FairEduNet Algorithm Design and Teaching Assessment Calibration Considering Equity in Higher Education

Summary

The research aims to provide a new approach for teaching assessment calibration and educational equity in intelligent education. Experimental results indicate that the proposed model significantly outperforms comparative models, effectively addressing issues of poor multi-source data integration and fairness bias in existing calibration methods.

Abstract

Current teaching assessment calibration methods face challenges such as low integration efficiency when processing multi-source heterogeneous assessment data, insufficient adaptability across different disciplines and teaching scenarios, and biases in results with weak fairness guarantees. The research aims to construct an interpretable, transferable, and scalable fairness-correction framework for deep modeling and dynamic correction of teaching assessment data. These issues make it difficult to meet the core requirement of balanced teaching in intelligent education. This study proposes a fairness-oriented educational network algorithm that integrates a generative adversarial network and a gradient boosting decision tree. The algorithm constructs a fairness calibration mechanism and combines a bidirectional encoder representation model with a graph attention network to optimize feature extraction and dynamic adaptability, improving multi-source data processing efficiency and fairness calibration accuracy. This approach achieves precise calibration and fairness assurance in teaching assessments. In the indicator test, the model achieved an accuracy of 98.76% in clustering evaluation features, 98.05% in fairness correction, and 0.968 in cross-scenario correction consistency in the validation set. In the loss value testing task, the total loss of the model decreased from 0.1237 to 0.1018 during the teaching evaluation correction task in the validation set. In the indicator test, the model achieved an accuracy of 98.76% in clustering evaluation features, 98.05% in fairness correction, and 0.968 in cross-scenario correction consistency in the validation set. Experimental results indicate that the proposed model significantly outperforms comparative models, effectively addressing issues of poor multi-source data integration and fairness bias in existing calibration methods. Furthermore, it provides innovative algorithmic frameworks for technology developers and reliable technical tools for educational administrators to promote teaching equity. The research contributions lie in: (i) integrating three core modules, including multi-source data preprocessing, dynamic fairness index verification, and explainability visualization; and (ii) proposing a fairness correction paradigm based on the collaborative optimization of generative adversarial networks and gradient boosting trees, which breaks through the limitations of traditional single-model correction and enables decoupled learning of deviation modeling and correction decision-making.

Introduction

Teaching evaluation and calibration form the core of ensuring teaching quality in intelligent education. Through dynamic analysis, error correction, and result calibration, they provide a basis for instructional improvement and personalized learning. Their accuracy and fairness directly affect whether intelligent education can break through technological biases and provide balanced educational support^1,2. However, existing methods have several shortcomings: they rely on a single data source, ignore the actual situation, and lead to data bias and correction bias. Lack of a dynamic fairness adaptation mechanism makes it difficult to meet the fairness needs of disciplines and scenarios^3,4. In addition, there are issues such as a lack of fairness indicators and bias in correction strategies when dealing with heterogeneous data from multiple sources. To this end, scholars have conducted research from multiple dimensions, such as evaluating tennis-assisted teaching based on an improved dense trajectory algorithm⁵. Yin et al. utilized the analytic hierarchy process and a fuzzy synthesis algorithm to monitor online teaching⁶. Chen analyzed educational data for older adults using an improved data mining algorithm to enhance instructional quality⁷.

Although these studies have made progress, issues such as calibration result bias and poor fairness coordination remain. Therefore, constructing a model that simultaneously provides accurate teaching assessment calibration and dynamic fairness assurance has become a research focus in intelligent education. Generative adversarial network (GAN) adapts to eliminate the implicit influence of sensitive attributes on results through real-time adversarial training between the generator and discriminator, ensuring outputs are determined by core factors rather than label-induced bias⁸. GAN shows advantages in handling group biases in assessment data and modeling non-linear fairness constraints. Cheng et al. proposed a GAN-based image enhancement algorithm to address low-resolution images. The generator output high-resolution images from low-resolution inputs, and the discriminator distinguished the generated images from real high-resolution images. Adversarial training optimizes parameters so the generator output approaches real images⁹. Kang et al. proposed a cross-modal GAN model to improve multimodal data processing efficiency in renewable energy fields. The generator receives single-modal data, the discriminator distinguishes generated from real multimodal data, and adversarial training optimizes the model for improved efficiency¹⁰. The gradient boosting decision tree (GBDT) algorithm is a classic algorithm for structured data processing with strong feature capture capability. By iteratively integrating multiple decision trees, GBDT precisely learns data error patterns and shows advantages in processing multi-source heterogeneous assessment information and correcting non-linear score deviations^11,12. Mizuno et al proposed a GBDT-based path prediction method for heavy rainfall disasters, learning path features and predicting disaster paths with real-time data, selecting optimal predictions through multiple decision trees¹³. Gao et al. developed a GBDT-based visual analysis method for advertising click-through rate prediction, learning the relationship between visual features and historical click data to predict click rates for new advertisements¹⁴. Based on the above challenges, this research aims to systematically answer the following core scientific question: How can an intelligent educational evaluation model be constructed to efficiently integrate multi-source heterogeneous data, dynamically adapt to different teaching scenarios, and simultaneously ensure calibration accuracy and result fairness. To answer this question, this study utilizes the synergistic effect of the fairness constraint mechanism of GAN and the precise error calibration ability of GBDT. Furthermore, research is conducted on how to introduce advanced technologies, such as natural language processing (BERT), reinforcement learning (RL), attention mechanisms (AM), long short-term memory networks (LSTM), graph attention networks (GAT), and knowledge distillation (KD), to compensate for the inherent limitations of a single model in feature extraction, dynamic adaptation, and reasoning efficiency. Ultimately, the goal is to provide a teaching assessment calibration solution that is both precise and fair, and has strong generalization capabilities for the field of intelligent education.

Protocol

FairEduNet algorithm design and optimization
The study combines GAN and GBDT to construct the FairEduNet algorithm, achieving the dual goals of fairness correction and accuracy correction, rather than the trade-off of one-way optimization. FairEduNet adopts a dual-channel collaborative architecture: the GBDT backbone channel is responsible for structured error modeling and precise correction, while the GAN auxiliary channel focuses on decoupling sensitive attributes and dynamically adjusting fairness. The FairEduNet algorithm architecture, combining GAN and GBDT is shown in Figure 1.

Figure 1: FairEduNet algorithm architecture combining GAN and GBDT. Please click here to view a larger version of this figure.

As shown in Figure 1, FairEduNet first collects and preprocesses multi-source assessment data. GBDT uses multiple decision trees to iteratively learn assessment error patterns, outputs initial calibration results, and passes them to the GAN module. GAN trains the discriminator to learn the relationship between initial calibration results and sensitive attributes, while the generator dynamically adjusts calibration parameters. Through multiple rounds of adversarial training and fairness verification, fairness bias is optimized. Finally, the fairness-calibrated results are fed back to the GBDT module, which adjusts for accuracy and fairness, outputting the teaching assessment calibration basis and report. GBDT calculates residuals as shown in Equation (1).

    (1)

In Equation (1), represents the residual of the i-th sample in the t-th iteration, y_i represents the true value, and represents the prediction value of the t-1-th iteration. The GAN adversarial process is shown in Equation (2).

    (2)

In Equation (2), x represents the initial calibration result, z represents sensitive attribute features, P_data(x) represents the true distribution of non-sensitive features, P_z(z) represents the distribution of sensitive attributes, D represents the discriminator, and G represents the generator¹⁵. The initial calibration output of GBDT is shown in Equation (3).

    (3)

In Equation (3), represents the prediction of the i-th sample by the initial decision tree, K represents the total number of decision trees, γ_k represents the weight of the k-th tree, and h_k(x_i) represents the prediction output of the k-th tree for the i-th sample. The FairEduNet algorithm can provide precise calibration and fairness assurance for teaching assessment data. However, when processing unstructured assessment data and adapting to dynamic scenarios, FairEduNet exhibits weak feature extraction capability for unstructured features, resulting in calibration bias. In addition, FairEduNet's fairness indicators and calibration parameters are statically set, limiting its adaptability to different teaching scenarios and affecting overall calibration quality. The combination of bidirectional encoder representations from transformers (BERT) and reinforcement learning (RL), the BERT-RL algorithm, can accurately extract deep features from unstructured text and dynamically adapt scenario parameters without manually setting static thresholds, further improving the reliability of algorithm outputs across scenarios^16,17. Traditional methods like TF-IDF rely on word frequency but lack deep contextual understanding and cannot distinguish specific directions of "fairness" in different scenarios. Word2Vec produces static word vectors, which struggle to adapt to complex contextual changes and recognize indirect bias. BERT uses multi-layer self-attention to encode bidirectional context, capturing both explicit biased terms and implicit biased logic. Traditional approaches require manually constructed bias word lists, depend on subjective experience, and cover limited scenarios. Through pre-trained model transfer learning, BERT automatically learns deep semantic features without extensive manual effort, offering stronger generalization in recognizing ambiguous bias. Moreover, traditional methods often lose information with long texts, while BERT supports up to 512 tokens, preserving contextual relationships, accurately locating bias features, and enabling fine-grained fairness correction. The operation process of BERT-RL is shown in Figure 2.

Figure 2: BERT-RL algorithm operation flow chart. Please click here to view a larger version of this figure.

As shown in Figure 2, BERT-RL first standardizes unstructured text assessment data and inputs it into the BERT model. BERT uses a Transformer encoder for bidirectional semantic modeling to extract key features and construct a feature matrix. The feature matrix is then input into the RL module, which learns the optimal strategy through multiple iterations. The optimized parameter configuration is finally fed into FairEduNet, improving its adaptability. The BERT self-attention feature weight calculation is shown in Equation (4).

   (4)

In Equation (4), d_k represents the dimension of vector K. The RL multi-objective reward function is expressed in Equation (5).

   (5)

In Equation (5), ω₁, ω₂, and ω₃ represent weight coefficients, represents calibration error, D_t is fairness deviation, D_target is target fairness deviation, and T_t is runtime¹⁸. This study combines BERT-RL with FairEduNet to form the BERT-RL-FairEduNet algorithm, referred to as BFEN. BFEN achieves precise calibration and fairness assurance of teaching assessment data through GBDT feature iteration and GAN real-time adversarial training, while BERT-RL optimizes FairEduNet in handling unstructured data and dynamically adapting to scenarios, addressing weaknesses in feature extraction and static parameter limitations. The study used the BERT-base-Chinese model, and pre-trained the corpus of real teaching logs, classroom recorded texts, and educational policy documents from the National Smart Education Platform for the academic year 2024 to 2025, with a vocabulary size of 21128. The hidden layer dimension of the model is 768, with 12 attention heads and 12 layers. The depth of the GBDT tree was set to 8, the number of trees was 200, and the learning rate was 0.1. The training cycle of GAN is 150 rounds, and the discriminator uses a 3-layer fully connected network with sizes of 512, 256, and 1. The generator uses a 4-layer fully connected network with sizes of 1024, 512, 256, and 1. The number of hidden units in LSTM is 128, and the sequence length is 512. GAT has 2 layers and 4 attention heads. The temperature for knowledge distillation is set to 3.0. RL reward weight coefficients ω₁ = 0.4, ω₂ = 0.35, and ω₃ = 0.25. The weight selection criteria are to balance the Pareto front equilibrium of multi-objective optimization with the interpretability requirements of educational equity practices. The experiments were completed using a 50% data split through cross-validation and hyperparameter tuning. The BFEN process for teaching assessment calibration is shown in Figure 3.

Figure 3: Correction process of the BFEN algorithm for intelligent education and teaching evaluation. Please click here to view a larger version of this figure.

As shown in Figure 3, BFEN first preprocesses multi-source teaching assessment data. BERT optimizes FairEduNet's feature extraction capability by calculating text semantic similarity and deep feature weights based on a semantic window, selecting important features to construct a high-dimensional semantic feature matrix. RL then sets a multi-objective reward function and computes scenario adaptability and parameter adjustment gradients to further optimize fairness thresholds and calibration parameters. FairEduNet calculates deviations between assessment data and error pattern models, iteratively updates decision tree weights, and adjusts calibration strategies until errors stabilize within preset thresholds. The final output includes the error calibration model and the fairness verification report. The calibration model is applied to teaching assessment data, generating pre- and post-calibration comparison tables, which are combined with the intelligent education fairness standard library to determine target calibration parameters, providing a scientific basis for teaching assessment calibration in intelligent education. This standard library covers three dimensions of educational opportunity fairness, process fairness, and outcome fairness, and includes 12 core indicators. These indicators include the balance of educational resource allocation between regions, the difference in digital infrastructure coverage between urban and rural schools, the tolerance threshold for delayed response to learning situation diagnosis (≤200 ms), the tolerance for missing data in multimodal assessment (≤3.5%), the sensitivity of algorithm bias detection (AUC deviation for gender/region/stage labeling ≤ 0.02), the KL divergence threshold for score distribution before and after correction (≤0.08), the interpretability score of teacher intervention suggestions (≥4.2/5.0), the timeliness of student profile updates (completed within T + 1 days), the consistency coefficient of cross-platform credit recognition (≥0.93), the accuracy of education policy semantic alignment (BERT score ≥ 0.86), and the completeness of fairness verification report generation (including bias attribution, confidence intervals, and reproducible parameter snapshots), as well as the dynamic scenario adaptation validation pass rate covering three typical scenarios: live classroom, asynchronous assignments, and AI teaching assistants. The semantic feature similarity calculation is shown in Equation (6).

     (6)

In Equation (6), f_i represents the value of the i-th semantic feature dimension, g_i represents the value of the i-th important assessment feature dimension, and n represents the total number of feature dimensions. The RL scenario adaptation parameter calculation is shown in Equation (7).

    (7)

In Equation (7), θ_t represents the parameter value at the t-th iteration, α represents the learning rate, and represents the parameter gradient based on reward function R(θ_t).

Teaching assessment calibration model construction
Although BFEN demonstrates certain advantages in teaching assessment calibration, it still faces low integration efficiency for multi-source heterogeneous assessment data and insufficient dynamic fairness adaptation in practical applications. The AM-LSTM algorithm, which combines Attention Mechanism (AM) and Long Short-Term Memory (LSTM), assigns weights to key features of multi-source assessment data through AM and captures the dynamic change patterns of assessment data over time using LSTM's sequential memory capability. This approach effectively improves multi-source data integration efficiency and dynamic feature learning^19,20. The operation process of AM-LSTM is shown in Figure 4.

Figure 4: AM-LSTM operation flow chart. Please click here to view a larger version of this figure.

As shown in Figure 4, AM-LSTM first preprocesses multi-source heterogeneous teaching assessment data to form a standardized dataset. AM calculates attention weights for features from different data sources to select important features and constructs a weighted feature matrix. The weighted feature matrix is then input into LSTM, which uses its gating mechanism to learn dynamic patterns over time, capturing relationships among assessment data at different stages and outputting dynamic feature vectors. Finally, these dynamic feature vectors are combined with fairness constraints to generate intermediate calibration results that adapt to multi-source data integration and dynamic scenarios, providing a basis for subsequent model optimization. Attention weight calculation is shown in Equation (8).

     (8)

In Equation (8), a_n represents the attention allocation for the n-th feature, x_n represents similarity, q represents the query vector, and softmax represents the activation function. The LSTM forget gate is expressed in Equation (9).

    (9)

In Equation (9), W_f is the forget gate weight, b_f is the forget gate bias, x_t is the input at time t, h_t-1 is the external state variable at time t-1, and σ represents the sigmoid function²¹. This study combines AM-LSTM with BFEN to construct the AM-LSTM-BFEN teaching assessment calibration model, referred to as FBFEN. In this model, AM-LSTM addresses BFEN's limitations in multi-source heterogeneous data integration efficiency and dynamic feature extraction. It also integrates fairness constraints to optimize intermediate calibration results, enabling BFEN to further achieve precise calibration and dynamic fairness assurance for teaching assessment data. The operation process of FBFEN is shown in Figure 5.

Figure 5: Operation process of the FBFEN teaching evaluation and correction model. Please click here to view a larger version of this figure.

As shown in Figure 5, FBFEN first preprocesses the original multi-source teaching assessment data and inputs the standardized dataset into the AM-LSTM module. AM calculates feature attention weights, while LSTM learns dynamic patterns, outputting intermediate calibration results that integrate multi-source information and dynamic features. The intermediate results are then input into the BFEN module. GBDT iteratively learns the error patterns of assessment data based on the intermediate results and the preset error model, outputting preliminary calibration results. Meanwhile, GAN analyzes fairness bias caused by sensitive attributes in the preliminary results through adversarial training between the generator and discriminator, dynamically adjusting calibration parameters to eliminate bias. Finally, the GAN-optimized calibration parameters are fed back to GBDT to update decision tree weights and calibration strategies, producing a secondary calibration result that balances precision and fairness. Data standardization is expressed in Equation (10).

    (10)

In Equation (10), z' represents the standardized value, z represents the original value, and z_min and z_max represent the minimum and maximum values. The final objective function of the GAN generator is expressed in Equation (11).

   (11)

In Equation (11), N represents the number of iterations, λ represents the loss weight factor, ωⁱ represents the weight factor of the i-th iteration, represents the adversarial loss function, and represents the binary cross-entropy loss²².

FBFEN teaching assessment calibration model optimization
Although FBFEN demonstrates strong multi-source data integration and dynamic fairness assurance in teaching assessment calibration, it still faces weak feature correlation and low training and inference efficiency due to the complex model structure in complex educational scenarios. The GAT-KD algorithm, which combines Graph Attention Network (GAT) and Knowledge Distillation (KD), dynamically constructs a semantic association graph among features through GAT's attention mechanism, enhancing the semantic alignment of multi-source data. KD compresses the knowledge of the complex model into a lightweight network, significantly improving inference efficiency while maintaining model performance^23,24. The operation process of GAT-KD is shown in Figure 6.

Figure 6: GAT-KD operation flow chart. Please click here to view a larger version of this figure.

As shown in Figure 6, GAT-KD constructs a semantic association graph among features through the GAT module, dynamically aggregates neighborhood feature information using the attention mechanism, and enhances the semantic representation of local features. KD then uses the output soft labels as supervision signals, minimizing divergence loss between output distributions and cross-entropy loss between its predictions and true labels, achieving knowledge transfer and model compression. The final output is a lightweight calibration model with high precision and efficiency. GAT attention coefficient calculation is shown in Equation (12).

     (12)

In Equation (12), and represent the feature vectors of nodes i and j, W represents the learnable weight matrix, a represents the attention weight vector, represents the concatenation operation, and LeakyReLU represents the activation function²⁵. KD loss function calculation is shown in Equation (13).

     (13)

In Equation (13), KL represents the divergence loss, T² represents the gradient scaling balance, p^student represents the soft label probability of the lightweight model, and p^teacher represents the soft label probability of the complex model²⁶. By combining attention-enhanced feature association and lightweight knowledge transfer, GAT-KD effectively addresses feature semantic bias and high computational complexity. This study integrates GAT-KD into FBFEN to form the GAT-KD-FBFEN teaching assessment calibration model, referred to as GFBFEN. GAT-KD optimizes FBFEN's shortcomings in incomplete multi-source feature correlation capture and low inference efficiency. The operation process of GFBFEN is shown in Figure 7.

Figure 7: GFBFEN teaching evaluation and correction model operation process. Please click here to view a larger version of this figure.

As shown in Figure 7, GFBFEN standardizes multi-source teaching assessment data. GAT models complex relationships among features and dynamically calculates semantic association weights between nodes using attention mechanisms. KD compresses the knowledge of the complex model into a lightweight network, significantly improving inference efficiency while maintaining precision. The AM-LSTM module assigns importance weights to multi-source features and captures temporal dynamic patterns of assessment data, outputting intermediate calibration results that integrate multi-source information. These results are input into the BFEN module for deep processing. Specifically, BERT extracts semantic features from unstructured text, RL dynamically optimizes fairness thresholds and calibration parameters, GBDT iteratively learns error patterns for preliminary calibration, and GAN eliminates biases caused by sensitive attributes through adversarial training. The dynamic parameter adjustment mechanism automatically calibrates the response sensitivity and fairness weights of each module by sensing the temporal changes and group distribution shifts of the teaching scene in real time, thereby solving the problem of insufficient adaptability caused by static parameter configuration, and ensuring that the model can maintain robustness and fairness in different teaching stages, student groups, and evaluation scenarios. The evaluation type directly affects the granularity and feedback cycle of time series modeling, while formative evaluation emphasizes the dynamic nature of the process. The differences in disciplines determine the allocation of features between BERT and GBDT modules. Therefore, adopting a dynamic adjustment mechanism can effectively adapt to different evaluation types and disciplinary characteristics. Finally, through multiple rounds of parameter feedback and iterative optimization, the model outputs precise and fair teaching assessment calibration results. The dynamic fairness constraint optimization function is expressed in Equation (14).

   (14)

In Equation (14), S represents the set of sensitive attributes, represents the predicted output calibration result, represents the variance of predictions within groups, γ represents inter-group parameters, and represents the overall expected value. The adaptive calculation of multi-source feature fusion weights is shown in Equation (15).

     (15)

In Equation (15), w_k represents the feature weight of the k-th data source, β represents the weight parameter, Sim represents the feature similarity measurement function, f_k represents the feature vector extracted from the k-th data source, f_global represents the global feature center, and K represents the total number of data sources. The study set the weights of sensitive attributes, including gender, ethnicity, region, family economic status, and mother tongue background. The weights are determined based on the results of correlation analysis. The study used the Pearson correlation coefficient and Spearman rank correlation coefficient as dual indicators for joint evaluation, combined with expert experience in the field of education for weight calibration. The final determination of the weights for each sensitive attribute yielded values of 0.18 for gender, 0.22 for ethnicity, 0.25 for region, 0.19 for family economic status, and 0.16 for mother tongue background. Gender: Gender identity per legal registration and self-identification, binary (male/female) with non-binary options; data field "gender". Ethnicity: Based on national ethnic identification of 56 groups, compatible with unidentified and cross-border groups; data field "ethnicity". Region: Administrative division of household registration or long-term residence, covering provincial, municipal, and county levels, considering urban-rural dual structure and migrant population; data field "region". Family economic status: According to national statistical standards and education assistance indicators, using a five-level classification; data field "economic_status". Mother tongue background: Primary teaching and family communication language during basic education, covering Mandarin, minority languages, dialects, and foreign languages, weighted by acquisition age and frequency; data field "mother_tongue."

The correction process adopted a three-stage dynamic weighting mechanism. The first stage implemented initial deviation identification based on the weight matrix of sensitive attributes, marking data points that significantly deviated from the global feature center in dimensions such as gender, ethnicity, and region. The second stage introduces educational context constraints to re-qualify the deviation intensity in a context-aware manner. The third stage verifies the correction stability through counterfactual perturbation tests, systematically replacing the values of sensitive attributes while keeping non-sensitive features unchanged, and observing whether the change in evaluation scores is within the ±±3.2% threshold.

Results

BFEN algorithm performance analysis
The indicators evaluated in the experiments included the following:

Evaluation correction accuracy (ECA) denotes the proportion of deviation items correctly identified and rectified by the model within the teaching assessment results, thereby reflecting the overall effectiveness of the correction mechanism. Higher values indicate superior correction accuracy.

Fairness deviation rate (FDR) measures the magnitude of inter-group score disparities that the model fails to eliminate during the calibration process. It is computed as the ratio of the standard deviation of each sensitive attribute (e.g., gender, region, ethnicity) within the score distribution to the overall standard deviation; lower values imply minimized inter-group variances and a more robust fairness guarantee.

Scenario adaptation rate (SAR) indicates the percentage of correction tasks successfully executed by the model across heterogeneous instructional contexts (e.g., sciences vs. humanities, online vs. offline environments).

Fairness maintenance degree (FMD) is defined as one minus the ratio of the variance of fairness indicators among sensitive attribute groups after correction to that observed before correction, reflecting the capability of the system to preserve its structural fairness during the calibration process.

Disciplinary feature recognition rate (DFRR) calculates the proportion of samples in which core features are correctly identified within the assessment data of each discipline relative to the total assessment sample size, demonstrating the model's capacity to discern and capture domain-specific variations.

Multi-source data fusion efficiency (MDFE) refers to the throughput efficiency of the model during feature alignment, weight allocation, and deviation correction when fusing multi-source heterogeneous evaluation data. This comprehensive metric is defined as the product of the number of assessment samples processed per second (samples/sec) and the compliance rate of correction consistency (at a >95% confidence level), where higher values signify a more efficient and stable fusion pipeline.

Correction result stability (CRS) indicates the consistency of calibration outputs across multiple independent runs, quantified by the reciprocal of the standard deviation of the Euclidean distance between the correction outputs of each run under identical inputs. Higher values signify enhanced system reliability.

Single data correction time (SDCT) represents the average computational latency consumed by the model to complete the calibration of a single assessment sample, measured in milliseconds (ms); lower values indicate a stronger real-time response capability.

Corrected absolute error (CAE) measures the mean absolute error between the calibrated predictions and the ground-truth values, representing the residual deviation of the model relative to the original uncalibrated data. Lower values signify that the calibrated outputs align more closely with true performance levels.

Corrected relative error (CRE) quantifies the average absolute error between the calibrated predictions and ground-truth values expressed as a percentage of the ground truth, with lower values indicating higher relative calibration precision.

Correction accuracy rate (CAR) represents the proportion of samples whose post-calibration absolute error is < 0.5 relative to the total sample size, demonstrating the reliability of the model in satisfying predefined precision constraints. High values substantiate the empirical utility and credibility of the outputs.

Total loss value (TL) represents the comprehensive optimization objective value derived from the weighted summation of individual sub-loss terms upon task completion. Specifically, seven metrics—DFRR, MDFE, CRS, SDCT, CAE, CRE, and CAR—are standardized via Z-score normalization and weighted according to the operational priority of the evaluation tasks. Given the weight vector [0.15, 0.12, 0.1, 0.08, 0.13, 0.12, 0.1], these seven normalized indicator values are aggregated to compute the final loss.

Evaluation feature clustering accuracy (EFCA) evaluates the degree of alignment between the unsupervised clustering configurations generated by the model and the ground-truth categories validated by domain experts.

High-dimensional data processing efficiency (HDPE) measures the number of samples subjected to feature dimensionality reduction and deviation correction per unit time when handling high-dimensional sparse vectors containing ≥50 evaluation features. Measured in samples/second, higher values reflect a more robust capability to process complex, large-scale assessment inputs in real time.

Fairness correction precision (FCP) evaluates the uniformity of the calibration effects across divergent student cohorts. This metric is quantified by calculating the mean of the distribution of the Kolmogorov-Smirnov distance for the corrected absolute errors across the sensitive attribute subgroups; lower values imply a more balanced cross-group performance and stronger fairness assurance.

Cross-scene correction consistency (CSCC) quantifies the distributional offset of calibration results across three typical scenarios—namely, classroom assessment, practical evaluation, and online testing—modeled as the Wasserstein distance ratio.

To verify the performance of the proposed BFEN teaching assessment calibration algorithm, the study compared it with the PRF algorithm, which combined principal component analysis (PCA) and random forest (RF); the EAB algorithm, which combined extreme learning machine (ELM) and adaptive boosting (AdaBoost); and the DBLR algorithm, which combined deep belief network (DBN) and logistic regression (LR). The experiments ran on a system equipped with an Intel Core i9-13900HX processor and NVIDIA RTX 4080 GPU, providing high parallel computing capability and large video memory to efficiently complete feature extraction and calibration computations for large-scale teaching assessment data. The operating system was Windows 11, and Python 3.9 was used for programming. The algorithms were implemented using the Scikit-learn library, deep learning modules were developed via the TensorFlow framework, and Pandas and NumPy libraries were employed for data preprocessing. The development environment used PyCharm Professional 2023.1, and Seaborn was applied for the visualization of calibration results to facilitate intuitive comparison of algorithm performance. To ensure the reliability of the data, the study utilized the Global Education Monitoring Report dataset¹ and the Spanish middle school student academic performance dataset². The dataset contains 12,486 records, covering multimodal teaching evaluation data such as university faculty teaching assessments, student academic performance, student teaching evaluations, and course feedback, encompassing 137 teaching dimension features, including classroom interaction frequency, assignment feedback timeliness, scoring consistency, language inclusivity, and cross-cultural sensitivity. The target variable is the teaching evaluation score, which in this study is mapped to a multidimensional weighted composite value calibrated by experts. The research integrates four types of information sources: student teaching evaluations, peer reviews, supervisory classroom observations, and teaching dossier reviews, with weights of 0.35, 0.25, 0.25, and 0.15, respectively. The training and validation sets are divided by chronological order from the two datasets. The study uses data from September 2024 as the training set and data from October 2024 to June 2025 as the validation set to align with the temporal progression of educational assessments. In the preprocessing stage, a modality-specific strategy is employed, with structured features standardized via Z-score normalization and outliers Winsorized, while textual features are encoded using the BERT-base-chinese model and reduced to 768 dimensions. Adversarial training is applied to enhance robustness against implicit bias expressions and mitigate semantic drift caused by cultural background differences.

To achieve cross-domain training and validation research, the model will be transferred to four heterogeneous teaching scenarios: K-12 basic education, vocational education, continuing education, and international Chinese education for zero or few sample validation. In these four scenarios, the study introduces domain adaptive regularization terms during algorithm training to constrain the consistency of feature distribution of the model in different teaching scenarios. Embedding a lightweight syntax-aware mask mechanism for short sentence noise in K-12 scenarios; To address the ambiguity of professional terminology in vocational education, a dynamic domain dictionary is constructed and integrated with the curriculum outline knowledge graph. Design an age group comparison learning module to address the intergenerational semantic gap in continuing education, aligning semantic anchors for different age groups' evaluation expressions in the latent space. To address the cultural load expression in international Chinese education, cross-language comparison prompts are used for fine-tuning to enhance the robustness of understanding non-literal educational concepts. Research on expert calibration and reliability testing of structured fields in original evaluation records, removing low-reliability entries with Krippendorff's alpha coefficient below 0.75. Double blind manual annotation was implemented on student evaluation texts, with 3 educational measurement experts independently completing deviation-type labeling, achieving Cohen's k = 0.86 consistency in annotation. Finally, the annotated results were cross-validated with the supervision attendance records and teaching file review conclusions to generate the final calibration labels.

The selected datasets originated from distinct populations, measurement systems, and educational backgrounds. This cross-domain validation paradigm rigorously tested the robustness and generalization boundaries of the model within an authentic educational ecosystem, preventing evaluation illusions caused by data homogenization. Both repositories contained substantial volumes of teaching assessment records, offering direct reference value for the deployment of fairness-aware intelligent education algorithms. Both datasets contained substantial teaching assessment data, providing direct reference values for the design of fairness-aware intelligent education algorithms and teaching assessment calibration. The study calibrated 1000 teaching assessment records with the four algorithms and calculated their ECA and Fairness FDR. The results are shown in Figure 8.

Figure 8: Comparison of test results between ECA and FDR. (A) BFEN. (B) PRF. (C) EAB. (D) DBLR. Please click here to view a larger version of this figure.

As shown in Figure 8A, BFEN initially achieved an ECA of 82.47% and an FDR of 5.71% in the assessment calibration task. As the number of calibrated records increased, the ECA rose to 98.75% and stabilized after calibrating 421 records, while the FDR decreased to 2.87% and stabilized after calibrating 514 records. As shown in Figure 8B, the ECA curve of the PRF algorithm gradually increased, while the FDR line slowly decreased. At the end of the calibration task, its ECA value was 92.30%, and PDR value was 6.72%. Figure 8C demonstrates that the ECA curve of the EAB algorithm rose rapidly before flattening, whereas its FDR trajectory displayed severe early-stage fluctuations prior to convergence, terminating with an ECA of 84.64% and an FDR of 8.93%. As delineated in Figure 8D, the DBLR algorithm displayed a sluggish upward ECA trajectory accompanied by marginal fluctuations and limited compression in its FDR line, concluding the calibration task with an ECA of 83.51% and an FDR of 8.42%. These experimental findings confirm that the BFEN algorithm significantly outperforms the baseline approaches in both evaluation correction accuracy and fairness bias control. This superior generalization capability is attributed to the integration of BERT within BFEN, which extracts deep semantic features from unstructured evaluation text to capture hidden bias expressions, while the RL module dynamically optimizes fairness thresholds and correction parameters via a multi-objective reward function to decouple sensitive attributes from the final outputs. The study then tested SAR and FMD for classroom assessment, final exams, practical evaluation, and online assessment, labeling the four scenarios as A, B, C, and D. The results are shown in Figure 9.

Figure 9: Comparison of SAR and FMD results in different scenarios. (A) SAR test results. (B) FMD test results. Please click here to view a larger version of this figure.

As shown in Figure 9A, BFEN achieved SAR above 98% across all teaching scenarios, with the highest SAR of 99.01% for classroom tests. The SAR of the comparison algorithm in different scenarios is lower than that of the BFEN algorithm, among which the DBLR algorithm has the lowest SAR for the final assessment scenario among all algorithms, with a SAR of 70.23%. As shown in Figure 9B, the FMD of the BFEN algorithm is still significantly higher than that of the comparison algorithm in different teaching scenarios, with FMDs of 98.47%, 98.05%, 97.81%, and 97.94% in different scenarios, respectively. The FMDs of the benchmark algorithm DBLR are 82.36%, 71.74%, 70.59%, and 69.12%, respectively. The FMDs of the EAB algorithm are 80.79%, 68.21%, 67.65%, and 66.03%, respectively, while the FMDs of the PRF algorithm are 86.42%, 82.17%, 80.98%, and 78.33%, respectively. These results demonstrated that BFEN outperformed comparison algorithms due to GBDT's iterative learning with multiple decision trees, which accurately captured error patterns across scenarios and ensured stable and fair calibration results. The study further evaluated disciplinary feature recognition rate (DFRR) for Chinese, Mathematics, and English using the four algorithms. The results are shown in Figure 10.

Figure 10: Comparison of DFRR test results in different disciplines. (A) Training set. (B) Validation set. Please click here to view a larger version of this figure.

As shown in Figure 10A, BFEN achieved DFRR of 99.23%, 98.94%, and 99.13% for Chinese, Mathematics, and English in the training set. Figure 10B indicated that BFEN also achieved higher DFRR in the validation set compared to other algorithms. Specifically, BFEN's DFRR for Chinese reached 98.41%, 10.8% higher than DBLR's 87.61%; for Mathematics, 97.54%, 9.07% higher than PRF's 88.47%; and for English, 98.13%, 13.34% higher than EAB's 84.79%. These results confirmed that BFEN efficiently adapted to disciplinary assessment calibration by combining BERT's deep capture of semantic differences with GBDT's personalized learning of error patterns. To comprehensively assess BFEN performance, the study evaluated MDFE, CRS, and SDCT for the four algorithms. The results are presented in Table 1.

Dataset	Algorithm	MDFE (%)	CRS	SDCT (s)
Training	BFEN	98.42 ± 0.28*&@	0.975 ± 0.28*&@	0.78 ± 0.04*&@
	PRF	83.85 ± 0.41	0.779 ± 0.36	1.32 ± 0.08
	EAB	85.91 ± 0.50	0.806 ± 0.40	1.15 ± 0.07
	DBLR	88.76 ± 0.47	0.753 ± 0.39	1.45 ± 0.08
Validation	BFEN	97.58 ± 0.25*&@	0.968 ± 0.27*&@	0.86 ± 0.03*&@
	PRF	81.62 ± 0.32	0.687 ± 0.50	1.51 ± 0.06
	EAB	84.37 ± 0.40	0.749 ± 0.42	1.28 ± 0.05
	DBLR	87.53 ± 0.30	0.728 ± 0.47	1.63 ± 0.07

Table 1: Comparison of MDFE, CRS, and SDCT test results. “*” indicates a significant difference (p < 0.05) between BFEN and PRF results. “&” indicates the difference between BFEN and EAB results was significant (p < 0.05). “@” indicates the difference between BFEN and DBLR results is significant (p < 0.05)

Table 1 indicates that BFEN achieved an MDFE of 98.42% in the training set, 14.57% higher than PRF's 83.85%, a CRS of 0.975, 0.222 higher than DBLR's 0.753, and an SDCT of 0.78 s, 0.67 s less than DBLR's 1.45 s. In the validation set, BFEN maintained superior performance, with MDFE, CRS, and SDCT of 97.58%, 0.968, and 0.86 s, respectively, outperforming comparison algorithms. Overall, the results validated the superior comprehensive performance of BFEN in teaching assessment calibration.

GFBFEN teaching assessment calibration model performance validation
Based on the BFEN performance validation, the study further evaluated the performance of the GFBFEN teaching assessment calibration model constructed on BFEN and compared it with calibration models built using PRF, EAB, and DBLR algorithms. To ensure consistent testing conditions and comparability, the global education monitoring report dataset served as the training set, while the university student academic performance dataset served as the validation set. The four models calibrated 500 teaching assessment records, and their CAE and CRE were calculated. The results are shown in Figure 11.

Figure 11: Comparison of CAE and CRE test results. (A) CAE test results. (B) CRE test results. Please click here to view a larger version of this figure.

As shown in Figure 11A, GFBFEN initially achieved a CAE of 0.1279. As calibration proceeded, the CAE decreased to 0.0641 and stabilized after 104 records. The comparison models had higher initial and final CAE than GFBFEN, with EAB having the highest initial and final CAE of 0.1858 and 0.1217, respectively. Figure 11B indicated that GFBFEN's initial and final CRE during the calibration task remained lower than the comparison models, with values of 0.1137 and 0.0912, respectively. These results demonstrated that GFBFEN exhibited strong fitting capability, benefiting from GAT's attention mechanism, which constructed semantic correlation graphs among features, enhanced semantic alignment of multi-source assessment data, and reduced ineffective calibration caused by feature bias. The study then evaluated CAR for student assessment data, teacher assessment data, school assessment data, and online assessment data, labeled as I, II, III, and IV. The results are shown in Figure 12.

Figure 12: Comparison of CAR results of different evaluation data. (A) GFBFEN. (B) PRF. (C) EAB. (D) DBLR. Please click here to view a larger version of this figure.

As shown in Figure 12A, GFBFEN achieved CAR of 99.34%, 98.93%, 99.01%, and 99.12% for student, teacher, school, and online assessment data in the training set. In the validation set, CAR values were 97.89%, 98.56%, 97.57%, and 98.46%, respectively. Figure 12B–D showed that comparison models had lower CAR in both training and validation sets. Among them, EAB performed the worst in the validation set, with CAR of 76.31% for teacher data and 78.29% for online data. These results confirmed that GFBFEN significantly outperformed the comparison models. Next, the study tested the TL in the teaching assessment calibration task to evaluate fitting ability and training stability. The results are shown in Figure 13.

Figure 13: Model fitting ability and training stability test results. (A) GFBFEN. (B) PRF. (C) EAB. (D) DBLR. Please click here to view a larger version of this figure.

As shown in Figure 13A, GFBFEN initially had a TL of 0.1021 in the training set. After 67 iterations, TL decreased to 0.0725 and stabilized. In the validation set, TL decreased from 0.1237 to 0.1018. Figure 13B–D indicated that PRF's final TL in the training set was 0.0824, EAB's final TL in the validation set was 0.1178, and DBLR's TL in both sets was unstable and significantly higher than other models. These results demonstrated that GFBFEN exhibited strong fitting ability and training stability, benefiting from KD-based knowledge transfer, which allowed the model to quickly learn effective features, accelerate loss convergence, and ensure generalization across datasets. To comprehensively validate GFBFEN's core performance, the study tested EFCA, HDPE, FCP, and CSCC for the four models. The results are shown in Figure 14.

Figure 14: Comprehensive indicator test results of teaching assessment and correction tasks. (A) GFBFEN test results. (B) PRF test results. (C) EAB test results. (D) DBLR test results. Please click here to view a larger version of this figure.

As shown in Figure 14A–D, the GFBFEN model has EFCA values of 99.35% and 98.76% in the training and validation sets, respectively. The EFCA of the PRF model are 88.53% and 87.04%, respectively. In the validation set, the EFCA of the GFBFEN model was 6.59% higher than that of EAB model at 92.17%, and 11.72% higher than that of the DBLR model at 87.04%. In addition, its HDPE, FCP, and CSCC indicators are also comprehensively leading: in the validation set, the HDPE of the GFBFEN model reached 98.05%, the FCP reached 97.93%, and the CSCC reached 0.968, all of which outperformed the PRF, EAB, and DBLR models. Overall, these results validated the superior comprehensive performance of GFBFEN, demonstrating that the model's coordinated optimization of GAT-KD, AM-LSTM, and BFEN effectively improved calibration accuracy, efficiency, and fairness in teaching assessment.

The research gradually constructed FairEduNet, BFEN, FBFEN, and GFBFEN, and the final GFBFEN includes components such as FairEduNet, BERT-RL, AM-LSTM, and GAT-KD. To verify the independent contribution of individual components, research and design ablation experiments are conducted, gradually removing each component and evaluating its impact on overall performance. Comparative indicators include ECA, FDR, SAR, and FMD. The results showed that the ECA value of the FairEduNet component alone was 87.32%, FDR was 12.45%, SAR was 89.67%, and FMD was 11.89%. The ECA of the complete GFBFEN model reached 99.21%, the FDR further dropped to 1.93%, the SAR remained stable at 99.45%, and the FMD was optimized to 7.28%. After removing BERT-RL, ECA decreased to 91.04%, FDR value was 6.23%, SAR value was 92.33%, and FMD increased to 7.85%. The ablation of AM-LSTM increased SAR fluctuations by 14.2%. The removal of GAT-KD led to an increase in FMD to 8.36%, and its suppression mechanism of graph structure bias propagation guided by knowledge distillation was significantly effective in alleviating implicit association bias between populations.

To further test the research methods, the study introduced established fairness standard indicators, namely the Fairness Correction Benchmark (FCB), which covers three types of rigid constraints widely recognized in educational scenarios: (1) The absolute value of the mean difference after correction between groups is ≤ 1.2 points (based on the percentage system); (2) The overlap rate of the corrected confidence intervals for each sensitive attribute subgroup is ≥ 95%; (3) In any single scenario, the joint feasible region of FDR and SAR must satisfy the Pareto frontier constraint (that is, it is impossible to improve SAR without deteriorating FDR, and vice versa). The GFBFEN model was compared with mainstream baseline models, such as EAB, PRF, DBLR, and GBDT, in the experimental evaluation. The experiment was conducted on 421 real assessment data collected from actual teaching scenarios, covering three typical educational scenarios: classroom assessment, practical evaluation, and online testing. The results are shown in Table 2.

Algorithm	Absolute value of mean score difference	Confidence interval overlap rate(%)	Joint feasible region satisfaction rate(%)	Comprehensive score
GFBFEN	0.87	98.3	100	97.2
EBA	1.52	98.4	82.6	78.5
PRF	1.68	85.1	74.3	71.2
DBLR	1.45	87.9	79.8	75.6
GBDT	1.33	91.2	86.4	79.9

Table 2: Evaluation results of fairness criteria indicators and practical application verification.

As shown in Table 2, GFBFEN achieved full compliance with all four FCB rigid constraints independently, and its comprehensive score significantly outperformed the other baseline models by a margin of +18.7 points, validating the robustness of the proposed multi-stage collaborative correction paradigm. Particularly, in the core indicator reflecting the fairness-efficiency trade-off capability, the joint feasible region coverage rate reached 100%, indicating that the model possesses deployable Pareto-optimal decision-making capabilities in real educational scenarios.

Fairness in teaching evaluation refers to the use of verifiable quantitative constraints as a benchmark while respecting individual differences and educational laws. The calculation logic and threshold setting of equity indicators are based on the theoretical framework of educational equity and empirical data verification standards. They are jointly reviewed by an expert committee composed of five scholars to ensure a high degree of unity between theoretical rigor and adaptability to educational practice. To further validate the adaptability of the model under different fairness standards, the study conducted further validation under multiple standards. Specifically, three fairness standards were selected for verification. Standard 1 is the acceptance rate balance between groups based on Demographic Parity. Standard 2 is the cross-group consistency of true positive rate and false positive rate based on Equalized Odds. Standard 3 is based on Predictive Parity for controlling inter-group bias in prediction accuracy. It was found that GFBFEN consistently met the rigid constraint requirements under all three standards. Standard 1 group acceptance rate deviation ≤1.2%, standard 2 true/false positive rate intergroup difference ≤0.85%, standard 3 prediction accuracy intergroup deviation controlled within ±±0.6%. In contrast, the other models showed a constraint breakthrough of ≥3.7% under at least one criterion, verifying the generalization compatibility of GFBFEN for multivariate fairness paradigms.

DATA AVAILABILITY:
To ensure data reliability, the study used the Global Education Monitoring Report dataset (https://www.education-progress.org/) and the Spanish middle school student academic performance dataset (https://archive.ics.uci.edu/dataset/320/student+performance). The training and validation sets are divided by chronological order from the two datasets. The study uses data from September 2024 as the training set and from October 2024 to June 2025 as the validation set, aligning with the temporal progression of educational assessments.

Discussion

The BFEN algorithm proposed in this study demonstrated superior performance in comparative experiments. From the perspectives of ECA and FDR, it significantly outperformed PRF, EAB, and DBLR algorithms. When calibrating 421 assessment records, BFEN increased ECA to 98.75% and maintained stability, while FDR decreased to 2.87%. This performance resulted from the integration of BERT-RL for feature optimization and dynamic adaptation. BERT precisely extracted deep semantic bias information from unstructured assessment texts, and RL dynamically adjusted the fairness thresholds and calibration parameters through a multi-objective reward function, eliminating the influence of sensitive attributes on results. This is similar to the work conducted by Valencia-Londoño et al., who used artificial intelligence algorithms to construct an educational inclusion model for adults with various neuromuscular disorders. Although the constructed educational inclusion model was verified for feasibility within specific neuromuscular disease groups, its fairness constraints only covered a single dimension (equal representation of diagnostic groups), and no rigid constraint system across multiple groups and multiple objectives was established^27. In tests across different teaching scenarios, BFEN achieved a SAR of 99.01% for classroom assessments and an FMD of 97.81% for practical evaluations, significantly higher than the comparison models. This advantage stemmed from GBDT's iterative learning of error patterns across multi-scenario assessment data, combined with the intelligent education fairness standards library to match target calibration parameters, effectively reducing calibration bias and fairness imbalance caused by scenario differences. Compared with the adaptive and interpretable fair artificial intelligence evaluation system proposed by Kumar et al., which incorporates human expert feedback loops and cross-language fairness verification modules, although it introduces these features, it fails to effectively model the implicit bias chains in the evaluation texts in the domain of educational assessment, which has high semantic density and strong context dependence²⁸. This study is more deeply integrated in terms of the rigid guarantee of fairness constraints and the deep coupling of educational semantic modeling compared to the literature²⁸.

In practical intelligent education assessment tasks, the GFBFEN model constructed on BFEN also demonstrated notable advantages. For 500 calibration records, GFBFEN achieved a final CAE of 0.0641 and CRE of 0.0912. Its CAR for multiple types of assessment data exceeded 97% in both training and validation sets. This performance resulted from GAT-KD's feature correlation optimization and lightweight processing. GAT constructed semantic correlation graphs among features to reduce multi-source data feature bias, while KD's knowledge transfer enhanced inference efficiency without compromising model accuracy, avoiding calibration delays caused by model complexity. Compared with the research conducted by Deho et al. on the impact of dataset drift on the fairness of learning analysis models, although their study focused on the mechanism of the influence of data drift on fairness, the correction strategies relied on static reweighting and post-event compensation, lacking the ability to perceive and respond dynamically to the real-time semantic deviations. It was difficult to deal with the evolution of implicit biases in educational assessment caused by subjective expressions of teachers and differences in students' language styles. However, this study, through the GFBFEN model, effectively perceives and responds dynamically to semantic deviations in assessment semantics, capturing the implicit value tendencies in teacher comments, language style deviations in students' response texts, and semantic drift in assessment expressions across grades/subjects, achieving a paradigm shift from "static compensation" to "dynamic calibration"²⁹. In core metric testing, GFBFEN achieved EFCA of 99.35% and 98.76% in training and validation sets, respectively, and FCP of 98.52% and 97.93%, significantly higher than comparison models. Similar to the open-ended examination system with automatic scoring based on artificial intelligence developed by the Dimari A team, although it has achieved a high degree of consistency in the scoring of open-ended responses, its fairness correction only relies on the preset weight of scoring dimensions and the biased sample library annotated by humans, without embedding an educational semantic dynamic evolution mechanism. This research has made substantive breakthroughs in the dynamic evolution mechanism of fairness correction and the deep modeling of educational semantics compared to the research results of the Dimari A team, especially demonstrating strong robustness and interpretability in addressing real-world challenges such as the drift in the value orientation of teacher comments, the generational differences in student language styles, and the ambiguity in cross-disciplinary assessment expressions³⁰.

This study contributed in three aspects. First, the BFEN algorithm integrated BERT-RL with FairEduNet, improving processing efficiency and multi-scenario adaptability for multi-source heterogeneous data through semantic feature optimization and dynamic parameter adjustment. Second, the GFBFEN model, based on GAT-KD and FBFEN, introduced semantic correlation learning and lightweight knowledge transfer, solving weak feature correlation and low inference efficiency issues in complex educational scenarios, enhancing practicality in assessment calibration. Third, the models were applied to calibrate different subjects and assessment types, providing technical support for precise teaching assessment and promoting educational fairness. These contributions offered a new approach for intelligent education assessment calibration and a reference for multi-algorithm collaboration in complex educational data processing.

Current intelligent education assessment calibration methods lack adaptability and fairness. This study proposes a GFBFEN model based on FairEduNet, using GAN and GBDT for fairness calibration. BERT-RL enhances unstructured data processing, AM-LSTM improves multi-source data fusion, and GAT-KD strengthens feature correlation and lightweighting. By integrating these algorithms, the model achieves precise calibration and dynamic fairness. Tests show excellent multi-source data processing and cross-scenario fairness calibration, meeting precision and fairness requirements. However, the model has limitations with special education groups, needing improved generalization. Future work will optimize parameters and adjust fairness constraints for diverse scenarios, supporting educational fairness and teaching quality.

Materials

List of materials used in this article

Name	Company	Catalog Number	Comments
Global Education Monitoring Report dataset	UNESCO	https://www.education-progress.org/	Dataset
NumPy	NumPy	https://numpy.org/	Data preprocessing
Pandas	Pandas, NumPy	https://pandas.pydata.org/	Data preprocessing
PyCharm Professional 2023.1	PyCharm	Version 2023.1	Development environment
Python 3.9	Python	Version 3.9	‌Programming language
Scikit-learn	Scikit-learn	https://scikit-learn.org/stable/index.html	Machine Learning
Seaborn	Seaborn	https://seaborn.pydata.org/	Visualization
Spanish middle school student academic performance dataset	UC Irvine Machine Learning Repository	https://archive.ics.uci.edu/dataset/320/student+performance	Dataset
TensorFlow	TensorFlow	https://www.tensorflow.org/	Deep Learning
Windows 11	Microsoft	Version 11	Operating System