A method for constructing a regional basic education comprehensive evaluation large model

Abstract

The application provides a method for constructing a regional basic education comprehensive evaluation large model, and relates to the technical field of data processing.The method comprises the following steps: acquiring multi-source data related to regional basic education; performing data cleaning, missing value filling and abnormal value detection; performing feature representation conversion on different data types in the preprocessed data set; calculating the weight relationship between different education factors; calculating the contribution degree of time series data according to multi-source feature data; performing weighted fusion on all data; and constructing a regional basic education comprehensive evaluation large model, which specifically comprises the following steps: performing layer-by-layer feature extraction on a comprehensive feature matrix; extracting the nonlinear relationship between data in the comprehensive feature matrix; fusing deep feature data and high-order feature data to generate model input data; and inputting the model input data into the trained regional basic education comprehensive evaluation large model.The application improves the autonomy and accuracy of the regional basic education comprehensive evaluation large model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a method for constructing a regional basic education comprehensive evaluation large model. BACKGROUND

[0002] At present, the comprehensive evaluation of regional basic education mainly relies on the method of manually designing the evaluation system. Specifically, the construction of the evaluation system is usually set by experts in the field of education according to experience to set multiple dimensions such as academic achievement, teacher level, school resources, etc., and the comprehensive evaluation score is calculated through fixed weight. This method has a certain guiding nature on the evaluation result, but its evaluation model is mostly linear structure, and the index setting and weight distribution have strong subjectivity, which is difficult to fully reflect the complexity and dynamic changes of the regional education system, especially in the era of big data when facing the integration and analysis of multi-source data, it is not enough.

[0003] Specifically, the traditional evaluation system mostly relies on simple quantitative data such as the number of teachers or the number of teaching equipment per student, and cannot effectively handle the correlation between multi-source heterogeneous data such as the actual use efficiency of teaching equipment, the professional development record of teachers, and the learning feedback of students. This leads to the evaluation result focusing on surface phenomena and ignoring deep-seated problems. SUMMARY

[0004] The purpose of the present application is to provide a method for constructing a regional basic education comprehensive evaluation large model, which aims to solve the problems mentioned in the background art.

[0005] To solve the above technical problems, the technical solution of the present application is as follows:

[0006] A method for constructing a regional basic education comprehensive evaluation large model, the method comprising:

[0007] Obtaining multi-source data related to regional basic education to form an original data set;

[0008] Performing data cleaning, missing value filling, and outlier detection on the original data set to obtain a preprocessed data set;

[0009] Converting the different data types in the preprocessed data set to feature representations to obtain multi-source feature data;

[0010] According to the multi-source feature data, the weight relationship between different education factors is calculated to generate a factor influence matrix;

[0011] According to the multi-source feature data, the contribution degree of the time series data is calculated to generate a time series influence matrix;

[0012] According to the factor influence matrix and the time series influence matrix, all data is weighted and fused to generate a comprehensive feature matrix;

[0013] According to the comprehensive feature matrix, a regional basic education comprehensive evaluation large model is constructed, specifically including:

[0014] The comprehensive feature matrix is extracted layer by layer to obtain deep feature data;

[0015] The nonlinear relationship between the data in the comprehensive feature matrix is extracted to obtain high-order feature data;

[0016] The deep feature data and the high-order feature data are fused to generate model input data;

[0017] The model input data is input into the trained regional basic education comprehensive evaluation large model for evaluation calculation to obtain an initial evaluation result;

[0018] Based on historical evaluation data, the initial evaluation result is adaptively weighted to generate a final evaluation result.

[0019] Preferably, the different data types in the preprocessed data set are converted into feature representations to obtain multi-source feature data, including:

[0020] The numerical data in the preprocessed data set is normalized to obtain normalized data;

[0021] The category mapping matrix is constructed for the category data in the preprocessed data set, the category features are converted into calculable numerical features, and the category feature data is obtained;

[0022] The time window of the time series data in the preprocessed data set is divided, the short-term trend features and the long-term trend features are extracted, and the time dependence relationship is constructed to obtain the time series feature data;

[0023] The text data in the preprocessed data set is segmented, the text data is mapped into a numerical feature vector, and the text feature data is obtained.

[0024] Preferably, the weight relationship between different education factors is calculated according to the multi-source feature data to generate a factor influence matrix, including:

[0025] The correlation analysis method based on the attention mechanism is used to calculate the weights of the multi-source feature data to obtain the basic correlation distribution between the education factors;

[0026] According to the fluctuation trend of the time series data, a time sensitive coefficient is set for each education factor, and the factors affected by time are filtered through the time sensitive coefficient to adjust the correlation distribution and obtain the correlation adjustment distribution;

[0027] The correlation adjustment distribution is normalized to generate the factor influence matrix; wherein,

[0028] , is the influence weight of the education factor in the factor influence matrix on the education factor , is the basic correlation score between the education factor and the education factor , is the modulation coefficient of the education factor on the education factor , is the fluctuation trend intensity of the education factor over time is the time sensitivity coefficient , if the education factor is affected by time, then ; if the education factor is not affected by time, then , is the total number of education factors is the basic correlation score between the education factor and the education factor , is the modulation coefficient of the education factor on the education factor .

[0029] Preferably, the contribution degree of the time series data is calculated according to the multi-source feature data, and a time series influence matrix is generated, comprising:

[0030] time window division is performed on the time series feature data to extract short-term trend features and long-term trend features;

[0031] The dynamic time warping method is used to calculate the data matching degree between different time periods, and the influence of abnormal values is reduced through an adaptive smoothing factor;

[0032] The comprehensive influence weight of each time window is calculated by combining the short-term influence weight and the long-term trend influence weight, and normalized to generate a time series influence matrix; wherein,

[0033] , wherein, is the time series influence weight, is the normalization coefficient, is the short-term influence weight, is the long-term trend influence weight, is the change rate of the education factor in the time window ; wherein,

[0034] wherein, is an education factor is an education factor is a dynamic alignment distance in the time dimension, and is an education factor is an education factor is an education factor is an education factor is a smoothing coefficient.

[0035] Preferably, the factor influence matrix and the time sequence influence matrix are used to weight and fuse all data to generate a comprehensive feature matrix, including:

[0036] The factor influence matrix and the time sequence influence matrix are standardized to have the same weight value in the same calculation range;

[0037] The factor influence matrix after standardization is used to weight and calculate the multi-source feature data to obtain factor weighted feature data;

[0038] The time sequence influence matrix after standardization is used to weight and calculate the time sequence feature data to obtain time sequence weighted feature data;

[0039] The factor weighted feature data and the time sequence weighted feature data are fused by using a nonlinear activation method to generate a final comprehensive feature matrix.

[0040] Preferably, the comprehensive feature matrix is subjected to layer-by-layer feature extraction to obtain deep feature data, including:

[0041] The comprehensive feature matrix is input into a multi-layer feature extraction network, and a deep feature learning mechanism is used to extract feature representations at different levels;

[0042] The features extracted at each layer are subjected to nonlinear transformation to obtain preliminary feature representations;

[0043] The preliminary feature representations are subjected to dimension reduction processing to remove redundant information to obtain deep feature data.

[0044] Preferably, the nonlinear relationship between data in the comprehensive feature matrix is extracted to obtain high-order feature data, including:

[0045] A feature interaction network is constructed for the comprehensive feature matrix to establish cross connections between different features;

[0046] The nodes in the feature interaction network are subjected to multi-layer propagation calculation to extract nonlinear influence relationships between features;

[0047] The nonlinear relationships are screened to remove low-contribution-degree interaction features to obtain high-order feature data.

[0048] Preferably, the depth feature data and high-order feature data are fused to generate model input data, including:

[0049] The depth feature data and high-order feature data are scaled and aligned.

[0050] The depth feature data and high-order feature data are fused layer by layer using a multi-scale feature fusion method.

[0051] The fused feature data is regularized to obtain model input data.

[0052] Preferably, the model input data is input into the trained regional basic education comprehensive evaluation large model to perform evaluation calculation to obtain an initial evaluation result, including:

[0053] The model input data is feature mapped to extract feature representation data for comprehensive evaluation.

[0054] The feature representation data is calculated using pre-trained weight parameters to generate basic scores in multiple dimensions.

[0055] The basic scores in multiple dimensions are corrected based on score bias of historical evaluation data to obtain normalized score data.

[0056] The overall score is calculated using a weighted fusion method to generate an initial evaluation result; wherein,

[0057] , wherein, is the initial evaluation result of the i-th educational unit, is the dynamic weight of the j-th evaluation factor, is the basic score of the i-th educational unit on the j-th evaluation factor, is the adaptive score correction coefficient, is the score bias of the j-th evaluation factor, is the total number of evaluation factors. Preferably, the initial evaluation result is adjusted based on historical evaluation data to generate a final evaluation result, including: Based on the score bias of historical evaluation data, an exponential decay correction method is used to dynamically adjust the score weight to obtain an adjusted time sequence influence weight. The initial evaluation result is adjusted according to the adjusted time sequence influence weight to obtain a final evaluation result; wherein,

[0058]

[0059]

[0060] ​​​​​

[0061] , is the final evaluation result, is the correction time sequence influence weight of the first evaluation factor, is the correction time sequence influence weight of the first evaluation factor, is the initial evaluation result; wherein,

[0062] wherein, is the time sequence influence weight of the first evaluation factor, is the time sequence influence weight of the first evaluation factor, is the learning rate, is the adjustment stability coefficient.

[0063] The above scheme of the present application at least includes the following beneficial effects:

[0064] In the comprehensive evaluation process of regional basic education, the traditional evaluation system relies on manual setting of weights and indicators, and calculates the evaluation score by linear model. Although this method can provide guidance for education evaluation to a certain extent, due to the linear characteristics of the evaluation system, it is difficult to fully depict the complex relationship of the regional education system. Especially when facing different types of data, the traditional method cannot efficiently handle the deep correlation between multi-source heterogeneous data, so that the final evaluation result may only reflect the surface data characteristics, and the dynamic change trend of the education system cannot be revealed, resulting in limitations of the evaluation result.

[0065] The present method can effectively overcome the limitations in the prior art by introducing a data-driven large model method combined with the integration analysis of multi-source data. First, the present method can collect education data including academic performance, teacher level, school resource allocation, teaching equipment usage, teacher professional development, student learning feedback and other aspects from the regional education management system to form an original data set. Compared with the traditional method which only relies on quantitative data such as the number of teachers and the number of teaching equipment, the present method can collect more comprehensive education data, and through data preprocessing means such as data cleaning, outlier detection and missing value filling, the data quality is improved to ensure that the data input into the model has high credibility. This way can avoid the influence of data quality problems on the evaluation result, and improve the stability and adaptability of the model.

[0066] Secondly, in terms of data processing, the method adopts specific feature conversion methods for different types of data, making it applicable to subsequent calculations. Numerical data is normalized to the same numerical range, avoiding the bias caused by different scales of data; categorical data is converted into calculable numerical data through the construction of a category mapping matrix, improving the feasibility of data calculation; time series data extracts short-term and long-term features through time window segmentation, ensuring that the model can capture the dynamic trends of the education system; and text data is converted into structured numerical data through word segmentation and word vector embedding, making student learning feedback and other information available to the model. This efficient integration of multi-source data solves the problem of traditional evaluation methods that cannot handle multi-source heterogeneous data, allowing the complexity and dynamic changes of the regional education system to be more accurately described.

[0067] In terms of data modeling, the method builds a deep learning-based regional basic education comprehensive evaluation large model, which can automatically learn the complex interaction between different education factors and improve the expression ability of the evaluation model. Traditional evaluation methods use fixed weight calculation methods, which are difficult to adapt to changes in education in different regions and at different times. This method builds a factor influence matrix and a time series influence matrix, allowing the model to automatically learn the weight relationship between different education factors and adjust the weights adaptively based on historical evaluation data, allowing the evaluation model to dynamically optimize over time. In particular, the model architecture based on deep feature extraction can automatically extract deep features from education data, allowing the model to not only identify explicit features such as academic performance and teacher quality, but also to uncover the impact of implicit factors such as teaching resource allocation and teacher development on education quality, providing more accurate comprehensive evaluation results.

[0068] In addition, the method can capture the non-linear relationship between different education factors by building high-order feature data, improving the adaptability of the model. For example, the impact of teacher training on academic performance is not linear, but there is a certain time lag effect, i.e. the effect of teacher training needs to be reflected in student performance after a period of time. Traditional linear evaluation methods are difficult to model this non-linear and time-lag characteristic, while the method can identify the long-term impact of education factors through feature interaction modeling and time series analysis, improving the scientificity and reliability of the evaluation. Through the optimization of model input data and deep feature learning mechanism, the method can effectively avoid the subjectivity of manually setting weights, making the scoring results more objective and accurate, and more truly reflecting the operation of the education system.

[0069] Finally, the method adjusts the initial evaluation results based on historical evaluation data to adaptively adjust the weights, further optimizing the stability and rationality of the evaluation results. Traditional methods often lead to scoring results that are easily affected by short-term data fluctuations due to the lack of dynamic feedback mechanisms for historical data, making it difficult to accurately reflect long-term trends in education quality. However, the method can dynamically optimize the scoring weights by combining the scoring bias of historical data, making the final evaluation results more consistent with long-term trends and improving the stability and interpretability of the scores. This not only improves the sensitivity of regional education management departments to changes in education quality, but also provides a more scientific basis for education decision-making, helping to promote high-quality development of regional basic education.

[0070] In summary, compared with traditional methods based on artificial design of evaluation systems, the method has improved in data collection, data processing, model construction, nonlinear feature extraction, time series analysis, adaptive optimization, and other aspects, overcoming the limitations of existing technologies in multi-source data integration, dynamic adjustment, nonlinear relationship modeling, and other aspects, making the comprehensive evaluation of regional basic education more accurate, scientific, dynamic, and objective. BRIEF DESCRIPTION OF DRAWINGS

[0071] Figure 1 is a flowchart architecture diagram of a construction method of a regional basic education comprehensive evaluation large model provided by an embodiment of the present application. DETAILED DESCRIPTION

[0072] Exemplary embodiments of the present disclosure will be described in greater detail below with reference to the accompanying drawings. Although exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the present disclosure to those skilled in the art.

[0073] As Figure 1 shown, an embodiment of the present application proposes a construction method of a regional basic education comprehensive evaluation large model, which comprises:

[0074] S100, acquiring multi-source data related to regional basic education to form an original data set, including academic achievement data, teacher level data, school resource allocation data, teaching equipment usage data, teacher professional development data, and student learning feedback data;

[0075] S200, data cleaning, missing value filling, and outlier detection are performed on the original data set to obtain a preprocessed data set;

[0076] S300, feature representation conversion is performed on different data types in the preprocessed data set to obtain multi-source feature data;

[0077] S400, calculating the weight relationship between different education factors according to the multi-source feature data, and generating a factor influence matrix;

[0078] S500, calculating the contribution degree of time series data according to the multi-source feature data, and generating a time series influence matrix;

[0079] S600, weighting and fusing all data according to the factor influence matrix and the time series influence matrix, and generating a comprehensive feature matrix;

[0080] S700, constructing a regional basic education comprehensive evaluation large model according to the comprehensive feature matrix, specifically including:

[0081] performing layer-by-layer feature extraction on the comprehensive feature matrix to obtain deep feature data;

[0082] extracting the nonlinear relationship between data in the comprehensive feature matrix to obtain high-order feature data;

[0083] fusing the deep feature data and the high-order feature data to generate model input data;

[0084] S800, inputting the model input data into the trained regional basic education comprehensive evaluation large model to perform evaluation calculation, and obtaining an initial evaluation result;

[0085] S900, based on historical evaluation data, adaptively adjusting the initial evaluation result to generate a final evaluation result.

[0086] In the embodiments of the present application, when performing regional basic education comprehensive evaluation, the integrity, diversity and scalability of the data need to be ensured first. By collecting various education data in the region, multiple aspects such as teacher quality, student performance, school resources and teaching methods can be covered. Academic achievement data can be extracted from various examinations, teacher level data can include teacher education, title, teaching experience and other information, and school resource allocation data can cover laboratory equipment, library resources, computer equipment, etc. In addition, the use data of teaching equipment can reflect the actual utilization rate of resources, the data of teacher professional development can be used to analyze the long-term impact of teacher growth on student performance, and the learning feedback data of students can reflect the adaptability of the education mode.

[0087] By standardizing the data preprocessing, including data cleaning, missing value filling, and outlier detection, the quality of the data can be improved, and the errors and data bias in model training can be reduced. For example, if some students' academic performance data is missing in a school, an interpolation filling method based on similar student groups can be used to ensure the continuity and integrity of the data. At the same time, outlier detection can be based on statistical methods or machine learning algorithms to filter out data points that may have been entered incorrectly or are extreme cases, avoiding a significant impact on the evaluation results.

[0088] After preprocessing the data, feature engineering is needed, which means converting different types of data to make them suitable for subsequent large model calculations. Since regional education data usually includes numerical data, categorical data, time series data, and text data, different methods need to be used for feature conversion according to different data types. For numerical data, normalization is used to normalize all feature values to the same numerical range, avoiding calculation bias caused by dimension differences. For categorical data, a category mapping matrix is constructed to convert categorical data into numerical representation so that it can participate in mathematical calculations. For example, for the category variable of teacher title, a mapping matrix can be constructed to convert "assistant", "lecturer", "associate professor", and "professor" into corresponding numerical representations. For time series data, a time window slicing method is used to extract short-term and long-term trend features, so that the model can identify the trend of educational quality over time. For text data, such as student feedback on courses, word segmentation processing is used, and word vector embedding methods are used to map text data into numerical features, so that they can participate in subsequent calculations.

[0089] After completing the feature conversion, a multi-layer data correlation matrix needs to be constructed to depict the relationship between different educational factors. In traditional education evaluation methods, the weights of different indicators are usually determined by expert experience, while this method uses a data-driven weight calculation method to make the evaluation model more dynamically reflect the complex relationships in the education system. Based on the feature matrix constructed from multiple sources of data, deep learning methods are used to extract features, which can obtain deep feature data with stronger expression ability. At the same time, in order to further enhance the non-linear expression ability of the model, a feature interaction network is constructed to automatically learn high-order features from the data, so that the model can identify complex factor influence relationships, such as the long-term impact of teacher training on student performance or the hidden role of teaching equipment utilization on learning effectiveness.

[0090] Finally, through the constructed regional basic education comprehensive evaluation model, multi-layer information fusion can be performed based on the input data to generate an initial evaluation result. To ensure the stability of the model evaluation, an adaptive weight adjustment mechanism is used, so that the model can dynamically optimize the initial evaluation result by combining historical evaluation data. For example, the education quality of a certain region may be affected by factors such as economic level, policy support, etc. Through feedback from historical evaluation data, the weights of various factors can be dynamically adjusted, making the final evaluation result more accurate and reasonable. In this way, the comprehensive evaluation of regional basic education can more accurately reflect the true situation of the education system, providing a scientific basis for decision-making for education managers.

[0091] In one case of the present embodiment, for the acquisition of regional basic education related multi-source data, the original data set is formed:

[0092] In the process of comprehensive evaluation of regional basic education, the comprehensiveness and diversity of data directly determine the accuracy and applicability of the evaluation system. Since regional basic education involves multiple dimensions, such as academic performance, teacher level, school resource allocation, teaching equipment use, teacher professional development, student learning feedback, etc., it is necessary to obtain a complete original data set from multiple data sources to ensure that the basic data for subsequent calculations have high coverage and high credibility.

[0093] First, academic performance data is usually stored in a structured table form, including student identity information such as student ID, class, subject name, test date, score, etc. To improve the applicability of the data, statistical features of the scores can be further extracted, such as average scores of each subject, score variance, progress rate, etc., to reflect the long-term academic trends of students. In addition, fine-grained data such as homework completion, classroom test scores, etc. can be collected from education management platforms to provide more comprehensive academic performance analysis.

[0094] Second, teacher level data mainly involves teachers' education, title, teaching experience, research achievements, teacher training, etc. These data are usually stored in teacher management systems or human resource systems, including teachers' basic information such as name, gender, teaching subject, education background such as master's and doctoral degrees, title evaluation such as lecturer, associate professor, professor, annual assessment results, etc. In addition to static data, dynamic data such as teacher training records, classroom teaching scores, teaching feedback, etc. should also be collected, which can reflect the growth trajectory of teachers' professional ability and the impact on students' learning effect.

[0095] In addition, school resource allocation data covers the school's infrastructure, teaching equipment, teacher-student ratio, class size, etc. It can usually be obtained from education statistics databases or school annual reports, such as the number of classrooms, laboratory equipment, computer classrooms, library holdings, sports facilities, etc. These data can not only be used to assess whether the school's basic resources are sufficient, but also combined with other data to calculate the utilization rate of teaching resources, so as to measure the rationality of school resource allocation.

[0096] During data collection, teaching equipment usage data is also crucial. Traditional education evaluation systems usually only focus on the total amount of equipment, ignoring the actual usage of the equipment. Therefore, data including equipment usage frequency, usage duration, usage efficiency, etc. need to be collected. For example, a school may have a large number of computers, but if its actual usage rate is low, the effectiveness of its educational resources may not be ideal. Therefore, the usage logs of equipment can be collected through Internet of Things devices or equipment management systems, and combined with teaching arrangements to calculate the utilization rate and efficiency of equipment.

[0097] In addition to physical resources, teacher professional development data can effectively measure the degree of support for teacher career growth in the education system. This part of data can include the number of trainings a teacher has attended, the content of the trainings, the outcomes of the trainings, etc. In addition, a teacher's research achievements, teaching research projects, and publication of academic papers can also be important indicators of a teacher's professional growth. For example, a teacher's training experience may have an impact on their teaching methods, which in turn indirectly affects students' learning performance, so reasonable collection of teacher training data can provide an important basis for education evaluation.

[0098] Finally, student learning feedback data can complement the shortcomings of traditional quantitative evaluation, making the evaluation system more comprehensive. This part of data usually comes from student questionnaires, parent feedback, classroom interaction data, etc. For example, data such as students' learning duration, homework submission, and classroom participation can be collected through online learning platforms. In addition, students' feedback on different teaching modes, such as online learning and flipped classroom, can also be analyzed. These data can be combined with academic performance data to analyze the impact of teaching methods on students' performance improvement.

[0099] In summary, the process of obtaining raw data involves multiple data sources, including the examination system, teacher management system, equipment management system, education statistics database, learning feedback platform, etc. The data collection methods can include database query, API interface acquisition, sensor data collection, questionnaire survey, etc. In order to ensure the integrity and consistency of the data, the time period of all data collection should be consistent, such as collecting according to the semester or academic year. In addition, in order to protect data privacy, data desensitization processing should be carried out during data collection, such as anonymizing the information of students and teachers, to ensure the safety of personal privacy.

[0100] In one case of the present embodiment, for the data cleaning, missing value filling, and outlier detection of the raw data set, a preprocessed data set is obtained:

[0101] After obtaining the raw data, the quality of the data directly determines the reliability of the subsequent analysis. Therefore, it is necessary to clean the data, fill in the missing values, and detect outliers to improve the quality of the data and reduce errors in the model calculation process.

[0102] Data cleaning is the first step of data preprocessing, and its goal is to remove invalid data, correct data format errors, and ensure data normalization. In the raw data set, there may be problems of inconsistent formats, such as academic achievement data that may have both percentage and GPA systems; teacher titles may be represented by words such as "associate professor" or numbers such as "2". Therefore, it is necessary to unify the data format so that all data use the same representation. In addition, data cleaning also includes removing duplicate data, for example, data collected at different times may have duplicate records, if not removed, it may cause calculation results to deviate. Usually, unique identifiers such as student ID, teacher ID can be used for deduplication to ensure the uniqueness of the data.

[0103] Missing value filling is a common challenge in data processing. Due to the complexity of the source of education data, some data may be missing due to information not being entered or system errors. For example, some students may lack the scores of a certain exam, or some teachers' training records may not be complete. For missing values of numerical data, methods such as mean filling, median filling, or K-nearest neighbor filling can be used. For example, in the case of missing scores in a certain subject, the average score of other subjects of that student can be used for filling, or the average score of the class can be used as the filling value. For categorical data, the mode filling method can be used, that is, the most common category value is filled. For example, if some teachers' title information is missing, the most common title category among teachers of the same subject in the school can be filled. In addition, for data with a high missing rate, the column can be directly deleted to avoid the uncertainty brought by filling data.

[0104] Outlier detection is an important step to ensure data reliability. In educational data, outliers may manifest as extremely high or low academic performance, unreasonable teacher teaching time, excessively high or low device usage, etc. For example, if a student's test score is much higher or lower than the normal range of the subject, it may be a data entry error or an actual abnormal situation. Therefore, statistical methods such as standard deviation method or machine learning methods such as Isolation Forest algorithm can be used for outlier detection. For numerical data, the mean and standard deviation of the data can be calculated, and data exceeding the preset threshold is considered as an outlier and is manually reviewed or filled. For example, if a student's test score exceeds the full score, it may be a data entry error and should be manually corrected. For categorical data, we can check if some categories are not logical, for example, if a teacher's title is "professor" but his working time is only one year, it may be a data error.

[0105] After completing data cleaning, missing value filling and outlier detection, the final preprocessed dataset is more complete and accurate, and all data has been standardized, suitable for subsequent analysis and calculation. This data preprocessing process not only improves the data quality, but also ensures the reliability of the model calculation results, providing a solid data foundation for regional basic education comprehensive evaluation.

[0106] The regional basic education comprehensive evaluation large model includes but is not limited to the following:

[0107] Model architecture:

[0108] Multi-layer feature extraction network for layer-by-layer deep feature extraction.

[0109] Feature interaction network for capturing high-order nonlinear relationships between education factors.

[0110] Time series influence modeling module for calculating the time contribution of education factors.

[0111] Adaptive weight adjustment mechanism for dynamic optimization combined with historical evaluation data.

[0112] Training method:

[0113] Supervised learning method, such as supervised training using known education evaluation data.

[0114] Loss function selection, such as calculating score error based on mean square error MSE or cross entropy.

[0115] Optimization algorithm, such as Adam, SGD for parameter optimization.

[0116] Data input features, such as academic performance, teacher level, teaching device usage, etc.

[0117] In a preferred embodiment of the present application, the feature representation conversion of different data types in the preprocessed data set obtains multi-source feature data, comprising:

[0118] The numerical data in the preprocessed data set is normalized to obtain normalized data;

[0119] The category mapping matrix is constructed for the category data in the preprocessed data set, the category features are converted into calculable numerical features, and the category feature data is obtained;

[0120] The time window is divided for the time series data in the preprocessed data set, the short-term trend features and the long-term trend features are extracted, and the time dependence relationship is constructed to obtain the time series feature data;

[0121] The text data in the preprocessed data set is segmented, the text data is mapped into a numerical feature vector to obtain text feature data.

[0122] In the embodiment of the present application, when performing regional basic education comprehensive evaluation, the type and processing method of data directly affect the precision and adaptability of the evaluation model. The education data usually contains multiple types, including numerical data such as academic performance, number of teachers, school resources, category data such as teacher title, student category, time series data such as teaching quality trend, and text data such as teaching evaluation feedback. If different types of data are not properly processed, information loss or information distortion may occur in the model calculation process, affecting the accuracy of the final evaluation result.

[0123] In order to ensure that the data can be correctly input into the evaluation model, feature conversion needs to be performed on different types of data. For numerical data, normalization method is adopted to map all data into the same numerical range, avoiding the scale difference between different variables to interfere with the calculation. For example, academic performance is usually expressed in percentage, while the number of teachers is an integer value. Directly using these data for calculation may cause the model to be biased towards variables with larger numerical range, and normalization can effectively alleviate this problem. For category data, mapping matrix method is needed to convert discrete category variables into numerical variables. For example, the teacher title can be converted to “assistant = 1, lecturer = 2, associate professor = 3, professor = 4”, so that different categories have mathematical calculation feasibility.

[0124] The processing of time series data is particularly important because the quality of education is a dynamic process that needs to capture trends in the time dimension. Therefore, the method of time window segmentation is adopted to divide continuous time series data into fixed time periods and extract short-term and long-term trend features. For example, when analyzing the education level of a certain area, the past five years of data can be used as a time window to extract the long-term trend of student performance and the short-term impact of teacher training activities on performance. In this way, the evaluation model can be time-sensitive and dynamically adjust the weights of different factors.

[0125] The processing of text data is also a key link. Student and teacher feedback data often contain important subjective evaluation information. In order to enable the model to use these text data, natural language processing techniques need to be used to convert text data into numerical features. Specifically, first, perform word segmentation to remove irrelevant words, and then use word vector embedding methods to convert text into computable numerical representations. For example, the Word2Vec method can be used to convert evaluations such as "good classroom interaction" into vectors so that they can be used for subsequent analysis and calculation.

[0126] Finally, through the feature representation conversion of multi-source data, it can ensure that the data input model is complete and reasonable, providing a basis for subsequent education factor weight calculation and comprehensive evaluation.

[0127] In a preferred embodiment of the present application, the weight relationship between different education factors is calculated according to the multi-source feature data to generate a factor influence matrix, which includes:

[0128] The correlation analysis method based on the attention mechanism is used to calculate the weights of multi-source feature data to obtain the basic correlation distribution between education factors; wherein,

[0129] , is the basic correlation score between education factors and education factors , and is the characteristic value of education factors and education factors at time , is the length of the time window;

[0130] According to the fluctuation trend of time series data, a time-sensitive coefficient is set for each education factor, and the time-affected factor is filtered through the time-sensitive coefficient to adjust its correlation distribution to obtain a correlation adjustment distribution;

[0131] The correlation adjustment distribution is normalized to generate a factor influence matrix; wherein,

[0132] , is the influence weight of the education factor in the factor influence matrix on the education factor , is the basic correlation score between the education factor and the education factor , is the modulation coefficient of the education factor on the education factor , is the fluctuation trend intensity of the education factor over time is the time sensitivity coefficient is the time sensitivity coefficient, if the education factor is affected by time ; if the education factor is not affected by time , is the total number of education factors is the basic correlation score between the education factor and the education factor , is the modulation coefficient of the education factor on the education factor ; wherein

[0133] , is the mean value of the education factor in the time window is the length of the time window

[0134] In the embodiments of the present application, in the calculation process of the education evaluation model, the influence of different education factors on the final score is not the same, so it is necessary to calculate the weight relationship between each factor and generate a factor influence matrix. The traditional education evaluation method usually adopts the way of manually setting the weight, for example, experts give "teacher level accounts for 30%, school resources accounts for 20%" according to experience. This way has strong subjectivity and is difficult to adapt to the complexity of different regional education systems. The present method adopts a data-driven way to automatically calculate the correlation between education factors and dynamically adjust the weight according to the correlation, so as to more accurately reflect the real situation of the education system.

[0135] By the correlation analysis method of the attention mechanism, the basic correlation distribution between multi-source data can be calculated, so that the influence relationship between different factors can be quantified. For example, when analyzing the relationship between teacher training and academic performance, the correlation score can be calculated to determine whether the increase in teacher training has a significant impact on the improvement of student performance. In addition, considering that the quality of education is a dynamic process, it is necessary to set a time-sensitive coefficient considering the time factor, so that the weight of the education factor can be dynamically adjusted over time. For example, the implementation of a certain education policy may take several years to see the effect, and the time-sensitive parameter can ensure that the model considers this time delay when calculating, so that the weight adjustment is more in line with the actual situation. Finally, through normalization processing, the factor influence matrix is obtained, which provides more accurate basic data support for comprehensive evaluation.

[0136] Among them, the education factor refers to various factors that affect the quality of regional basic education, that is, the data in the multi-source feature data set, including academic performance data, teacher level data, school resource allocation data, teaching equipment use data, teacher professional development data, and student learning feedback data.

[0137] In a preferred embodiment of the present application, the contribution degree of the time series data is calculated according to the multi-source feature data, and a time series influence matrix is generated, comprising:

[0138] The time window is divided for the time series feature data, and short-term trend features and long-term trend features are extracted;

[0139] The dynamic time warping method is used to calculate the data matching degree between different time periods, and the influence of abnormal values is reduced through an adaptive smoothing factor;

[0140] The short-term influence weight and the long-term trend influence weight are combined to calculate the comprehensive influence weight of each time window, and normalization processing is performed to generate a time series influence matrix; wherein,

[0141] , wherein, is the time series influence weight, is the normalization coefficient, is the short-term influence weight, is the long-term trend influence weight, is the education factor The change rate within the time window ; wherein,

[0142] , wherein, is the education factor is the dynamic alignment distance of the education factor in the time dimension, and are the time upper education factor with the education factor value, is a smoothing coefficient.

[0143] In the embodiment of the present application, in the comprehensive evaluation process of regional basic education, the contribution degree calculation of time series data is the key link to ensure the accuracy and dynamic adaptability of the evaluation system. Traditional education evaluation methods often only focus on single time data, ignoring the change trend of education quality in different time periods. This static evaluation method is difficult to reflect the dynamic adjustment process of the education system. By calculating the contribution degree of time series data, the time series influence matrix is generated, so that the evaluation model can capture the change law of education factors in the time dimension, and ensure that the evaluation result can accurately reflect the long-term development trend of the education system.

[0144] In order to calculate the contribution degree of time series data, it is necessary to divide the time window of time series characteristic data first, and extract short-term trend features and long-term trend features. Short-term trend features are used to analyze the fluctuations of education factors in the short term, such as the immediate impact of teacher training on student performance, while long-term trend features are used to analyze the long-term impact of resource adjustment on overall education quality. For example, in the evaluation of basic education in a certain region, the past five years can be selected as the time window to extract time series features of student performance, teacher level, school resources and other factors, so as to identify the change trend of education quality in different time periods.

[0145] In calculating the matching degree of time series data, the dynamic time warping method is adopted to reduce the interference of data noise and outliers. Traditional time series analysis methods usually assume that the time step is fixed, but in education evaluation, the data collection period of different schools or regions may not be consistent, resulting in errors when directly calculating the correlation of time series. The dynamic time warping method adjusts the time alignment path dynamically to ensure that the data in different time periods can be reasonably matched, thereby improving the accuracy of the calculation. For example, when analyzing the teacher training data of a school, the school may organize training once a year, while another school may organize training once every half year. The dynamic time warping method can ensure that these data of different frequencies can be aligned when calculating, making the calculation result more reliable.

[0146] In addition, in order to further improve the calculation stability of the time series influence matrix, an adaptive smoothing factor is used to reduce the influence of outliers. In actual data, some time education data may have extreme values due to sudden events, policy adjustments and other factors. If not handled, it may have a great impact on the final calculation of the time series influence matrix. By introducing an adaptive smoothing factor, outliers can be reasonably processed during calculation, ensuring that the calculated time series influence matrix can truly reflect the time series change law of education factors.

[0147] Finally, the comprehensive influence weight of each time window is calculated by combining the short-term influence weight and the long-term trend influence weight, and the result is normalized to generate the final time sequence influence matrix. This matrix not only quantifies the contribution of different time education factors, but also provides a more adaptive time sequence weight adjustment mechanism in subsequent comprehensive evaluation calculation, thereby ensuring the reliability and dynamic adaptability of the final evaluation result.

[0148] In a preferred embodiment of the present application, the comprehensive feature matrix is generated by weighting and fusing all data according to the factor influence matrix and the time sequence influence matrix, which includes:

[0149] The factor influence matrix and the time sequence influence matrix are standardized to make their weight values in the same calculation range;

[0150] The factor weighted feature data is obtained by weighting calculation of multi-source feature data according to the standardized factor influence matrix;

[0151] The time sequence weighted feature data is obtained by weighting calculation of time sequence feature data according to the standardized time sequence influence matrix;

[0152] The factor weighted feature data and the time sequence weighted feature data are fused by using a nonlinear activation method to generate the final comprehensive feature matrix; wherein,

[0153] , wherein, is the comprehensive feature matrix, is the factor influence matrix weight, is the time sequence influence matrix weight, is the factor weighted feature data, is the time sequence weighted feature data, is a bias term for correcting the deviation of the causal relationship, is a bias correction coefficient, is a nonlinear activation function, such as ReLU or Sigmoid.

[0154] In the embodiment of the present application, in the comprehensive evaluation process of basic education, the influence of education factors and time sequence factors is often cross-acting, and separately calculating the education factor weight or the time sequence factor weight may cause result deviation. Therefore, it is necessary to weight and fuse all data based on the calculated factor influence matrix and time sequence influence matrix to generate a comprehensive feature matrix, so that the influence of different factors can work together to provide a more complete evaluation result.

[0155] In order to ensure the rationality of data fusion, the factor influence matrix and the time sequence influence matrix need to be standardized first, so that the weight values of different data sources are within the same calculation range. In educational data from different regions or different schools, the dimensions of various factors may differ greatly. For example, academic performance may be in percentage, while the level of teaching staff may be represented by the number of teachers. Direct weighting calculation may cause some factors to be overemphasized, while others are ignored. Therefore, standardization processing can ensure that all data are calculated on the same scale, avoiding weight deviation problems in the calculation process.

[0156] After completing the standardization processing, the multi-source characteristic data is weighted calculated according to the standardized factor influence matrix, to obtain factor weighted characteristic data. At the same time, the time sequence characteristic data is weighted calculated according to the standardized time sequence influence matrix, to obtain time sequence weighted characteristic data. The factor weighted characteristic data is mainly used to measure the contribution of different educational factors to the comprehensive evaluation, while the time sequence weighted characteristic data is used to measure the influence weight of different time periods. The combination of the two can ensure that the comprehensive characteristic matrix calculated can not only reflect the static influence of educational factors, but also consider the dynamic effect of time factors.

[0157] In order to enhance the expression ability of the comprehensive characteristic matrix, the factor weighted characteristic data and the time sequence weighted characteristic data are fused by using a nonlinear activation method. For example, a nonlinear activation function such as ReLU or Sigmoid is used, so that the comprehensive characteristic matrix can learn more complex factor interaction relationships and improve the expression ability of the evaluation model for nonlinear relationships. The finally generated comprehensive characteristic matrix can not only more accurately reflect the overall situation of regional basic education, but also provide high-quality input data in subsequent large model training and score calculation.

[0158] In a preferred embodiment of the present application, the comprehensive characteristic matrix is subjected to layer-by-layer feature extraction to obtain deep feature data, comprising:

[0159] The comprehensive characteristic matrix is input into a multi-layer feature extraction network, and a deep feature learning mechanism is used to extract feature representations at different levels;

[0160] The features extracted at each layer are subjected to nonlinear transformation to obtain preliminary feature representations;

[0161] The preliminary feature representations are subjected to dimension reduction processing to remove redundant information, to obtain deep feature data.

[0162] In the embodiment of the present application, in the comprehensive evaluation of basic education, the complexity and high dimensionality of the data make it difficult for simple linear feature extraction methods to meet the high-precision modeling requirements. Therefore, it is necessary to perform layer-by-layer feature extraction on the comprehensive feature matrix to obtain deep feature data, so that the model can automatically learn the high-dimensional features in the data and improve the prediction ability and generalization ability of the model.

[0163] In the feature extraction process, the comprehensive feature matrix is first input into the multi-layer feature extraction network, and a deep feature learning mechanism is used to extract features at different levels. Traditional shallow models can only capture low-level patterns in data, while deep feature learning mechanisms can extract high-level features from data layer by layer. For example, in educational data, high-level features may include "indirect influence of teaching resources on teacher level" or "long-term effect of student feedback on teaching method optimization". Through deep feature extraction, the model can automatically learn these implicit relationships from the data to improve evaluation accuracy.

[0164] In the feature extraction process, further nonlinear transformation is performed on the features extracted at each layer to enhance the expression ability of the data. Since the relationships between educational factors often have complex nonlinear characteristics, such as the influence of teacher training on academic performance, which may have different effects due to differences in region, school type, etc., the use of nonlinear transformation methods can make the model better adapt to these complex relationships, thereby improving the generalization ability of the model.

[0165] In addition, in order to reduce computational complexity and avoid data redundancy, a dimensionality reduction processing method is used to reduce the preliminary feature representation. For example, principal component analysis (PCA) method can be used to remove redundant information and retain the most representative features, so that the final extracted deep feature data can not only maintain information integrity, but also reduce computational burden and improve computational efficiency. The final deep feature data not only accurately represents the overall structure of the education system, but also provides high-quality feature input for subsequent score calculation and decision optimization.

[0166] The construction of the multi-layer feature extraction network: first, a multi-layer feature extraction network suitable for educational data needs to be established, which can be based on deep learning frameworks such as multi-layer perceptron MLP or convolutional neural network CNN. The role of the feature extraction network is to learn different levels of information in educational data in layers, so that the model can not only extract the influence of individual features, but also learn the relevance between different educational factors. For example, when evaluating student academic performance, simply focusing on the performance itself may not be able to obtain a complete education quality assessment, and information such as teacher's teaching level and school's resource allocation needs to be combined. Therefore, deep feature extraction methods are needed to gradually uncover the influence of these implicit factors on education quality.

[0167] The application of the nonlinear transformation: In the traditional linear evaluation system, the contribution of different education factors is usually a fixed weight coefficient, but in actual situations, there may be a nonlinear relationship between these factors, for example, teacher training may have less impact on performance in the short term, but may have more significant effects in the long term. Therefore, in the feature extraction process, nonlinear transformations such as activation functions ReLU, Sigmoid, Tanh, etc. are needed to enhance the expression ability of data, so that the model can adapt to complex education systems. For example, the impact of student feedback on teaching quality may be nonlinear, i.e. teachers with good classroom interaction may significantly improve student performance, while teachers with lower interaction may only have a slight impact, and this nonlinear feature needs to be expressed through nonlinear transformation.

[0168] In a preferred embodiment of the present application, the nonlinear relationship between data in the comprehensive feature matrix is extracted to obtain high-order feature data, including:

[0169] A feature interaction network is constructed for the comprehensive feature matrix to establish cross-connections between different features;

[0170] Multi-layer propagation calculation is performed on the nodes in the feature interaction network to extract the nonlinear influence relationship between the features;

[0171] The nonlinear relationship is screened to remove low-contribution interaction features to obtain high-order feature data.

[0172] In the embodiment of the present application, in the calculation process of regional basic education comprehensive evaluation, in addition to directly extracting deep feature data, it is also necessary to further mine the complex interaction relationship between different education factors. Therefore, a feature interaction network needs to be constructed to extract the nonlinear relationship between data in the comprehensive feature matrix to obtain high-order feature data, so that the model can more accurately express the complex influence relationship of different education factors.

[0173] The construction of the feature interaction network is mainly based on the comprehensive feature matrix, and by establishing cross-connections between different features, the interaction relationship between different education factors can be explicitly modeled. For example, when analyzing academic performance, not only the single factor of teacher level needs to be considered, but also the synergistic effect between factors such as teacher training and school resource allocation. Through the feature interaction network, these complex interaction relationships can be automatically learned in the model, so that the scoring calculation can more comprehensively consider the mutual influence between different factors.

[0174] In the feature interaction network, the nonlinear influence relationship between features is extracted through multi-layer propagation calculation. For example, the graph neural network (GNN) method can be used to enable the model to perform information propagation on the feature graph structure, so as to learn high-order interaction features between different factors. In addition, in order to improve the calculation efficiency and stability of the feature interaction network, a feature screening method can be used to remove low-contribution interaction features, so as to ensure that the calculated high-order feature data can truly reflect the key influence relationship of the education system. The final high-order feature data will be used as the model input to improve the accuracy of score calculation and the scientificity of evaluation.

[0175] The feature interaction network is used to learn the interaction between different education factors, which can be explicitly modeled by establishing cross-connections between features. For example, when analyzing academic performance, not only the single factor of teacher level needs to be considered, but also the synergistic effect between factors such as teacher training and school resource allocation. Through the feature interaction network, these complex interaction relationships can be automatically learned in the model, so that the score calculation can more comprehensively consider the mutual influence between different factors.

[0176] The multi-layer propagation mechanism is established to extract high-order features. In traditional education evaluation models, only first-order feature relationships are considered, such as the direct impact of teachers' education on academic performance. However, in the actual education system, the influence of many factors is transmitted through multi-level relationships. For example, teacher professional development not only affects classroom teaching quality, but also indirectly affects students' learning interest, which is ultimately reflected in academic performance. Therefore, a multi-layer propagation mechanism needs to be established to enable the feature interaction network to learn second-order, third-order, or even higher-order feature relationships. For example, graph neural networks (GNN) or attention mechanisms can be used to enable the model to identify indirect influences between different education factors, improving the model's ability to model the complexity of the education system.

[0177] The low-contribution interaction features are removed. In the feature interaction modeling process, not all feature combinations have important influence on the final evaluation result. For example, the student's residence may have little relationship with the teacher's title, while the teacher's teaching style may have a strong interaction with the student's performance. Therefore, a feature screening method such as information gain, feature importance evaluation, etc. is needed to screen low-contribution interaction features, remove irrelevant features, improve model calculation efficiency, and reduce calculation complexity.

[0178] In a preferred embodiment of the present application, the deep feature data and the high-order feature data are fused to generate model input data, which includes:

[0179] aligning scales of the deep feature data and the high-order feature data;

[0180] performing layer-by-layer fusion of the deep feature data and the high-order feature data by using a multi-scale feature fusion method;

[0181] performing regularization processing on the fused feature data to obtain model input data.

[0182] In the embodiment of the present application, in the comprehensive evaluation process of regional basic education, the complexity and high-dimensionality of the data require the model to fully utilize the complementary information of the deep feature data and the high-order feature data, so as to improve the evaluation accuracy and generalization ability. Since the deep feature data is extracted from the comprehensive feature matrix layer by layer, and the high-order feature data is obtained by mining the complex interaction relationship between different education factors through the feature interaction network, it is necessary to further fuse the two types of feature data to ensure the completeness and stability of the model in the representation ability.

[0183] In the feature fusion process, first, the scales of the deep feature data and the high-order feature data need to be aligned. Since the two types of data are from different sources, their data distribution, dimension and information density may have large differences, and direct feature fusion may cause one type of feature to have too large or too small contribution to the final score. Therefore, a scale alignment method is used to standardize or normalize different types of feature data, so that they are calculated in the same scale range, improving the reliability of feature fusion. For example, in actual application, the deep feature data may include high-dimensional features extracted by a deep learning model, and the high-order feature data may include an interaction relationship matrix calculated based on an attention mechanism, the numerical ranges of the two types of data are different, and normalization processing is needed to ensure the numerical consistency during fusion.

[0184] After the scale alignment is completed, a multi-scale feature fusion method is used to perform layer-by-layer fusion of the deep feature data and the high-order feature data. The core idea of the multi-scale feature fusion method is that in different scale feature spaces, different levels of weighted calculation are used to make the deep feature data and the high-order feature data complement each other and constrain each other, improving the generalization ability of the model. For example, a weighted feature splicing method can be used to adjust the contribution of different feature data by using dynamic weights, so that the fused feature data contains both deep information and high-order interaction information, so that the model can learn a more comprehensive education factor influence pattern.

[0185] To improve the model's stability and generalization ability, regularization is applied to the fused feature data after feature fusion. Since feature fusion may introduce noisy or redundant information, regularization can constrain the data distribution, reducing the risk of overfitting and improving the model's adaptability to different datasets. For example, L2 regularization can be used to constrain the fused features, ensuring a more uniform distribution of the final feature data and improving model stability.

[0186] Ultimately, after scale alignment, multi-scale feature fusion, and regularization, the resulting model input data not only more comprehensively reflects the overall situation of basic education in the region, but also enhances the model's adaptability to different educational factors and improves the accuracy and reliability of score calculation.

[0187] The multi-scale feature fusion method employed here aims to ensure the effective combination of features at different scales during data fusion, thereby fully leveraging the strengths of each feature. The core idea of ​​this method is to use different weighting strategies in feature spaces at different scales, allowing deep and high-order feature data to complement and constrain each other. For example, feature concatenation, weighted summation, or attention mechanisms can be used to combine different feature data, ensuring that the final model input data covers the most important educational influencing factors.

[0188] In a preferred embodiment of the present invention, the step of inputting model input data into the trained regional basic education comprehensive evaluation model, performing evaluation calculations, and obtaining initial evaluation results includes:

[0189] Feature mapping is performed on the model input data to extract feature representation data for comprehensive evaluation;

[0190] Pre-trained weight parameters are used to calculate the feature representation data and generate a multi-dimensional base score.

[0191] Based on the scoring deviation of historical evaluation data, the basic scores of multiple dimensions are corrected to obtain normalized scoring data;

[0192] A weighted fusion method is used to calculate the overall score, generating initial evaluation results; among which,

[0193] ,in, For the first The initial evaluation results of each educational unit, For the first The dynamic weights of each evaluation factor For the first The first educational unit in the The base score for each evaluation factor is an adaptive scoring correction coefficient, is the score deviation of the first evaluation factor, is the score deviation of the first evaluation factor, is the total number of evaluation factors; wherein,

[0194] , is the original value of the first educational unit on the first evaluation factor, is the highest value of all educational units on the first evaluation factor, is the lowest value of all educational units on the first evaluation factor; , is the number of historical data samples, is the average score of the first evaluation factor in the historical evaluation data.

[0195] In the embodiment of the present application, in the comprehensive evaluation calculation of regional basic education, the model input data needs to be input into the trained evaluation large model after feature extraction and feature fusion to perform the final scoring calculation. The traditional scoring calculation method usually relies on the rules set by experts, lacks the dynamic adjustment ability of data-driven, and leads to the deviation of the scoring result from the actual education quality. The present method can automatically learn the influence relationship of different education factors by constructing a comprehensive evaluation large model based on data-driven, and improve the accuracy and rationality of scoring calculation. In the scoring calculation process, the model input data needs to be first mapped to feature representation data for comprehensive evaluation. Since the model input data has been processed by feature extraction and feature fusion, it already contains deep feature data and high-order feature data. However, in the final scoring calculation, these data need to be further converted to adapt to the mathematical model of scoring calculation. For example, a multi-layer perceptron (MLP) structure can be used to map the input data to a low-dimensional feature space to reduce the computational complexity and improve the stability of scoring calculation.

[0196]

[0197]

[0198] ​​​​​After feature mapping is completed, pre-trained weight parameters are used to calculate the feature representation data, generating multi-dimensional basic scores. These basic scores correspond to different educational evaluation factors, such as academic performance, teacher quality, school resources, and teaching quality. The calculation method for each scoring dimension is based on a data-driven weight adjustment mechanism to ensure the scientific and reasonable nature of the scoring results. For example, academic performance scores can be calculated based on a weighted average of exam scores, teacher quality scores can be calculated based on a comprehensive consideration of factors such as teachers' qualifications, professional titles, and teaching experience, while teaching quality scores can be calculated by combining student feedback and teacher evaluations. This approach ensures that the scoring calculation comprehensively covers the key influencing factors in the education system.

[0199] After generating the base scores, to improve their adaptability, the base scores across multiple dimensions are revised based on the scoring deviation from historical evaluation data. The scoring deviation is calculated using the mean squared error (MSE) of historical data to measure the difference between the current and historical scores, and is adjusted through a scoring correction mechanism. For example, if a school's academic performance scores over the past three years have significantly declined compared to the current scores, the scoring correction mechanism can dynamically adjust the school's scoring weights, making the final scores more consistent with long-term trends and improving the stability and interpretability of the scores.

[0200] Finally, after revising the scores, a weighted fusion method is used to calculate the overall score, generating the initial evaluation result. This score not only reflects the overall situation of the current education system but also allows for dynamic adjustments based on historical data, improving the rationality and reliability of the score and providing education administrators with more accurate decision-making support.

[0201] In a preferred embodiment of the present invention, the step of adaptively adjusting the weights of the initial evaluation results based on historical evaluation data to generate the final evaluation results includes:

[0202] Based on the scoring deviation of historical evaluation data, the exponential decay correction method is used to dynamically adjust the scoring weights and obtain the adjustment time-series impact weights.

[0203] The initial evaluation results are adjusted based on the weighting of the time-series influence to obtain the final evaluation results; among which...

[0204] , For the final evaluation result, For the first The weight of the adjusted time series effect of each evaluation factor. This is the initial evaluation result; among which,

[0205] ,in, For the first a timing influence weight of an evaluation factor, is a learning rate, is an adjustment stability coefficient.

[0206] In the embodiment of the present application, in the comprehensive evaluation of regional basic education, after the score calculation is completed, further adaptive weight adjustment is needed to ensure the stability of the score result and reduce the influence of short-term data fluctuation on the final evaluation. Since the education system is a long-term evolving system, short-term fluctuations in scores may occur due to policy adjustments, changes in test difficulty and other factors, so an adaptive weight adjustment mechanism based on historical evaluation data is needed to make the final evaluation result more consistent with long-term trends and improve the robustness and interpretability of the score.

[0207] In the weight adjustment process, first, based on the score deviation of historical evaluation data, an exponential decay correction method is used to dynamically adjust the score weight to obtain the adjusted timing influence weight. The core idea of the exponential decay correction method is that when the score deviation is large, the weight adjustment amplitude is large to quickly converge to a reasonable range, and when the score deviation is small, the weight adjustment amplitude is small to maintain the stability of the score. For example, if the education evaluation result of a certain region fluctuates sharply in a short period of time, the weight adjustment mechanism will automatically reduce the score fluctuation amplitude, so that the final score will not change sharply due to short-term data anomalies.

[0208] After the weight adjustment is completed, the initial evaluation result is adjusted according to the adjusted timing influence weight to obtain the final evaluation result. In this way, the final score not only accurately reflects the true situation of the current regional education system, but also considers the long-term trend to avoid the interference of short-term data fluctuation on the final evaluation, making the evaluation system more stable and reliable.

[0209] The above is the preferred embodiment of the present application. It should be noted that for ordinary skilled persons in the technical field, without departing from the principles of the present application, several improvements and refinements can be made, which should also be considered as the protection scope of the present application.

Claims

1. A method for constructing a regional basic education comprehensive evaluation large model, characterized in that, The method comprises: Obtain multi-source data related to regional basic education to form an original data set; Data cleaning, missing value filling and outlier detection are performed on the original data set to obtain a preprocessed data set; Different data types in the preprocessed data set are converted to feature representations to obtain multi-source feature data; According to the multi-source feature data, the weight relationship between different education factors is calculated to generate a factor influence matrix; According to the multi-source feature data, the contribution degree of time series data is calculated to generate a time series influence matrix; According to the factor influence matrix and the time series influence matrix, all data are weighted and fused to generate a comprehensive feature matrix; According to the comprehensive feature matrix, a regional basic education comprehensive evaluation large model is constructed, specifically including: Performing layer-by-layer feature extraction on the comprehensive feature matrix to obtain deep feature data; Extracting the nonlinear relationship between data in the comprehensive feature matrix to obtain high-order feature data; Fusing the deep feature data and the high-order feature data to generate model input data; Inputting the model input data into the trained regional basic education comprehensive evaluation large model to perform evaluation calculation to obtain an initial evaluation result; Based on historical evaluation data, the initial evaluation result is adaptively weighted and adjusted to generate a final evaluation result; According to the multi-source feature data, the weight relationship between different education factors is calculated to generate a factor influence matrix, including: Using a correlation analysis method based on an attention mechanism, the weight of the multi-source feature data is calculated to obtain the basic correlation distribution between the education factors; According to the fluctuation trend of the time series data, a time sensitivity coefficient is set for each education factor, and the factors affected by time are filtered through the time sensitivity coefficient to adjust the correlation distribution and obtain the correlation adjustment distribution; The correlation adjustment distribution is normalized to generate the factor influence matrix; wherein, , Educational factors in the factor influence matrix Educational factors Influence weight, For educational factors Educational factors The basic correlation score between them For educational factors Educational factors The adjustment coefficient, For educational factors The intensity of the fluctuation trend over time, The time sensitivity coefficient, For time-sensitive switch parameters, if educational factors... Due to the influence of time, If educational factors If unaffected by time, then , For the total number of educational factors, For educational factors Educational factors The basic correlation score between them For educational factors Educational factors The adjustment coefficient; According to the multi-source feature data, the contribution degree of time series data is calculated to generate a time series influence matrix, including: Time window division is performed on the time series feature data to extract short-term trend features and long-term trend features; The dynamic time warping method is used to calculate the data matching degree between different time periods, and the influence of abnormal values is reduced through an adaptive smoothing factor; The comprehensive influence weight of each time window is calculated by combining the short-term influence weight and the long-term trend influence weight, and is normalized to generate the time series influence matrix; wherein, wherein, is a timing influence weight, is a normalization coefficient, is a short-term influence weight, is a long-term trend influence weight, is an educational factor is a rate of change over a time window ; wherein, wherein, is the education factor is the education factor is the dynamic alignment distance in the time dimension, and is the value of the education factor is the value of the education factor is the value of the education factor is the value of the education factor is the smoothing coefficient; The model input data is input into the trained regional basic education comprehensive evaluation large model to perform evaluation calculation to obtain an initial evaluation result, including: Feature mapping is performed on the model input data to extract feature representation data for comprehensive evaluation; Using pre-trained weight parameters, the feature representation data is calculated to generate a plurality of dimension basis scores; Based on the score deviation of the historical evaluation data, the plurality of dimension basis scores are corrected to obtain normalized score data; The overall score is calculated by using a weighted fusion method to generate an initial evaluation result; wherein, wherein, is the initial evaluation result of the th educational unit, is the dynamic weight of the th evaluation factor, is the basic score of the th educational unit on the th evaluation factor, is the adaptive score correction coefficient, is the score deviation of the th evaluation factor, is the total number of evaluation factors; Based on the historical evaluation data, the initial evaluation result is adaptively weighted and adjusted to generate a final evaluation result, including: Based on the score deviation of the historical evaluation data, an exponential decay correction method is used to dynamically adjust the score weight to obtain an adjusted time series influence weight; According to the adjustment timing influence weight, the initial evaluation result is adjusted to obtain the final evaluation result; wherein , is the final evaluation result, is the correction timing influence weight of the first evaluation factor, is the correction timing influence weight of the first evaluation factor, is the initial evaluation result; wherein, wherein, is the time series influence weight for the th evaluation factor, is the learning rate, is the adjustment stability coefficient.

2. The method of claim 1, wherein the method further comprises: The feature representation conversion of different data types in the preprocessed data set is performed to obtain multi-source feature data, including: The numerical data in the preprocessed data set is normalized to obtain normalized data; The category type data in the preprocessed data set is constructed to build a category mapping matrix, and the category features are converted into calculable numerical features to obtain category feature data; The time window of the time series data in the preprocessed data set is divided, the short-term trend features and long-term trend features are extracted, and the time dependence relationship is constructed to obtain the time series feature data; The text data in the preprocessed data set is segmented, and the text data is mapped into a numerical feature vector to obtain text feature data.

3. The method of claim 1, wherein the method further comprises: The factor influence matrix and the timing influence matrix are standardized to make the weight values in the same calculation range; According to the factor influence matrix after standardization, the multi-source feature data is weighted and calculated to obtain the factor weighted feature data; According to the timing influence matrix after standardization, the timing feature data is weighted and calculated to obtain the timing weighted feature data; The factor weighted feature data and the timing weighted feature data are fused by using a nonlinear activation method to generate a final comprehensive feature matrix. The comprehensive feature matrix is extracted layer by layer to obtain deep feature data, including:

4. The method of claim 1, wherein the method further comprises: The comprehensive feature matrix is input into a multi-layer feature extraction network, and a deep feature learning mechanism is used to extract feature representations at different levels; The features extracted by each layer are nonlinearly transformed to obtain preliminary feature representations; The preliminary feature representations are processed by dimension reduction to remove redundant information to obtain deep feature data. The nonlinear relationship between the data in the comprehensive feature matrix is extracted to obtain high-order feature data, including:

5. The method of claim 1, wherein the method further comprises: A feature interaction network is constructed for the comprehensive feature matrix to establish cross connections between different features; The nodes in the feature interaction network are calculated by multi-layer propagation to extract the nonlinear influence relationship between the features; The nonlinear relationship is screened to remove low-contribution-degree interaction features to obtain high-order feature data. The deep feature data and the high-order feature data are fused to generate model input data, including:

6. The method of claim 1, wherein the method further comprises: The deep feature data and the high-order feature data are scaled and aligned; A multi-scale feature fusion method is used to layer-by-layer fuse the deep feature data and the high-order feature data; The fused feature data is regularized to obtain model input data. ​

Images