A feature generation method based on symbol language numerical perspective fusion

Abstract

The application belongs to the technical field of feature engineering, and discloses a feature generation method based on symbol language numerical value perspective fusion; the method comprises the following steps: collecting observation data of a target field, and constructing a multi-perspective feature generator group; generating expression trees corresponding to each candidate feature through the multi-perspective feature generator group, and predicting the effectiveness index of each candidate feature and planning the verification execution sequence of each candidate feature; sequentially executing feature verification on each candidate feature, real-time evaluating the feature quality index of each candidate feature, and judging whether the feature generation is completed; if the feature generation is not completed, dynamically adjusting the exploration parameters of each feature generator, and regenerating the expression trees corresponding to each candidate feature; if the feature generation is completed, screening key features from each candidate feature; the application can realize the transformation from single perspective exploration to multi-perspective collaborative optimization, and significantly improve the feature generation efficiency, prediction accuracy and generalization ability.

Description

Technical Field

[0001] This invention relates to the field of feature engineering technology, and more specifically, to a feature generation method based on symbolic language numerical perspective fusion. Background Technology

[0002] In practical applications across various modern professional fields, the core issue for accurately predicting key dependent variables lies in constructing advanced features with strong explanatory and predictive capabilities from a large number of independent variables. For example, in the financial sector, factor analysis of the secondary market requires constructing factors with significant information content based on financial market data to accurately predict key market indicators. In the medical field, it is necessary to explore the unique judgment logic of specific diseases based on physical indicators to achieve early and accurate diagnosis. In the manufacturing sector, it is necessary to integrate a large amount of information on raw materials, process parameters, environmental factors, etc., to determine the final output indicators. It is evident that different industries urgently need an efficient method to extract key features from raw observation data to improve the accuracy of dependent variable prediction and the interpretability of features.

[0003] In existing technologies, three main technical approaches are typically used to solve this type of problem: First, research based on disciplinary knowledge, which constructs features through professional research results and manual analysis. Its advantage lies in its rigor and completeness, but it is limited by the discipline, has a narrow perspective, and a high threshold. Second, automated feature engineering, which explores features in a broad feature space using evolutionary computation or reinforcement learning. It can break free from disciplinary dependence, but it is prone to getting stuck in local optima, and the features are complex and have poor interpretability. Third, symbolic regression, which generates features through function modeling and deconstruction. It is convenient to use functions to characterize high-level features, but it is difficult to solve, prone to non-convergence, or difficult to obtain effective solutions. Therefore, the common problem faced by these three technical approaches is that they rely on a single feature exploration perspective and lack cross-perspective information fusion, resulting in a single exploration path, low feature generation efficiency, and difficulty in obtaining key features with high interpretability, good stability, and generalization ability under limited experimental resources.

[0004] In view of this, the present invention proposes a feature generation method based on symbolic language numerical perspective fusion to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned shortcomings of the prior art and achieve the above objectives, the present invention provides the following technical solution: a feature generation method based on symbolic language numerical perspective fusion, comprising:

[0006] Step S1: Collect observation data of the target domain, fine-tune and optimize the preset large language model based on the pre-built domain knowledge base, and set exploration parameters to form a multi-view feature generator group;

[0007] Step S2: Based on the observation data, each feature generator in the multi-view feature generator group generates an expression tree corresponding to each candidate feature, collects the feature index of each expression tree, and calculates the initial contribution value of each candidate feature in turn according to the feature index.

[0008] Step S3: Based on the initial contribution value and the prior distribution of features in the exploration parameters, dynamically predict the effectiveness index of each candidate feature, and calculate the convergence of each candidate feature for different exploration paths in turn. Based on the effectiveness index and convergence, dynamically plan the verification execution order of each candidate feature.

[0009] Step S4: Perform feature verification on each candidate feature in sequence according to the verification execution order, evaluate the feature quality index of each candidate feature in real time, and determine whether feature generation is complete based on the feature quality index;

[0010] Step S5: If feature generation is not completed, dynamically adjust the exploration parameters of each feature generator according to the feature quality index, and return to step S2; if feature generation is completed, select key features from each candidate feature according to the feature quality index.

[0011] Furthermore, the observational data includes multiple sets of validation data, which include independent variable data and dependent variable data; the exploration parameters include exploration tendency, constraint parameters, feature prior distribution, and generation scale.

[0012] Methods for forming multi-view feature generator sets include:

[0013] Based on the domain knowledge base, the model parameters in the large language model are fine-tuned and optimized to obtain the language perspective feature generator; a numerical perspective feature generator is constructed, and the language perspective feature generator and the numerical perspective feature generator are integrated to form a multi-perspective feature generator group.

[0014] Methods for generating expression trees corresponding to each candidate feature include:

[0015] The observation data is input into each feature generator in the multi-view feature generator group; the language perspective feature generator generates features based on the input observation data and the corresponding exploration tendency and generation scale. The expression tree corresponding to each candidate feature; the numerical perspective feature generator generates an expression tree based on the input observation data and the corresponding exploration parameters. The expression tree corresponding to each candidate feature.

[0016] Furthermore, the method for sequentially calculating the initial contribution value of each candidate feature includes:

[0017] Feature metrics include tree depth, number of nodes, number of operations, and number of functions;

[0018] The feature indices of each candidate feature are normalized sequentially to obtain standard indices; different index weights are assigned to different feature indices, and the different standard indices corresponding to the same candidate feature are weighted and summed based on the index weights to obtain the feature complexity of each candidate feature; the feature complexity of each candidate feature is then appropriately adjusted sequentially to obtain the appropriate complexity of each candidate feature.

[0019] Calculate the sum of the number of operations and the number of functions corresponding to the same candidate feature to obtain the number of operations for each candidate feature; calculate the total penalty value for each candidate feature based on the tree depth and the number of operations corresponding to each candidate feature; calculate the type weights corresponding to different feature generators, and calculate the effective complexity of each candidate feature based on the type weights and appropriate complexity; calculate the difference between the effective complexity and the total penalty value for each candidate feature to obtain the initial contribution value of each candidate feature.

[0020] Furthermore, methods for dynamically predicting the effectiveness indicators of each candidate feature include:

[0021] Based on the expression tree of each candidate feature, the prior probability of each candidate feature is obtained from the prior distribution of the features, and the basic potential of each candidate feature is calculated by combining the initial contribution value of each candidate feature; based on the basic potential of each candidate feature and the type weight, the modified potential of each candidate feature is calculated.

[0022] Based on the corrected potential energy and initial contribution value of each candidate feature, calculate the feature quality of each candidate feature; construct the feature vector of each candidate feature, and calculate the feature distance between each pair of candidate features in turn; based on the feature distance between each pair of candidate features and the corresponding feature quality, calculate the feature attraction between each pair of candidate features in turn; select candidate features with feature quality greater than a preset quality threshold as high-quality features, and add the feature attraction between the same candidate feature and each high-quality feature in turn to obtain the total attraction of each candidate feature;

[0023] Among the modified potential energies of each candidate feature, the one with the largest value is taken as the maximum potential energy. The ratio of the modified potential energy to the maximum potential energy of each candidate feature is calculated to obtain the static effectiveness of each candidate feature. The hyperbolic tangent function is used to convert each total gravitational force into a corresponding gravitational gain. Combined with the static effectiveness of each candidate feature, the effectiveness index of each candidate feature is calculated.

[0024] Furthermore, methods for sequentially calculating the convergence of different exploration paths include:

[0025] For all candidate features, they are grouped according to the corresponding feature generator and feature type to obtain multiple path sets, and each path set corresponds one-to-one with an exploration path. The generation time corresponding to each candidate feature is obtained, and the candidate features in each path set are sorted from earliest to latest according to their corresponding generation time to generate a feature sequence corresponding to each path set. Based on the feature sequence of each path set and the effectiveness index of each candidate feature in the feature sequence, the convergence of different exploration paths is calculated in turn.

[0026] Furthermore, methods for dynamically planning the verification execution order of each candidate feature include:

[0027] From the feature sequences of each path set, the first and last features are obtained respectively; based on the effectiveness index of the first and last features of each path set, the index difference of each path set is calculated; the feature interval of each path set is obtained, and combined with the index difference, the changing trend of each exploration path is calculated; the changing trend of each exploration path is standardized in turn to obtain the standard trend of each exploration path.

[0028] Different validation weights are set for convergence and standard trend, and the convergence and standard trend of each exploration path are weighted and summed based on the validation weights to obtain the comprehensive index of each exploration path. The candidate features are sorted from largest to smallest according to the comprehensive index of the corresponding exploration path to generate a preliminary sequence. Among them, candidate features with the same comprehensive index are in parallel positions in the preliminary sequence. The candidate features in parallel positions in the preliminary sequence are sorted from largest to smallest according to the corresponding validity index to generate a comprehensive sequence. The validation execution order of each candidate feature is determined according to the ascending order of the comprehensive sequence.

[0029] Furthermore, methods for real-time evaluation of feature quality metrics for each candidate feature include:

[0030] The observation data is divided into multiple sets of validation data to obtain training and test sets; a benchmark model is constructed and trained based on the validation data in the training set; the independent variable data in each set of validation data in the test set is input into the trained benchmark model to obtain the first prediction data corresponding to each set of validation data; based on the dependent variable data and the first prediction data corresponding to each set of validation data in the test set, the performance index of the benchmark model is calculated and marked as the benchmark index.

[0031] Define the enhancement index calculation process, and perform the enhancement index calculation process for each candidate feature in sequence according to the verification execution order to obtain the enhancement index of each candidate feature; calculate the difference between the baseline index and the enhancement index of each candidate feature in sequence to obtain the feature quality index of each candidate feature.

[0032] Furthermore, methods for determining whether feature generation is complete based on feature quality metrics include:

[0033] Preset indicator thresholds and quantity thresholds, compare the feature quality index of each candidate feature with the indicator thresholds respectively, and mark the candidate features whose feature quality index is greater than the indicator threshold as high-quality features;

[0034] The number of high-quality features is counted and marked as high-quality; the number of high-quality features is compared with a threshold; if the number of high-quality features is greater than the threshold, feature generation is considered complete; if the number of high-quality features is less than or equal to the threshold, feature generation is considered incomplete.

[0035] Furthermore, methods for dynamically adjusting the exploration parameters of each feature generator include:

[0036] The average quality index of each feature generator is obtained by averaging the feature quality indices of the candidate features corresponding to the same feature generator; the average quality index of each path set is obtained by averaging the feature quality indices of the candidate features corresponding to each path set.

[0037] Based on the feature quality index of each candidate feature, the average quality index of each feature generator, the path quality index of each path set, and the number of high-quality features, a heuristic optimization algorithm is used to formulate parameter adjustment strategies for different feature generators; and the exploration parameters of each feature generator are dynamically adjusted according to the parameter adjustment strategies for each feature generator.

[0038] Furthermore, methods for selecting key features from candidate features include:

[0039] Each pair of high-quality features is combined sequentially to obtain multiple feature sets. Based on the candidate feature values ​​of the two high-quality features corresponding to each feature set, the feature correlation of each feature set is calculated sequentially. Each feature correlation is compared with a preset correlation threshold, and feature sets with feature correlation greater than the correlation threshold are marked as high similarity sets. For each high similarity set, the feature quality index of the two high-quality features is compared, and the high-quality feature with the smallest feature quality index is deleted.

[0040] Based on the feature quality index of each high-quality feature, calculate the contribution ratio of each high-quality feature in turn; sort each high-quality feature in descending order of its corresponding contribution ratio to generate a contribution sequence; calculate the cumulative contribution of each high-quality feature in ascending order of the contribution sequence; compare the cumulative contribution of each high-quality feature with a preset contribution threshold, delete the high-quality features whose cumulative contribution is greater than the contribution threshold from the contribution sequence in turn, and use the high-quality features retained in the contribution sequence as key features.

[0041] The technical effects and advantages of the feature generation method based on symbolic language numerical perspective fusion of the present invention are as follows:

[0042] By integrating linguistic and numerical perspectives into feature generation mechanisms, this approach achieves comprehensive utilization of domain knowledge-driven and automated feature exploration. It efficiently extracts candidate features with both interpretive and predictive capabilities from observational data. In the feature evaluation stage, a dynamic potential-gravitational field evaluation mechanism is introduced, considering not only the intrinsic potential of individual candidate features but also capturing the synergistic effects between different candidate features, thus enabling more accurate and comprehensive prediction of the effectiveness indicators of each candidate feature. Through dynamic planning of the stability and changing trends of the exploration path and full utilization of the effectiveness indicators of each candidate feature, intelligent scheduling of the verification execution order is achieved, thereby improving the efficiency of feature verification and the utilization rate of computing resources. Furthermore, it addresses the dynamic adjustment of exploration parameters and related aspects during feature generation. This paper addresses the key feature selection problem by employing heuristic optimization algorithms and feature contribution analysis methods to achieve adaptive optimization of feature generator parameters and efficient identification of important features, further enhancing the adaptability and robustness of the entire feature generation process. It shifts from a single-perspective exploration to multi-perspective collaborative optimization, effectively overcoming the problems of traditional methods such as single exploration paths, susceptibility to local optima, and insufficient feature interpretability. While ensuring appropriate feature complexity and structural simplicity, it significantly improves feature generation efficiency, prediction accuracy, and generalization ability. The final selected key features not only possess strong interpretability and high predictability but also effectively remove redundancy and maintain feature diversity, providing an efficient and reliable feature engineering solution for accurate prediction in various professional fields. Attached Figure Description

[0043] Figure 1 This is a flowchart of a feature generation method based on symbolic language numerical perspective fusion according to Embodiment 1 of the present invention. Detailed Implementation

[0044] 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 embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] Example 1

[0046] Please see Figure 1 As shown in this embodiment, a feature generation method based on symbolic language numerical perspective fusion includes:

[0047] Step S1: Collect observation data of the target domain, fine-tune and optimize the preset large language model based on the pre-built domain knowledge base, set exploration parameters, and form a multi-view feature generator group.

[0048] The target domain refers to the specific professional field where features are to be generated, such as the medical field or the financial field. The observation data includes multiple sets of validation data, which include independent variable data and dependent variable data. The independent variable data refers to the raw data used to generate features, such as trading volume and turnover rate in the financial field, and blood pressure and blood sugar in the medical field. The dependent variable data refers to the target that is to be predicted or explained, used to evaluate the effectiveness of the features, such as the future volatility of the investment portfolio and the future return of stocks in the financial field, and whether one has a specific disease or the probability of postoperative complications in the medical field.

[0049] A language-perspective feature generator is obtained by fine-tuning and optimizing the model parameters in large language models (such as LLaMA and Claude) based on a domain knowledge base. The domain knowledge base is a collection of knowledge specific to the target domain, used to guide feature generation and enhance the domain awareness of the large language model. This domain knowledge base is pre-constructed by individuals skilled in the art through the collection and filtering of relevant professional literature, standards, research reports, and empirical data. Model parameters refer to various weights within the large language model used to represent word meanings, step relationships, and generated outputs, such as word embedding weights, attention weights, and feedforward network weights.

[0050] It should be noted that the method of fine-tuning and optimizing the model parameters in the large language model based on the domain knowledge base is an existing technology, such as full parameter fine-tuning, incremental learning, and efficient parameter fine-tuning. The specific process will not be elaborated on here.

[0051] The exploration parameters include, but are not limited to, exploration tendency, constraint parameters, feature prior distribution and generation scale. These parameters are comprehensively set by those skilled in the art based on factors such as professional knowledge of the target domain, computational resource constraints, and historical feature engineering experience. They are used to guide and constrain the feature generator to efficiently explore valuable features within a reasonable search space.

[0052] The exploration tendency indicates the ability or strength of the feature generator to explore diverse feature structures when generating expression trees corresponding to candidate features. The higher the exploration tendency, the more diverse the generated results; the lower the exploration tendency, the more homogeneous the generated results.

[0053] Constraint parameters represent the boundary conditions or rule restrictions that constrain the expression tree generation process, preventing the feature generator from generating expression trees that do not conform to the logic or numerical rules of the target domain. Constraint parameters include, but are not limited to, expression tree structure constraints, symbolic constraints, and variable constraints. Expression tree structure constraints include, for example, maximum depth (the longest path length from the root node to the deepest leaf node), minimum depth (the shortest path length from the root node to the shallowest leaf node), maximum number of nodes (the upper limit of the total number of nodes in the expression tree), and minimum number of nodes (the lower limit of the total number of nodes in the expression tree). Nodes include operator nodes (such as addition, subtraction, multiplication, and division) and function nodes (such as trigonometric functions). The expression tree is divided into several parts: number, exponential function, logarithmic function, variable nodes (independent variable data), and constant nodes (fixed values). Symbolic constraints restrict the operators or functions used in the expression tree and their combinations, ensuring that the generated feature's computation process is legal and conforms to the target domain logic. For example, trigonometric functions are not allowed, logarithmic nodes cannot be nested under addition nodes, and square roots can appear at most once. Variable constraints restrict the data types of independent variables used in the expression tree and their combinations, ensuring that the variables used in the generated features are reasonable and conform to the target domain requirements. For example, independent variable data 1 and independent variable data 2 cannot be used as denominators simultaneously, and independent variable data 3 and independent variable data 4 cannot be multiplied.

[0054] The prior distribution of features represents the prior probability of the feature generator for different feature types when generating expression trees corresponding to candidate features. Feature types include univariate features (containing only one independent variable data), linear combination features (one or more independent variable data are combined according to a linear relationship), nonlinear function features (one or more independent variable data are generated through nonlinear function transformation), and compound nested features (one or more independent variable data are combined through multi-level operations).

[0055] The generation scale represents the number of expression trees generated by the feature generator at one time, and is used to control feature diversity and computational resource consumption.

[0056] A numerical perspective feature generator is constructed, which is an automated feature engineering generator that integrates the function deconstruction symbolic regression method. The language perspective feature generator and the numerical perspective feature generator are integrated to form a multi-perspective feature generator group. It should be noted that the function deconstruction symbolic regression method and the automated feature engineering generator are existing technologies, and their details will not be elaborated here.

[0057] Step S2: Based on the observation data, each feature generator in the multi-view feature generator group generates an expression tree corresponding to each candidate feature, collects the feature index of each expression tree, and calculates the initial contribution value of each candidate feature in turn according to the feature index.

[0058] Methods for generating expression trees corresponding to each candidate feature include:

[0059] The observation data is input into each feature generator in the multi-perspective feature generator group (language perspective feature generator and numerical perspective feature generator); the language perspective feature generator generates features based on the input observation data and the corresponding exploration tendency and generation scale. The expression tree corresponding to each candidate feature; the numerical perspective feature generator generates an expression tree based on the input observation data and the corresponding exploration parameters. The expression tree corresponding to each candidate feature; where... The generation scale corresponding to the language perspective feature generator. This represents the generation scale corresponding to the numerical perspective feature generator.

[0060] Methods for calculating the initial contribution value of each candidate feature sequentially include:

[0061] The feature metrics include tree depth, number of nodes, number of operations, and number of functions; where tree depth represents the length of the longest path from the root node to the deepest leaf node in the expression tree; number of nodes represents the total number of nodes in the expression tree; number of operations represents the number of operator types in the expression tree; and number of functions represents the number of function types in the expression tree.

[0062] The feature indices of each candidate feature are normalized sequentially to obtain standard indices; different index weights are set for different feature indices, and the different standard indices corresponding to the same candidate feature are weighted and summed based on the index weights to obtain the feature complexity of each candidate feature; the feature complexity corresponding to each candidate feature is appropriately adjusted sequentially to obtain the appropriate complexity of each candidate feature; wherein, the index weights are preset by those skilled in the art according to the actual situation.

[0063] The process involves several steps: First, calculate the sum of the number of operations and the number of functions corresponding to the same candidate feature to obtain the number of operations for each candidate feature. Second, calculate the total penalty value for each candidate feature based on its tree depth and the number of operations. Third, calculate the sum of the generation scale of the language-perspective feature generator and the generation scale of the numerical-perspective feature generator to obtain the total scale value. Fourth, calculate the ratio between the generation scale of the language-perspective feature generator and the total scale value to obtain the type weight of the language-perspective feature generator. Fifth, calculate the ratio between the generation scale of the numerical-perspective feature generator and the total scale value to obtain the type weight of the numerical-perspective feature generator. Sixth, calculate the product of the appropriate complexity and the type weight for each candidate feature to obtain the effective complexity of each candidate feature. Finally, calculate the difference between the effective complexity and the total penalty value for each candidate feature to obtain the initial contribution value of each candidate feature.

[0064] The method for obtaining the appropriate complexity corresponding to each candidate feature is as follows: preset the optimal complexity, calculate the difference between the feature complexity corresponding to each candidate feature and the optimal complexity in turn, take the absolute value, and then divide it by the optimal complexity to obtain the complexity deviation of each candidate feature; calculate the difference between the optimal complexity deviation and the complexity deviation of each candidate feature to obtain the appropriate complexity of each candidate feature; wherein, the optimal complexity is preset by those skilled in the art based on the feature complexity calculated in the historical feature generation process;

[0065] The method for calculating the total penalty value corresponding to each candidate feature is as follows: A depth penalty threshold and an operation penalty threshold are preset. The depth penalty threshold is preset by a person skilled in the art based on the interpretability of the expression tree and computational resource limitations. The operation penalty threshold is preset by a person skilled in the art based on the reasonable range of operators and functions in the expression tree. The depth penalty value for each candidate feature is obtained by subtracting the depth penalty threshold from the tree layer depth corresponding to each candidate feature and then dividing by the depth penalty threshold. The operation penalty value for each candidate feature is obtained by subtracting the operation penalty threshold from the number of operations corresponding to each candidate feature and then dividing by the operation penalty threshold. The total penalty value for each candidate feature is obtained by summing the depth penalty value and the operation penalty threshold. It should be noted that if the tree layer depth is less than or equal to the depth penalty threshold, the depth penalty value is 0; if the number of operations is less than or equal to the operation penalty threshold, the operation penalty value is 0.

[0066] It should be understood that the initial contribution value comprehensively considers the appropriateness of feature complexity, structural simplicity, and generator reliability. By quickly evaluating the structural features of the expression tree, it provides a low-cost pre-screening score for a large number of candidate features, thereby avoiding the lack of expressive power of overly simple features and preventing overly complex features from being difficult to interpret. This effectively identifies high-potential features and improves the efficiency of the entire feature generation process.

[0067] Step S3: Based on the initial contribution value and the prior distribution of features in the exploration parameters, dynamically predict the effectiveness index of each candidate feature, and calculate the convergence of each candidate feature for different exploration paths in turn. Based on the effectiveness index and convergence, dynamically plan the verification execution order of each candidate feature.

[0068] Methods for dynamically predicting the effectiveness index of each candidate feature include:

[0069] Based on the expression tree corresponding to each candidate feature, determine the feature type of each candidate feature; based on the feature type of each candidate feature, obtain the prior probability of each candidate feature from the feature prior distribution; based on the initial contribution value and prior probability of each candidate feature, calculate the basic potential of each candidate feature; calculate the product of the basic potential of each candidate feature and the type weight to obtain the modified potential of each candidate feature.

[0070] The product of the corrected potential energy and the initial contribution value of each candidate feature is calculated to obtain the feature quality of each candidate feature. Different numerical labels are assigned to different feature types and marked as type labels. Based on the standard index, initial contribution value, prior probability, and type label of each candidate feature, a feature vector of each candidate feature is constructed. Based on the feature vector of each candidate feature, the Euclidean distance between each pair of candidate features is calculated sequentially and marked as feature distance. Based on the feature distance between each pair of candidate features and the corresponding feature quality, the feature attraction between each pair of candidate features is calculated sequentially. The feature quality of each candidate feature is compared with a preset quality threshold. Candidate features with feature quality greater than the quality threshold are considered high-quality features, and candidate features with feature quality less than or equal to the quality threshold are considered low-quality features. The feature attraction between the same candidate feature and each high-quality feature is added sequentially to obtain the total attraction of each candidate feature. The quality threshold is preset by those skilled in the art according to the actual situation.

[0071] The corrected potential energy of each candidate feature is compared, and the corrected potential energy with the largest value is taken as the maximum potential energy. The ratio of the corrected potential energy to the maximum potential energy of each candidate feature is calculated to obtain the static effectiveness of each candidate feature. The hyperbolic tangent function is used to convert each total gravitational force into a corresponding gravitational gain. The gravitational gain of each candidate feature is incremented by one and then multiplied by the static effectiveness of the corresponding candidate feature to obtain the effectiveness index of each candidate feature.

[0072] The method for calculating the basic potential energy is as follows: A preset energy decay coefficient is used; the ratio of the energy decay coefficient to the prior probability is calculated and its negative value is taken to obtain the potential energy factor; using the natural constant as the base and the potential energy factor as the exponent, exponential operations are performed to obtain the decay factor; the product of the decay factor and the initial contribution value is calculated to obtain the basic potential energy; wherein, the energy decay coefficient is preset by those skilled in the art according to the actual situation, and in this embodiment, the preferred range of the energy decay coefficient is [insert range here]. ;

[0073] The method for calculating the characteristic gravity is as follows: A pre-set gravity constant is used; the gravity constant is then multiplied sequentially by the characteristic masses of the two candidate features to obtain the potential gravity; the potential gravity is then divided by the square of the feature distance to obtain the characteristic gravity. The gravity constant is pre-set by those skilled in the art based on actual conditions; in this embodiment, the preferred range of the gravity constant is [insert range here]. .

[0074] It should be understood that this embodiment employs a dual mechanism of potential energy and gravity field to dynamically evaluate the effectiveness index of candidate features, transforming feature evaluation from an isolated static analysis into a dynamic process that considers interactions. Specifically, the potential energy mechanism utilizes the prior distribution and initial contribution value of features to evaluate the intrinsic potential of each candidate feature, ensuring that candidate features with good historical performance receive reasonable attention. Simultaneously, the gravity mechanism captures the synergistic effect between different candidate features, allowing high-quality features to attract and enhance the effectiveness index of similar features. This avoids underestimating high-quality features due to their similarity in form and also reveals the potential value of feature combinations. Ultimately, through the dual effects of static potential energy (individual potential) and dynamic gravity (group synergy), a more accurate and comprehensive prediction of the effectiveness of candidate features is achieved.

[0075] Methods for calculating the convergence of different exploration paths sequentially include:

[0076] For all candidate features, they are grouped according to their corresponding feature generators and feature types to obtain multiple path sets; among them, the candidate features in the same path set have the same feature generator and feature type, and the path set corresponds one-to-one with the exploration path;

[0077] The generation time of each candidate feature is obtained through a feature generator; the generation time is the time when the expression tree corresponding to the candidate feature is generated. The candidate features in each path set are sorted from earliest to latest according to their corresponding generation time, generating feature sequences for each path set. Based on the feature sequences of each path set and the effectiveness index of each candidate feature in the feature sequences, the convergence of different exploration paths is calculated sequentially. The expression for convergence is:

[0078] ;

[0079] In the formula, Indicates the degree of convergence. This indicates the number of candidate features in the path set. Indicates the number of the characteristic sequence. The effectiveness index of candidate features of position, Indicates the number of the characteristic sequence. The effectiveness index of candidate features of position, .

[0080] Methods for dynamically planning the validation execution order of candidate features include:

[0081] From the feature sequences corresponding to each path set, obtain the first and last features respectively; where the first feature is the candidate feature ranked first in the feature sequence, and the last feature is the candidate feature ranked last in the feature sequence; calculate the difference between the effectiveness index of the last feature and the effectiveness index of the first feature for each path set, and obtain the index difference for each path set; obtain the feature interval for each path set, where the feature interval is the number of candidate features in the path set minus one, specifically... ; Calculate the ratio between the index difference of each path set and the corresponding feature interval in turn to obtain the changing trend of each exploration path; Standardize the changing trend of each exploration path in turn to obtain the standard trend of each exploration path.

[0082] Different verification weights are set for convergence and standard trend, and the convergence and standard trend of each exploration path are weighted and summed based on the verification weights to obtain the comprehensive index of each exploration path; wherein, the verification weights are set by those skilled in the art according to the actual situation; each candidate feature is sorted from largest to smallest according to the comprehensive index of the corresponding exploration path to generate a preliminary sequence; wherein, candidate features with the same comprehensive index are in parallel positions in the preliminary sequence; the candidate features in parallel positions in the preliminary sequence are sorted from largest to smallest according to the corresponding validity index to generate a comprehensive sequence; the verification execution order of each candidate feature is determined according to the ascending order of the comprehensive sequence.

[0083] It should be understood that dynamic planning of the verification execution order of candidate features aims to balance the stability and trend changes of the exploration path, while making full use of the effectiveness indicators of each candidate feature. First, a comprehensive index is calculated based on the convergence and trend of change, which can prioritize the verification of stable and trend-clear exploration paths. For candidate features with the same comprehensive index, they are then ranked according to the effectiveness index to ensure that high-quality features are verified first. This can improve the efficiency of feature verification, avoid waste of resources, and improve the quality and representativeness of the key features selected later.

[0084] Step S4: Perform feature verification on each candidate feature in sequence according to the verification execution order, evaluate the feature quality index of each candidate feature in real time, and determine whether feature generation is complete based on the feature quality index.

[0085] Methods for real-time evaluation of feature quality metrics for each candidate feature include:

[0086] The observation data is divided into multiple sets of validation data to obtain training and test sets, each containing multiple sets of validation data. A baseline model is constructed and trained based on the validation data in the training set. Specifically, the baseline model is trained using the independent variable data from each set of validation data as input and the dependent variable data from each set of validation data as the prediction target. The independent variable data from each set of validation data in the test set is input into the trained baseline model to obtain the first prediction data corresponding to each set of validation data. Based on the dependent variable data and the first prediction data corresponding to each set of validation data in the test set, the performance index of the baseline model is calculated and labeled as the baseline index.

[0087] Among them, performance metrics are used to measure the accuracy and reliability of the model's prediction results on a given task, such as mean squared error (MSE) and root mean square error (RMSE). It should be noted that the calculation process of performance metrics is based on existing technology, and the specific process will not be elaborated on here.

[0088] The baseline model is a deep neural network model, which includes an input layer, hidden layers, and an output layer. Each hidden layer contains multiple neurons, and each neuron is connected to the neurons in the next layer. The connections contain weights that determine the importance and influence of data transmitted in the neural network. An activation function is applied to each neuron between the hidden and output layers. The activation function introduces non-linearity, allowing the network to learn more complex patterns and features. It should be noted that the deep neural network model is existing technology, and the specific training process will not be elaborated on here.

[0089] The enhancement index calculation process is defined, and the process is executed sequentially for each candidate feature according to the verification execution order to obtain the enhancement index for each candidate feature. The difference between the baseline index and the enhancement index of each candidate feature is calculated sequentially to obtain the feature quality index for each candidate feature, thus completing the feature verification of each candidate feature. The enhancement index calculation process includes:

[0090] A candidate feature is selected as the current feature, and the candidate feature value (i.e., the numerical value of the candidate feature) is calculated sequentially for each set of independent variable data in the observed data and the expression tree corresponding to the current feature. Each candidate feature value is added to the corresponding independent variable data to obtain the independent variable augmentation data, and the validation data with added candidate feature values ​​is marked as validation augmentation data. An augmentation model is constructed and trained based on the validation augmentation data in the training set. Specifically, the augmentation model is trained using the independent variable augmentation data in each set of validation augmentation data as input and the dependent variable data in each set of validation augmentation data as the prediction target. The independent variable augmentation data in each set of validation augmentation data in the test set is input into the trained augmentation model to obtain the second prediction data corresponding to each set of validation augmentation data. Based on the dependent variable data and the second prediction data corresponding to each set of validation augmentation data in the test set, the performance index of the augmentation model is calculated and marked as the benchmark index. It should be noted that the augmentation model is also a deep neural network model.

[0091] Methods for determining whether feature generation is complete based on feature quality metrics include:

[0092] Preset indicator thresholds and quantity thresholds are set by those skilled in the art based on actual conditions. The feature quality index of each candidate feature is compared with the indicator threshold. Candidate features with feature quality indexes greater than the indicator threshold are marked as high-quality features, while candidate features with feature quality indexes less than or equal to the indicator threshold are not marked.

[0093] The number of high-quality features is counted and marked as high-quality; the number of high-quality features is compared with a threshold; if the number of high-quality features is greater than the threshold, feature generation is considered complete; if the number of high-quality features is less than or equal to the threshold, feature generation is considered incomplete.

[0094] Step S5: If feature generation is not completed, dynamically adjust the exploration parameters of each feature generator according to the feature quality index, and return to step S2; if feature generation is completed, select key features from each candidate feature according to the feature quality index.

[0095] Methods for dynamically adjusting the exploration parameters of each feature generator include:

[0096] The average quality index of each feature generator is obtained by averaging the feature quality indices of the candidate features corresponding to the same feature generator; the average quality index of each path set is obtained by averaging the feature quality indices of the candidate features corresponding to each path set; based on the feature quality indices of each candidate feature, the average quality index of each feature generator, the path quality index of each path set, and the number of high-quality features, a heuristic optimization algorithm is used to formulate parameter adjustment strategies for different feature generators.

[0097] Among them, the parameter adjustment strategies corresponding to the language perspective feature generator include the bias adjustment strategy, the prior adjustment strategy, and the scale adjustment strategy; the parameter adjustment strategies corresponding to the numerical perspective feature generator include the bias adjustment strategy, the constraint adjustment strategy, the prior adjustment strategy, and the scale adjustment strategy; heuristic optimization algorithms include ant colony algorithm, genetic algorithm, gray wolf algorithm, etc.

[0098] The exploration parameters of each feature generator are dynamically adjusted according to the parameter adjustment strategy corresponding to each feature generator. Among them, the tendency adjustment strategy is used to adjust the exploration tendency, the constraint adjustment strategy is used to adjust the constraint parameters, the prior adjustment strategy is used to adjust the prior distribution of features, and the scale adjustment strategy is used to adjust the generation scale.

[0099] Methods for selecting key features from candidate features include:

[0100] Each pair of high-quality features is combined sequentially to obtain multiple feature sets. Based on the candidate feature values ​​of the two high-quality features corresponding to each feature set, the mutual information of each feature set is calculated sequentially and marked as feature relevance. It should be noted that the method for calculating mutual information is existing technology, and the specific calculation process will not be elaborated upon here. Each feature relevance is compared with a preset relevance threshold, which is preset by those skilled in the art based on actual conditions. Feature sets with relevance greater than the relevance threshold are marked as high-similarity sets, while feature sets with relevance less than or equal to the relevance threshold are not marked. For each high-similarity set, the feature quality index of the two corresponding high-quality features is compared, and the high-quality feature with the smallest feature quality index is deleted.

[0101] Based on the feature quality index of each high-quality feature, the contribution ratio of each high-quality feature is calculated sequentially; the high-quality features are sorted from largest to smallest according to their corresponding contribution ratios to generate a contribution sequence; the cumulative contribution of each high-quality feature is calculated sequentially according to the ascending order of the contribution sequence; the cumulative contribution of each high-quality feature is compared with a preset contribution threshold, which is preset by those skilled in the art according to the actual situation; high-quality features with cumulative contributions greater than the contribution threshold are deleted from the contribution sequence in turn, and the high-quality features retained in the contribution sequence are used as key features.

[0102] The contribution percentage is calculated as follows: the sum of the feature quality indicators of all high-quality features is calculated to obtain the comprehensive quality indicator; the ratio between the feature quality indicator of each high-quality feature and the comprehensive quality indicator is calculated to obtain the contribution percentage of each high-quality feature.

[0103] The cumulative contribution is calculated as follows: based on the ascending order of the contribution sequence, obtain the preceding set corresponding to each high-quality feature; the preceding set includes all high-quality features that are preceding the corresponding high-quality feature in the contribution sequence; the contribution ratio of each high-quality feature is added to the contribution ratio of each high-quality feature in the corresponding preceding set to obtain the cumulative contribution of each high-quality feature.

[0104] This embodiment integrates linguistic and numerical perspectives in feature generation, achieving a comprehensive utilization of domain knowledge-driven and automated feature exploration. It efficiently extracts candidate features with both interpretive and predictive capabilities from observational data. In the feature evaluation stage, a dynamic potential-gravitational field evaluation mechanism is introduced, considering not only the intrinsic potential of individual candidate features but also capturing the synergistic effects between different candidate features, thus enabling more accurate and comprehensive prediction of the effectiveness indicators of each candidate feature. By dynamically planning the stability and changing trends of the exploration path and fully utilizing the effectiveness indicators of each candidate feature, intelligent scheduling of the verification execution order is achieved, thereby improving the efficiency of feature verification and the utilization rate of computing resources. The dynamic adjustment of exploration parameters during feature generation is also addressed. This paper addresses the key feature selection problem by employing heuristic optimization algorithms and feature contribution analysis methods to achieve adaptive optimization of feature generator parameters and efficient identification of important features, further enhancing the adaptability and robustness of the entire feature generation process. It shifts from single-perspective exploration to multi-perspective collaborative optimization, effectively overcoming the problems of single exploration paths, susceptibility to local optima, and insufficient feature interpretability in traditional methods. While ensuring appropriate feature complexity and structural simplicity, it significantly improves feature generation efficiency, prediction accuracy, and generalization ability. The final selected key features not only possess strong interpretability and high predictability but also effectively remove redundancy and maintain feature diversity, providing an efficient and reliable feature engineering solution for accurate prediction in various professional fields.

[0105] Example 2

[0106] This application also provides an electronic device. The electronic device may include one or more processors and one or more memories. The memories store computer-readable code, which, when executed by the one or more processors, can perform a feature generation method based on symbolic language numerical perspective fusion as described above.

[0107] The method or system according to the embodiments of this application can also be implemented using the architecture of the electronic device shown in this application. The electronic device may include a bus, one or more CPUs, ROM, RAM, a communication port connected to a network, input / output, a hard disk, etc. The storage device in the electronic device, such as a ROM or hard disk, may store the feature generation method based on symbolic language numerical perspective fusion provided in this application. Furthermore, the electronic device may also include a user interface. Of course, the architecture shown in this application is merely exemplary; when implementing different devices, one or more components in the electronic device shown in this application may be omitted according to actual needs.

[0108] Example 3

[0109] Please refer to the accompanying drawings. One embodiment of this application discloses a computer-readable storage medium. The computer-readable storage medium stores computer-readable instructions. When executed by a processor, the computer-readable instructions can perform a feature generation method based on symbolic language numerical perspective fusion according to an embodiment of this application, as described above. The storage medium includes, but is not limited to, volatile memory and / or non-volatile memory. Volatile memory may include, for example, random access memory (RAM) and cache memory. Non-volatile memory may include, for example, read-only memory (ROM), hard disk, flash memory, etc.

[0110] Furthermore, according to embodiments of this application, the processes described in the above-referenced flowcharts can be implemented as computer software programs. For example, this application provides a non-transitory machine-readable storage medium storing machine-readable instructions that can be executed by a processor to perform instructions corresponding to the method steps provided in this application, such as a feature generation method based on symbolic language numerical perspective fusion. When this computer program is executed by a central processing unit (CPU), it performs the functions defined in the method of this application.

[0111] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0112] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0113] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A feature generation method based on symbol language numerical perspective fusion, characterized in that, The application relates to a method for generating candidate features in a target field. The method comprises the following steps: S1: collecting observation data of the target field, fine-tuning and optimizing a preset large language model according to a pre-constructed field knowledge base, setting exploration parameters, and forming a multi-view feature generator group; The observation data comprises multiple sets of verification data, the verification data comprises independent variable data and dependent variable data, and the exploration parameters comprise an exploration tendency, constraint parameters, a feature prior distribution, and a generation scale; The method for forming the multi-view feature generator group comprises the following steps: fine-tuning and optimizing model parameters in the large language model according to the field knowledge base to obtain a language view feature generator; constructing a numerical view feature generator, which is an automatic feature engineering generator of a symbolic regression method based on a fusion function; and integrating the language view feature generator and the numerical view feature generator to form the multi-view feature generator group; S2: based on the observation data, generating expression trees corresponding to each candidate feature through each feature generator in the multi-view feature generator group, collecting feature indicators of the expression trees, and sequentially calculating initial contribution values of the candidate features according to the feature indicators; The method for sequentially calculating the initial contribution values of the candidate features comprises the following steps: The feature indicators comprise tree layer depth, node quantity, operation type number and function type number; The feature indicators of the candidate features are sequentially normalized to obtain standard indicators; different index weights are set for different feature indicators, different standard indicators corresponding to the same candidate feature are weighted and summed based on the index weights to obtain feature complexity of the candidate features; and the feature complexity corresponding to the candidate features is sequentially moderately processed to obtain moderate complexity of the candidate features; The sum of the operation type number and the function type number corresponding to the same candidate feature is calculated to obtain operation type numbers of the candidate features; the total penalty value of each candidate feature is calculated according to the tree layer depth and the operation type number corresponding to the candidate feature; the type weight corresponding to each feature generator is calculated, and the effective complexity of each candidate feature is calculated according to the type weight and the moderate complexity; and the difference between the effective complexity and the total penalty value corresponding to each candidate feature is calculated to obtain the initial contribution value of the candidate feature; S3: dynamically predicting the effectiveness indicators of the candidate features based on the initial contribution values and the feature prior distribution in the exploration parameters, sequentially calculating the convergence degrees of different exploration paths corresponding to the candidate features, and dynamically planning the verification execution order of the candidate features according to the effectiveness indicators and the convergence degrees; The method for dynamically predicting the effectiveness indicators of the candidate features comprises the following steps: The prior probability of each candidate feature is obtained from the feature prior distribution according to the expression tree of the candidate feature, and the initial contribution value of each candidate feature is combined to calculate the basic potential energy of each candidate feature; and the type weight and the basic potential energy of each candidate feature are combined to calculate the modified potential energy of each candidate feature. According to the modified potential energy and the initial contribution value of each candidate feature, the feature quality of each candidate feature is calculated; the feature vector of each candidate feature is constructed, and the feature distance between each two candidate features is calculated in turn; according to the feature distance between each two candidate features and the corresponding feature quality, the feature attractive force between each two candidate features is calculated in turn; the candidate feature with a feature quality greater than a preset quality threshold is taken as a high-quality feature, and the feature attractive force between the same candidate feature and each high-quality feature is added in turn to obtain the total attractive force of each candidate feature; The maximum potential energy is taken as the maximum potential energy from the modified potential energy of each candidate feature; the ratio of the modified potential energy to the maximum potential energy of each candidate feature is calculated respectively to obtain the static effectiveness of each candidate feature; the hyperbolic tangent function is used to convert each total attractive force into a corresponding attractive force gain, and the effectiveness index of each candidate feature is calculated by combining the static effectiveness of each candidate feature; Step S4: performing feature verification on each candidate feature in turn according to the verification execution sequence, evaluating the feature quality index of each candidate feature in real time, and judging whether the feature generation is completed based on the feature quality index; Step S5: if the feature generation is not completed, the exploration parameters of each feature generator are dynamically adjusted according to the feature quality index, and the step S2 is returned; if the feature generation is completed, the key features are selected from each candidate feature according to the feature quality index.

2. The feature generation method based on symbol language numerical perspective fusion according to claim 1, characterized in that, The method for generating the expression tree corresponding to each candidate feature comprises: The observation data are respectively input into each feature generator in the multi-view feature generator set; the language view feature generator generates an expression tree corresponding to each candidate feature based on the input observation data and the corresponding exploration tendency and generation scale; The numerical view feature generator generates an expression tree corresponding to each candidate feature based on the input observation data and the corresponding exploration parameter. The numerical view feature generator generates an expression tree corresponding to each candidate feature based on the input observation data and the corresponding exploration parameter.

3. The feature generation method based on symbol language numerical perspective fusion according to claim 2, characterized in that, The method for calculating the convergence degree of different exploration paths comprises: For all candidate features, the corresponding feature generator and feature type are grouped to obtain a plurality of path sets, and the path set corresponds to one exploration path; the generation time corresponding to each candidate feature is obtained, and the candidate features in each path set are sorted from early to late according to the corresponding generation time to generate a feature sequence corresponding to each path set; based on the feature sequence of each path set and the effectiveness index of each candidate feature in the feature sequence, the convergence degree of different exploration paths is calculated in turn.

4. The feature generation method based on symbol language numerical perspective fusion according to claim 3, characterized in that, The method for dynamically planning the verification execution sequence of each candidate feature comprises: The first feature and the last feature are obtained from the feature sequence of each path set; the index difference of each path set is calculated according to the effectiveness index of the first feature and the last feature corresponding to each path set; the feature interval of each path set is obtained, and the change trend of each exploration path is calculated by combining the index difference; the change trend of each exploration path is standardized in turn to obtain the standard trend of each exploration path; Different verification weights are set for the convergence degree and the standard trend, and the convergence degree and the standard trend of each exploration path are weighted and summed based on the verification weights to obtain the comprehensive index of each exploration path; the candidate features are sorted from large to small according to the comprehensive index of the corresponding exploration path to generate a preliminary sequence; wherein, the candidate features with the same comprehensive index are in parallel positions in the preliminary sequence; the candidate features in the parallel positions in the preliminary sequence are sorted from large to small according to the corresponding effectiveness index to generate a comprehensive sequence; the verification execution sequence of each candidate feature is determined according to the positive sequence of the comprehensive sequence.

5. The feature generation method based on symbol language numerical perspective fusion according to claim 4, characterized in that, The method for evaluating the feature quality indicators of each candidate feature in real time comprises: dividing a plurality of sets of validation data in the observation data to obtain a training set and a test set; constructing a baseline model and training the baseline model based on the validation data in the training set; inputting the independent variable data in each set of validation data in the test set into the trained baseline model to obtain first prediction data corresponding to each set of validation data; calculating the performance indicators of the baseline model based on the dependent variable data and the first prediction data corresponding to each set of validation data in the test set, and marking the performance indicators as baseline indicators; defining an enhanced indicator calculation process, and sequentially performing the enhanced indicator calculation process on each candidate feature according to a validation execution sequence to obtain the enhanced indicators of each candidate feature; sequentially calculating the differences between the baseline indicators and the enhanced indicators of each candidate feature to obtain the feature quality indicators of each candidate feature.

6. The feature generation method based on symbol language numerical perspective fusion according to claim 5, characterized in that, The method for determining whether the feature generation is completed based on the feature quality indicators comprises: presetting an indicator threshold and a quantity threshold, comparing the feature quality indicators of each candidate feature with the indicator threshold, and marking the candidate features with the feature quality indicators greater than the indicator threshold as high-quality features; counting the number of high-quality features and marking the number as a high-quality quantity; comparing the high-quality quantity with the quantity threshold; if the high-quality quantity is greater than the quantity threshold, determining that the feature generation is completed; if the high-quality quantity is less than or equal to the quantity threshold, determining that the feature generation is not completed.

7. The feature generation method based on symbol language numerical perspective fusion according to claim 6, characterized in that, The method for dynamically adjusting the exploration parameters of each feature generator comprises: performing mean calculation on the feature quality indicators of the candidate features corresponding to the same feature generator to obtain the average quality indicators of each feature generator; performing mean calculation on the feature quality indicators of the candidate features corresponding to each path set to obtain the path quality indicators of each path set; adopting a heuristic optimization algorithm to formulate the parameter adjustment strategies corresponding to different feature generators according to the feature quality indicators of each candidate feature, the average quality indicators of each feature generator, the path quality indicators of each path set, and the high-quality quantity; and dynamically adjusting the exploration parameters of each feature generator according to the parameter adjustment strategies corresponding to each feature generator.

8. The feature generation method based on symbol language numerical perspective fusion according to claim 7, characterized in that, The method for screening key features from the candidate features comprises: combining each two high-quality features in turn to obtain a plurality of feature sets; calculating the feature correlations of each feature set in turn according to the candidate feature values of the two high-quality features corresponding to each feature set; comparing each feature correlation with a preset correlation threshold, and marking the feature set with the feature correlation greater than the correlation threshold as a high-similarity set; for each high-similarity set, comparing the feature quality indicators of the two high-quality features corresponding thereto, and deleting the high-quality feature with the smallest feature quality indicator. According to the feature quality indexes of the respective high-quality features, the contribution proportions of the respective high-quality features are sequentially calculated; the respective high-quality features are sorted according to the respective corresponding contribution proportions from large to small to generate a contribution sequence; according to the positive sequence of the contribution sequence, the cumulative contributions of the respective high-quality features are sequentially calculated; the cumulative contributions of the respective high-quality features are respectively compared with a preset contribution threshold value, the high-quality features with the cumulative contributions greater than the contribution threshold value are sequentially deleted from the contribution sequence, and the high-quality features remaining in the contribution sequence are taken as key features.

Images