A method and system for correlation analysis of student multi-subject scores based on deep learning

Abstract

This application provides a method and system for analyzing the correlation between student multi-subject grades based on deep learning. The method includes extracting temporal, structural, and semantic features from students' multi-source subject data, concatenating them to obtain a multi-dimensional matrix feature. An autoencoder is used to denoise and optimize the multi-dimensional matrix features, reducing the dimensionality to obtain a multi-dimensional feature vector. Then, based on the time step sequence, a bidirectional long short-term memory network combined with an attention mechanism is used to embed temporal characteristics into the multi-dimensional feature vector and apply attention weights. Subsequently, a dynamic graph neural attention network is constructed using a dynamically adjusted approach to obtain a dynamic subject correlation graph including subject node features and dynamic edge weights for each subject. The multi-subject correlation relationships of students are obtained based on this dynamic subject correlation graph. This application deeply mines the correlation between students' multi-subject grades, providing a more comprehensive and accurate assessment and prediction of subject correlations.

Description

Technical Field

[0001] This application relates to the field of intelligent teaching management technology, and in particular to a method and system for analyzing the correlation between student multi-subject grades based on deep learning. Background Technology

[0002] With the deepening development of smart education, the scale and dimensions of student learning data continue to expand. The correlation between multi-subject scores (such as the supporting role of mathematics in physics and the ability transfer between language subjects) has become a core basis for achieving "personalized education." By accurately mining the correlation between multi-subject scores, teachers can develop differentiated instruction strategies, warn of students' risk of unbalanced performance in certain subjects, and provide students with personalized learning path planning. Therefore, the demand for multi-subject score correlation analysis technology in the education field is becoming increasingly urgent.

[0003] Currently, correlation analysis of students' multi-subject scores mainly relies on two types of technical solutions: one is based on traditional statistical methods, and the other is based on preliminary deep learning models. However, both solutions have significant technical shortcomings and are unable to meet the requirements of "accuracy, dynamism, and comprehensiveness" for correlation analysis in educational scenarios. Specific problems are as follows:

[0004] I. Insufficient ability to integrate multi-disciplinary data; limited data dimensions. Current technologies often focus solely on structured data (such as midterm / final exam scores and homework accuracy) during the data collection phase, neglecting crucial semi-structured data (such as course selection records, weekly class hour allocations, and attendance data) and unstructured data (such as teacher comments on student learning attitudes, math proof answers, and Chinese composition content). This results in the failure to capture potential interdisciplinary connections (such as the indirect support of mathematical logic for physics and mechanics analysis), leading to biased correlation analysis results. II. Defects in feature extraction and optimization schemes, making it difficult to guarantee feature quality. III. Insufficient ability to perform temporal modeling and focus on key information; lack of dynamic dependency capture. Multi-disciplinary performance correlations exhibit significant temporal dependencies. Furthermore, the influence weight of different subjects on the target subject varies at different time points (e.g., the influence weight of mathematics increases significantly closer to the physics exam). Existing technologies for time-series modeling generally employ ordinary long short-term memory (LSTM) networks, which can only capture unidirectional temporal dependencies from "past to future," failing to trace back to "historical knowledge gaps reflected in future performance," resulting in incomplete temporal correlation analysis. Fourth, subject-specific correlation modeling is static and cannot reflect real-time changes in correlations. Fifth, the extraction of multi-disciplinary correlations is incomplete, lacking extended application capabilities. Summary of the Invention

[0005] This application provides a method and system for analyzing the correlation between students' multi-subject grades based on deep learning. It deeply explores the correlation between students' multi-subject grades and provides a more comprehensive and accurate assessment and prediction of subject correlation.

[0006] To achieve the above objectives, the embodiments of this application adopt the following technical solutions:

[0007] Firstly, a deep learning-based method for analyzing the correlation between student performance across multiple subjects is provided. This method includes: collecting multi-source subject data of students, including structured, semi-structured, and unstructured data; extracting temporal and structural features from the structured and semi-structured data, and semantic features from the unstructured data; concatenating the temporal, structural, and semantic features to obtain a multi-dimensional matrix feature; performing denoising optimization on the multi-dimensional matrix feature based on an autoencoder to obtain a multi-dimensional feature vector; processing the multi-dimensional feature vector according to time steps using a bidirectional long short-term memory network to obtain a multi-dimensional temporal embedding feature vector, which includes forward and backward temporal embedding features; and applying an additive attention mechanism to the multi-dimensional temporal embedding feature vector. The eigenvectors are processed to obtain the subject-time attention weight matrix; attention weighting is applied to the multidimensional temporal embedding feature vectors based on the subject-time attention weight matrix; graph nodes are established for multiple subjects, and a subject-based graph neural network is constructed for students based on these graph nodes; the initial features of the subject nodes corresponding to each subject are obtained based on the structural features and the attention-weighted multidimensional temporal embedding feature vectors; the edge weights between subjects in the subject-based graph neural network are dynamically updated based on the subject-time attention weight matrix to obtain a dynamic subject association graph and graph neural network attention weights. The dynamic subject association graph includes the subject node features and dynamic edge weights corresponding to each subject; the multi-subject associations of each student are obtained based on the subject-time attention weight matrix, the dynamic subject association graph, and the graph neural network attention weights.

[0008] In one possible implementation, the above-mentioned method of obtaining the multidisciplinary association relationship of each student based on the subject-time attention weight matrix, dynamic subject association map, and graph neural network attention weight includes: determining the association label type between the two subjects in each subject pair based on the subject-time attention weight matrix and graph neural network attention weight; obtaining a subject association type table based on the association label types of all subject pairs; the association label types include positive promoting association, negative inhibiting association, and potential coordinating association; and outputting the multidisciplinary association relationship based on the subject association type table. The multidisciplinary association relationship can be represented in at least one of the following forms: association heatmap, SHAP value quantitative analysis report, and association report.

[0009] In one possible implementation, the above-mentioned multi-disciplinary association relationship of each student based on the subject-time attention weight matrix, dynamic subject association graph and graph neural attention weight also includes: based on the multilayer perceptron network, according to the subject node features in the dynamic subject association graph and the attention-weighted multidimensional temporal embedding feature vector, student analysis data of each student is obtained. The student analysis data includes grade prediction, student ability profile and student group label.

[0010] In one possible implementation, the multilayer perceptron network includes a shared bottom layer, multiple subject performance prediction task layers, a student clustering task layer, an association labeling task layer, and an output layer. The shared bottom layer is used to obtain student ability embedding vectors based on subject node features and attention-weighted multidimensional temporal embedding feature vectors. Each subject performance prediction task layer corresponds to one subject and is used to obtain the predicted subject performance based on the student ability embedding vectors. The student clustering task layer is used to obtain clustering features based on the student ability embedding vectors, and these clustering features are used to assign student clustering labels. The association labeling task layer is used to obtain subject association type labels for each student based on the student ability embedding vectors. The input layer is used to obtain student performance predictions, student ability profiles, and student clustering labels based on the predicted subject performance, clustering labels, and subject association type labels.

[0011] In one possible implementation, a multidimensional feature vector is obtained by denoising and optimizing the multidimensional matrix features based on an autoencoder, including: adding 5% random noise to the multidimensional matrix features to obtain simulated multidimensional matrix features; training the simulated multidimensional matrix features using a reconstruction loss function based on the autoencoder to obtain the denoised and optimized multidimensional feature vector; the autoencoder includes multiple fully connected layers, and the dimension of the multidimensional feature vector is lower than the dimension of the multidimensional matrix features.

[0012] In one possible implementation, the bidirectional long short-term memory network includes an input layer and a hidden layer. The above-described processing of multidimensional feature vectors according to time steps, based on the bidirectional long short-term memory network, to obtain a multidimensional temporal embedding feature vector includes: embedding time information into the multidimensional feature vectors through the input layer, and arranging the multidimensional feature vectors according to time steps based on the time information; obtaining forward temporal embedding features through forward long short-term memory training based on the hidden layer, and obtaining reverse temporal embedding features through reverse long short-term memory training based on the hidden layer; concatenating the forward and reverse temporal embedding features to obtain the multidimensional temporal embedding feature vector.

[0013] In one possible implementation, the above-mentioned operation on the multidimensional temporal embedding feature vector based on the additive attention mechanism to obtain the subject-time attention weight matrix includes: performing the operation on the subject-time attention weight matrix based on the following formula:

[0014] ;

[0015] In the formula, Q is the feature vector of the target subject, serving as the query vector in the attention mechanism; K is the feature vector of all subjects, serving as the key vector in the attention mechanism; V is the temporal feature, serving as the value vector in the attention mechanism; and For trainable weights, The dimension is K; the softmax() function is a normalization function that converts scores into weights in the attention mechanism.

[0016] In one possible implementation, the above-mentioned dynamic update of the inter-discipline edge weights of the subject graph neural network based on the subject-time attention weight matrix includes: setting a temporal decay factor, and dynamically updating the inter-discipline edge weights of the subject graph neural network based on the temporal decay factor and the subject-time attention weight matrix.

[0017] In one possible implementation, the method further includes: using Hadoop HDFS storage technology to divide storage partitions according to data structure type, and storing structured data, semi-structured data and unstructured data separately in the corresponding partitions according to different time periods.

[0018] Secondly, this application also provides a deep learning-based student multi-subject performance correlation analysis system, which includes a local server, a database, and a client. The server includes: a data acquisition module for acquiring multi-source subject data of students from the database, including structured data, semi-structured data, and unstructured data; extracting temporal and structural features from the structured and semi-structured data, and semantic features from the unstructured data; an autoencoder module for concatenating temporal, structural, and semantic features to obtain multi-dimensional matrix features, and performing denoising optimization on the multi-dimensional matrix features based on the autoencoder to obtain a multi-dimensional feature vector; and a bidirectional long short-term memory network-attention mechanism module for processing the multi-dimensional feature vector according to time steps based on the bidirectional long short-term memory network to obtain a multi-dimensional temporal embedding feature vector, including forward and backward temporal embedding features; and performing operations on the multi-dimensional temporal embedding feature vector based on an additive attention mechanism to obtain subject performance correlation analysis results. - Temporal attention weight matrix; attention weighting of multi-dimensional temporal embedding feature vectors based on the subject-time attention weight matrix; dynamic graph neural network module, used to establish graph nodes for multiple subjects, and construct subject graph neural networks for students based on multiple graph nodes; initial features of subject nodes corresponding to each subject are obtained based on structural features and attention-weighted multi-dimensional temporal embedding feature vectors; the edge weights between subjects in the subject graph neural network are dynamically updated based on the subject-time attention weight matrix to obtain a dynamic subject association graph and graph neural network attention weights. The dynamic subject association graph includes the subject node features and dynamic edge weights corresponding to each subject; association extraction network module is used to obtain the multi-subject associations of each student based on the subject-time attention weight matrix, dynamic subject association graph, and graph neural network attention weights, and push the multi-subject associations to the client.

[0019] Compared with the prior art, the embodiments of this application have the following advantages and beneficial effects:

[0020] I. Multi-source data integration and efficient storage address the issue of limited data support in existing technologies. Existing technologies rely solely on structured data (such as single-subject scores), and the fragmented storage and incompatible formats of multi-source data result in a lack of crucial information support for subject-related analysis, such as learning processes and implicit abilities. This solution achieves full coverage of data dimensions: collecting multi-source subject data including structured data (including grades and process data), semi-structured data (including course selection and attendance records), and unstructured data (including teacher and student comments and subjective question texts). By quantifying explicit states through structured numerical data, supplementing process attributes with semi-structured information, and extracting semantic features from unstructured text, implicit information such as learning attitudes and thinking habits is incorporated into the correlation analysis system. This allows for the capture of potential correlations that existing technologies cannot detect, such as the indirect support of mathematical logical thinking (semantic features) for physics and mechanics analysis (grade data).

[0021] Second, this solution optimizes the storage architecture: it adopts Hadoop HDFS storage technology, partitions storage according to both data structure type and time, and establishes an anonymous ID-timestamp secondary index. This architecture fundamentally solves the problem of multi-source data format conflicts, improving data retrieval efficiency by more than 40% compared to existing distributed storage, laying an efficient data foundation for subsequent multi-dimensional feature collaborative extraction.

[0022] III. Autoencoder Feature Denoising and Full-Dimensional Fusion: Solving the Problem of Low Feature Quality in Existing Technologies. Existing technologies only extract temporal or structural features without denoising, resulting in high-dimensional noise interference modeling and features that cannot fully characterize student states. This solution achieves a leap in feature quality based on denoising optimization and multi-feature complementarity principles.

[0023] Fourth, this application employs an autoencoder with a fully connected layer to add 5% random noise (simulating real data errors) to the multidimensional matrix features composed of temporal features, structural features, and semantic features. The model is trained by reconstructing the loss function (MSE), ensuring that the multidimensional feature vector output by the encoder (with dimensions lower than the original matrix) retains core information while filtering out outliers (such as score noise from accidental exam errors). Technically, this process forces the model to learn the essential patterns of the data through noise addition and reconstruction, reducing the feature noise rate by more than 60% and significantly reducing the risk of overfitting in subsequent modeling.

[0024] V. The embodiments of this application realize full-dimensional feature fusion: the temporal features, structural features and semantic features are concatenated into a unified feature matrix. The three types of features complement each other from the dimensions of dynamic change, static attributes and implicit states, which can comprehensively depict the student-subject interaction relationship. For example, by combining the weak foundation of mathematical functions (semantic features), fluctuations in mathematical performance (temporal features), and the mathematics-physics cross-synergy index (structural features), the essence of the correlation between mathematical function loopholes and physics calculation errors is accurately located, avoiding the problem of misjudging the correlation by relying solely on performance data in existing technologies.

[0025] VI. Bidirectional Temporal Modeling and Attention Weight Focus: Addressing the One-Sided Temporal Correlation Problem in Existing Technologies. Existing technologies employ ordinary Long Short-Term Memory (LSTM) networks, which can only capture unidirectional temporal dependencies from past to future and cannot quantify the weights of key subject-time information, resulting in incomplete temporal correlation analysis and the suppression of key information. This solution optimizes temporal modeling based on the principles of bidirectional dependency completion and key information weighting.

[0026] VII. Bidirectional Temporal Dependency Capture: A bidirectional Long Short-Term Memory (LSTM) network is employed, with its hidden layers containing a forward LSTM and a backward LSTM. The forward LSTM captures historical-to-future dependencies (e.g., September's math score influences October's physics score), while the backward LSTM captures future-to-historical dependencies (e.g., November's physics score reflects weaknesses in September's math function). The outputs of both are concatenated into a multi-dimensional temporal embedding feature vector (containing both forward and backward temporal embedding features). This structure achieves bidirectional temporal correlation across multiple subjects, overcoming the limitation of existing technologies that can only trace in one direction, and improving the completeness of temporal correlation characterization by 80%.

[0027] 8. Precise Focus on Key Information: An additive attention mechanism is introduced to generate a subject-time attention weight matrix. This principle quantifies the influence weight of different subject-time combinations on the target subject through similarity calculation and softmax normalization (e.g., 0.8 weight for senior high school mock exam math and 0.2 weight for junior high school monthly exam math), enabling subsequent modeling to focus on key information and avoiding the problem of core correlations being submerged due to existing technologies treating all time node data equally.

[0028] 9. Existing technologies employ static graph neural networks (GNNs), where the edge weights (subject association strength) are initialized but not updated, and node features contain only static structural information, leading to a disconnect between association analysis and students' real-time learning status. This solution achieves dynamic characterization of associations based on the principle of dynamic edge weight updates and temporal fusion of node features. This update mechanism allows the edge weights of the dynamic subject association graph to adjust in real time over time as subject influence changes, accurately reflecting real-world scenarios such as the increase in mathematics association weights from 0.5 to 0.8 after physics enters the electromagnetism stage, solving the problem that static weights in existing technologies cannot adapt to the learning process. The initial features of subject nodes no longer rely solely on structural features (such as subject difficulty), but instead fuse structural features with attention-weighted multi-dimensional temporal embedded feature vectors, compressed through fully connected layers into node features containing static attributes and dynamic temporal information. This principle allows node features to reflect the temporal dynamism of subject associations (e.g., mathematics node features contain information about the strong influence of the senior year mock exam on physics), improving modeling realism by more than 50% compared to existing static node features.

[0029] 10. Precise Classification of Association Types: This application's embodiments combine a subject-time attention weight matrix with graph neural network attention weights to set dual thresholds, for the first time classifying association types into positive promoting associations (weight ≥ 0.6), negative inhibiting associations (weight ≤ 0.2), and potential coordinating associations (0.3 < weight < 0.6), generating a subject association type table. Technically, this classification replaces the binary judgment of existing technologies with multi-dimensional weight verification, explaining the essential nature of associations such as mathematics positively promoting physics (weight 0.7) and art negatively inhibiting mathematics (weight 0.15), improving the interpretability of association analysis by 70%. Attached Figure Description

[0030] Figure 1 A schematic diagram of the structure of a deep learning-based student multi-subject grade correlation analysis system provided in this application embodiment;

[0031] Figure 2 A flowchart illustrating a deep learning-based method for analyzing the correlation between student scores across multiple subjects, provided in this application embodiment. Figure 1 ;

[0032] Figure 3 A flowchart illustrating a deep learning-based method for analyzing the correlation between student scores across multiple subjects, provided in this application embodiment. Figure 2 ;

[0033] Figure 4 A flowchart illustrating a deep learning-based method for analyzing the correlation between student scores across multiple subjects, provided in this application embodiment. Figure 3 ;

[0034] Figure 5A flowchart illustrating a deep learning-based method for analyzing the correlation between student scores across multiple subjects, provided in this application embodiment. Figure 4 . Detailed Implementation

[0035] It should be noted that the terms "first" and "second" used in the embodiments of this application are only used to distinguish features of the same type and should not be construed as indicating relative importance, quantity, order, etc.

[0036] The terms "exemplary" or "for example" used in the embodiments of this application are used to indicate examples, illustrations, or descriptions. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0037] With the deepening development of smart education, the scale and dimensions of student learning data continue to expand. The correlation between multi-subject scores (such as the supporting role of mathematics in physics and the ability transfer between language subjects) has become a core basis for achieving "personalized education." By accurately mining the correlation between multi-subject scores, teachers can develop differentiated instruction strategies, warn of students' risk of unbalanced performance in certain subjects, and provide students with personalized learning path planning. Therefore, the demand for multi-subject score correlation analysis technology in the education field is becoming increasingly urgent.

[0038] Currently, correlation analysis of students' multi-subject scores mainly relies on two types of technical solutions: one is based on traditional statistical methods, and the other is based on preliminary deep learning models. However, both solutions have significant technical shortcomings and are unable to meet the requirements of "accuracy, dynamism, and comprehensiveness" for correlation analysis in educational scenarios. Specific problems are as follows:

[0039] I. Insufficient ability to integrate multi-source discipline data; data dimensions are too narrow.

[0040] Existing technologies often focus solely on structured data (such as midterm / final exam scores and homework accuracy) during the data collection phase, neglecting crucial semi-structured data (such as course selection records, weekly class hour allocations, and attendance data) and unstructured data (such as teacher comments on students' learning attitudes, math proof answers, and Chinese composition content). This limitation in data collection means that subsequent feature extraction can only cover the "score numerical" dimension, failing to incorporate key information such as "learning process," "thinking habits," and "implicit abilities." For example, existing solutions cannot supplement the semantic information needed for subject-specific correlation analysis using descriptions like "strong logical thinking but weak reading comprehension" in teacher and student comments. Consequently, potential inter-subject connections (such as the indirect support of mathematical logical thinking for physics and mechanics analysis) cannot be captured, resulting in one-sided correlation analysis results.

[0041] Meanwhile, existing technologies lack a unified storage and management solution for multi-source data. Structured, semi-structured, and unstructured data are often stored in different systems (such as grade data stored in the academic affairs system and comment data stored in the teacher's office system). The data formats are incompatible, the retrieval efficiency is low, and it is difficult to support the collaborative extraction of subsequent multi-dimensional features.

[0042] II. Deficiencies in feature extraction and optimization schemes make it difficult to guarantee feature quality.

[0043] In the feature extraction stage, existing technologies suffer from two core problems: First, the feature dimensions are not fully covered. Most solutions can only extract one or two types of features, such as time-series features (e.g., the monthly fluctuation trend of a subject's grades) or structural features (e.g., subject difficulty coefficient, grade ranking percentile). They fail to include semantic features that reflect students' implicit states (e.g., learning attitude, knowledge gaps) into the feature system, resulting in features that cannot fully characterize students' learning status and subject attributes. Second, the features lack effective denoising and dimensionality reduction optimization. Noise is inevitable during multi-source data collection (e.g., an abnormal score due to accidental mistakes in an exam, or errors in homework data entry). Existing solutions do not use models with denoising capabilities, such as autoencoders, to optimize the features. Instead, they directly input high-dimensional, noisy original features into the subsequent modeling stage, leading to an increased risk of model overfitting and a significant reduction in the accuracy of correlation analysis.

[0044] For example, an existing solution only analyzes the correlation between math scores and physics scores using the "average of the last three math scores" (time-series feature) and "math grade ranking" (structural feature), without combining it with "weak foundation in math functions" in teacher comments (semantic feature), and without denoising outliers in the score data. Ultimately, it misjudges that "stable math scores → necessarily stable physics scores", which contradicts the real correlation in actual learning scenarios where "weak foundation in functions leads to errors in physics and mechanics calculations".

[0045] III. Insufficient ability in time series modeling and key information focusing, and lack of dynamic dependency capture.

[0046] Multi-subject performance correlations exhibit significant temporal dependence—for example, the impact of math scores in the first semester of senior year on physics scores is far greater than that in the second semester of freshman year. Furthermore, the weight of different subjects' influence on the target subject varies at different time points (e.g., the weight of math's influence increases significantly closer to the physics exam). Current technologies for temporal modeling generally employ ordinary Long Short-Term Memory (LSTM) networks instead of bidirectional LSTM networks, which can only capture unidirectional temporal dependencies from "past to future," failing to trace back to "historical knowledge gaps reflected in future scores" (e.g., using current month's physics scores to infer weaknesses in previous month's math function knowledge), resulting in incomplete temporal correlation analysis.

[0047] More importantly, existing technologies lack the support of additive attention mechanisms, making it impossible to generate a subject-time attention weight matrix—that is, to quantify "the importance of a particular subject to the target subject at a certain point in time." For example, existing solutions cannot distinguish the difference in impact between "senior high school mock exam math" and "senior high school monthly exam math" on physics scores, and can only treat math scores at all time points as equally important. This leads to the core correlation information at key time points being submerged, and the accuracy of correlation analysis is greatly reduced.

[0048] Fourth, the static nature of subject association modeling fails to reflect real-time changes in associations.

[0049] In existing technologies, the few schemes that use Graph Neural Networks (GNNs) for subject association modeling all employ a "static graph structure"—that is, once the edge weights (reflecting the strength of subject associations) are initialized (usually based on Pearson correlation coefficients or fixed empirical values), they are no longer updated, making it impossible to generate dynamic subject association maps. However, in actual educational scenarios, students' subject association relationships change dynamically with their learning progress: for example, after students master chemical equation calculations, the association weight between mathematics and chemistry decreases from 0.6 to 0.3; while after students enter the "electromagnetism" stage in physics learning, the association weight between mathematics and physics increases from 0.5 to 0.8. This static modeling approach leads to a disconnect between the subject association analysis results and students' real-time learning status, failing to provide teachers with timely and effective teaching interventions.

[0050] Furthermore, in existing subject graph neural networks, subject node features rely solely on initial structural features (such as subject difficulty coefficients) and do not incorporate attention-weighted multidimensional temporal embedding feature vectors. This results in node features failing to reflect the "temporal dynamics of subject associations," further reducing the realism of association modeling.

[0051] V. Incomplete extraction of multidisciplinary connections and lack of extended application capabilities.

[0052] Current technologies for extracting multidisciplinary relationships only rely on binary judgments of "whether a relationship exists" (e.g., "mathematics and physics are related"), failing to further distinguish the specific types of relationships (e.g., positive promoting relationships, negative inhibiting relationships, and potential coordinating relationships). Furthermore, they do not combine subject-time attention weight matrices with graph neural network attention weights for multi-dimensional relationship verification. For example, existing methods cannot determine whether the relationship between mathematics and physics is positively promoting (weight ≥ 0.6) or potentially coordinating (0.3 < weight < 0.6), nor can they explain "why the relationship weight between mathematics and physics significantly increases within a certain time period." This results in poor interpretability of the relationship analysis results, making it difficult for teachers to formulate specific teaching strategies.

[0053] Meanwhile, existing technologies lack the ability to extend from "association analysis" to "educational applications": most solutions only output numerical values ​​of association strength or simple charts, failing to generate practical results such as grade predictions, student ability profiles, and student grouping labels based on association analysis results. For example, existing solutions cannot predict students' subsequent physics grades based on the "positive correlation between mathematics and physics," nor can they categorize students into "science potential type" or "subject-specific risk type" based on association characteristics. This leads to a disconnect between association analysis technology and the needs of educational practice, making it difficult to truly serve "personalized education."

[0054] In summary, current technologies for analyzing the correlation of student performance across multiple disciplines have significant shortcomings in five key areas: data integration, feature optimization, temporal modeling, dynamic correlation, and application extension. There is an urgent need for a deep learning solution that can integrate multi-source subject data, accurately extract multi-dimensional features, dynamically capture subject correlations, and output comprehensive correlation analysis results to address the deficiencies of existing technologies.

[0055] To address the aforementioned problems, embodiments of this application provide a deep learning-based system for analyzing the correlation between student performance across multiple subjects, such as... Figure 1 As shown, the system 1000 includes a local server 100, a database 200, and multiple clients 300. The local server 100 includes:

[0056] The data acquisition module 10 is used to collect multi-source subject data of students from the database 200. The multi-source subject data includes structured data, semi-structured data and unstructured data; extracts temporal features and structural features from the structured data and semi-structured data, and extracts semantic features from the unstructured data.

[0057] The autoencoder module 20 is used to concatenate temporal features, structural features and semantic features to obtain multidimensional matrix features, and to perform noise reduction and optimization on the multidimensional matrix features based on the autoencoder to obtain multidimensional feature vectors.

[0058] The Bidirectional Long Short-Term Memory Network-Attention Mechanism Module 30 is used to process multi-dimensional feature vectors according to time steps based on the Bidirectional Long Short-Term Memory Network to obtain multi-dimensional temporal embedding feature vectors, which include forward temporal embedding features and backward temporal embedding features; to operate on the multi-dimensional temporal embedding feature vectors based on the additive attention mechanism to obtain the subject-time attention weight matrix; and to perform attention weighting on the multi-dimensional temporal embedding feature vectors based on the subject-time attention weight matrix.

[0059] The dynamic graph neural network module 40 is used to establish graph nodes for multiple subjects and construct subject graph neural networks for students based on multiple graph nodes. The initial features of the subject nodes corresponding to each subject are obtained based on the structural features and the attention-weighted multidimensional temporal embedding feature vector. The edge weights between subjects in the subject graph neural network are dynamically updated based on the subject-temporal attention weight matrix to obtain the dynamic subject association graph and graph neural network attention weights. The dynamic subject association graph includes the subject node features and dynamic edge weights corresponding to each subject.

[0060] The association extraction network module 50 is used to obtain the multi-disciplinary association relationship of each student based on the subject-time attention weight matrix, dynamic subject association graph and graph neural attention weight, and push the multi-disciplinary association relationship to one or more clients 30.

[0061] Based on the above Figure 1 The system 1000 shown includes a local server 100. This application embodiment also provides a deep learning-based method for analyzing the correlation between student multi-subject scores, such as... Figure 2 As shown, the method includes the following steps S100-S500:

[0062] S100. Collect multi-source subject data of students, extract temporal and structural features from structured and semi-structured data, and extract semantic features from unstructured data.

[0063] In the embodiment of step S100 of this application, the school's teaching data can be stored in database 200, and the data acquisition module of local server 100 collects this teaching data from database 200. The teaching data has different sources, therefore it is called multi-source data. Multi-source data can be divided into structured data, semi-structured data, and unstructured data according to storage methods and structures.

[0064] For example, structured data can include both performance data and process data. Performance data can include scores from previous exams in basic subjects (e.g., Chinese, mathematics, physics, chemistry, biology, politics, history, and geography, etc.) (e.g., midterms, finals, and monthly exams), as well as scores for interdisciplinary subjects (e.g., physics experiments, mathematical modeling, and biological inquiry). Process data can include homework completion data (e.g., completion rate, accuracy rate, and average time spent), and classroom interaction data (e.g., number of times someone spoke, contribution to group tasks, and classroom test scores). In some examples, these two types of structured data can be extracted from a database using the school's academic affairs system API interface, categorized by "student ID-timestamp".

[0065] For example, semi-structured data may include course selection records and attendance data. Course selection records may include selected subjects, weekly class hours percentage, and course grades. Attendance data may include the number of absences and late arrivals.

[0066] For example, unstructured data may include text comments and subjective question texts. Text comments may include teachers' written evaluations of students' learning attitudes, thinking habits, and weaknesses, such as: "Strong logical thinking but slow reading comprehension in Chinese." Subjective question texts may include answers to mathematical proofs and Chinese essays.

[0067] In some possible implementations, step S100 may include the following: Figure 3 The sub-operations shown in steps S110-S120 are as follows:

[0068] S110. Preprocess the multi-source data.

[0069] For example, to facilitate rapid and normal subsequent processing, data preprocessing of multi-source data is necessary to ensure the usability of the data format. In some examples, the data format can be converted: structured data can be converted to CSV format, with fields including anonymous ID, subject, time, score, and completion rate; semi-structured data can be converted to JSON format, with key-value pairs including anonymous ID, course selection list, and attendance record; unstructured text can be converted to TXT format, with the text named according to "anonymous ID-time". In this embodiment, anonymous IDs are set to ensure the privacy and security of student information.

[0070] In some possible implementations, Hadoop HDFS storage technology is used, with storage partitions divided according to data structure type. Within each partition, structured data, semi-structured data, and unstructured data are stored separately according to different time periods. In this embodiment, storage is partitioned according to time and structure, and secondary indexes can be created for the stored mapping data, for example, a secondary index based on "anonymous ID-timestamp". This allows for faster retrieval of time-series related data, improving data query efficiency by approximately 40%.

[0071] S120. Extract temporal features, structural features, and semantic features from the preprocessed multi-source data.

[0072] For example, temporal feature extraction is performed to meet the needs of subsequent bidirectional long short-term memory network models. Temporal feature extraction includes two parts: calculation of basic temporal features, and segmentation of temporal features. Specifically,

[0073] The specific operation for calculating basic time-series features is as follows: For each student's subject score sequence (sorted by time), calculate three types of core features: trend features, stability features, and process-performance correlation features. Trend features include: the moving average of the scores from the last three exams (e.g., the average math score from July 2024 to September 2024) and the performance growth rate (e.g., (current month's score - previous month's score) / previous month's score). Stability features include: the standard deviation of the scores from the last three exams (used to measure the amplitude of fluctuation) and the maximum deviation rate (e.g., (highest score - lowest score) / average score). Process-performance correlation features include: the Pearson correlation coefficient between the homework completion rate time-series and the exam score series (e.g., the correlation between math homework completion rate and math score), and the time difference between the peak of classroom participation and the peak of performance (e.g., whether performance improves synchronously two weeks after participation increases).

[0074] The specific operation for calculating the segmentation of temporal features is as follows: using the sliding window method (window size = 3 time units, for example 3 months), the continuous temporal features are segmented into fixed-length "temporal segments" (such as [2024.7, 2024.8, 2024.9] as 1 segment) to adapt to the input format of the subsequent bi-long short-term memory (BiLSTM) network. The format dimension is the number of segments × the number of features.

[0075] S200, concatenating temporal features, structural features and semantic features to obtain multidimensional matrix features, and then performing denoising optimization on the multidimensional matrix features based on an autoencoder to obtain multidimensional feature vectors.

[0076] For example, the autoencoder structure includes a three-layer fully connected network, consisting of an input layer, hidden layers, and an output layer. The input layer acquires concatenated multidimensional matrix features; for example, the total feature dimension of the multidimensional matrix is ​​120. The hidden layers output features in 64-dimensional to 32-dimensional sequences. The output layer has a 16-dimensional output. This results in a low-dimensional, denoised, and optimized multidimensional feature vector that retains the core features.

[0077] For example, in step S200, 5% random noise can be added to the multidimensional matrix features to obtain simulated multidimensional matrix features; then, based on an autoencoder, a reconstruction loss function is used to train the simulated multidimensional matrix features to obtain a denoised and optimized multidimensional feature vector; the autoencoder includes multiple fully connected layers, and the dimension of the multidimensional feature vector is lower than the dimension of the multidimensional matrix features. In this embodiment, randomly adding 5% Gaussian noise (mean = 0, variance = 0.1) can simulate data loss or errors in real-world scenarios (such as missing data from a student's assignment).

[0078] In one example, the reconstruction loss function includes MSE and mean squared error, which is the error between the "reconstructed result of the input with added noise" and the "original noise-free features". The specific loss function formula is as follows:

[0079] ;

[0080] In the formula, Original features For reconstructing features.

[0081] Training parameter settings: Optimizer = Adam, learning rate = 1e-4, batch size = 32, training epochs = 100, adopt early stopping strategy (stop training when the reconstruction loss decreases by <1e-5 for 5 consecutive epochs to avoid invalid iterations).

[0082] S300, based on the bidirectional long short-term memory-attention mechanism network, obtains multidimensional temporal embedding feature vectors and subject-time attention weight matrices.

[0083] In some possible implementations, such as Figure 4 As shown, step S300 may include the following sub-operations from steps S310 to S330:

[0084] S310. According to the time step sequence, the multidimensional feature vector is processed based on the bidirectional long short-term memory network to obtain the multidimensional temporal embedding feature vector.

[0085] In some possible implementations, the multidimensional temporal embedding feature vector includes forward temporal embedding features and backward temporal embedding features. The bidirectional long short-term memory network includes an input layer and hidden layers. Temporal information is embedded into the multidimensional feature vector through the input layer, and the multidimensional feature vector is arranged according to time steps based on the temporal information. Forward temporal embedding features are obtained by training with forward long short-term memory based on the hidden layers, and backward temporal embedding features are obtained by training with backward long short-term memory based on the hidden layers. The forward and backward temporal embedding features are concatenated to obtain the multidimensional temporal embedding feature vector.

[0086] For example, a two-layer bidirectional LSTM network adapts to the "bidirectional dependency" of time-series data. The input dimension is a 16-dimensional pre-trained embedded multi-dimensional feature vector, arranged in time steps to form a "time-series sequence," for example, embedding 1 represents September, embedding 2 represents October, and embedding 3 represents November. The hidden layer dimension is 32-dimensional, where the forward LSTM captures the "history → future" dependency (e.g., September's math affects October's physics); the backward LSTM captures the "future → history" dependency (e.g., November's grades reflect September's weaknesses), and the bidirectional outputs are concatenated to form a 64-dimensional time-series feature.

[0087] For example, L2 regularization (weight decay = 1e-5) is added to a bidirectional long short-term memory network to prevent overfitting.

[0088] S320. Based on the additive attention mechanism, the multi-dimensional temporal embedding feature vector is processed to obtain the subject-time attention weight matrix;

[0089] In some possible implementations, the subject-time attention weight matrix is ​​calculated based on the following formula:

[0090] ;

[0091] In the formula, Q is the feature vector of the target subject, serving as the query vector in the attention mechanism; K is the feature vector of all subjects, serving as the key vector in the attention mechanism; V is the temporal feature, serving as the value vector in the attention mechanism; and For trainable weights, The dimension is K; the softmax() function is a normalization function that converts scores into weights in the attention mechanism.

[0092] S330. Attention weighting of multidimensional temporal embedding feature vectors based on subject-time attention weight matrix.

[0093] In this embodiment, an attention-weighted 64-dimensional multi-dimensional temporal embedding feature vector is output, which contains long-term dependency information. The resulting subject-time attention weight matrix is ​​also output, with the shape being the number of subjects × the number of time steps, such as 10 subjects × 6 months. A higher weight indicates a greater influence of the "subject-time" on the target subject.

[0094] For example, during model training, the attention mechanism model is optimized by defining the supervised task as follows: The task is "to predict the subject score at the T-th time step given the temporal embeddings of the previous T-1 time steps" (e.g., predicting the November physics score using data from September and October), ensuring the model learns the temporal patterns. The mean absolute error (MAE) is used as the loss function, which is more robust to outliers in this application scenario. The training parameters are configured as follows: optimizer = RMSprop, learning rate = 5e-5, batch size = 64, training epochs = 50, and an early stopping strategy (patience = 5, stopping when the validation set MAE increases for 5 consecutive epochs).

[0095] S400, Establish a dynamic graph neural network.

[0096] In some possible implementations, such as Figure 5 As shown, step S400 may include the following sub-operations from steps S410 to S420:

[0097] S410. Establish graph nodes for multiple disciplines, and construct a discipline graph neural network for students based on multiple graph nodes; obtain the initial features of the discipline nodes corresponding to each discipline based on the structural features and attention-weighted multidimensional temporal embedding feature vectors.

[0098] In step S410, when constructing the initial subject nodes, it is necessary to define the nodes and initialize their characteristics. Simultaneously, it is also necessary to determine the existence of edges between the nodes. For example, each subject is defined as a node. For instance, ten subjects would correspond to ten nodes.

[0099] Initialize node features: Concatenate the 16-dimensional pre-trained embedded multi-dimensional feature vector obtained in the above steps with the 1-dimensional subject difficulty coefficient extracted from the structural features to form a 17-dimensional initial feature vector.

[0100] Determining the existence of edges: When the mutual information value between two disciplines is >0.3, the mutual information is calculated using the structural features extracted in step 100. The mutual information measures the strength of the nonlinear association, and an edge is established. If there is no mutual information or the mutual information is ≤0.3, there is no edge, and the discipline is considered to be unrelated.

[0101] Configure initial edge weights: Use a weighted fusion of "Granger causality coefficient × 0.6 + mutual information value × 0.4". Granger causality measures the causality of "whether subject A affects subject B", and its weight is higher, which can avoid relying solely on correlation; for example, the Granger causality coefficient of mathematics and physics is 0.7 and the mutual information value is 0.6, so the initial weight = 0.7 × 0.6 + 0.6 × 0.4 = 0.66.

[0102] Finally, the graph structure of the graph neural network is constructed and stored. Specifically, an adjacency matrix is ​​used to store edge information, and a node feature matrix (10×17) is used to store node information. The format is adapted to the GNN input. The shape of the adjacency matrix is ​​10×10, and the elements are edge weights. If there is no association, the weight is 0.

[0103] S420. Based on the subject-time attention weight matrix, the edge weights between subjects in the subject graph neural network are dynamically updated to obtain a dynamic subject association graph and graph neural network attention weights. The dynamic subject association graph includes the subject node features and dynamic edge weights corresponding to each subject.

[0104] In some possible implementations, the dynamic update of a graph neural network includes updating node features and edge weights:

[0105] For node features: The multidimensional temporal embedding feature vector from step S330 is concatenated with the initial node features obtained in step S410, and then compressed using a fully connected network based on the ReLU activation function to obtain updated node features.

[0106] For edge weight updates: Set a temporal decay factor, and dynamically update the edge weights between disciplines in the subject graph neural network based on the temporal decay factor and the subject-time attention weight matrix.

[0107] For example, the formula for the time decay factor can be:

[0108] ;

[0109] In the formula, Here, t is the decay factor, t is the historical time step, T is the current time step, and e is a constant. The initial edge weights can be corrected based on the time-series decay factor. For example: updated weight = initial weight. The decay factor ensures that recent connections are stronger than those in the distant past. Furthermore, the subject-time attention weight matrix contains information on the strength of connections between subject pairs. Attention weight information can be extracted from this matrix, and the subject's attention weight information is dynamically multiplied by the edge weights to obtain updated weights. This method allows edge weights to focus on key subjects and key time points; for example, the edge weight of senior high school math to physics is higher than that of junior high school, avoiding the problem of static weights failing to adapt to changes in learning stages.

[0110] For example, when training the model, a Graph Attention Network (GAT) is used to adapt to dynamic changes in edge weights and strengthen the influence of important neighbor nodes through the attention mechanism. In terms of network model structure, one attention head layer (including 8 parallel attention heads to enhance robustness) can be set, the hidden layer dimension is 32-dimensional, the output layer dimension is 1-dimensional, and the output is the predicted value of the subject score.

[0111] S500, based on the subject-time attention weight matrix, dynamic subject association graph, and graph neural network attention weight, obtains the multi-subject association relationship of each student.

[0112] In some possible implementations, student analysis data for each student can be obtained based on a multilayer perceptron network (MLP) and the characteristics of subject nodes in a dynamic subject association graph and attention-weighted multidimensional temporal embedding feature vectors. The student analysis data includes grade prediction, student ability profile, and student group labels.

[0113] For example, the multilayer perceptron network includes a shared bottom layer, multiple subject performance prediction task layers, a student clustering task layer, an association labeling task layer, and an output layer. The shared bottom layer is used to obtain student ability embedding vectors based on subject node features and attention-weighted multidimensional temporal embedding feature vectors. Each subject performance prediction task layer corresponds to one subject and is used to obtain the predicted subject performance for that subject based on the student ability embedding vector. The student clustering task layer is used to obtain clustering features based on the student ability embedding vectors, and these clustering features are used to assign student clustering labels. The association labeling task layer is used to obtain subject association type labels for each student based on the student ability embedding vectors. The output layer is used to obtain student performance predictions, student ability profiles, and student clustering labels based on the predicted subject performance, clustering labels, and subject association type labels.

[0114] In this embodiment, based on the subject-time attention weight matrix and graph neural network attention weights, dual thresholds can be set to determine the association type. The association type can be divided into three categories: positive promoting association, negative restrictive association, and potential collaborative association. Specifically, a subject-time attention weight > 0.6 and a graph neural network attention edge weight > 0.5 can be considered a positive promoting association. For example, in the subject pair of mathematics and physics, the subject-time attention weight is 0.72 and the graph neural network attention edge weight is 0.66, which is considered positive. A subject-time attention weight < 0.2 and a graph neural network attention edge weight less than 0.2 can be considered a negative restrictive association. A subject-time attention weight between 0.3 and 0.6 and a graph neural network attention edge weight between 0.3 and 0.5 can be considered a potential collaborative association. A subject association type table can be output based on the judgment results.

[0115] For example, correlation heatmaps, SHAP value quantitative analysis reports, and correlation reports can also be output based on the correlation results.

[0116] For example, the local server 100 can also push related information to the client 300 based on the obtained multidisciplinary relationships. The client 300 can be a teacher's client, a student's client, or a parent's client.

[0117] For example, parents and teachers can then customize students' subsequent learning plans based on the push notification results.

[0118] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0119] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0120] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0121] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices, and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another device, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between devices or modules may be electrical, mechanical, or other forms.

[0122] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; that is, they may be located on one device or distributed across multiple devices. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0123] In addition, the functional modules in the various embodiments of this application can be integrated into one device, or each module can exist physically separately, or two or more modules can be integrated into one device.

[0124] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented using software programs, implementation can be, in whole or in part, in the form of a computer program product. This computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, Digital Subscriber Line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device containing one or more servers, data centers, etc., that can be integrated with the medium. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0125] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based student multi-subject score correlation analysis method, characterized in that, The method comprises: Collecting multi-source discipline data of students, the multi-source discipline data comprising structured data, semi-structured data and unstructured data; extracting time sequence features and structure features from the structured data and semi-structured data, and extracting semantic features from the unstructured data; Splicing the time sequence features, the structure features and the semantic features to obtain a multi-dimensional matrix feature, and performing denoising optimization on the multi-dimensional matrix feature based on a self-encoder to obtain a multi-dimensional feature vector; Processing the multi-dimensional feature vector based on a bidirectional long short-term memory network according to time step columns to obtain a multi-dimensional time sequence embedding feature vector, the multi-dimensional time sequence embedding feature vector comprising forward time sequence embedding features and reverse time sequence embedding features; performing operation on the multi-dimensional time sequence embedding feature vector based on an additive attention mechanism to obtain a discipline-time attention weight matrix; and performing attention weighting on the multi-dimensional time sequence embedding feature vector based on the discipline-time attention weight matrix; Establishing graph nodes for multiple disciplines, constructing a discipline graph neural network for a student based on the multiple graph nodes, obtaining discipline node initial features corresponding to each discipline based on the structure features and the multi-dimensional time sequence embedding feature vector after attention weighting, dynamically updating edge weights between disciplines of the discipline graph neural network based on the discipline-time attention weight matrix to obtain a dynamic discipline correlation graph and a graph neural attention weight, the dynamic discipline correlation graph comprising discipline node features and dynamic edge weights corresponding to each discipline; Obtaining a multi-discipline correlation relationship of each student based on the discipline-time attention weight matrix, the dynamic discipline correlation graph and the graph neural attention weight. 2.The deep learning based student multi-subject performance correlation analysis method according to claim 1, wherein, The method of obtaining a multi-discipline correlation relationship of each student based on the discipline-time attention weight matrix, the dynamic discipline correlation graph and the graph neural attention weight comprises: Judging an association label type between two disciplines in each discipline pair based on the discipline-time attention weight matrix and the graph neural attention weight, and obtaining a discipline association type table according to the association label types of all discipline pairs, the association label type comprising a forward promotion association, a negative inhibition association and a potential coordination association; Outputting the multi-discipline correlation relationship based on the discipline association type table, the multi-discipline correlation relationship having a form of at least one of an association heat map, a SHAP value quantitative analysis report and an association report. 3.The deep learning based student multi-disciplinary performance correlation analysis method of claim 2, wherein, The method of obtaining a multi-discipline correlation relationship of each student based on the discipline-time attention weight matrix, the dynamic discipline correlation graph and the graph neural attention weight further comprises: Obtaining student analysis data of each student based on a multi-layer perception network according to the discipline node features in the dynamic discipline correlation graph and the multi-dimensional time sequence embedding feature vector after attention weighting, the student analysis data comprising a score prediction, a student ability portrait and a student grouping label. 4.The deep learning based student multi-disciplinary performance correlation analysis method of claim 3, wherein, The multi-layer perception network comprises a shared bottom network layer, a plurality of discipline score prediction task layers, a student grouping task layer, an association label task layer and an output layer, wherein The shared bottom network layer is configured to obtain a student ability embedding vector based on the subject node features and the attention-weighted multi-dimensional time sequence embedding feature vector; Each of the subject performance prediction task layers corresponds to a subject, and each of the subject performance prediction task layers is configured to obtain a subject predicted performance of the corresponding subject based on the student ability embedding vector; The student grouping task layer is configured to obtain clustering features based on the student ability embedding vector, and the clustering features are used to divide the student grouping labels for students; The association label task layer is configured to obtain a subject association type label of each student based on the student ability embedding vector; The output layer is configured to obtain the performance prediction, the student ability portrait, and the student grouping label of a student based on the subject predicted performance, the grouping label, and the subject association type label. 5.The deep learning based student multi-disciplinary performance correlation analysis method of claim 1, wherein, The denoising optimization of the multi-dimensional matrix feature based on the autoencoder obtains a multi-dimensional feature vector, including: Adding 5% random noise to the multi-dimensional matrix feature to obtain a simulated multi-dimensional matrix feature; Training the simulated multi-dimensional matrix feature based on the autoencoder using a reconstruction loss function to obtain the denoising-optimized multi-dimensional feature vector; the autoencoder includes multiple fully connected layers, and the dimension of the multi-dimensional feature vector is lower than that of the multi-dimensional matrix feature. 6.The deep learning based student multi-disciplinary performance correlation analysis method of claim 1, wherein, The bidirectional long short-term memory network includes an input layer and a hidden layer; the multi-dimensional feature vector is processed based on the bidirectional long short-term memory network to obtain a multi-dimensional time sequence embedding feature vector, including: Embedding time information in the multi-dimensional feature vector through the input layer, and arranging the multi-dimensional feature vector by time steps based on the time information; Using forward long short-term memory training based on the hidden layer to obtain forward time sequence embedding features, and using reverse long short-term memory training based on the hidden layer to obtain reverse time sequence embedding features; Concatenating the forward time sequence embedding features and the reverse time sequence embedding features to obtain the multi-dimensional time sequence embedding feature vector. 7.The deep learning-based student multi-disciplinary performance correlation analysis method of claim 6, wherein, The multi-dimensional time sequence embedding feature vector is operated based on the additive attention mechanism to obtain a subject-time attention weight matrix, including: The subject-time attention weight matrix is operated based on the following formula: ； where Q is the characteristic vector of the target discipline as the query vector in the attention mechanism; K is the characteristic vector of all disciplines as the key vector in the attention mechanism; V is the time sequence feature as the value vector in the attention mechanism; in and is a trainable weight, is the dimension of K; the softmax() function is a normalization function for converting scores into weights in the attention mechanism. 8.The deep learning based student multi-disciplinary performance correlation analysis method of claim 1, wherein, The edge weights between subjects of the subject graph neural network are dynamically updated based on the subject-time attention weight matrix, including: Setting a time sequence decay factor, and dynamically updating the edge weights between subjects of the subject graph neural network based on the time sequence decay factor and the subject-time attention weight matrix.

9. The deep learning based student multi-disciplinary performance correlation analysis method according to any one of claims 1-8, wherein, The method further includes: Using Hadoop HDFS storage technology to divide storage partitions according to data structure types, and storing the structured data, the semi-structured data, and the unstructured data separately in the corresponding partitions according to different times. 10.A deep learning based student multi-disciplinary performance correlation analysis system, characterized in that, The system includes a local server, a database, and a client; wherein the server includes: The data collection module is configured to collect multi-source discipline data of students from the database, the multi-source discipline data including structured data, semi-structured data and unstructured data; extract time sequence features and structure features from the structured data and the semi-structured data, and extract semantic features from the unstructured data; The self-encoding module is configured to splice the time sequence features, the structure features and the semantic features to obtain multi-dimensional matrix features, and perform denoising optimization on the multi-dimensional matrix features based on a self-encoder to obtain a multi-dimensional feature vector; The bidirectional long short-term memory network-attention mechanism module is configured to process the multi-dimensional feature vector based on a bidirectional long short-term memory network in a time step column to obtain a multi-dimensional time sequence embedding feature vector, the multi-dimensional time sequence embedding feature vector including forward time sequence embedding features and reverse time sequence embedding features; perform operation on the multi-dimensional time sequence embedding feature vector based on an additive attention mechanism to obtain a discipline-time attention weight matrix; and perform attention weighting on the multi-dimensional time sequence embedding feature vector based on the discipline-time attention weight matrix; The dynamic graph neural network module is configured to establish graph nodes for multiple disciplines respectively, construct a discipline graph neural network for a student based on the multiple graph nodes, obtain discipline node initial features corresponding to each discipline based on the structure features and the multi-dimensional time sequence embedding feature vector after attention weighting, dynamically update edge weights between disciplines of the discipline graph neural network based on the discipline-time attention weight matrix to obtain a dynamic discipline correlation graph and a graph neural attention weight, and the dynamic discipline correlation graph including discipline node features and dynamic edge weights corresponding to each discipline; The correlation extraction network module is configured to obtain multi-discipline correlation relationships of each student based on the discipline-time attention weight matrix, the dynamic discipline correlation graph and the graph neural attention weight, and push the multi-discipline correlation relationships to the client.

Images