An education state evaluation and intervention method based on multi-source data fusion

Abstract

The application provides an education state evaluation and intervention method based on multi-source data fusion, and belongs to the technical field of education management based on artificial intelligence; academic depth, industrial value and popular science propaganda data of a target object are collected, preprocessed to form a standardized text set, subjected to text vectorization, semantic analysis and vector aggregation to obtain a final feature set, the feature set is input into an index reconstruction network, subjected to feature re-calibration mapping, and a core index vector covering theoretical originality, method breakthrough and interdisciplinary is output; a high-dimensional index embedding vector is generated through a feature expansion module, subjected to parallel sub-network coupling processing and full-connection normalization to obtain a state evaluation matrix; matrix diagnosis features are extracted, compared with a standard education track to calculate an Euclidean distance deviation, combined with a remedial coefficient operation, and corresponding intervention strategies are output through a strategy determination function. The application improves the comprehensiveness and objectivity of the evaluation system, and constructs an evaluation and intervention integrated closed-loop system.

Description

Technical Field

[0001] The present invention belongs to the technical field of education management based on artificial intelligence, and particularly relates to an education status evaluation and intervention method based on multi-source data fusion. Background Art

[0002] Under the background of the current knowledge economy and innovation-driven development strategy, high-level education has gradually shifted from the traditional experience-driven mode to a data-driven and intelligent decision-making mode. Whether in the university scientific research system or in the industry-university-research collaborative innovation platform, how to scientifically and objectively evaluate the education status and implement precise intervention based on this has become one of the key issues restricting the improvement of education quality. Therefore, constructing an education status evaluation and intervention method that can integrate multi-source information, achieve dynamic evaluation and intelligent feedback has important theoretical value and practical significance. Most of the existing education evaluation systems rely on single or limited-dimensional data sources, such as the number of published papers, the scale of project funds, or the situation of achievement transformation. Such evaluation methods have obvious limitations: on the one hand, their data structure is relatively single, making it difficult to comprehensively reflect the development status of talents in multiple dimensions such as academic depth, technological innovation, and social impact; on the other hand, traditional evaluation methods are mostly based on static index statistics, lacking the ability to dynamically track and continuously describe the cultivation process, and it is difficult to support the whole-process and refined education process management. In addition, the existing methods usually rely on manual experience to set evaluation rules, with strong subjectivity and lack of a unified quantitative standard, resulting in insufficient objectivity and comparability of evaluation results.

[0003] With the rapid development of big data technology and natural language processing technology, the value of unstructured data such as academic papers, technical reports, patent texts, and popular science communication content has become increasingly prominent. These data contain rich talent ability characteristic information, such as theoretical innovation level, method breakthrough ability, and interdisciplinary integration degree. However, how to uniformly model and effectively integrate multi-source heterogeneous unstructured data is still a major difficulty in the current technology. When dealing with such data, existing technologies often only perform shallow feature extraction, making it difficult to mine deep semantic information, resulting in limited expression ability of evaluation indicators. At the same time, in terms of evaluation models, traditional methods mostly use linear weighting or simple machine learning models, making it difficult to describe complex high-dimensional feature relationships and unable to achieve high-precision modeling of the talent education status. In terms of intervention strategies, the existing talent cultivation management mostly relies on empirical judgment and lacks a data-driven personalized decision-making mechanism. There are significant differences in the development paths, ability structures, and stage characteristics of different talents, but traditional methods are difficult to dynamically generate targeted cultivation strategies according to individual differences, resulting in low resource allocation efficiency and difficult to optimize the cultivation effect. Summary of the Invention

[0004] To address the above problems, this invention provides a method for assessing and intervening in educational status based on multi-source data fusion, comprising the following steps: S1: Acquire academic depth data, industry value data, and science popularization data of the target object, and preprocess them to obtain a standardized text set; vectorize the standardized text set, and call a natural language processing model to map the unstructured text data into a semantic feature set; after semantic vector aggregation, obtain the final output feature set. ; S2, Input metrics to reconstruct the network, perform feature recalibration and mapping to obtain the core metric vector. These include indicators of theoretical originality, methodological breakthroughs, and interdisciplinary collaboration. S3, will The input feature extension mapping module generates high-dimensional index embedding vectors. ,Will After inputting three parallel sub-network branches, coupled feature representations are generated through feature mapping. ;Will By performing a fully connected mapping and normalization, the normalized state evaluation matrix is ​​obtained. ; S4, from Diagnostic features are extracted and spatially mapped with preset standard educational trajectory features. The deviation is obtained by calculating the Euclidean distance between the two. After calculating the remedial coefficient of the deviation, it is input into the strategy determination function to obtain the intervention strategy instruction.

[0005] Preferably, in step S1, academic depth data is acquired through an academic database interface, and the collected content includes text information such as paper abstracts, keywords, and research direction descriptions; industrial value data is acquired through a patent database interface and a technology transfer system, and the collected content includes texts such as patent specifications, technical solution descriptions, and application scenario descriptions; and popular science data is acquired through an open interface of a media platform, and the collected content includes news reports, popular science articles, and public communication content related to the target audience.

[0006] Preferably, the natural language processing model adopts a natural language processing model based on the Transformer structure; for each vector matrix Perform a linear mapping transformation to generate a query vector, a key vector, and a value vector: ; They represent respectively by The query matrix, key matrix, and value matrix obtained by linear transformation; Represented as a trainable weight matrix; Subsequently, based on Calculate the attention weight matrix according to the standard formula for attention. The value vectors are then weighted and summed to obtain the fused weighted intermediate semantic representation. ; A multi-head attention mechanism is used to concatenate weighted intermediate semantic representations from multiple attention heads in parallel. , obtain the output result Then through a feedforward neural network The function is subjected to a nonlinear mapping to obtain the first... i semantic feature matrix of the text Ultimately, a set of semantic features is obtained. .

[0007] Preferably, the S2 process specifically includes: The final output feature set The input feature recalibration layer performs a weighted transformation on each dimension of the features through a single-layer fully connected network to generate a recalibrated feature matrix. Next, the feature matrix will be recalibrated. The input similarity modeling layer generates a similarity matrix by calculating the cosine similarity between samples. The similarity weight matrix was obtained after softmax normalization. This characterizes the correlation strength between different samples in the feature space; subsequently, the similarity weight matrix is... With recalibrated feature matrix Perform matrix multiplication to generate a fused feature matrix. Next, the fused feature matrix... Perform global average pooling to obtain the global fused feature vector. Finally, the globally fused feature vector is... The input indicator mapping layer generates core indicator vectors through fully connected mapping. , ; Indicators of theoretical originality measure the ability of a target object to contribute to fundamental theories. Breakthrough indicators are used to measure the degree of innovation in technical methods or implementation paths. Indicators representing interdisciplinary characteristics are used to measure the ability to integrate and apply knowledge across disciplines. The values ​​are all within the range of [0,1]. The larger the index value, the stronger the corresponding capability dimension.

[0008] Preferably, in step S3, there are three parallel sub-network branches, each consisting of two cascaded fully connected layers. This enhances feature representation through two-level mapping, and residual connections are set within the fully connected layers of each branch. The three branches correspond to the nonlinear expansion paths of three core indicators. Simultaneously, the features from each branch are concatenated to obtain... Then, a gating coupling mechanism is introduced to perform weighted fusion of the spliced ​​features: First, Gating vectors are generated through gating mapping. The sigmoid activation function is used to constrain the gate value to the interval [0,1]. Subsequently, based on the gate vector... right Weighting is performed to obtain the gated modulation features. Finally, a linear mapping is performed on the modulated features to obtain the coupled feature representation. .

[0009] Preferably, the S4 process specifically includes: S41, Educational Trajectory Feature Extraction and Standard Educational Trajectory Mapping: Applying Normalized State Evaluation Matrix Local neighborhood feature extraction is performed to generate the current educational trajectory structure feature matrix. Introducing a standard educational trajectory matrix Generate a standard educational trajectory structure feature matrix ; S42, Spatial Deviation Calculation: For and Perform difference operations to generate state deviation distribution And calculate the overall deviation. At the same time, combined with the state vector With standard state vector Euclidean distance generates positional deviation Further fusion yields the joint deviation. ; S43, Remedial Coefficient Calculation and Intervention Strategy Generation: Input the Joint Deviation The remedial coefficient generation module obtains the global remedial coefficient. And generate a local remedial distribution matrix. ;right Divide the region into multiple sub-regions Calculate the intervention intensity in each region and the intensity of intervention The modified strategy strength coefficient is obtained by making corrections. Finally, the strategy strength coefficient will be adjusted. The input strategy determination function performs piecewise mapping to generate corresponding discrete intervention strategies. Ultimately, personalized intervention strategy instructions are constructed and output. .

[0010] Preferably, the specific process of S41 includes: Use size The convolution kernel in The algorithm performs sliding calculations, weighted summations on each local region, and extracts the state change trends and distribution patterns within that local region to obtain the current educational trajectory structure feature matrix of the current object. ; The set of state matrices of historical high-quality objects in stage t is: Then the standard education trajectory matrix ;in, Indicates the first l The state matrix of a high-quality object at stage t, where L represents the number of samples; and the standard educational trajectory matrix. Perform two-dimensional convolution operations to obtain the standard educational trajectory structure feature matrix. .

[0011] Preferably, the specific process of S42 includes: Calculate the state deviation distribution Then, calculate the overall deviation. : ; Where p represents the number of rows in the current educational trajectory structure feature matrix, and q represents the number of columns in the diagnostic feature matrix; The larger the value, the greater the degree of deviation; For the state vector With standard state vector Perform Euclidean distance calculation ; where, state vector From the normalized state evaluation matrix Expand to obtain; For the standard education trajectory matrix Through state assessment matrix The standard state vector is obtained through the same vectorization process; Indicates the positional deviation; Constructing joint deviation ;in and These are the weighting coefficients.

[0012] Preferably, the specific process of S43 includes: Based on joint deviation Generate global remediation coefficient , The adjustment parameter; the remedial coefficient is used to control the overall strength of the intervention strategy; Generate a local salvage distribution matrix: ; Describe the intensity of intervention demand at different locations; The region is divided into regions using a regular grid method to distribute the local remedial distribution matrix. Divide into K sub-regions according to a fixed size; Calculate the arithmetic mean of all element values ​​within each region sequentially, and use this arithmetic mean as the intervention intensity for the corresponding region. The intervention intensity for each region is denoted as ; The direction of the intervention intensity is adjusted for each region; for the first region... Intervention intensity in each region Calculate the state vector With standard state vector The difference vector is calculated, and its projection value on the corresponding policy dimension is extracted. The projection value is then compared with a preset adjustment coefficient. Multiplying them gives the direction correction amount, which is then used to correct the direction. Adding the directional correction amount to the formula yields the correction strategy strength coefficient. ; Construct a policy decision function to map continuous intensity to discrete policy categories: ; in, and The threshold is set based on the statistical distribution of historical data and is used to divide the intervention intensity into three levels: low, medium, and high. , and These represent different types of intervention strategies.

[0013] Preferred, As a mild intervention strategy, a suggestive adjustment approach is adopted, providing the target subjects with a list of minor optimization suggestions, including a list of recommended readings, pointing out the shortcomings of the current research, and requiring minor improvements based on the original research, without changing the overall research path; As a moderate intervention strategy, structural optimization measures are adopted to partially restructure the current research protocol, including adjusting research methods, redesigning experimental procedures, or adding key validation steps. Improvements are required to be completed within a specified period and a progress report must be submitted. To strengthen intervention strategies, key resources are allocated and pathways are restructured. Overall adjustments are made to research directions or technical routes, including redefining research questions, introducing new methodologies, or changing research pathways. Additional resource support is provided, including expert guidance and specialized training tasks.

[0014] The beneficial effects of this invention include: 1) Multi-source heterogeneous data fusion modeling mechanism. This invention breaks through the traditional single indicator system and for the first time unifies the modeling of academic depth data, industry value data, and popular science data to achieve collaborative fusion and expression of multi-source unstructured data.

[0015] 2) Deep semantic vectorization method based on natural language processing. A pre-trained natural language processing model is introduced to perform high-dimensional semantic encoding on unstructured text, significantly enhancing feature representation capabilities compared to traditional shallow feature extraction methods.

[0016] 3) Attention-Driven Indicator Reconstruction Network. An indicator reconstruction network with a feature attention layer is constructed. Through an adaptive weight allocation mechanism, key features are dynamically extracted to accurately characterize the originality of the theory, the breakthrough of the method, and the interdisciplinary nature.

[0017] 4) Intelligent evaluation model based on high-dimensional space mapping. A nonlinear modeling method based on deep neural networks is used to map the educational status to an evaluation matrix in a high-dimensional space, thereby improving the expressiveness and discriminative power of the evaluation results.

[0018] 5) Generate targeted educational strategies based on individual differences to improve the adaptability and effectiveness of intervention measures. Achieve a complete process from data collection and status assessment to strategy output, effectively improving the intelligence level and decision-making efficiency of talent education management. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the following description is only one embodiment of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart illustrating the overall technical process of the present invention.

[0021] Figure 2 A flowchart for reconstructing and fusing features for core indicators.

[0022] Figure 3 A flowchart of the overall process for intelligent assessment of educational status.

[0023] Figure 4 The diagram shows the feature reconstruction and similarity weight matrix in the example.

[0024] Figure 5 This is a distribution map of the regional levels of the intervention strategy in the embodiment.

[0025] Figure 6 This is a bar chart comparing the evaluation and intervention effects of indicators at different educational stages in the implementation example. Detailed Implementation

[0026] This invention provides a method for assessing and intervening in educational status based on multi-source data fusion. The overall process is as follows: Figure 1 As shown: S1, Multi-source data acquisition and feature vectorization processing. First, acquire academic depth data, industry value data, and science popularization data of the target object. After correlation and filtering processing, obtain the raw text data set. Secondly, regarding the original text data set... Preprocessing and standardization are performed to obtain a standardized text set. And then vectorize it to obtain a set of text vector representations. Finally, a natural language processing model is invoked to map the aforementioned unstructured text data into a set of semantic features. The vectors from different data sources are fused to obtain the final output feature set. .

[0027] S2, Core Indicator Reconstruction and Feature Fusion Extraction. The final output feature set... The input feature recalibration layer performs a weighted transformation on each dimension of the features through a single-layer fully connected network to generate a recalibrated feature matrix. Next, the feature matrix will be recalibrated. The input similarity modeling layer generates a similarity matrix by calculating the cosine similarity between samples. The similarity weight matrix was obtained after softmax normalization. This characterizes the correlation strength between different samples in the feature space; subsequently, the similarity weight matrix is... With recalibrated feature matrix Perform matrix multiplication to generate a fused feature matrix. Next, the fused feature matrix... Perform global average pooling to obtain the global fused feature vector. Finally, the globally fused feature vector is... The input indicator mapping layer generates core indicator vectors through fully connected mapping. It is characterized by three core dimensions: theoretical originality, methodological breakthrough, and interdisciplinary nature.

[0028] S3, intelligent assessment and calculation of educational status. Input core indicator vector. The input feature extension mapping module generates a high-dimensional index embedding vector. Next, the high-dimensional index is embedded into the vector. The inputs are fed into three parallel sub-network branches, and then a gating coupling mechanism is introduced to generate gating modulation features. Then, a coupled feature representation is generated through linear mapping and nonlinear transformation. This enables dynamic interaction and fusion among multiple indicators; next, the coupled feature representation is... Further, through fully connected mapping and normalization, the normalized state evaluation matrix is ​​obtained. And then vectorize and expand it to generate a state vector. The final output is the normalized state evaluation matrix. and the corresponding state vector .

[0029] S4, Dynamic Generation and Output of Intervention Strategies. First, the normalized state evaluation matrix... Local neighborhood feature extraction is performed to generate the current educational trajectory structure feature matrix. Introducing a standard educational trajectory matrix Generate a standard educational trajectory structure feature matrix Subsequently, regarding and Perform difference operations to generate state deviation distribution And calculate the overall deviation. At the same time, combined with the state vector With standard state vector Euclidean distance generates positional deviation Further fusion yields the joint deviation. Next, input the joint deviation. The remedial coefficient generation module obtains the global remedial coefficient. And generate a local remedial distribution matrix. ;right Divide the region into multiple sub-regions Calculate the intervention intensity in each region and the intensity of intervention The modified strategy strength coefficient is obtained by making corrections. Finally, the strategy strength coefficient will be adjusted. The input strategy determination function performs piecewise mapping to generate corresponding discrete intervention strategies. Ultimately, personalized intervention strategy instructions are constructed and output. .

[0030] The specific implementation process of the present invention will be further described below with reference to specific embodiments.

[0031] S1. Multi-source data acquisition and feature vectorization processing S1-1 Determine the identification information of the target object to be evaluated: Target object identification information Defined as a set of structured information used to uniquely identify a target object, including the target object's name, the name of its affiliated institution, and a set of keywords related to its research field; S1-2 Multi-source Text Data Acquisition and Input Construction. Multi-source data is collected in different application scenarios, including in-depth academic data, industry-value data, and science popularization data.

[0032] Specifically, automated data acquisition is achieved by configuring web crawler interfaces. Academic depth data is preferably acquired through academic database interfaces, collecting textual information including paper abstracts, keywords, and research direction descriptions. Industry value data is preferably acquired through patent database interfaces and technology transfer systems, collecting textual information including patent specifications, technical solution descriptions, and application scenario descriptions. Popular science data is collected through open interfaces of media platforms, acquiring news reports, popular science articles, and public communication content relevant to the target audience.

[0033] Obtain a candidate text data set from the above data sources, denoted as . ,in, Indicates the first number obtained from the initial collection. Candidate text data, This represents the total amount of candidate data.

[0034] Then, from the set Select valid text data that is relevant to the target object. Specifically, for each candidate text... Entity recognition processing is performed to extract personal names, organization names, and keywords related to research fields, thereby obtaining text entity information. Then calculate the text entity information. Identification information of the target object The degree of matching between them yields the entity matching degree. The entity matching degree is specifically calculated using cosine similarity, denoted as: .

[0035] Next, set the preset filtering threshold. ,when When the text is determined to be relevant to the target object, the corresponding candidate text data is added to the original text data set. .

[0036] Therefore, the original text data set . This indicates the number of items retained after filtering. Each piece of valid text data; All of them are complete semantic text units directly related to the target object. This indicates the number of valid texts after filtering, satisfying... .

[0037] S1-3 Text Preprocessing and Standardization. This involves processing the original text data set. Preprocessing is performed to eliminate differences in format, length, and noise between texts from different sources.

[0038] Specifically, for each text Perform preprocessing operations: First, for The continuous text is divided into discrete word units through word segmentation. Then, stop words, punctuation marks, and characters that do not contribute semantically are removed, and abnormal characters are cleaned to obtain the preprocessed text. Furthermore, for the preprocessed text Length standardization is performed to map all text to a fixed length. The processing rules are as follows: when the text length is less than... Padding is performed when the text length is greater than 10 ... Truncation is performed at specific times to obtain standardized text. .

[0039] Thus, a standardized text set is obtained. .

[0040] S1-4 Word vector embedding representation construction based on the standardized text set obtained in S1-2. For sets The text in the document is vectorized.

[0041] Specifically, through word embedding functions Mapping each word unit in the standardized text yields the vector representation of the j-th word in the i-th text. Therefore, the i-th standardized text It can be mapped to a vector matrix. .

[0042] Through the above processing, we obtain the set of text vector representations. .

[0043] S1-5 uses Transformer-based semantic feature encoding. Deep semantic modeling is employed for each vector matrix. Semantic feature extraction is performed to obtain a deep semantic representation containing contextual information. A natural language processing model based on the Transformer architecture is preferred.

[0044] Specifically, for each vector matrix Perform a linear mapping transformation to generate a query vector, a key vector, and a value vector: . They represent respectively by The query matrix, key matrix, and value matrix obtained by linear transformation. It is represented as a trainable weight matrix.

[0045] Subsequently, based on Calculate the attention weight matrix according to the standard formula for attention. The value vectors are then weighted and summed to obtain the fused weighted intermediate semantic representation. .

[0046] Furthermore, a multi-head attention mechanism is employed to concatenate weighted intermediate semantic representations from multiple attention heads in parallel. , obtain the output result Then through a feedforward neural network The function performs a non-linear mapping to obtain the semantic feature matrix of the i-th text. Ultimately, a set of semantic features is obtained. This representation can effectively capture the potential semantic information and contextual relationships in the text.

[0047] S1-6 Semantic Vector Aggregation and Feature Set Construction. To transform the variable-length semantic representation into a fixed-length vector, the semantic feature set obtained in S1-4 is... Each semantic feature matrix Aggregation processing is performed. Average pooling is preferred to generate fixed-dimensional text representation vectors. Each text line is mapped to a single vector representation. All text vectors construct the final output feature set. .

[0048] Furthermore, for each vector Perform normalization processing This yields the final output feature set after normalization. .

[0049] S2. Core Indicator Reconstruction and Feature Extraction The overall process is as follows Figure 2 As shown: S2-1 will use the final output feature set obtained in S1-5. Subsequently, to enhance the contribution of different feature dimensions to the indicator construction, a feature recalibration layer is introduced. This layer uses a single-layer fully connected neural network structure to perform weighted transformation on the input features: .in, For a trainable weight matrix, The recalibrated feature matrix, It is the ReLU nonlinear activation function.

[0050] S2-2 Obtaining the recalibrated characteristic matrix Subsequently, a similarity modeling layer based on the feature space structure is constructed to characterize the correlation between features of different samples. This similarity modeling layer is a parameter-free computation layer, which is implemented through geometric relationships between vectors.

[0051] Specifically, for matrix Any two row vectors in and Calculate cosine similarity Thus, the similarity matrix is ​​obtained. ; Then, for the matrix Normalization is performed. The weight normalization layer is a softmax normalization layer, which converts the similarity values ​​into weight coefficients in the form of a probability distribution. , making Thus, the similarity weight matrix is ​​obtained. This matrix is ​​used to describe the contribution ratio of different samples in the overall index calculation.

[0052] S2-3 Obtaining the similarity weight matrix Then, feature fusion is achieved through matrix multiplication: . This is the fused feature matrix.

[0053] Furthermore, in the indicator mapping layer, in order to map the fused features into interpretable indicator results, global average pooling is used to map the fused matrix. Perform global aggregation processing to obtain a global fused feature vector. .

[0054] Then, the global fused feature vectors are obtained through fully connected mapping. Mapping to a low-dimensional index space: , For the index mapping matrix, This is the core indicator vector. Indicators of theoretical originality measure the ability of a target object to contribute to fundamental theories. Breakthrough indicators are used to measure the degree of innovation in technical methods or implementation paths. This indicates the interdisciplinary nature of the subject, used to measure its ability to integrate and apply across disciplines.

[0055] in, The values ​​range from [0,1]. A larger value indicates stronger performance in the corresponding capability dimension; a value closer to 0 indicates relatively weaker capability in that dimension. Specifically: when... A larger value indicates that the target object has a high level of contribution in theoretical innovation and basic research; when... A larger value indicates a stronger ability in methodological innovation or technological breakthroughs; when... A larger value indicates that it performs well in the interdisciplinary integration and comprehensive application of multiple disciplines.

[0056] Through the aforementioned range constraints and semantic definitions, each indicator possesses a unified dimension and a clear magnitude pattern, thus facilitating quantitative calculations for subsequent state assessments and intervention decisions. Ultimately, a core indicator vector is output. This serves as the input for subsequent assessments of the educational status.

[0057] S3, Intelligent Assessment Calculation of Educational Status The overall process is as follows Figure 3 As shown: S3-1 Feature Extension Mapping Module: To enhance the expressive power of low-dimensional indicator vectors, the core indicator vectors are extended and mapped. Perform feature expansion mapping. Obtain a high-dimensional index embedding vector through fully connected mapping. . To embed the weight matrix, The ReLU function is used as the activation function unit. To improve model stability, a batch normalization layer is added after the fully connected mapping unit, thereby accelerating model convergence.

[0058] S3-2 constructs a multi-branch parallel network structure to describe the differentiated impact of different indicators on educational status. High-dimensional indicators are embedded into vectors. The inputs are fed into three parallel sub-network branches respectively: .in, Let be the weight matrix of the k-th branch.

[0059] Specifically, each branch subnetwork consists of two cascaded fully connected layers. Feature representation is enhanced through two-level mapping, and residual connections are implemented within the fully connected layers of each branch. The three branches correspond to the non-linear expansion paths of three core metrics, enabling independent modeling of different capability dimensions.

[0060] S3-3 introduces a state coupling mechanism to model the interaction relationships between different indicator branches. First, the features of each branch are concatenated to obtain... Furthermore, to achieve dynamic interaction between features from different branches, a gating coupling mechanism is introduced to perform weighted fusion of the spliced ​​features.

[0061] Specifically, firstly, Gating vectors are generated through gating mapping. The sigmoid activation function is used to constrain the gate values ​​to the interval [0,1]. Then, based on the gate vector... right Weighting is performed to obtain the gated modulation features. Finally, a linear mapping is performed on the modulated features to obtain the coupled feature representation. .

[0062] Through the above gating mechanism, the adaptive selection and dynamic fusion of different indicator branch features are realized, enabling the model to automatically adjust the influence weight of each indicator according to different educational statuses.

[0063] S3-4 State Distribution Generation Layer and Determination of State Normalized Coordinates. To construct a high-dimensional state space representation, the coupled features are further mapped to a state distribution matrix. Specifically, the coupled feature representation... Reconstructed into matrix form: . This is the mapping matrix for the fully connected layer. This is the state evaluation matrix. The fully connected layer is used to represent the coupling features. Mapped to a higher-dimensional space, The operation is used to rearrange a one-dimensional vector into a two-dimensional structure, thereby forming a state evaluation matrix.

[0064] Furthermore, to enhance spatial expressiveness, a two-dimensional convolutional smoothing layer is introduced to smooth the state evaluation matrix. Local neighborhood information is fused, followed by a normalization layer to obtain the normalized state evaluation matrix. .

[0065] The normalized state evaluation matrix Expand into vector form: .in, This represents the state vector. The vectorization operation... It is implemented using tensor expansion units, which are used to map two-dimensional state structures to one-dimensional state coordinates.

[0066] Finally, the output state evaluation matrix is ​​obtained. and its corresponding state vector .

[0067] S4. Dynamic Generation and Output of Intervention Strategies S4-1 Educational Trajectory Feature Extraction and Standard Educational Trajectory Mapping

[0068] Based on the relevant information of the target object, from the normalized state evaluation matrix Extract diagnostic information that reflects the characteristics of the educational status.

[0069] Specifically, using a size of The convolution kernel in The algorithm performs sliding calculations, weighted summations on each local region, and extracts the state change trends and distribution patterns within that local region to obtain the current educational trajectory structure feature matrix of the current object. .

[0070] Next, in order to construct a benchmark, the set of state matrices of historical high-performing objects at stage t is recorded as follows: Then the standard education trajectory matrix .in, Indicates the first l The state matrix of a high-quality object at stage t, where L represents the number of samples. Similarly, the standard educational trajectory matrix... Perform two-dimensional convolution operations to obtain the standard educational trajectory structure feature matrix. .

[0071] S4-2 Spatial Deviation Calculation and Deviation Structure Analysis. This involves obtaining the current educational trajectory structure feature matrix. and standard education trajectory structure feature matrix Then, calculate the state deviation distribution. Then, the overall deviation was calculated. : ; Where p represents the number of rows in the current educational trajectory structure feature matrix, and q represents the number of columns in the diagnostic feature matrix. The larger the value, the greater the degree of deviation.

[0072] Simultaneously, in order to introduce overall state position information, the state vector is... With standard state vector Perform Euclidean distance calculation .in, For the standard education trajectory matrix Through state assessment matrix The standard state vector is obtained through the same vectorization process. This indicates the amount of positional deviation.

[0073] Therefore, a joint deviation is constructed. .in and These are weighting coefficients used to balance the effects of structural and positional deviations. (Joint deviation amount) It comprehensively reflects the overall degree of deviation of the current educational status from the standard trajectory.

[0074] Next, a deviation structure analysis is performed. To avoid relying solely on a single deviation value for decision-making, the deviation matrix is ​​analyzed in this step. Perform structural decomposition analysis. Specifically, statistically analyze the deviations along two directions: ; in, This indicates the distribution of deviations along the row direction. This indicates the distribution of deviations along the column direction. This represents the row number of the current educational trajectory structure feature matrix. This indicates the number of columns in the diagnostic feature matrix. The above analysis is used to identify the distribution of deviations in different regions from the perspective of diagnostic features, thereby providing a basis for the targeted generation of subsequent intervention strategies.

[0075] S4-3 Remedial Coefficient Calculation and Intervention Strategy Generation. Based on Joint Deviation. Generate global remediation coefficient , The adjustment parameter is used to control the overall strength of the intervention strategy, allowing the intervention behavior to adaptively adjust to the degree of deviation.

[0076] Furthermore, a local remedial distribution matrix is ​​generated. The local remedial distribution matrix It describes the intensity of intervention needs at different locations, thereby enabling the expansion from overall control to fine-grained local control.

[0077] Next, after obtaining the salvage distribution matrix... Then, it is divided into regions. .in, This represents the m-th region.

[0078] The region division is based on the spatial structure of the current educational trajectory structural feature matrix, specifically using a regular grid division method. Under this method, the matrix is ​​divided into K sub-regions of a fixed size.

[0079] Further calculate the intervention intensity for each region. And make directional corrections: . This represents the projection function of the state offset vector onto the corresponding policy dimension. For adjustment coefficients, This is to correct the policy strength coefficient. This mechanism is used to directionally adjust the intervention policy based on the offset direction of the state in high-dimensional space, rather than relying solely on local structural differences.

[0080] Obtaining the strength coefficient of the modified strategy Next, a policy decision function is constructed to map continuous intensity to discrete policy categories. Specifically, discrete intervention policies are defined: . This represents the policy mapping function, the policy mapping function Implemented using a segmented decision mechanism: ; in, and This is a preset threshold. The threshold is set based on the statistical distribution of historical data; preferably, it is set to a value that is... , It is used to classify intervention intensity into three levels: low, medium, and high. , and These represent different types of intervention strategies. Specifically: This is a mild intervention strategy, applicable to cases with small state deviations, employing a suggestive adjustment approach. It provides the target group with a list of minor optimization suggestions, including a recommended reading list, pointing out shortcomings in the current research, and requiring small improvements based on the existing research, without altering the overall research path.

[0081] A moderate intervention strategy was adopted, but significant deviations were observed. Structural optimization measures were implemented. This involved a partial restructuring of the current research protocol, such as adjusting research methods, redesigning experimental procedures, or adding key validation steps. The improvements were required to be completed within a specified timeframe, with a progress report submitted.

[0082] Strengthen intervention strategies to address significant deviations, and implement targeted resource allocation and pathway restructuring. This involves a comprehensive adjustment of research directions or technical approaches, including redefining research questions, introducing new methodologies, or changing research pathways, while providing supplementary resource support, including expert guidance and specialized training tasks.

[0083] Secondly, based on the core indicator vector of S2 The three components determine the dominant capability dimension: ; in, The dominant capability dimension of region m is represented by k=1, where k=2 represents theoretical originality, k=2 represents methodological breakthrough, and k=3 represents interdisciplinary collaboration. A corresponding intervention action library is constructed based on this. Finally, based on the dominant capability dimension From the intervention action library Select the corresponding intervention action as the intervention content. Specifically, the intervention action library The three corresponding ability dimensions are: This is a theoretically original intervention library. It includes execution: 1) Select 10 representative papers from authoritative journals in the field, break them down into three stages: "problem definition - methodological framework - theoretical conclusions", and write a structured analysis table and a unified theoretical summary report for each paper. 2) For the core model in the existing research methods, the target object is required to reproduce the complete derivation link step by step, and provide symbolic interpretation and verification of the assumptions for each mathematical transformation, while conducting boundary perturbation analysis on the key assumptions. This is a breakthrough intervention library. It includes execution: 1) To address the performance bottlenecks in the current method flow, we identify three specific constraint steps and propose three independent improvement paths for each. Each path must correspond to an executable algorithm adjustment scheme. 2) Redefine the experimental variable system, including the set of input variables, the set of control variables, and the set of evaluation indicators, and construct a standardized experimental procedure document, while completing the verification of the reproducible experiment; 3) Select a heterogeneous model structure, replace the original method with a new structure, and build a control experimental system to evaluate the accuracy, convergence speed and computational complexity in three dimensions.

[0084] This is an interdisciplinary intervention database. It includes implementation: 1) Embed the current research problem into a uniquely defined application scenario, construct an application model that includes an input layer, a decision layer, and an output layer, and complete a real-world case simulation and verification. 2) Select a classic data-driven model from different disciplinary method libraries to construct a combined fusion architecture, such as a fusion structure of physical constraint models and deep neural networks, and conduct quantitative comparative analysis of the performance differences before and after fusion.

[0085] Finally, personalized intervention strategy instructions are formulated. Among them, the period , This represents the maximum intervention period. The frequency is... . and These are the preset minimum and maximum number of weekly tasks, respectively.

[0086] Finally, personalized intervention strategy instructions are output. This enables the transformation from continuous decision variables to discrete intervention instructions.

[0087] Simulation experiment: Figure 4The diagram shows the weight matrix obtained after similarity modeling and normalization. The left figure shows the result before feature reconstruction, and the right figure shows the result after feature reconstruction. After feature reconstruction, the distribution of samples from different categories in the feature space is more compact and the separation is higher, with clearer inter-class boundaries. This indicates that the index reconstruction network constructed in this invention can effectively enhance feature representation capabilities and improve the discriminative performance of subsequent evaluation models.

[0088] Figure 5 The spatial distribution of intervention strategies is illustrated. Different colored areas correspond to different intervention levels, with green areas indicating mild intervention needs, yellow areas indicating moderate intervention needs, and red areas indicating areas requiring intensive intervention. It can be seen that the intervention strategies can be spatially differentiated based on the distribution of deviations, demonstrating the invention's ability to transform from overall assessment to precise local intervention.

[0089] Figure 6 This paper demonstrates the effectiveness of the method of this invention in improving core competency indicators at different educational stages. As shown in the figures, after implementing the intervention strategy, the three indicators—theoretical originality, methodological breakthrough, and interdisciplinary relevance—all show a significant upward trend. The improvement is particularly significant in the early stages of education, indicating that this invention can effectively identify early deviations and provide targeted interventions. In the middle and later stages, each indicator continues to optimize, demonstrating the method's good continuous regulatory capacity. Overall, this invention can achieve stable and significant competency improvements at different stages, verifying the effectiveness of the proposed educational status assessment and intervention mechanism.

[0090] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0091] While the above description illustrates specific embodiments of the present invention, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for assessing and intervening in educational status based on multi-source data fusion, characterized in that, The process includes the following: S1: Acquire academic depth data, industry value data, and science popularization data of the target object, and preprocess them to obtain a standardized text set; vectorize the standardized text set, and call a natural language processing model to map the unstructured text data into a semantic feature set; after semantic vector aggregation, obtain the final output feature set. ; S2, Input metrics to reconstruct the network, perform feature recalibration and mapping to obtain the core metric vector. These include indicators of theoretical originality, methodological breakthroughs, and interdisciplinary collaboration. S3, will The input feature extension mapping module generates high-dimensional index embedding vectors. ,Will After inputting three parallel sub-network branches, coupled feature representations are generated through feature mapping. ;Will By performing a fully connected mapping and normalization, the normalized state evaluation matrix is ​​obtained. ; S4, from Diagnostic features are extracted and spatially mapped with preset standard educational trajectory features. The deviation is obtained by calculating the Euclidean distance between the two. After calculating the remedial coefficient of the deviation, it is input into the strategy determination function to obtain the intervention strategy instruction.

2. The method for assessing and intervening in educational status based on multi-source data fusion as described in claim 1, characterized in that: In S1, academic depth data is acquired through an academic database interface, and the collected content includes text information such as paper abstracts, keywords, and research direction descriptions; industrial value data is acquired through a patent database interface and a technology transfer system, and the collected content includes texts such as patent specifications, technical solution descriptions, and application scenario descriptions; popular science data is acquired through an open interface of a media platform, and the collected content includes news reports, popular science articles, and public communication content related to the target audience.

3. The method for assessing and intervening in educational status based on multi-source data fusion as described in claim 1, characterized in that: The natural language processing model adopts a Transformer-based natural language processing model; for each vector matrix Perform a linear mapping transformation to generate a query vector, a key vector, and a value vector: ; They represent respectively by The query matrix, key matrix, and value matrix obtained by linear transformation; Represented as a trainable weight matrix; Subsequently, based on Calculate the attention weight matrix according to the standard formula for attention. The value vectors are then weighted and summed to obtain the fused weighted intermediate semantic representation. ; A multi-head attention mechanism is used to concatenate weighted intermediate semantic representations from multiple attention heads in parallel. , obtain the output result Then through a feedforward neural network The function is subjected to a nonlinear mapping to obtain the first... i semantic feature matrix of the text Ultimately, a set of semantic features is obtained. .

4. The method for assessing and intervening in educational status based on multi-source data fusion as described in claim 1, characterized in that: The specific process of S2 includes: The final output feature set The input feature recalibration layer performs a weighted transformation on each dimension of the features through a single-layer fully connected network to generate a recalibrated feature matrix. Next, the feature matrix will be recalibrated. The input similarity modeling layer generates a similarity matrix by calculating the cosine similarity between samples. The similarity weight matrix was obtained after softmax normalization. This characterizes the correlation strength between different samples in the feature space; subsequently, the similarity weight matrix is... With recalibrated feature matrix Perform matrix multiplication to generate a fused feature matrix. Next, the fused feature matrix... Perform global average pooling to obtain the global fused feature vector. Finally, the globally fused feature vector is... The input indicator mapping layer generates core indicator vectors through fully connected mapping. , ; Indicators of theoretical originality measure the ability of a target object to contribute to fundamental theories. Breakthrough indicators are used to measure the degree of innovation in technical methods or implementation paths. Indicators representing interdisciplinary characteristics are used to measure the ability to integrate and apply knowledge across disciplines. The values ​​are all within the range of [0,1]. The larger the index value, the stronger the corresponding capability dimension.

5. The method for assessing and intervening in educational status based on multi-source data fusion as described in claim 1, characterized in that: In S3, there are three parallel sub-network branches, each consisting of two cascaded fully connected layers. Two-level mapping enhances feature representation capabilities, and residual connections are set within the fully connected layers of each branch. The three branches correspond to the non-linear expansion paths of three core indicators. Simultaneously, the features from each branch are concatenated to obtain... Then, a gating coupling mechanism is introduced to perform weighted fusion of the spliced ​​features: First, Gating vectors are generated through gating mapping. The sigmoid activation function is used to constrain the gate value to the interval [0,1]. Subsequently, based on the gate vector... right Weighting is performed to obtain the gated modulation features. Finally, a linear mapping is performed on the modulated features to obtain the coupled feature representation. .

6. The method for assessing and intervening in educational status based on multi-source data fusion as described in claim 1, characterized in that: The specific process of S4 includes: S41, Educational Trajectory Feature Extraction and Standard Educational Trajectory Mapping: Applying Normalized State Evaluation Matrix Local neighborhood feature extraction is performed to generate the current educational trajectory structure feature matrix. Introducing a standard educational trajectory matrix Generate a standard educational trajectory structure feature matrix ; S42, Spatial Deviation Calculation: For and Perform difference operations to generate state deviation distribution And calculate the overall deviation. At the same time, combined with the state vector With standard state vector Euclidean distance generates positional deviation Further fusion yields the joint deviation. ; S43, Remedial Coefficient Calculation and Intervention Strategy Generation: Input the Joint Deviation The remedial coefficient generation module obtains the global remedial coefficient. And generate a local remedial distribution matrix. ;right Divide the region into multiple sub-regions Calculate the intervention intensity in each region and the intensity of intervention The modified strategy strength coefficient is obtained by making corrections. Finally, the strategy strength coefficient will be adjusted. The input strategy determination function performs piecewise mapping to generate corresponding discrete intervention strategies. Ultimately, personalized intervention strategy instructions are constructed and output. .

7. The method for assessing and intervening in educational status based on multi-source data fusion as described in claim 6, characterized in that: The specific process of S41 includes: Use size The convolution kernel in The algorithm performs sliding calculations, weighted summations on each local region, and extracts the state change trends and distribution patterns within that local region to obtain the current educational trajectory structure feature matrix of the current object. ; The set of state matrices of historical high-quality objects in stage t is: Then the standard education trajectory matrix ;in, Indicates the first l The state matrix of a high-quality object at stage t, where L represents the number of samples; and the standard educational trajectory matrix. Perform two-dimensional convolution operations to obtain the standard educational trajectory structure feature matrix. .

8. The method for assessing and intervening in educational status based on multi-source data fusion as described in claim 6, characterized in that: The specific process of S42 includes: Calculate the state deviation distribution Then, calculate the overall deviation. : ； Where p represents the number of rows in the current educational trajectory structure feature matrix, and q represents the number of columns in the diagnostic feature matrix; The larger the value, the greater the degree of deviation; For the state vector With standard state vector Perform Euclidean distance calculation ; where, state vector From the normalized state evaluation matrix Expand to obtain; For the standard education trajectory matrix Through state assessment matrix The standard state vector is obtained through the same vectorization process; Indicates the positional deviation; Constructing joint deviation ;in and These are the weighting coefficients.

9. The method for assessing and intervening in educational status based on multi-source data fusion as described in claim 6, characterized in that: The specific process of S43 includes: Based on joint deviation Generate global remediation coefficient , The adjustment parameter; the remedial coefficient is used to control the overall strength of the intervention strategy; Generate a local salvage distribution matrix: ; Describe the intensity of intervention demand at different locations; The region is divided into regions using a regular grid method to distribute the local remedial distribution matrix. Divide into K sub-regions according to a fixed size; Calculate the arithmetic mean of all element values ​​within each region sequentially, and use this arithmetic mean as the intervention intensity for the corresponding region. The intervention intensity for each region is denoted as ; The direction of the intervention intensity is adjusted for each region; for the first region... Intervention intensity in each region Calculate the state vector With standard state vector The difference vector is calculated, and its projection value on the corresponding policy dimension is extracted. The projection value is then compared with a preset adjustment coefficient. Multiplying them gives the direction correction amount, which is then used to correct the direction. Adding the directional correction amount to the formula yields the correction strategy strength coefficient. ; Construct a policy decision function to map continuous intensity to discrete policy categories: ； in, and The threshold is set based on the statistical distribution of historical data and is used to divide the intervention intensity into three levels: low, medium, and high. , and These represent different types of intervention strategies.

10. The method for assessing and intervening in educational status based on multi-source data fusion as described in claim 9, characterized in that: As a mild intervention strategy, a suggestive adjustment approach is adopted, providing the target subjects with a list of minor optimization suggestions, including a list of recommended readings, pointing out the shortcomings of the current research, and requiring minor improvements based on the original research, without changing the overall research path; As a moderate intervention strategy, structural optimization measures are adopted to partially restructure the current research protocol, including adjusting research methods, redesigning experimental procedures, or adding key validation steps. Improvements are required to be completed within a specified period and a progress report must be submitted. To strengthen intervention strategies, key resources are allocated and pathways are restructured. Overall adjustments are made to research directions or technical routes, including redefining research questions, introducing new methodologies, or changing research pathways. Additional resource support is provided, including expert guidance and specialized training tasks.

Images