differences in technical details.

[0028] Among them, the distance between the vectors of technical words in the semantic similarity dimension is calculated by the cosine distance between the vectors of patent and standard secondary technical words. The smaller the distance, the higher the matching degree. For example, the vector distance between "dynamic simulation modeling" and "real-time simulation modeling" is close, which directly improves the matching degree.

[0029] The deviation rate, a quantitative indicator of parameter fit, is calculated by the relative difference between the patented technical parameters and the standard requirements. For example, when the patent efficiency is 92% and the standard requirement is 90%, the deviation rate is +2%. The smaller the deviation rate, the higher the matching degree, which is the core characteristic for measuring the consistency of hard indicators.

[0030] The topic overlap of LDA in the topic consistency dimension is calculated by the intersection and union ratio of patents and standards in the distribution of potential topics, reflecting the consistency between the two in the macro technical direction. The higher the overlap, the higher the matching score.

[0031] The functional goal matching rate of efficacy achievement is quantified by the proportion of the fit between the specific efficacy of the patent (such as "response time ≤ 40ms") and the standard core functional goal (such as "response time ≤ 50ms"). The higher the proportion, the more points are added to the matching degree.

[0032] The overlap ratio of IPC classifications in the technology field dimension is reflected by the number of overlaps between the IPC classifications of patents and standards / the total number of classifications. The higher the ratio, the more solid the matching foundation.

[0033] The terminology system compatibility dimension, specifically the terminology mapping rate, measures the interoperability of terminology systems by the number of mappings between patent-specific technical terms (such as "layered energy coordination") and the standard terminology system, divided by the total number of patent terms. A low mapping rate may lead to a decrease in matching degree.

[0034] The technology lifecycle matching dimension of development trend alignment reflects the synchronicity of technology and standard update rhythms by measuring the time difference between the patent application year and the standard revision cycle (e.g., patent application is 2 years before the standard revision). The smaller the time difference, the higher the matching degree.

[0035] See Table 1 for key features and their business significance.

[0036] Table 1 Key Features and Business Significance Step S2 involves performing feature preprocessing and feature interaction engineering on the core feature system to obtain preprocessed features and feature interaction items.

[0037] (1) Pretreatment As a key step in improving the accuracy of matching metrics and probability distribution, the core objective of this step is to strengthen the correlation between features and the degree of matching with technology and the orientation of the model through targeted processing. This ensures that the probability distribution output by the model accurately reflects the orientation of the model and that changes in the matching degree accurately reflect the differences in technical details, thus providing clear feature evidence for the interpretation of the results.

[0038] For numerical features, piecewise normalization is used to process the deviation rate of quantitative indicators in each direction. When the deviation rate is ≤0, meaning the patent parameter is better than the standard for that direction, it is mapped to the interval [0.8, 1]; when 0 < deviation rate ≤10%, it is considered slightly out of standard and mapped to [0.5, 0.8); when the deviation rate >10%, it is considered significantly out of standard and mapped to [0, 0.5). This processing method retains the business logic of parameter superiority and inferiority while amplifying the differences in key intervals, making the model more sensitive to the core differences in whether the parameters meet the standard, thereby more accurately affecting the matching degree and probability of each direction.

[0039] In text feature processing, the focus is on optimizing the distance between word vectors for each technical direction. Using a technical lexicon for each technical direction in the power industry (including a synonym library, such as "energy storage" and "energy storage" in the "energy storage technology" direction), semantic expansion is performed. Patent and standard technical terms for each direction are mapped to a unified lexicon before distance calculation, avoiding excessively large distances due to differences in terminology. For example, the distance between "battery balancing" and "cell balancing" in the "battery management" direction is reduced after mapping. This improves the accuracy of semantic similarity calculation and ensures the reliability of probability and matching degree calculations for each direction.

[0040] For proportional features, such as "matching rate of functional objectives in each direction" and "intersection ratio of IPC classification in each direction", Box-Cox transformation is used to eliminate distribution skewness, making the correlation between feature values ​​and matching degree and probability distribution more significant. For example, after transforming the original matching rate in a certain direction, the value range is more uniform, and the model can more accurately capture the technical differences corresponding to the improvement in matching degree and probability changes.

[0041] (2) Feature Interaction Engineering As a crucial step in matching multi-dimensional synergistic effects and directional attribution in deep characterization technology, its core objective is to capture the joint influence of "semantics-parameters-topics" in various technical directions by constructing feature interaction terms. This allows the matching degree value to reflect both single-dimensional differences and multi-dimensional synergistic effects, while also enabling the probability distribution to more accurately reflect directional attribution and enhancing the interpretability of the results.

[0042] Specifically, three types of feature interaction terms were constructed: First, "distance between technical word vectors in a certain direction × deviation rate of quantitative indicators in that direction." This interaction term highlights the contradictory characteristic of "semantic matching in that direction but parameters exceeding the standard." For example, a distance of 0.1 (high semantic matching) × a deviation rate of 15% (significantly exceeding the standard) = 0.015. Although such patents match the terminology in that direction, the parameters do not meet the standard, and the overall matching degree in that direction needs to be lowered and the probability reduced to avoid misjudgment of direction due to a single semantic feature. Second, "overlapping degree of LDA topics in a certain direction × matching rate of functional goals in that direction." This strengthens the synergistic signal of "consistent macro-themes and matching functional goals in that direction." For example, an overlap of 0.9 × a matching rate of 0.8 = 0.72. Such patents are highly matched in both the topic direction and specific functions in that direction, and the matching degree in that direction should be significantly improved and the probability increased. Third, "mapping rate of professional terms in a certain direction × distance between technical word vectors in that direction." To correct the "semantic distance deviation caused by incompatibility of terminology in this direction", for example, the mapping rate 0.5 (low compatibility) × distance 0.2 (medium semantic matching) = 0.1, indicating that even if the technical terms are similar on the surface, if the terminology system of this direction is incompatible, the actual matching degree of this direction needs to be discounted and the probability reduced.

[0043] Step S3: Construct an XGBoost multi-classification and regression combined model. Train the XGBoost multi-classification and regression combined model based on the pre-acquired patent-standard type-technology direction matching training set, preprocessed features and feature interaction terms. Optimize the model hyperparameters through grid search and multi-fold cross-validation. For abnormal samples during the training process, adopt a weighted training strategy, add new abnormal label features and direction cross features.

[0044] This step uses a GBoost multi-class classification and regression combined model, focusing on Top-1 accuracy, mean probability bias, and R-value. 2 All meet business requirements and can output the specific contribution value of each feature to the probability and matching degree of each direction, providing a quantitative basis for technical difference analysis and direction attribution judgment.

[0045] The model training process includes the following steps: (1) Data partitioning A stratified sampling strategy was employed to divide the matching sample set for "patent-standard type-technology direction." The data was stratified according to the technology direction and matching degree range under each standard type: 0-0.2, 0.2-0.4, 0.4-0.6, 0.6-0.8, and 0.8-1.0. This ensured consistency in the distribution of technology directions and matching degrees across subsets. Specifically, the proportions were: 70% for the training set (used to learn the matching patterns and probability of attribution for each technology direction), 20% for hyperparameter optimization and early stopping detection, and 10% for the test set (used to evaluate the model's generalization ability and interpretability). This partitioning method avoids model bias caused by uneven distribution of technology directions or matching degrees, ensuring accurate quantification and attribution determination for each technology direction and for samples with high, medium, and low matching degrees.

[0046] (2) Training process Model training follows a closed-loop logic of "initial parameter setting, iterative optimization, and early stopping convergence," focusing on ensuring the accuracy of matching metrics and the accuracy and stability of probability distributions. Initial hyperparameters are based on default values ​​for XGBoost hybrid tasks, such as learning_rate=0.1, max_depth=7, and subsample=0.8. During iterative optimization, Top-1 accuracy, average probability bias, and R-squared are calculated on the validation set after every 100 iterations. 2 The model parameters and performance changes are recorded along with the MAE (mean absolute error) to form a dynamic curve of "iteration rounds - performance index". The early stopping mechanism is set as follows: when the Top-1 accuracy on the validation set does not improve for 5 consecutive rounds (fluctuation ≤ 0.005), training is terminated, and the current optimal parameters are used as the final model parameters to prevent the problem of "memorizing the matching degree and probability distribution of the training set but poor generalization" caused by overfitting.

[0047] (3) Model hyperparameter optimization With the goal of "maximizing Top-1 accuracy, minimizing average probability bias, and R0", 2 With "maximization" as the core optimization objective, key parameters were adjusted using "grid search + 5-fold cross-validation" to significantly improve the model's ability to capture differences in technical details and the accuracy of direction attribution. The final determined core hyperparameters are as follows: learning_rate controls the iteration step size, ensuring that the model learns the correlation between features and matching degree, probability distribution in detail; `max_depth` limits the tree depth to avoid overfitting to noisy features, such as accidental terminological differences, and to ensure the stability of the results. subsample, the sampling ratio, improves the model's adaptability to different "patent-standard type-technology direction" pairs; colsample_bytree, the feature sampling ratio, focuses on core features, such as the influence of technical word distance in different directions and parameter deviation rate; gamma, the minimum splitting loss, splits only when feature splitting reduces the combined loss (i.e., multi-class loss and regression loss) by ≥0.1, thus strengthening the role of effective features; `scale_pos_weight` (by technology direction) sets weights based on the sample proportion of each technology direction. Directions with fewer samples have higher weights, thus improving the recognition accuracy of niche technology directions.

[0048] See Table 2 for the core hyperparameters.

[0049] Table 2 Core Hyperparameters (4) Handling of abnormal samples For samples with extreme matching scores and abnormal direction attribution (e.g., those actually belonging to direction A but whose features show a stronger fit to direction B, accounting for approximately 3%), a "weighted training" strategy is adopted: abnormal samples are given double the weight, forcing the model to focus on learning the feature patterns of such samples, such as hidden technical term synonym associations or cross-directional parameter mapping relationships. At the same time, the reasons for the anomalies are manually labeled, such as "ambiguous description of standard parameters" and "intersection of technical directions". New "anomaly marker features" are added, where 1 = abnormal and 0 = normal, as well as "direction intersection features", where 1 = intersection exists and 0 = no intersection. These features help the model identify special cases and reduce the prediction error and direction misjudgment rate of extreme value samples.

[0050] (5) Convergence analysis The model convergence process is achieved through iteration rounds - Top-1 accuracy - R. 2 Curve tracing reached a stable state after 500 iterations. The first 200 iterations were a rapid improvement phase, mainly because the model quickly captured the patterns of core features such as technical term distances and parameter deviation rates in various directions.

[0051] Rounds 200-450 constitute the fine-tuning phase. During this phase, the model learns the synergistic influence of semantics and parameters in various directions through feature interaction terms. For example, it learns the magnitude of the reduction in matching degree and the degree of probability decrease in a certain direction for "high semantic matching but parameter overshoot". After 450 rounds, it enters a plateau period, with the Top-1 accuracy and R-value on the validation set reaching a certain level. 2 Fluctuations ≤ 0.005 trigger the early stopping mechanism. The final model shows no overfitting, and its quantification accuracy and direction attribution judgment are stable for samples with different technical directions and matching degrees. Furthermore, the feature contribution of a single sample is highly consistent with the actual technical differences.

[0052] Step S4: Obtain the target patent and its corresponding standard type after the previous conversion potential prediction and standard type recommendation screening. Input the trained XGBoost multi-classification and regression combined model, output the probability distribution of all potential technical directions under the standard type and the matching degree between the target patent and the standard of each technical direction, and determine the recommended technical direction of the target patent based on the preset threshold.

[0053] Example 2 Based on Example 1, this example provides a patent technology direction matching system based on a combination of multi-classification and regression, used to implement the patent technology direction matching method of Example 1. See [link to example]. Figure 2 The system includes: (1) Feature preprocessing and interaction module, used to construct the core feature system for matching patent technology direction, perform feature preprocessing and feature interaction engineering, and obtain preprocessed features and feature interaction items; (2) Model building module, used to build XGBoost multi-class and regression combined model. Based on the pre-acquired patent-standard type-technology direction matching training set, pre-processed features and feature interaction terms, the XGBoost multi-class and regression combined model is trained, and the model hyperparameters are optimized by grid search and multi-fold cross-validation. Abnormal samples in the training process are processed by weighted training strategy, new abnormal label features and direction cross features. (3) The target patent recommendation technology direction recommendation module is used to obtain the target patent and corresponding standard type after the previous transformation potential prediction and standard type recommendation screening, input the trained XGBoost multi-classification and regression combined model, output the probability distribution of all potential technology directions under the standard type and the matching degree between the target patent and each technology direction standard, and determine the recommended technology direction of the target patent based on the preset threshold.

[0054] 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 patent technology direction matching method based on multi-classification combined with regression, characterized in that, include: Construct a core feature system for matching patent technology directions; The core feature system is subjected to feature preprocessing and feature interaction engineering to obtain preprocessed features and feature interaction items; An XGBoost multi-class and regression combined model was constructed. The XGBoost multi-class and regression combined model was trained based on a pre-acquired patent-standard type-technology direction matching training set, preprocessed features and feature interaction terms. The model hyperparameters were optimized by grid search and multi-fold cross-validation. A weighted training strategy, new abnormal label features and direction cross features were added to the abnormal samples during the training process. Obtain the target patent and its corresponding standard type after prior conversion potential prediction and standard type recommendation screening. Input the trained XGBoost multi-classification and regression combined model, output the probability distribution of all potential technical directions under the standard type and the matching degree between the target patent and the standard of each technical direction, and determine the recommended technical direction of the target patent based on the preset threshold.

2. The patent technology direction matching method based on a combination of multi-classification and regression as described in claim 1, characterized in that, The core feature system includes: technical word vector distance in semantic similarity dimension, quantitative indicator deviation rate in parameter fit dimension, LDA topic overlap in topic consistency dimension, functional goal matching rate in efficacy achievement dimension, IPC classification intersection ratio in technical field overlap dimension, professional terminology mapping rate in terminology system compatibility dimension, and technology life cycle matching in development trend fit dimension.

3. The patent technology direction matching method based on a combination of multi-classification and regression as described in claim 2, characterized in that, The LDA topic overlap is the intersection-union ratio of the patent and standard LDA topic distributions.

4. The patent technology direction matching method based on a combination of multi-classification and regression as described in claim 2, characterized in that, The functional target matching rate is the number of patent efficacy functions that meet the standard functional targets / the total number of standard functional targets.

5. The patent technology direction matching method based on a combination of multi-classification and regression as described in claim 1, characterized in that, The feature preprocessing includes the following steps: The deviation rate of the quantitative index for numerical features is processed by piecewise normalization. Specifically, when the deviation rate is ≤0, it is mapped to the interval [0.8,1]; when 0 < deviation rate ≤10%, it is mapped to the interval [0.5,0.8); and when the deviation rate >10%, it is mapped to the interval [0,0.5]. The distance between technical word vectors of text features is semantically expanded and optimized using a technical vocabulary list and thesaurus for different technical directions in the power industry. The Box-Cox transformation is used to eliminate the distribution skewness of the functional target matching rate and the intersection ratio of IPC classification for proportional features.

6. The patent technology direction matching method based on a combination of multi-classification and regression according to claim 1, characterized in that, The feature interaction items include: Technical term vector distance of a technical direction × deviation rate of the quantitative indicator of that technical direction; LDA topic overlap of technical directions × functional goal matching rate of that technical direction; Technical term mapping rate × distance between technical word vectors in that technical field.

7. The patent technology direction matching method based on a combination of multi-classification and regression according to claim 1, characterized in that, The hyperparameters include learning rate, maximum tree depth, sample sampling ratio, feature sampling ratio, minimum split loss, and negative / positive sample ratio.

8. The patent technology direction matching method based on a combination of multi-classification and regression according to claim 1, characterized in that, The process of obtaining the training set includes: A stratified sampling strategy is adopted to divide the pre-obtained patent-standard type-technology direction matching sample set into a training set, a validation set, and a test set.

9. A patent technology direction matching method based on a combination of multi-classification and regression as described in claim 8, characterized in that, During the partitioning process, the technology direction and matching degree intervals under the standard type are divided into layers to ensure that the distribution of technology direction and matching degree of each subset is consistent.

10. A patent technology direction matching system based on a combination of multi-classification and regression, characterized in that, The system for implementing the patent technology direction matching method based on a combination of multi-classification and regression as described in any one of claims 1-9 includes: The feature preprocessing and interaction module is used to construct a core feature system for matching patent technology directions, perform feature preprocessing and feature interaction engineering, and obtain preprocessed features and feature interaction items. The model building module is used to build an XGBoost multi-class and regression combined model. It trains the XGBoost multi-class and regression combined model based on a pre-acquired patent-standard type-technology direction matching training set, preprocessed features and feature interaction terms, and optimizes the model hyperparameters through grid search and multi-fold cross-validation. It also uses a weighted training strategy, adds new anomaly label features and direction cross features to handle abnormal samples during the training process. The target patent recommendation technology direction recommendation module is used to obtain the target patent and its corresponding standard type after the previous transformation potential prediction and standard type recommendation screening. The module inputs the trained XGBoost multi-classification and regression combined model, outputs the probability distribution of all potential technology directions under the standard type and the matching degree between the target patent and the standard of each technology direction, and determines the recommended technology direction of the target patent based on the preset threshold.

Images