Education attribution analysis method and system based on dynamic causal network

Abstract

The invention relates to the technical field of educational data mining and causal inference, and discloses an educational attribution analysis method and system based on a dynamic causal network, and the method comprises the steps: collecting structured educational data and unstructured educational data; mapping the unstructured education data into a causal semantic feature vector containing numerical confidence, and splicing the causal semantic feature vector with the structured education data to form a semantic multi-modal feature vector; outputting candidate causal paths and confidence thereof through a prompt template to form an initial causal hypothesis set; performing conditional independence test on the initial causal hypothesis set by adopting a Hilbert-Schmidt independence criterion, and constructing a dynamic Bayesian network structure; performing incremental updating on the conditional probability table of the dynamic Bayesian network based on a sliding window mechanism; calculating a causal effect of an intervention variable on a result variable by using a backdoor adjustment method under the condition of controlling a hybrid factor; personalized intervention strategies of direct intervention, intermediary intervention and fundamental intervention are generated according to the causal effect, and executable natural language suggestions are output.

Description

Technical Field

[0001] This invention relates to the field of educational data mining and causal inference technology, and in particular to an educational attribution analysis method and system based on dynamic causal networks. Background Technology

[0002] With the deepening of educational informatization, schools have accumulated a large amount of structured data (grades, attendance, online behavior logs) and unstructured text data (teacher comments, student notes, and transcripts of classroom recordings). How to accurately identify the key factors affecting students' academic performance based on this data and generate feasible intervention plans is a core issue that smart education urgently needs to address.

[0003] The existing technologies mainly fall into the following three categories: 1. Prediction and early warning based on structured data, such as CN109710876A which uses random forests to integrate multi-semester grades and attendance, and achieves high-risk group early warning through SMOTE and OOB optimization. However, this method only processes numerical data and cannot utilize textual information such as teacher comments, resulting in a sparse feature space and insufficient interpretability.

[0004] 2. Expert-rule-driven Bayesian networks, WO202015678A1, propose a hybrid causal discovery mechanism of "rule constraints + statistical verification." Initial edges are generated by an expert-defined rule base, then filtered using Fisher-Z tests to finally output intervention strategies. Its drawbacks are: ① Rules require manual maintenance, leading to high deployment costs across campuses; ② It cannot parse implicit causal relationships in unstructured text, such as the causal chain of "strong logic → calculation error" implied in "active thinking but careless"; ③ Intervention suggestions remain at the superficial level of "increasing study time," failing to distinguish the root cause.

[0005] 3. Large-scale text analysis: US202100426A uses BERT+BiLSTM to extract features from educational texts, fuses structured data through gating cross-attention, and finally generates natural language strategies using LLM. However, it only performs correlation analysis and lacks causal modeling, making it difficult to answer counterfactual questions such as "If I practice two more questions every day, my score will improve by how many points?" Furthermore, the model is statically updated and cannot adapt to changes in education policies or curriculum guidelines.

[0006] Therefore, existing technologies generally suffer from drawbacks such as limited data dimensions, heavy reliance on human intervention, lack of causal modeling, poor interpretability, and shallow and static intervention strategies. Summary of the Invention

[0007] To solve the above problems, the technical solution adopted by the present invention is as follows: The educational attribution analysis method and system based on dynamic causal networks includes the following steps: Includes the following steps: Multimodal data acquisition is used to collect structured and unstructured educational data. The structured educational data includes at least student grades and attendance records, and the unstructured educational data includes at least teacher comments and class notes. Large-scale model semantic enhancement utilizes a large language model to map the unstructured educational data into causal semantic feature vectors containing numerical confidence, and concatenates them with the structured educational data to form semantically enhanced multimodal feature vectors. Causal hypothesis generation involves using prompt templates to drive the large language model to output candidate causal paths and their confidence levels, thereby forming an initial set of causal hypotheses. Statistical and semantic joint verification is performed by using the Hilbert-Schmidt independence criterion to test the conditional independence of the initial causal hypothesis set, retaining causal edges that simultaneously satisfy the statistical significance threshold and the semantic confidence threshold, and constructing a dynamic Bayesian network structure. Dynamic parameter updates are performed by incrementally updating the conditional probability table of the dynamic Bayesian network based on a sliding window mechanism to adapt to the time-varying characteristics of the educational scenario. The counterfactual reasoning step uses the backdoor adjustment method to calculate the causal effect of the intervention variable on the outcome variable while controlling for confounding factors; The three-tiered intervention strategy generation process generates personalized intervention strategies for direct intervention, mediated intervention, and fundamental intervention based on the causal effects, and outputs actionable natural language suggestions.

[0008] Furthermore, the large model semantic enhancement step specifically includes: Construct an education domain dictionary, map educational descriptive phrases in unstructured educational data to standard causal feature labels, and output causal semantic feature vectors containing numerical confidence based on the standard causal feature labels; An attention-weighted cross-modal fusion mechanism is adopted to dynamically and weight the causal semantic feature vector with the features of structured educational data to generate a unified multimodal feature vector.

[0009] Furthermore, the causal hypothesis generation includes inputting a set of factors of multimodal feature vectors into a prompt template, the prompt template being configured to guide the large language model to generate candidate causal paths in the form of cause and effect; assigning a semantic confidence score output by the large language model to each candidate causal path; and including candidate causal paths with semantic confidence scores higher than a first threshold into an initial causal hypothesis set.

[0010] Furthermore, in the statistical semantic joint verification, the HSIC statistic and the corresponding p-value are calculated for each candidate causal edge; if the p-value is less than 0.05 and the semantic confidence is greater than 0.8, the causal edge is retained; if the p-value is greater than or equal to 0.05 but the semantic confidence is greater than 0.9, the expert review interface is activated to decide whether to retain it.

[0011] Furthermore, in the dynamic parameter update, a time decay factor γ = 0.3 / week is set to exponentially decay the weights of historical data; when the cumulative number of newly collected data records reaches 200, the expectation-maximization algorithm is triggered to incrementally update the conditional probability table; when an educational policy change event is detected, a rapid reconstruction mode is initiated to complete the reconstruction of the network structure and parameters within 72 hours.

[0012] Furthermore, in the generation of the three-level intervention strategy, the direct intervention strategy is used for behavior adjustment that can be implemented immediately; the mediating intervention strategy is used for mediating variables that affect learning performance; the fundamental intervention strategy is used for deep-seated family or economic reasons; and the intervention strategies are prioritized using an effect-cost weighted algorithm, where the effect accounts for 70% and the implementation cost accounts for 30%.

[0013] Furthermore, it also includes implementation feedback and continuous optimization, collecting student performance data in real time through API interfaces after the strategy is implemented, calculating the difference between the actual improvement value and the predicted value every 30 days, and triggering the rule base update process when the difference exceeds 30%.

[0014] An educational attribution analysis system based on dynamic causal networks includes: A multimodal data acquisition module is used to acquire structured and unstructured educational data. The large model semantic enhancement module is used to transform unstructured educational data into causal semantic feature vectors and fuse them with structured educational data; The causal hypothesis generation module is used to drive the large language model to output candidate causal paths through prompt templates; The statistical semantic joint validation module is used to combine the Hilbert-Schmidt independence criterion test with semantic confidence to screen causal edges in order to construct a dynamic Bayesian network. The dynamic parameter update module is used for incremental updates of network parameters based on a sliding window mechanism. The counterfactual reasoning module is used to calculate the causal effect of the intervention variable on the outcome variable; The three-level intervention strategy generation module is used to output personalized strategies for direct, intermediate, and fundamental interventions. The implementation feedback and continuous optimization module is used to track the effect of the strategy and dynamically optimize the model.

[0015] The beneficial effects of this invention are: 1. Comprehensive data integration: The introduction of educational fine-tuning LLM transforms text into causal semantic features, which can capture implicit traits such as "carelessness" and "learning by analogy," thereby improving the utilization rate of information.

[0016] 2. Highly efficient causal discovery: The system automatically generates hypotheses and uses the Hilbert-Schmidt independence criterion for statistical verification, reducing manual rule configuration by 90%; it also performs semantic and data checks to reduce false positives.

[0017] 3. Dynamic adaptation: Sliding window + time decay γ=0.3 / week, updates are triggered when 200 new data entries are added, and the network is reconstructed within 72 hours of policy changes, so as to realize the synchronous evolution of the model and the teaching scenario.

[0018] 4. Precise intervention: The backdoor adjustment counterfactual reasoning provides a quantitative effect, such as "30 minutes of extra practice per day → 5.8 points improvement in math score (confidence level 87%)"; the three-level strategy covers the surface to the root cause, improving the adoption rate.

[0019] 5. Closed-loop optimization: Automatic evaluation every 30 days, rule updates are triggered if the deviation is >30%, and continuous iteration improves the attribution accuracy to 89.7%, with a recall rate of 93% for high-risk students. Attached Figure Description

[0020] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of the invention.

[0021] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a schematic diagram of the overall process; Figure 2 Flowchart for generating and jointly verifying causal hypotheses; Figure 3 Update the timing diagram for dynamic parameters; Figure 4 A schematic diagram of the three-level intervention strategy is generated. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the described embodiments of this disclosure without creative effort are within the scope of protection of this disclosure.

[0024] Unless otherwise defined, the technical or scientific terms used in this disclosure shall have the ordinary meaning understood by one of ordinary skill in the art to which this disclosure pertains. The terms "first," "second," and similar terms used in this disclosure do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed following the word and their equivalents, without excluding other elements or objects. Terms such as "upper," "lower," "left," and "right" are used only to indicate relative positional relationships, and these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0025] The educational attribution analysis method based on dynamic causal networks includes the following steps: Multimodal data acquisition is used to collect structured and unstructured educational data. The structured educational data includes at least student grades and attendance records, and the unstructured educational data includes at least teacher comments and class notes.

[0026] Specifically, an ETL tool is used daily at midnight to extract structured data such as student grades, attendance, and assignment submission timestamps from the academic affairs database, and the data is anonymized using the SHA-256 hash algorithm. Unstructured data includes end-of-term comments entered by teachers in the student's electronic academic file, text transcribed from classroom recordings using ASR, and text obtained by OCR from photos of students' uploaded paper notes. It is important to note that all text is uniformly encoded in UTF-8 and filtered for sensitive words. In the academic affairs system scenario, it is necessary to simultaneously process: mixed Chinese and English text in teacher comments, dialect pronunciation text transcribed from ASR, and mathematical formula symbols recognized by OCR, etc. UTF-8 uses a variable-length encoding mechanism of 1-4 bytes to ensure lossless transmission of this heterogeneous data between MySQL databases, NoSQL storage, and distributed computing frameworks, and to ensure data consistency.

[0027] Large-scale model semantic enhancement utilizes a large language model to map the unstructured educational data into causal semantic feature vectors containing numerical confidence, and concatenates them with the structured educational data to form semantically enhanced multimodal feature vectors. The large model semantic enhancement step specifically includes: An education domain dictionary is constructed, mapping educational descriptive phrases in unstructured educational data to standard causal feature labels. Based on these standard causal feature labels, a causal semantic feature vector containing numerical confidence is output. The EduBERT-LLaMA model, with 7B parameters, is further pre-trained on an education domain dataset of 50 million tokens, including educational dictionaries, textbooks, and lesson evaluation records. For example, "learning by analogy" is mapped to "transfer learning ability," and "carelessness" is mapped to standardized feature labels such as "lack of focus." Unstructured educational data is input into the model; for example, a teacher's comment "This student is quick-witted but careless in calculations" outputs a structured feature vector: {"Logical thinking ability": 0.82, "Calculation rigor": -0.76, "Learning attitude": 0.63}. Structured data is standardized: continuous numerical features (such as exam scores) are Z-score standardized, and categorical data (such as subject type) is one-hot encoded. An attention-weighted cross-modal fusion mechanism is adopted to dynamically and weight the causal semantic feature vector with the features of structured educational data to generate a unified multimodal feature vector.

[0028] Causal hypothesis generation involves using prompt templates to drive the large language model to output candidate causal paths and their confidence levels, thereby forming an initial set of causal hypotheses. The causal hypothesis generation specifically includes inputting a set of factors from multimodal feature vectors into a prompt template, which is configured to guide the large language model to generate candidate causal paths in the form of causes and effects. For example, inputting a set of factors ["classroom participation", "homework quality", "family support"], outputting possible causal paths ["family support → homework quality", "classroom participation → concept understanding depth"]. Each hypothesis is accompanied by a confidence score (0-1) generated by the LLM. An example returned by the LLM is: "family support → homework quality, 0.85; classroom participation → concept understanding depth, 0.78". Each candidate causal path is assigned a semantic confidence score output by the large language model; candidate causal paths with semantic confidence scores higher than a first threshold are included in the initial causal hypothesis set.

[0029] Statistical and semantic joint verification is performed by using the Hilbert-Schmidt independence criterion to test the conditional independence of the initial causal hypothesis set, retaining causal edges that simultaneously satisfy the statistical significance threshold and the semantic confidence threshold, and constructing a dynamic Bayesian network structure. In the statistical semantic joint verification, the HSIC statistic and the corresponding p-value are calculated for each candidate causal edge; if the p-value is less than 0.05 and the semantic confidence is greater than 0.8, the causal edge is retained; if the p-value is greater than or equal to 0.05 but the semantic confidence is greater than 0.9, the expert review interface is activated to decide whether to retain it.

[0030] The Bayesian network is constructed using a dynamic structure learning algorithm. Nodes are categorized into three types: environment nodes (family income, school resources), behavior nodes (learning time, classroom participation), and outcome nodes (subject scores, overall rating). The structure learning employs a modified EM-MAP algorithm with a smoothing factor α=1 / N to avoid zero-probability scenarios. The network is stored in JSON format in the graph database Neo4j, supporting edge-level version control.

[0031] Dynamic parameter updates are performed by incrementally updating the conditional probability table of the dynamic Bayesian network based on a sliding window mechanism to adapt to the time-varying characteristics of the educational scenario. In the dynamic parameter update, a time decay factor γ = 0.3 / week is set to exponentially decay the weights of historical data; when the cumulative number of newly collected data records reaches 200, the expectation-maximization algorithm is triggered to incrementally update the conditional probability table; when an educational policy change event is detected, a rapid reconstruction mode is initiated to complete the reconstruction of the network structure and parameters within 72 hours.

[0032] The counterfactual reasoning process utilizes a backdoor adjustment method to calculate the causal effect of the intervention variable on the outcome variable while controlling for confounding factors. First, the intervention variable (e.g., "increased study time") and the outcome variable (e.g., "math score") are identified. Then, the set of confounding variables requiring adjustment (e.g., family income, learning environment) is automatically located. The effect values ​​for each confounding layer are calculated stratified, and finally, a weighted summation is performed to obtain the adjusted causal effect estimate.

[0033] The three-tiered intervention strategy generation process generates personalized intervention strategies for direct intervention, mediated intervention, and fundamental intervention based on the causal effects, and outputs actionable natural language suggestions.

[0034] In the generation of the three-level intervention strategies, direct intervention strategies are used for immediate behavioral adjustments, such as "extending daily review time to 1.5 hours"; mediating intervention strategies are used for mediating variables that affect learning performance, such as "improving conceptual association ability through mind mapping training"; fundamental intervention strategies are used for deep-seated family or economic reasons, such as "applying for scholarships to alleviate family financial pressure"; and an effectiveness- and implementation cost-weighted algorithm is used to prioritize the intervention strategies, with effectiveness accounting for 70% and implementation cost accounting for 30%.

[0035] The system implements feedback and continuous optimization, collecting student performance data in real time via API. Every 30 days, it calculates the difference between actual and predicted improvement. When the difference exceeds 30%, a rule base update process is triggered. The strategy is pushed to homeroom teachers via WeChat. After teachers implement the strategy, the system retrieves the latest grades, homework submission timeliness, and classroom interaction frequency daily. After 30 days, the difference Δ between actual and predicted improvement is calculated. If |Δ|>30%, the confidence level of the corresponding rule is automatically reduced by 20%, and this is logged. If the difference exceeds the threshold for two consecutive periods, expert review is triggered, and causal edges are added or deleted, or the attenuation coefficient γ is adjusted, if necessary.

[0036] Deployed on Kubernetes in a distributed manner, it supports multi-campus data isolation and federated causal learning. Each campus maintains a local causal graph and periodically uploads encrypted parameter gradients to the central node. The central node aggregates the data and sends it back, achieving shared optimization while protecting privacy. A single campus can process 1.2 million records per day with a P99 latency of 480 milliseconds, meeting real-time alert requirements.

[0037] (1) Unless otherwise defined, the same reference numerals in the embodiments and drawings of this disclosure have the same meaning.

[0038] (2) The accompanying drawings of the embodiments of this disclosure only involve the structures involved in the embodiments of this disclosure. Other structures can be referred to the general design.

[0039] (3) For clarity, components or areas are enlarged in the drawings used to describe embodiments of the present disclosure. It will be understood that when an element is referred to as being “above” or “below” another element, the element may be “directly” located “above” or “below” the other element, or there may be an intermediate element.

[0040] The above description is merely a specific embodiment of this disclosure, but the scope of protection of this disclosure 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 disclosure should be included within the scope of protection of this disclosure. Therefore, the scope of protection of this disclosure should be determined by the scope of the claims.

Claims

1. A method for educational attribution analysis based on dynamic causal network, characterized in that: The method comprises the following steps: Multi-modal data acquisition, for collecting structured education data and unstructured education data, the structured education data at least including student performance, attendance records, the unstructured education data at least including teacher comments, classroom notes; Large model semantic enhancement, using a large language model to map the unstructured education data into a causal semantic feature vector containing a numerical confidence, and splicing with the structured education data to form a semantic enhanced multi-modal feature vector; Causal hypothesis generation, prompting the large language model to output candidate causal paths and their confidence through a prompt template to form an initial causal hypothesis set; Statistical semantic joint verification, using the Hilbert-Schmidt independence criterion to conduct conditional independence test on the initial causal hypothesis set, retaining causal edges that meet both statistical significance threshold and semantic confidence threshold, and constructing a dynamic Bayesian network structure; Dynamic parameter updating, based on a sliding window mechanism, incrementally updating the conditional probability table of the dynamic Bayesian network to adapt to the time-varying characteristics of the education scene; Counterfactual reasoning step, using a backdoor adjustment method to calculate the causal effect of the intervention variable on the result variable while controlling for confounding factors; Three-level intervention strategy generation, generating personalized intervention strategies for direct intervention, intermediate intervention and root intervention according to the causal effect, and outputting executable natural language recommendations. 2.The dynamic causal network-based educational attribution analysis method according to claim 1, characterized in that: The large model semantic enhancement step specifically includes: Building an education domain dictionary, mapping education description phrases in unstructured education data into standard causal feature labels, and outputting causal semantic feature vectors containing numerical confidence based on standard causal feature labels; Using an attention-weighted cross-modal fusion mechanism, dynamically weighting and fusing the causal semantic feature vectors and the features of structured education data to generate a unified multi-modal feature vector.

3. The dynamic causal network-based educational attribution analysis method according to claim 2, characterized in that: The causal hypothesis generation includes inputting a factor set of the multi-modal feature vector into a prompt template, the prompt template being configured to guide the large language model to generate candidate causal paths in the form of cause and effect; assigning a semantic confidence score output by the large language model to each candidate causal path; including candidate causal paths with a semantic confidence score higher than a first threshold in the initial causal hypothesis set.

4. The dynamic causal network-based educational attribution analysis method according to claim 3, characterized in that: In the statistical semantic joint verification, the HSIC statistic and the corresponding p-value are calculated for each candidate causal edge; if the p-value is less than 0.05 and the semantic confidence is greater than 0.8, the causal edge is retained; if the p-value is greater than or equal to 0.05 but the semantic confidence is greater than 0.9, an expert review interface is started to decide whether to retain.

5. The dynamic causal network-based educational attribution analysis method according to claim 4, characterized in that: In the dynamic parameter updating, a time decay factor γ = 0.3 / week is set to exponentially decay the weight of historical data; when the cumulative number of newly collected data records reaches 200, the expectation maximization algorithm is triggered to incrementally update the conditional probability table; when an education policy change event is detected, a fast reconstruction mode is started to complete the reconstruction of the network structure and parameters within 72 hours.

6. The dynamic causal network-based educational attribution analysis method according to claim 1, characterized in that: In the generation of the three-level intervention strategy, the direct intervention strategy is used for immediate behavior adjustment; the intermediate intervention strategy is used for influencing the intermediate variable of learning performance; the root intervention strategy is used for deep family or economic reasons; an effect and implementation cost weighting algorithm is adopted to prioritize the intervention strategies, wherein the effect weight accounts for 70%, and the implementation cost weight accounts for 30%.

7. The dynamic causal network-based educational attribution analysis method according to any one of claims 1-6, characterized in that: It also includes implementation feedback and continuous optimization, which collects student performance data after the implementation of the strategy in real time through the API interface, calculates the difference between the actual improvement value and the predicted value every 30 days, and triggers the rule base update process when the difference exceeds 30%.

8. A dynamic causal network-based educational attribution analysis system, characterized by, It includes: A multi-modal data acquisition module for obtaining structured education data and unstructured education data; A large model semantic enhancement module for converting unstructured education data into causal semantic feature vectors and fusing with structured education data; A causal hypothesis generation module for driving a large language model to output candidate causal paths through a prompt template; A statistical semantic joint verification module for constructing a dynamic Bayesian network by combining Hilbert-Schmidt independence criteria test and semantic confidence screening causal edges; A dynamic parameter update module for incrementally updating network parameters based on a sliding window mechanism; A counterfactual reasoning module for calculating the causal effect of intervention variables on result variables; A three-level intervention strategy generation module for outputting personalized strategies of direct, intermediate and root intervention; An implementation feedback and continuous optimization module for tracking strategy effects and dynamically optimizing the model.

Images