A multi-source heterogeneous educational data driven student intelligent evaluation method

Abstract

The application discloses a multi-source heterogeneous education data driven student intelligent evaluation method, and relates to the technical field of artificial intelligence. The method aims to solve the problems of data semantic fragmentation, uninterpretable model and application deviation from the essence of education in the prior art. The method constructs a three-layer architecture: first, based on the educational ontology, the multi-source heterogeneous data is subjected to semantic unified modeling to form a student knowledge graph; second, an interpretable multi-modal model integrating educational prior knowledge is designed to realize academic risk early warning and ability portrait, and to synchronously generate attribution explanation; finally, through a man-machine collaborative mechanism, an operable developmental intervention suggestion is output, and a privacy protection and algorithm fairness review mechanism is embedded. The application realizes the organic unification of data deep fusion, decision transparency and credibility, and education value regression.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and specifically relates to a student intelligent evaluation method driven by multi-source heterogeneous educational data. Background Technology

[0002] With the deepening of digital transformation in education, building intelligent student evaluation systems based on multi-source heterogeneous data has become an important development direction in the field of educational technology. Driven by the dual concepts of "educational modernization" and "personalized instruction," utilizing multi-dimensional data such as academic affairs systems, learning behavior logs, classroom videos, homework texts, and psychological assessments to conduct dynamic, comprehensive, and accurate integrated evaluations of students not only helps improve the timeliness and pertinence of teaching interventions but also provides technical support for educational equity and personalized development. Against this backdrop, intelligent evaluation methods integrating artificial intelligence, big data analysis, and educational science have emerged and are gradually evolving from single-indicator assessments to multimodal, full-process, and developmental evaluations.

[0003] Existing technical solutions generally adopt a basic paradigm of data collection—feature extraction—model prediction. This involves integrating raw data from different information systems to construct machine learning or deep learning models to generate student ability profiles or academic risk warnings. Such methods have demonstrated some effectiveness in specific scenarios, particularly in processing structured academic data (such as grades and attendance), where a relatively mature technical path has been established. Some advanced systems further incorporate natural language processing and computer vision technologies, attempting to analyze unstructured text feedback or classroom behavior videos to expand evaluation dimensions. These explorations, to some extent, alleviate the problems of traditional evaluation relying on subjective experience and having limited coverage, demonstrating the initial success of technology-enabled educational evaluation.

[0004] However, as application scenarios place higher demands on the accuracy, fairness, and development orientation of evaluation results, existing methods reveal deep-seated structural contradictions at the principle level. Specifically, the core dilemma does not stem from the insufficient performance of a single technical module, but from the lack of a coordinated and consistent logical closed loop between the three levels of data, model, and application. First, at the data level, although multi-source data is formally aggregated, the lack of a unified semantic framework means that data from systems such as academic management, online learning platforms, and campus IoT remain in a state of "physical concentration but logical fragmentation." For example, a student's "classroom participation" is reflected in the frequency of raising hands in a video analysis system, while in a learning platform it is reflected in the number of times they speak in the discussion area. The lack of semantic alignment at the ontological level prevents them from being effectively correlated, resulting in superficial fusion analysis that cannot support a deep characterization of students' cognitive status and development potential. Second, at the model level, even when using multimodal neural networks for joint modeling, their internal decision-making mechanisms often exhibit high nonlinearity and uninterpretability. When a model outputs a "high-risk warning," teachers cannot trace whether this conclusion is based on declining academic performance, abnormal social behavior, or changes in text sentiment. This "black box" characteristic not only weakens educators' trust in algorithmic results but also may lead to misjudgments due to attribution bias, thus affecting the scientific validity of intervention strategies. Furthermore, at the application level, existing systems generally tend to "emphasize prediction over guidance," simplifying the complex process of student development into static labels or risk levels, ignoring the fundamental purpose of educational evaluation: to promote growth. This transgression of technological logic may not only solidify students' identity perception and inhibit their potential for self-breakthrough but also raise ethical risks such as privacy leaks and algorithmic discrimination due to the centralized processing of large amounts of sensitive data. Especially in the absence of transparent review mechanisms, this can easily exacerbate educational inequality.

[0005] Ultimately, the root of these problems lies in the failure of existing technological architectures to establish an organic unity of "understandable data, interpretable models, and interventionizable applications." The essence of data silos is semantic deficiency, not fragmented storage; the crux of model black boxes lies in purely data-driven approaches detached from prior educational knowledge; and the risk of technological overreach stems from the neglect of the subjective role of "human beings" in system design. Therefore, how to construct an intelligent evaluation system with educational ontology as its semantic foundation, interpretable multimodal fusion models as its core, and developmental feedback through human-machine collaboration as its exit point, achieving synergistic optimization among the depth of data fusion, the transparency of model reasoning, and the return to educational value, has become a key challenge and an urgent technical problem facing those skilled in the art. Summary of the Invention

[0006] This invention provides a student intelligent evaluation method driven by multi-source heterogeneous educational data, aiming to resolve the structural contradictions in existing technologies, such as data semantic fragmentation, uninterpretable models, and application orientations deviating from the essence of education. To achieve the above-mentioned objectives, this invention constructs a three-layer technical architecture with educational ontology as the semantic foundation, an interpretable multimodal fusion model as the core, and developmental feedback through human-machine collaboration as the exit point, ensuring the synergistic unity of data fusion depth, model reasoning transparency, and the return to educational value.

[0007] The method first establishes a standardized, semantic data foundation. This foundation uses an education domain ontology to uniformly model and structure multi-source heterogeneous educational data. The education domain ontology defines core concepts such as students, courses, ability dimensions, behavioral events, and evaluation indicators, along with their interrelationships, forming a knowledge system with formal semantic constraints. Structured data from the academic affairs management system, behavioral logs from online learning platforms, classroom video streams, homework texts, and psychological assessment questionnaires are all converted into instantiated triples conforming to the education domain ontology through preset semantic mapping rules. These semantic mapping rules are jointly formulated by domain experts and knowledge engineers to ensure logical alignment of data from different sources at the ontology level. For example, the "number of times hands are raised" event extracted by the classroom video analysis module and the "number of times speaking in the discussion area" event recorded by the learning platform are both mapped to the "classroom participation behavior" class in the ontology, and cross-modal associations are achieved through attributes such as timestamps, course units, and teacher identifiers. All semantically transformed triple data are stored in a graph database, forming a dynamically updated student knowledge graph. The graph database adopts an attribute graph model, where nodes represent entities or events and edges represent semantic relationships, supporting deep association analysis and data tracing based on path queries.

[0008] Furthermore, this invention develops an accurate and interpretable multimodal fusion model. This model takes the student knowledge graph as input and combines it with original multimodal features to perform multi-task joint prediction and attribution explanation. The original multimodal features include: facial expression sequences, posture change trajectories, and vocal emotion features extracted from classroom video streams; semantic vectors, logical coherence indices, and error pattern encodings extracted from homework texts; and clickstream sequences, dwell time distributions, and resource access paths extracted from behavior logs. The multimodal fusion model comprises three functional sub-modules: a multimodal feature encoder, an educational prior knowledge guidance layer, and an interpretable decision outputter.

[0009] The multimodal feature encoder processes each modal input separately, generating high-dimensional feature representations. The visual modality employs a joint modeling approach combining a 3D convolutional neural network and a temporal attention mechanism to capture the temporal evolution of classroom behavior. The text modality uses a pre-trained language model fine-tuned to output context-aware semantic embeddings. The behavior log modality uses a graph neural network to perform message passing on the student-resource interaction graph, generating behavioral intent representations. After normalization, the features from each modality are fed into a hybrid attention fusion unit, which calculates the correlation weights between modalities and aggregates them to generate a unified student state vector.

[0010] The educational prior knowledge guidance layer formalizes educational theories into constraints and embeds them into the model reasoning process. Specifically, it introduces the Q-matrix from cognitive diagnostic theory as a priori knowledge link between ability and knowledge point. The rows of the Q-matrix correspond to student ability dimensions, the columns to course knowledge points, and the element values ​​represent the necessity of a certain ability for mastering a certain knowledge point. During model training, the student state vector and the Q-matrix are multiplied by a tensor to generate a probability distribution of knowledge point mastery. This distribution serves as an intermediate supervision signal, guiding the multimodal feature encoder to learn representations consistent with educational goals. Simultaneously, a regularization term based on the Q-matrix is ​​added to the loss function to constrain the model output to conform to the laws of cognitive development.

[0011] The interpretable decision outputter, based on student state vectors and knowledge point mastery distribution, performs two core tasks: academic risk warning and ability profile generation. For the academic risk warning task, a gated recurrent unit is used to model the historical state vector sequence to predict the probability of a decline in grades within a future time period; when this probability exceeds a preset threshold, an warning is triggered. For the ability profile task, a multi-head self-attention mechanism is used to perform fine-grained scoring on each ability dimension. The key innovation lies in the fact that the interpretable decision outputter simultaneously generates an attribution explanation report. This report calculates the contribution of each original input feature to the final decision through backpropagation and, combined with the semantic path in the educational ontology, maps high-contribution features to explanatory terms with clear educational significance. For example, if the warning is mainly driven by text sentiment features, the attribution explanation report points out that "the frequency of negative emotion words has increased significantly in recent assignment texts, which is associated with the mental health dimension"; if the "logical reasoning" score in the ability profile is low, it traces back to instances of broken argument structures in specific assignments. All attribution results are presented in natural language form, accompanied by links to original data fragments and ontology paths, ensuring that they are verifiable and understandable for teachers.

[0012] In a preferred embodiment of the present invention, the training of the multimodal fusion model adopts a two-stage strategy. The first stage involves end-to-end pre-training on the anonymized aggregated dataset to optimize the overall architecture parameters. The second stage involves personalized fine-tuning on local devices using a federated learning framework, where each school node only uploads the model gradients rather than the original data, and the central server aggregates the gradients and updates the global model, thereby improving the model's generalization ability while protecting data privacy.

[0013] Furthermore, this invention establishes a human-machine collaborative, development-oriented evaluation and feedback mechanism. Based on the principle of "artificial intelligence assistance, human leadership," this mechanism transforms model outputs into actionable educational intervention suggestions, rather than static labels. The system automatically generates personalized learning path recommendations based on academic risk warnings and ability profiles. These recommendations include three categories: supplementary learning resource sequences targeting weak knowledge points, peer collaboration matching suggestions based on social behavior analysis, and psychological support intervention prompts based on psychological state assessments. All recommendations are labeled with confidence levels and sources for teacher review and adjustment.

[0014] Teachers receive system suggestions through a dedicated interactive interface and can perform the following actions: confirm adoption, modify parameters and regenerate, or completely reject and manually develop an intervention plan. Each teacher's action is recorded as a new feedback event and added to the student's knowledge graph, forming a closed-loop learning cycle. The system regularly analyzes teacher adoption rates and intervention effects, dynamically adjusting the model's output strategy to ensure that technical suggestions remain aligned with educational practice.

[0015] To mitigate algorithmic bias and ethical risks, this invention incorporates a dual safeguard mechanism. The first is a privacy protection mechanism: all sensitive personal information undergoes differential privacy perturbation or k-anonymization at the collection end; multimodal feature extraction and model inference are completed at edge computing nodes, with only encrypted intermediate features uploaded to the central server; student nodes in the graph database use virtual identifiers, and their real identity information is stored by an independent identity management service with strict access control. The second is an algorithmic fairness review mechanism: the system has a built-in bias detection module that periodically calculates the differences in warning trigger rates among different genders, ethnicities, and regional groups at the same ability level; when the difference exceeds a preset tolerance threshold, model updates are automatically frozen, and a manual audit process is triggered. Auditors can access the complete decision-making chain and training data distribution to locate the source of bias and correct ontology rules or training samples.

[0016] In summary, this invention achieves deep correlation and traceability of multi-source heterogeneous educational data by constructing a semantically unified data foundation; it ensures the accuracy of evaluation results and the transparency of decision-making by embedding an interpretable multimodal fusion model with prior educational knowledge; and it strictly limits the technical tools to an auxiliary role in education through a human-machine collaborative feedback mechanism and an embedded ethical protection system, effectively preventing algorithmic overreach and educational distortion. The technical solution of this invention comprehensively covers the entire chain of data collection, semantic modeling, feature fusion, decision reasoning, attribution explanation, intervention generation, privacy protection, and bias review, forming a logically closed-loop, value-aligned, and engineering-feasible intelligent evaluation system, fundamentally resolving the structural contradictions of existing technologies at the data, model, and application levels.

[0017] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, embodiments of the present invention are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic diagram of the overall architecture of a student intelligent evaluation method driven by multi-source heterogeneous educational data according to the present invention.

[0020] Figure 2 This is a schematic diagram of the semantic modeling and data fusion process for constructing a student knowledge graph based on an ontology in the education field in this invention.

[0021] Figure 3 This is a structural block diagram of the interpretable multimodal fusion model in this invention.

[0022] Figure 4 This is a schematic diagram of the workflow of the human-machine collaborative developmental feedback mechanism in this invention.

[0023] Figure 5 This is a schematic diagram of the dual ethical protection mechanism of embedded privacy protection and algorithm fairness review in this invention. Detailed Implementation

[0024] This invention provides a student intelligent evaluation method driven by multi-source heterogeneous educational data. Its technical implementation encompasses a complete technical chain, from raw educational data collection, semantic modeling, multimodal feature fusion, interpretable decision reasoning to human-computer collaborative feedback and ethical safeguards (overall architecture as follows). Figure 1(As shown). The technical solution of the present invention will be systematically and structurally described in detail below with reference to the accompanying drawings and specific engineering implementation details, to ensure that those skilled in the art can fully reproduce all the technical contents of the present invention based on this specification.

[0025] First, during the data acquisition and preprocessing phase, the system accesses multi-source heterogeneous educational data streams. These multi-source heterogeneous educational data include, but are not limited to: structured data output from the academic affairs management system (such as student registration information, course registration records, exam scores, and course selection history); behavioral logs generated by the online learning platform (such as page click sequences, video playback progress, quiz submission timestamps, and resource download records); classroom audio and video streams (including synchronous recordings of teacher lectures and student interactions); student-submitted assignment texts (including handwritten scans digitized using OCR); and structured results from psychological assessment questionnaires (such as anxiety scale scores and self-efficacy scores). Before entering the system, all raw data undergoes format validation and preliminary cleaning via standardized interface protocols, removing invalid or incorrectly formatted data items, and performing context-based interpolation to complete missing fields.

[0026] Furthermore, the system constructs a semantic data foundation centered on the ontology of the education domain (the semantic modeling and data fusion process is as follows...). Figure 2 (As shown). The education domain ontology is formally defined using the OWL (Web Ontology Language). Its top-level classes include core concepts such as "Student," "Course," "Ability Dimension," "Behavioral Event," "Evaluation Indicator," and "Teaching Resources." Semantic relationships are established between these core concepts through object attributes. For example, there is an "enrolledIn" relationship between "Student" and "Course," an "indicates" relationship between "Behavioral Event" and "Ability Dimension," and a "covers" relationship between "Course" and "Knowledge Point." The ontology also defines data attributes to describe the quantitative characteristics of entities, such as "timestamp" and "duration" for "Behavioral Event," and "gradeLevel" and "schoolID" for "Student."

[0027] All raw data is transformed into RDF (Resource Description Framework) triples conforming to the ontology of the education domain through preset semantic mapping rules. These semantic mapping rules are collaboratively developed by education experts and knowledge engineers and deployed as a rule engine at the data access layer. For example, for the behavior logs of an online learning platform, the system defines the following mapping rule: if a log entry contains the event type "discussion_post", then a triple <student ID, performs, classroom participation behavior> is generated, with the attributes "platform=MOOC (Massive Open Online Course)", "courseUnit=U102", and "timestamp=2023-10-05T14:30:00Z". For the classroom video stream, the video analysis module first uses the YOLOv5 object detection model (a real-time object detection algorithm model used to quickly identify face regions in images) to locate the student's face region, and then uses the OpenFace toolkit (an open-source face behavior analysis toolkit) to extract facial action unit sequences. Simultaneously, the MediaPipe Pose model (a machine learning solution model developed by Google for high-fidelity human pose tracking) is used to extract the upper body keypoint trajectory. The speech signal is processed using the Wav2Vec 2.0 model (a self-supervised speech pre-training model) to extract acoustic embedding vectors. After the above multimodal features are identified as a "raise hand to ask a question" event by a behavior recognition model (such as an LSTM (Long Short-Term Memory) classifier), the system generates a triple <student ID, performs, classroom participation behavior>, and attaches the attributes "modality=video", "classroom=C301", "teacherID=T205", and "timestamp=2023-10-05T10:15:22Z".

[0028] All generated triples are written in real-time to a student knowledge graph built on the Neo4j graph database. This graph database uses an attribute graph model, where node types correspond to classes in the ontology, and edge types correspond to object attributes in the ontology. Each node contains several attribute key-value pairs, corresponding to data attributes in the ontology. For example, a "classroom participation behavior" node might contain the attributes: {eventID: "E789", timestamp: "2023-10-05T10:15:22Z", courseUnit: "U102", modality: "video", confidence: 0.92}. Edges carry relation type and direction information, such as (student S123)-[:performs]->(event E789). This knowledge graph supports the Cypher query language (a declarative graph query language specific to the Neo4j graph database), allowing cross-modal association queries through path patterns, for example:

[0029] To query "all events related to 'critical thinking' ability for student S123 in course U102", you can use the following expression: MATCH (s:Student {id: 'S123'})-[:performs]->(e:Event)-[:indicates]->(c:Competency {name: 'critical thinking'}) WHERE e.courseUnit = 'U102' RETURN e.

[0030] In a preferred embodiment of the present invention, the student knowledge graph supports a dynamic update mechanism. When new data flows in, the system first checks whether there are existing triples with the same subject-predicate combination; if they exist and their timestamps are close, a conflict resolution strategy is executed (e.g., selecting the one with higher confidence); if they do not exist, new nodes and edges are directly inserted. In addition, the system periodically performs graph consistency checks, detecting logical contradictions based on axioms in the ontology (e.g., "a student cannot be in both 'absent' and 'actively participating' states in the same course unit") and triggering a manual review process.

[0031] After completing semantic modeling, the system calls an interpretable multimodal fusion model to assess student status (model structure as follows). Figure 3 (As shown). This model takes the historical event sequences extracted from the student's knowledge graph and the corresponding original multimodal features as input. Specifically, the input data is divided into two categories: one is the structured event sequences extracted from the graph, which are sorted by time to form an event stream; the other is the original multimodal feature tensors associated with these events.

[0032] The multimodal feature encoder comprises three parallel sub-networks. The visual sub-network receives classroom video clips (sampled at 5fps, each clip lasting 30 seconds), first extracting spatiotemporal features using an I3D (Dilated 3D Convolutional Neural Network) model, outputting a 1024-dimensional video embedding. This embedding sequence is then input into a Transformer encoder, where positional encoding employs learnable temporal positional embeddings to capture the temporal evolution patterns of behavior. The text sub-network receives homework text (preprocessed by sentence and word segmentation), generating a classification tag vector for each sentence using a fine-tuned BERT-base model (a bidirectional encoder representation model based on a transformer), and then aggregating it into a document-level semantic vector with a dimension of 768 using a BiLSTM (Bidirectional Long Short-Term Memory) network. The behavior log sub-network models the student-resource interaction sequence as a temporal graph, with nodes representing resources (such as videos, exercises, and documents) and edge weights representing access frequencies. Two layers of GraphSAGE (graph sampling and aggregation algorithm) message passing are run on this graph to generate resource intent embeddings, which are then processed through attention pooling to obtain a student behavior intent vector with a dimension of 512.

[0033] After layer normalization, the feature vectors of each modality are input into the hybrid attention fusion unit. This unit first calculates the inter-modal similarity matrix:

[0034]

[0035] in Indicates the first The modality and the first The correlation score between modalities and The first The and the first Feature vectors of each modality , For learnable projection matrices, The feature dimension is used. Then, attention weights are generated using the Softmax function. And by weighted summation, a unified student state vector is obtained. :

[0036]

[0037] The educational prior knowledge guidance layer incorporates the Q-matrix from cognitive diagnostic theory. (Q-matrix) Defined Each capability dimension and The necessity relationship between the knowledge points, among which Representational ability Mastering knowledge points It is necessary. During the training phase, the student state vector Perform tensor operations with the Q matrix to generate a probability distribution of knowledge point mastery. :

[0038]

[0039] in For the Sigmoid function, For learnable weight matrix, This represents element-wise multiplication. It is a vector consisting entirely of 1s. This distribution... As an intermediate supervisory signal, a binary cross-entropy loss is calculated using the actual knowledge point mastery labels (from unit tests). Simultaneously, the total loss function includes a regularization term based on the Q-matrix.

[0040]

[0041] in For the capability dimension The predicted score, Let k be the probability of mastering knowledge point k. This is the regularization coefficient. This regularization term penalizes situations where "a certain knowledge point is not mastered, but the necessary ability is overestimated," forcing the model to follow the logic of cognitive development.

[0042] The interpretable decision outputter is based on a sequence of student state vectors. It performs two tasks. For academic risk early warning, a single-layer GRU (Gated Recurrent Unit) is used. Model and output the probability of a decline in grades in a future time period (such as next month). :

[0043]

[0044] in, Let be the hidden state at time t. and These represent the weights and biases of the output layer, respectively. An alert is triggered when the value reaches a preset threshold, such as 0.7. For capability profile generation, the final state vector... Input a multi-head self-attention layer and calculate fine-grained scores for each capability dimension:

[0045]

[0046] The query matrix Key / value matrix It is obtained by encoding the ability description text (such as "can identify logical fallacies in arguments") using BERT. This is the scaling factor. The output is a score vector for each capability dimension. .

[0047] The key is that the system synchronously generates an attribution explanation report. This report calculates the pairs of each original input feature using an ensemble gradient method. or The contribution of the input feature. The baseline is (e.g., zero vector), then features The attribution value is:

[0048]

[0049] in This is the model's output function. High-contribution features are mapped to semantic paths in the educational ontology. For example, if the sentiment features (such as negative word frequency) of a certain assignment text are highly relevant to the semantic path... If the contribution is the highest, the system queries the event node for that assignment in the knowledge graph, following the path (Assignment) - [:submittedBy] -> (Student) - [:hasAssessment] -> (Mental Health Dimension), and generates an explanation statement: "The frequency of negative emotion words has increased significantly in recent assignment texts, which is associated with the mental health dimension." This explanation includes the original text fragment (such as "I feel like I'll never learn this") and the graph path link for teachers to verify.

[0050] In a preferred embodiment of the present invention, the multimodal fusion model employs a two-stage training strategy. In the first stage, end-to-end pre-training is performed on a central server using anonymized aggregated datasets from multiple partner schools. Data anonymization includes: removing direct identifiers such as names and ID numbers; generalizing encoding of schools and classes; and adding Laplace noise to continuous variables to satisfy differential privacy. In the second stage, each school node performs personalized fine-tuning on its local device using a federated learning framework. The specific process is as follows: the central server distributes global model parameters; each node calculates gradients using local student data; the gradients are encrypted (e.g., Paillier homomorphic encryption) and uploaded; the central server aggregates the gradients (using a federated averaging algorithm) and updates the global model. This process iterates until convergence, ensuring that the model is both generalizable and adaptable to local educational contexts.

[0051] Furthermore, the system establishes a developmental feedback mechanism for human-machine collaboration (workflow such as...). Figure 4 (As shown). After the model outputs an academic risk warning or ability profile, the intervention generation module automatically constructs a personalized learning path recommendation. This recommendation includes three categories: First, for knowledge points where the probability of mastery is below a threshold... For knowledge points with a score of 0.6 or higher, relevant micro-lecture videos, exercises, and reading materials are retrieved from the teaching resource database and sorted by cognitive difficulty to form a supplementary learning sequence; secondly, based on student social behavior analysis (such as the frequency of interaction in discussion forums and group project collaboration records), peer similarity is calculated (using cosine similarity), and 3-5 peers with complementary abilities or matching styles are recommended to form a learning group; thirdly, if the score in the mental health dimension is lower than 0.6, If the value is 0.4, a psychological support suggestion will be generated, such as "It is recommended to arrange an initial interview with a psychological counselor", and the contact information of the school's psychological service center will be attached.

[0052] All recommendations are labeled with confidence levels (based on model-predicted probabilities) and sources of information (e.g., "based on sentiment analysis of three assignments from October 2023"). Teachers receive these suggestions through a web-based interactive interface, which displays warning information, a competency radar chart, recommended content, and attribution explanations in a card-style layout. Teachers can perform three actions: clicking the "Accept" button, which adds the recommended content to the student's personal learning plan; clicking the "Adjust" button, which modifies the order of recommended resources, adds or removes peers, or adjusts the level of psychological intervention, and the system regenerates recommendations accordingly; and clicking the "Reject" button, which allows teachers to manually input intervention plans, such as "Arrange one-on-one tutoring twice a week." All actions are recorded as new feedback events and written into the student's knowledge graph, forming a (Teacher) - [:providesFeedback] -> (Intervention Event) relationship for subsequent model optimization.

[0053] The system periodically (e.g., monthly) analyzes teachers' operation logs and calculates the adoption rate of various types of suggestions. Compared with post-intervention effect indicators (such as the increase in the probability of mastering knowledge points) If a certain type of suggestion (e.g., 0.3) or If the value is 0.1, the model output strategy adjustment mechanism will be triggered, such as lowering the default confidence threshold of this type of suggestion, or adding more contextual information to the attribution explanation.

[0054] To mitigate ethical risks, this invention incorporates a dual safeguard mechanism, see [link / reference]. Figure 5 As shown. Regarding privacy protection mechanisms, all sensitive personal information is processed at the data collection end: identity information is hashed using SHA-256 (a 256-bit secure hash algorithm), and the first 16 bits are truncated as a virtual identifier; continuous behavioral data (such as dwell time) is processed using k-anonymization (k=50) to ensure that any record in the dataset has at least 49 indistinguishable neighbors; multimodal feature extraction and model inference are completed on edge computing nodes deployed in the school's computer lab, encrypting only intermediate features (such as student state vectors). The data is uploaded to the central server via the TLS 1.3 protocol (Transport Layer Security version 1.3); student nodes in the graph database do not store any real identity information, and their real identities are kept by an independent identity management service (deployed on the education bureau's intranet), requiring two-factor authentication and operation tracking for access.

[0055] Regarding the algorithmic fairness review mechanism, the system has a built-in bias detection module that calculates the differences in warning trigger rates among different groups (divided by gender, ethnicity, and place of origin) at the same ability level each month. Specifically, students are binned according to their ability profile scores (e.g., 0.0-0.2, 0.2-0.4, ..., 0.8-1.0), and the warning rate for each group is calculated within each bin. And calculate the maximum difference. .like If a preset tolerance threshold (e.g., 0.05) is set, model updates will be automatically frozen, and an alert will be sent to the auditor. The auditor can log into the dedicated audit platform to retrieve the complete decision chain (including original features, intermediate activation values, and attribution paths) and training data distribution statistics (e.g., sample proportions of each group, label distribution) to pinpoint the source of bias. If a bias is found in the ontology rules (e.g., "classroom participation" is overly associated with "communication skills" while ignoring other manifestations of introverted students), the semantic mapping rules in the ontology will be corrected; if an imbalance in the training samples is found, resampling or cost-sensitive learning strategies will be initiated.

[0056] This invention constructs a student intelligent evaluation system that is technically rigorous, educationally appropriate, and value-aligned by engineering four major technical modules: a semantic data foundation, an interpretable multimodal model, human-computer collaborative feedback, and ethical safeguards. All technical components are implemented using industrial-grade open-source frameworks (such as Neo4j graph database, PyTorch deep learning framework, and TensorFlow Federated learning framework), exhibiting good deployability and scalability. Those skilled in the art can reproduce and adapt the technical solutions of this invention in different educational scenarios based on this specification.

[0057] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A student intelligent evaluation method driven by multi-source heterogeneous educational data, characterized in that, include: A semantic data foundation based on an education domain ontology is established. This ontology defines the core concepts and their semantic relationships, including students, courses, competency dimensions, behavioral events, and evaluation indicators. The ontology is formally defined using the OWL language. Its top-level classes include students, courses, competency dimensions, behavioral events, evaluation indicators, and teaching resources. Semantic relationships between core concepts are established through object attributes, including the relationship between students and courses, behavioral events and competency dimensions, and courses and knowledge points. Quantitative characteristics of entities are described through data attributes, including the timestamps and durations of behavioral events, and the grade and school identifiers of students. Multi-source heterogeneous raw educational data from the academic affairs management system, online learning platform, classroom audio and video streams, homework texts, and psychological assessment questionnaires are processed by a rule engine deployed in the data access layer to execute preset semantic mapping rules. The raw educational data is converted into RDF triples, and the generated RDF triples are written in real time to a graph database constructed using an attribute graph model to form a student knowledge graph. In this graph, node types correspond to classes in the ontology, edge types correspond to object attributes in the ontology, and nodes contain key-value pairs of attributes corresponding to data attributes. The graph database supports cross-modal association queries and data tracing through a path pattern query language. Based on the student knowledge graph and the original multimodal features associated with events in the knowledge graph, an interpretable multimodal fusion model is used for student state assessment and decision reasoning. The interpretable multimodal fusion model includes a multimodal feature encoder, an educational prior knowledge guidance layer, and an interpretable decision outputter. The multimodal feature encoder processes the original features of the visual, textual, and behavioral log modalities respectively, and generates a unified student state vector through a hybrid attention fusion unit. The educational prior knowledge guidance layer embeds educational theoretical priors, formalized as Q-matrices, into the model training process, generates a knowledge point mastery probability distribution through tensor operations, and uses this as an intermediate supervision signal and regularization constraint to guide the model to learn representations that conform to the laws of cognitive development. The specific working process of the educational prior knowledge guidance layer includes: rows of the Q-matrix correspond to ability dimensions, columns correspond to knowledge points, and matrix element values ​​represent corresponding... The ability dimension determines whether mastering the corresponding knowledge point is necessary; during the model training phase, the student state vector is multiplied by the Q matrix using a tensor product operation, and a knowledge point mastery probability distribution is generated through an activation function; the knowledge point mastery probability distribution is used to calculate the loss with the actual knowledge point mastery labels from unit tests, and a regularization term based on the Q matrix is ​​added to the total loss function; the regularization term is used to penalize the situation where the predicted scores of all ability dimensions defined as necessary by the Q matrix are overestimated when a certain knowledge point is not mastered, thereby constraining the model output to conform to the logic of cognitive development; the interpretable decision outputter is used to perform academic risk warning and ability profile generation tasks based on the student state vector, and simultaneously generate an attribution explanation report; the attribution explanation report maps high-contribution features to educational natural language explanations by calculating the contribution of input features to the decision and combining the semantic path of the education domain ontology; A human-machine collaborative developmental feedback mechanism is established. Based on the academic risk warning results and ability profile results output by the explainable decision outputter, a personalized learning path recommendation is automatically generated, which includes a sequence of supplementary learning resources, peer collaboration matching suggestions, and psychological support intervention prompts. The recommendation and attribution explanation report are provided to teachers through a web-based interactive interface, and the teacher's adoption, adjustment, or rejection of the recommendation is received. The teacher's actions are recorded as new feedback events and written into the student knowledge graph to form a closed-loop learning process. The model output strategy is dynamically adjusted based on the teacher's operation log.

2. The student intelligent evaluation method driven by multi-source heterogeneous educational data according to claim 1, characterized in that, The student knowledge graph supports dynamic updates and consistency checks. When new data flows in, the system checks whether there are existing triples with the same subject-predicate combination. If they exist and the timestamps are similar, a conflict resolution strategy based on confidence is executed. If they do not exist, new nodes and edges are directly inserted. The system periodically performs graph consistency checks based on the axioms defined in the ontology, detects logical contradictions, and triggers a manual review process.

3. The student intelligent evaluation method driven by multi-source heterogeneous educational data according to claim 1, characterized in that, The multimodal feature encoder comprises three parallel sub-networks: The visual subnetwork is used to receive classroom video clips, extract spatiotemporal features through a 3D convolutional neural network, and input the output video embedding sequence into a Transformer encoder with learnable temporal position embeddings to capture the temporal evolution patterns of classroom behavior. The text sub-network receives the pre-processed assignment text, generates sentence-level semantic vectors using a fine-tuned pre-trained language model, and aggregates them into document-level semantic vectors through a bidirectional long short-term memory network. The behavior log subnetwork is used to model the student-resource interaction sequence as a time sequence graph, run a graph neural network on the time sequence graph to perform message passing, generate resource intent embeddings, and obtain student behavior intent vectors through attention pooling. The hybrid attention fusion unit is used to calculate the similarity matrix between the feature vectors of each modality, generate intermodal attention weights through the Softmax function, and perform weighted summation on the feature vectors of each modality to output the unified student state vector.

4. The student intelligent evaluation method driven by multi-source heterogeneous educational data according to claim 1, characterized in that, The specific methods for the interpretable decision output device to perform academic risk warning and competency profile generation tasks include: For academic risk warning, a gated loop unit is used to model the historical student state vector sequence and output the probability of grade decline within a specified future time period; when the probability of grade decline exceeds a preset first threshold, an academic risk warning is triggered. For ability profile generation, the current student state vector is input into a multi-head self-attention layer. The query matrix is ​​obtained by linear transformation of the student state vector, and the key matrix and value matrix are obtained by encoding the description text of each ability dimension through a pre-trained language model. Fine-grained score vectors for each ability dimension are output through attention calculation.

5. The student intelligent evaluation method driven by multi-source heterogeneous educational data according to claim 4, characterized in that, The specific methods for generating the attribution explanation report include: The contribution of each original input feature to the academic risk warning probability or the target ability dimension score is calculated using the integral gradient method; the integral gradient method obtains the attribution value of each feature by integrating the gradient of the model along the straight path between the input feature and the baseline feature. Features with a contribution level higher than a preset threshold are identified as high-contribution features, and the student knowledge graph is queried to obtain the semantic path of the entity or event node corresponding to the high-contribution feature in the education domain ontology. Based on the semantic path, high-contribution features are mapped into natural language explanation statements with clear educational meanings, and linked with original data fragments and graph path links to form the attribution explanation report.

6. The student intelligent evaluation method driven by multi-source heterogeneous educational data according to claim 1, characterized in that, The training of the interpretable multimodal fusion model employs a two-stage strategy: In the first stage, the model is pre-trained end-to-end on a central server using anonymized aggregated datasets from multiple schools. The anonymization process includes removing direct identifiers, generalizing the encoding of institutional information, and adding noise to continuous variables to meet differential privacy requirements. In the second phase, each school node uses the federated learning framework on its local device to fine-tune the model in a personalized way. The central server distributes global model parameters, each node uses local data to calculate gradients, uploads the encrypted gradients, the central server aggregates the gradients and updates the global model, and this process is iterated until the model converges.

7. The student intelligent evaluation method driven by multi-source heterogeneous educational data according to claim 1, characterized in that, The specific content of generating personalized learning path recommendations in the aforementioned human-machine collaborative developmental feedback mechanism includes: For knowledge points whose mastery probability is below a preset second threshold, relevant micro-lecture videos, exercises, and reading materials are retrieved from the teaching resource database and sorted by cognitive difficulty to form a supplementary learning resource sequence; Based on students’ interaction frequency in the discussion area and group project collaboration records, the similarity between students’ peers is calculated, and several peers with complementary abilities or matching learning styles are recommended to form peer collaboration matching suggestions. If the score of the mental health dimension in the ability profile is lower than the preset third threshold, a psychological support intervention prompt will be generated, which includes a suggestion to arrange an interview with a mental health counselor and provide the contact information of a mental health service center. Each item in the personalized learning path recommendation is labeled with the confidence level based on the model's predicted probability and the specific source of the evidence.

8. The student intelligent evaluation method driven by multi-source heterogeneous educational data according to claim 1, characterized in that, The method also includes an embedded ethical safeguard mechanism, which includes a privacy protection mechanism and an algorithm fairness review mechanism: The privacy protection mechanism specifically includes: hashing identity information at the data collection end to generate virtual identifiers, and k-anonymizing continuous behavioral data; completing multimodal feature extraction and model inference at edge computing nodes, and uploading only encrypted intermediate features to the central server; storing only student virtual identifiers in the graph database, while real identity information is stored by an independent identity management service and subject to strict access control. The algorithm's fairness review mechanism specifically includes: periodically dividing students into groups based on gender, ethnicity, and place of origin, and calculating the academic risk warning trigger rate for each group within bins of the same ability level; when the maximum difference in warning trigger rates between different groups exceeds a preset tolerance threshold, automatically freezing model updates and triggering a manual audit process; auditors can retrieve complete decision-making link data and training data distribution statistics to locate the source of bias and correct semantic mapping rules or adjust training samples.

Images