A method for missing perception fusion and cost-sensitive decision-making for academic early warning

Abstract

This invention provides a missing information perception fusion and cost-sensitive decision-making method for academic early warning in the field of educational data mining technology, comprising: step S1, constructing a training dataset; step S2, constructing a missing information indication matrix and performing feature grouping mapping to obtain a joint input; step S3, inputting the input into an academic early warning model; step S4, conditionally modulating the numerical features and missing information; step S5, selecting a feature subset from the enhanced input representation and performing feature transformation; step S6, generating the final decision representation based on the feature transformation output; step S7, inputting the output into a classification head to obtain the posterior probability of each academic early warning level; step S8, calculating the expected cost of decisions for different early warning levels based on the cost matrix and the posterior probability of the classes, and outputting the academic early warning level corresponding to the minimum expected cost. The advantages of this invention are: significantly improving the accuracy and generalization of early warnings, effectively suppressing underreporting of high-risk students, and optimizing model computational efficiency.

Description

Technical Field

[0001] This invention relates to the field of educational data mining technology, and in particular to a method for missing perception fusion and cost-sensitive decision-making for academic early warning. Background Technology

[0002] Academic risk assessment in higher education is a crucial topic in the field of educational data mining. Its core objective is to utilize multi-source student behavior and academic data to identify students at risk of academic lapses in a timely and accurate manner, thereby providing data support for precise intervention and assistance decisions by educational management departments. With the deepening of information technology development in higher education, multi-source heterogeneous data from systems such as student basic information, campus card transactions, course exam scores, library borrowing, and access control are continuously accumulating, providing a rich data foundation for building data-driven academic risk assessment models.

[0003] However, despite the improvement in data availability, existing data mining methods for early warning of academic performance in universities still face many challenges in practical applications, which limit the accuracy, interpretability, and practicality of early warning models. These challenges are specifically reflected in the following aspects: First, there's the issue of missing values ​​in multi-source, heterogeneous data. University student behavior data comes from diverse sources, and observations are often incomplete, resulting in a large number of missing values. Existing methods typically use the mean, median, or mode for static imputation. While this makes the data format more regular, it completely masks the important risk information that the behavioral pattern of "missing data" itself may contain. For example, a student's long-term lack of spending or library entry records could be a high-risk signal, and simple numerical imputation would lose this semantic meaning.

[0004] Second, there's the issue of intra-group correlation and selection stability among features. Indicators such as student spending, academic performance, book borrowing, and library entry behavior typically exhibit significant correlations within their respective behavioral domains (feature groups). Traditional feature selection methods often operate on a feature-by-feature basis, which can easily lead to unstable selection results among strongly correlated similar variables. That is, the selected key features may repeatedly switch between different training cycles or inferences about similar samples. This instability not only reduces the reliability of the model's explanations but may also affect the model's generalization performance.

[0005] Third, there is the issue of the asymmetry between the ordered nature of early warning labels and the costs of erroneous decisions. Academic early warning labels (such as no risk, low risk, and high risk) have a natural hierarchical relationship, and the costs of misjudgment at different levels are severely asymmetrical. In actual operations, the management consequences of misjudging a high-risk student as low-risk (missed reporting) (such as the student potentially failing academically due to lack of timely intervention) are far greater than misjudging a low-risk student as high-risk (false positive). Most existing methods use standard classification loss functions (such as cross-entropy) for training and directly select the category corresponding to the highest probability during inference, failing to explicitly incorporate this ordered and asymmetric operational cost into the model's learning and decision-making process. This makes it difficult to meet the urgent need in practical deployments to suppress high-risk missed reporting.

[0006] Fourth, there is the issue of balancing the differences in sample discrimination difficulty with computational efficiency. Different student samples exhibit varying degrees of clarity in their academic risk patterns, resulting in differences in discrimination difficulty. Existing multi-step decision-making models (such as the original TabNet) typically allocate a fixed number of decision steps to all samples. This may lead to computational redundancy for "easy samples" with clear patterns, affecting inference efficiency; while for "difficult samples" with complex patterns or those at the decision boundary, the fixed number of steps may limit their feature representation and discrimination capabilities, resulting in insufficient model expressive power.

[0007] In summary, existing technologies have significant shortcomings in handling academic early warning tasks in universities, particularly in areas such as utilizing missing information, feature selection, cost-sensitive decision-making, and computational efficiency.

[0008] Therefore, how to provide a missing perception fusion and cost-sensitive decision-making method for academic early warning, so as to improve the accuracy and generalization of early warning, effectively suppress the underreporting of high-risk students, and optimize the model's computational efficiency, has become an urgent technical problem to be solved. Summary of the Invention

[0009] The technical problem to be solved by this invention is to provide a method for missing perception fusion and cost-sensitive decision-making for academic early warning, so as to improve the accuracy and generalization of early warning, effectively suppress the underreporting of high-risk students, and optimize the model calculation efficiency.

[0010] This invention is implemented as follows: a method for missing information fusion and cost-sensitive decision-making for academic early warning, comprising the following steps: Step S1: Obtain multi-source behavioral data of college students and construct an initial dataset. Perform data preprocessing on the initial dataset to obtain a training dataset. Step S2: Construct a missing indicator matrix based on the training dataset and perform feature grouping mapping to obtain a joint input containing numerical features, missing indicator information, and feature group priors; Step S3: Input the joint input into the academic early warning model constructed based on the missing information fusion module, the dual-mask group-level feature selection module, the sample-level adaptive decision step size module, the cost-order-level sensitive decision alignment module, the feature transformer, and the classification head; Step S4: Through the missing information fusion module, conditional modulation is performed on the numerical features and missing information to obtain an enhanced input representation; Step S5: In each decision step of the academic early warning model, a feature subset is selected from the enhanced input representation through the dual-mask group-level feature selection module, and the feature subset is input to the feature transformer to obtain the feature transformation output. Step S6: Generate the final decision representation of the sample based on the feature transformation output through the sample-level adaptive decision step size module. Step S7: Input the final decision representation into the classification head to obtain the posterior probability of the category corresponding to each academic warning level; Step S8: Through the cost-order sensitive decision alignment module, based on the preset cost matrix and the category posterior probability, calculate the expected cost of different warning level decisions, output the academic warning level corresponding to the minimum expected cost, and perform academic warning operation based on the academic warning level.

[0011] Furthermore, step S1 specifically includes: Acquire multi-source behavioral data from university students, including basic student information, campus card transaction data, exam score data, book borrowing data, and library access control data; Students with valid score records are selected based on the exam score data and used as the modeling subject to create a subject table. The multi-source behavioral data is then left-joined to the subject table at the student granularity to construct the initial dataset. The initial dataset is preprocessed to obtain the training dataset. The preprocessing includes outlier handling, light imputation of missing values, encoding of categorical features, and standardization of continuous features.

[0012] Furthermore, in step S2, the feature grouping mapping divides the features into basic information group, consumption feature group, academic feature group, borrowing feature group, and library entry feature group.

[0013] Furthermore, in step S3, the academic early warning model is based on a multi-step decision table deep learning structure.

[0014] Furthermore, in step S4, the conditional modulation of the numerical features and the missing information specifically involves: The missing indication information is encoded, and the global missing pattern context is extracted; Based on the global missing pattern context, dimension-wise modulation parameters are generated that act on the numerical features. The numerical features are subjected to an affine transformation using the dimension-wise modulation parameters and then added to the embedded representation of the missing indication information to obtain an enhanced input representation.

[0015] Furthermore, step S5 specifically includes: In each decision step of the academic early warning model, the dual-mask group-level feature selection module first performs group-level feature selection on the enhanced input representation based on the feature group prior, then performs feature-level selection within the selected group, generates the feature selection mask for the current decision step, and uses the feature selection mask to filter out a feature subset from the enhanced input representation, inputs the feature subset to the feature transformer, and obtains the feature transformation output for the current decision step. The specific steps for performing group-level feature selection on the enhanced input representation based on the prior of the feature group, followed by feature-level selection within the selected group to generate the feature selection mask for the current decision step are as follows: Based on the feature group prior, a feature group attribution matrix is ​​constructed. The enhanced input representation is then mapped to the group space through the feature group attribution matrix to obtain a group-level representation. Based on the attention mechanism and the group-level representation, a group-level selection mask is generated; The group-level selection mask is mapped back to the original feature space to obtain the group-level gating signal; Based on the attention mechanism and the enhanced input representation, a feature-level selection mask is generated; The group-level gating signal is multiplied by the feature-level selection mask to obtain the final feature selection mask for the current decision step.

[0016] Furthermore, during the training process of the academic early warning model, group-level constraints are applied to the group-level selection mask. The group-level constraints include regularization terms that control the sparsity of group selection and diversity regularization terms that suppress the repeated selection of the same feature group in adjacent decision steps.

[0017] Furthermore, step S6 specifically includes: The sample-level adaptive decision step size module calculates the halt probability of a sample in the current decision step based on the feature transformation output and learnable parameters and activation functions. It iteratively updates the remaining computational weights of the sample based on the halt probability and calculates the effective step weights for the current decision step, thereby dynamically allocating differentiated effective decision steps to different samples. Based on the effective step weights, the feature transformation outputs of each decision step are weighted and aggregated to obtain the final decision representation of the sample. The sample is a feature vector representing a single student after data preprocessing.

[0018] Furthermore, during the training process of the academic early warning model, a computational budget constraint is applied to the effective step weights. This computational budget constraint is used to encourage the academic early warning model to make decisions at earlier decision steps.

[0019] Furthermore, in step S8, the formula for the cost matrix is: C(y,a)=η(ya)²+α·max(0,ya)+β·max(0,ay); Where C(y,a) represents the cost matrix; y represents the actual academic warning level; a represents the predicted warning action; η, α, and β all represent preset cost coefficients, and α>β, which are used to make the cost of underestimating the high-risk level higher than the cost of overestimating the low-risk level.

[0020] The advantages of this invention are: 1. By constructing a missing data indication matrix and utilizing a missing data perception fusion module, the missing data pattern itself is transformed into a risk signal, thereby improving the ability to utilize incomplete data and enhancing the accuracy of early warning. By introducing a dual-mask group-level feature selection mechanism based on feature group priors, the reasoning process of experts first locating abnormal behavior domains and then examining specific indicators is simulated, ensuring the stability and interpretability of feature selection, and thus improving the generalization of the model. By designing a cost matrix with asymmetric cost coefficients and constructing a cost-order sensitive decision alignment module, the model minimizes the expected cost when making decisions, thereby explicitly injecting the business logic of "the higher cost of omitting high-risk students" into the model, effectively suppressing the underreporting of high-risk students. Finally, through a sample-level adaptive decision step size module, the number of calculation steps is dynamically allocated to samples with different discrimination difficulties, ensuring the ability to discriminate complex samples while reducing computational redundancy for simple samples, thus optimizing the overall computational efficiency of the model.

[0021] 2. By constructing a missing indicator matrix and designing a "missing information fusion module", we do not simply fill in missing values, but treat the "missing information" itself as an important contextual signal. This module extracts global missing patterns and generates modulation parameters to conditionally enhance numerical features, enabling the model to actively perceive and utilize the distribution patterns of missing data, thereby enhancing the robustness and generalization ability of the model in the case of widespread missing data from multiple sources in the real world.

[0022] 3. Through the dual-mask group-level feature selection module, a two-stage selection strategy of "group-level-feature-level" is implemented. First, the selection is carried out at the group level based on business logic (feature group prior), and then fine-grained feature selection is performed within the selected group. Constraints that promote the sparsity and diversity of group selection are applied. This mechanism not only makes the model's feature selection process more in line with domain knowledge and more interpretable, but also effectively avoids the model from repeatedly relying on a few feature groups in continuous decision steps, thus promoting the breadth and balance of feature utilization.

[0023] 4. An innovative sample-level adaptive decision step size module is introduced. By dynamically calculating the halt probability of each sample at each step, differentiated effective decision steps are allocated to samples of different complexities. This means that for "simple" samples, the model can make decisions earlier to save computing resources; while for "complex" or "difficult" samples, the model can use more computing steps for in-depth analysis. This design can achieve optimal allocation of computing resources, taking into account both efficiency and accuracy.

[0024] 5. The cost-order-level sensitive decision alignment module transforms the model's probability output into the final warning level. The core of this approach is to pre-set an asymmetric cost matrix, explicitly defining that the cost of "underestimating high-risk levels" (missed alerts) is far higher than the cost of "overestimating low-risk levels" (false alarms). This design directly encodes the "better to be strict than lenient" business logic in academic warnings into the model's decision-making process, guiding the model to make more conservative (higher warning level) decisions under uncertainty, thereby directly optimizing the practical application value and risk control capabilities of the technical solution.

[0025] 6. To address the core challenges in academic warning scenarios, such as missing data from multiple sources, complex features, and asymmetric decision costs, a systematic end-to-end deep learning solution is proposed. This solution effectively utilizes missing patterns to enhance model robustness through a missing data perception fusion module, achieves hierarchical filtering that balances business interpretability and selection diversity through dual-mask group-level feature selection, dynamically allocates computational resources through sample-level adaptive decision step size to balance efficiency and accuracy, and finally embeds the key business logic of "underestimating high-risk levels incurs higher costs" into the model output layer through a cost-order sensitive decision alignment module. This ensures that the entire technical solution not only demonstrates advanced predictive performance but also significantly improves its application value in real-world educational environments in terms of practicality, interpretability, and risk control. Attached Figure Description

[0026] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0027] Figure 1 This is a flowchart of a method for missing perception fusion and cost-sensitive decision-making for academic early warning, based on the present invention.

[0028] Figure 2 This is the architecture diagram of the academic early warning model of this invention. Detailed Implementation

[0029] The overall approach of the technical solution in this application is as follows: An end-to-end academic risk warning model is constructed to systematically address challenges such as missing data from multiple sources, unstable feature selection, asymmetric costs of misjudgment, and computational efficiency. First, a missing data perception fusion module transforms missing data patterns into enhanced signals to improve robustness. Second, a dual-mask group-level feature selection module performs hierarchical feature selection from coarse to fine based on prior knowledge of feature groups, ensuring stability and interpretability. Third, a sample-level adaptive decision step size module dynamically allocates computational steps to samples of different complexities, balancing model expressiveness and efficiency. Finally, a cost-order-level sensitive decision alignment module selects the warning level with the lowest expected cost based on a preset asymmetric cost matrix, thereby embedding the core business logic of "suppressing high-risk underreporting" into the model, ultimately achieving accurate, interpretable, and practical academic risk warning.

[0030] Please refer to Figures 1 to 2 As shown, a preferred embodiment of the present invention, a method for missing information fusion and cost-sensitive decision-making for academic early warning, includes the following steps: Step S1: Obtain multi-source behavioral data of college students and construct an initial dataset. Perform data preprocessing on the initial dataset to obtain a training dataset. In practice, multi-source behavioral data can be derived from the student-level wide table output by the university student big data platform. Considering that academic warning labels need to be based on course grades and failed credits, students with valid grade records are used as the modeling subjects, and incomplete observation samples formed after left join of multi-domain data are retained. For the 45 input fields and 1 warning label in the original wide table, the subject key id is removed to form 44 effective predictive features, which are divided into basic information group, consumption feature group, academic feature group, borrowing feature group and library entry feature group according to business semantics, providing priors for group-level structured selection in subsequent models. Step S2: Construct a missing indicator matrix based on the training dataset and perform feature grouping mapping to obtain a joint input containing numerical features, missing indicator information, and feature group priors. The missing indicator matrix transforms the missing state in the data from a kind of "noise" that needs to be masked or simply filled into an "information signal" that can be perceived and utilized by the academic warning model, so that the academic warning model can understand the input data more comprehensively, thereby improving the accuracy of academic risk warning. Step S3: Input the joint input into the academic early warning model constructed based on the missing information fusion module, the dual-mask group-level feature selection module, the sample-level adaptive decision step size module, the cost-order-level sensitive decision alignment module, the feature transformer, and the classification head; The missing information perception fusion module generates dimensionally modulated parameters by encoding missing information, enabling the model to learn a joint representation of "observation value + missing pattern" to obtain an enhanced input representation; the dual-mask group-level feature selection module first forms a group-level mask using the feature group affiliation matrix, and then works together with the feature-level mask to complete the feature selection for each decision step; the sample-level adaptive decision step size module uses the stopping probability to control the remaining computation budget (remaining computation weight) of different samples, stopping early for easy samples and retaining more decision steps for difficult samples; the cost-order-level sensitive decision alignment module uses the cost matrix to distinguish between underestimating high risk and overestimating low risk, and outputs the academic warning level corresponding to the minimum expected cost during the inference stage; Step S4: Through the missing information fusion module, conditional modulation is performed on the numerical features and missing information to obtain an enhanced input representation; Step S5: In each decision step of the academic early warning model, a feature subset is selected from the enhanced input representation through the dual-mask group-level feature selection module, and the feature subset is input to the feature transformer to obtain the feature transformation output. Step S6: Generate the final decision representation of the sample based on the feature transformation output through the sample-level adaptive decision step size module. Step S7: Input the final decision representation into the classification head to obtain the posterior probability of the category corresponding to each academic warning level; Step S8: Through the cost-order sensitive decision alignment module, based on the preset cost matrix and the category posterior probability, calculate the expected cost of different warning level decisions, output the academic warning level corresponding to the minimum expected cost, and perform academic warning operation based on the academic warning level.

[0031] Warning level decision and academic warning level are two closely related but different concepts. Warning level decision refers to the various options tried during the calculation process; academic warning level is the calculation result, that is, the final option determined.

[0032] The missing information-aware fusion module constructed in this invention can preserve the semantics of missing patterns; the dual-mask group-level feature selection module can stably filter key features in behavioral domains such as academics, consumption, borrowing, and library entry; the introduced sample-level adaptive decision step size module can reduce redundant inference computation; and the constructed cost-order-level sensitive decision module can effectively suppress the underreporting of high-risk students. The constructed academic early warning model shows good performance in university academic early warning tasks. In specific experiments, compared with the original TabNet model, Macro-F1 is improved by 7.2 percentage points, QWK is improved by 0.067, and Expected-Cost is relatively reduced by 26.1%.

[0033] Step S1 specifically involves: Acquire multi-source behavioral data from university students, including basic student information, campus card transaction data, exam score data, book borrowing data, and library access control data; Based on the exam score data, students with valid score records are selected and used as the modeling subject to create a subject table. The multi-source behavioral data is left-joined to the subject table at the student granularity to construct the initial dataset. And according to the preset rules, academic risk is mapped to an ordered label of level 0-2 based on the total failed credits. The initial dataset is preprocessed to obtain the training dataset. The preprocessing includes outlier handling, light imputation of missing values, encoding of categorical features, and standardization of continuous features.

[0034] In step S2, the feature grouping mapping divides the features into basic information group, consumption feature group, academic feature group, borrowing feature group, and library entry feature group.

[0035] In step S3, the academic early warning model is based on a multi-step decision table deep learning structure.

[0036] TabNet is a representative deep learning model for tabular data. Its core idea is to progressively select features through a multi-step decision-making mechanism and leverage sparse attention to improve the model's interpretability. In each decision step, the model first generates a feature selection mask based on the attention context and prior vectors from the previous step. Then, the gated features are fed into the Feature Transformer. Finally, the decision representations from each step are accumulated as the input to the classification head. Its core mathematical expression is as follows: ; in, represents the feature selection mask generated at the t-th decision step, which is a sparse vector used to select the features to be processed in the current step; sparsemax() represents the activation function used to transform the input vector into a sparse probability distribution (i.e., some elements in the output are 0, some are non-zero and the sum is 1). This represents the prior vector updated at step t-1, used to suppress the repeated selection of the same feature in subsequent decision steps; Let represent the attention mechanism function at step t, used to calculate the attention score of the feature; This indicates element-wise multiplication (Hadamard product). This represents the attention context output from the previous step (step t-1); After obtaining the mask, the model only transforms the selected features, and the process is represented as follows: ; ; in, This represents the feature representation after mask selection at step t; Show the feature selection mask at step t; This represents the original input feature vector, which is the enhanced feature representation after processing by the missing feature perception fusion module. The decision output at step t represents the contribution of that step to the final prediction. This represents the attention context output at step t that will be passed to the next step. The FeatureTransformer at step t is a learnable neural network module used to process the selected features. After completing the feature transformation of the current decision step, TabNet updates the prior terms and accumulates the decision outputs of each step to obtain the final representation: ; ; in, This represents the updated prior vector used in step t+1; This represents the relaxation parameter (usually slightly less than 1), which controls the rate at which the prior value of a feature decays after it has been suppressed. The final aggregate output of all decision steps is the feature representation input to the classification head; This represents the summation of the results from step 1 to step T (total number of steps); ReLU() represents the rectified linear unit activation function, ensuring that the output is non-negative; Specifically, data on basic student information, campus card transactions, exam scores, book borrowing, and library access control are first aggregated and aligned at the student level. Continuous features are processed for outliers and standardized using training set statistics; categorical features are encoded with integers and represented by embeddings as input to the model; missing values ​​are lightly imputed using the median or mode, while a missing value indication matrix is ​​constructed simultaneously. Through these processes, the model receives not just a single imputed numerical matrix, but a combined input of "numerically computable input + missing value pattern information + group structure priors".

[0037] To preserve the risk semantics carried by missing patterns during the input phase, a missing-aware fusion module is designed. This module first forms a computable matrix from the lightly padded input, then encodes the missing indicator vector, extracts the global missing pattern context, and generates dimension-wise modulation parameters that are applied to the original input representation. The implementation process of the missing-aware fusion module is as follows: ; ; ; in, The input feature matrix is ​​represented by lightly padded (e.g., with median or mode); Impute() represents the imputation function; X represents the original input feature matrix; M represents the original missing value matrix (indicating which values ​​are missing); γ and β represent the scaling and shift parameter vectors generated by the missing value-aware fusion module for modulating the input features. Used to generate modulation parameters based on the encoding context of the missing pattern; It is used to extract global missing pattern context information from the missing indicator matrix, and is an encoder function; This represents the enhanced feature representation modulated by the missing perception fusion module, which serves as the input to the model backbone. Represents a mapping function (such as a linear transformation); This represents the embedding function, which learns an embedding vector for each feature's missing state; In this way, the model can learn a joint representation of "observations + missing patterns" before entering the TabNet backbone, instead of just learning static imputations.

[0038] Considering the significant correlations among consumption features, academic features, borrowing features, and library entry features, a dual-mask group-level feature selection module is introduced in the TabNet attention selection stage. This module first uses the group affiliation matrix to form a group-level representation and a group-level mask in the group space, then maps the group-level gating back to the original feature space, and works together with the feature-level mask to obtain the final selection result of the current decision step. The mathematical expression is as follows: ; ; in, The group-level representation of features is obtained by mapping features to a group space; This represents the transpose of the group affiliation matrix; The group-level selection mask generated at step t represents the attention weights for each feature group; sparsemax() represents the sparse-maximizing activation function. This represents the learnable weight matrix for the group-level selection at step t; Subsequently, the group-level gating is mapped back to the original feature space, where it works in conjunction with the feature-level mask to obtain the final decision-making step selection result: ; ; in, G represents the intermediate feature mask obtained after mapping the group-level mask back to the original feature space; G represents the group affiliation matrix. This represents the final feature-level selection mask at step t; This represents the learnable weight matrix selected at the feature level in step t; To avoid repeated decision steps concentrating on the same group, group-level constraints are further added: ; in, This represents the group-level constraint loss term, used to optimize group selection behavior; This represents a hyperparameter used to control the weight of the second term (diversity constraint); The first term controls the effective sparsity of group selection, while the second term suppresses repetitive focusing in adjacent decision steps. Through this "group-first, feature-later" approach, the model can first determine which behavioral domain primarily determines the current sample, and then locate specific indicators within that behavioral domain.

[0039] To avoid budget imbalance caused by using a uniform decision step size for all samples, a sample-level adaptive decision step size module is further designed. This module calculates the sample-level halt probability and effective step weight based on the marginal contribution of the current decision step, stopping subsequent calculations as early as possible for easy samples with clear patterns, and reserving more decision steps for samples with ambiguous boundaries or high risks. Its core expression is as follows: ; ; ; Further results were obtained: ; ; ; in, σ represents the halting probability calculated at step t for the current sample, i.e., the probability that the model decides to stop computation after this step; σ() represents the Sigmoid activation function, used to map the output to the (0,1) interval; w represents the learnable weight vector. represents the decision output at step t; b represents the bias term; This represents the remaining budget (remainder) after step t, which is the "computational weights" that the model has not yet assigned to any decision steps; This indicates the remaining budget from the previous step; This represents the effective weights assigned to the decision output at step t; d represents the final remaining unallocated budget; d represents the final decision representation after sample-level adaptive weight aggregation. This represents the calculation of the budget constraint loss term, used to encourage the model to stop as early as possible; This represents the effective weight assigned to the nth sample at step t; t represents the index of the decision step. The weights mentioned above, together with the final remaining weights, constitute a complete allocation, enabling the model to improve the performance-efficiency tradeoff while ensuring predictive performance.

[0040] For the output side, a cost-order-sensitive decision alignment module is designed. Academic warning labels have a clear order from low to high, and the management consequences of predicting a high-risk student as low-risk are significantly greater than those of predicting a low-risk student as high-risk. A cost matrix is ​​used to characterize the business costs between the actual risk level and the predicted action.

[0041] Based on this, given the posterior probability of the model, the conditional risk and Bayes minimum risk decision for any action 'a' can be written as: ; in, This represents the expected risk (conditional risk) of taking action a given sample feature x; a represents the predicted action; x represents the input student feature; K represents the total number of categories; This represents the summation over all possible true labels y; This represents the posterior probability predicted by the model that the true label is y given x; Instead of using the simple argmax rule during the inference phase, the action that minimizes the expected cost is selected: ; in, This represents the model's final predicted output (warning level) for sample x. This indicates the operation of taking the parameter 'a' that minimizes the value of the subsequent function; During the training phase, softmin is used to transform the minimum-risk decision for discrete actions into a differentiable form: ; ; in, This represents the (soft) policy probability of selecting action 'a' calculated using the softmin function based on the expected risk during the training phase; exp() represents the exponential function; τ represents the temperature parameter, used to control the smoothness of the policy. Indicates a predicted action; This indicates a cost-sensitive risk loss item; This represents the probability that the model policy assigns to the real action y for the current sample (true label y) during the training phase. During training, preprocessed data is input into the model. The classification head outputs the posterior probability distributions of warning levels for classes 0, 1, and 2, and performs joint optimization by combining group-level constraints, computational budget constraints, and cost-sensitive objectives. After model training is complete, the final warning level, class probability, key feature group contribution, and weight allocation results for different decision steps can be output simultaneously. This provides teachers and administrators with interpretive evidence on "which students need warnings" and "which behavioral domains primarily trigger warnings."

[0042] In step S4, the conditional modulation of the numerical features and the missing information specifically involves: The missing indication information is encoded, and the global missing pattern context is extracted; Based on the global missing pattern context, dimension-wise modulation parameters are generated that act on the numerical features. The numerical features are subjected to an affine transformation using the dimension-wise modulation parameters and then added to the embedded representation of the missing indication information to obtain an enhanced input representation.

[0043] Step S5 specifically involves: In each decision step of the academic early warning model, the dual-mask group-level feature selection module first performs group-level feature selection on the enhanced input representation based on the feature group prior, then performs feature-level selection within the selected group, generates the feature selection mask for the current decision step, and uses the feature selection mask to filter out a feature subset from the enhanced input representation, inputs the feature subset to the feature transformer, and obtains the feature transformation output for the current decision step. The specific steps for performing group-level feature selection on the enhanced input representation based on the prior of the feature group, followed by feature-level selection within the selected group to generate the feature selection mask for the current decision step are as follows: Based on the feature group prior, a feature group attribution matrix is ​​constructed. The enhanced input representation is then mapped to the group space through the feature group attribution matrix to obtain a group-level representation. Based on the attention mechanism and the group-level representation, a group-level selection mask is generated; The group-level selection mask is mapped back to the original feature space to obtain the group-level gating signal; Based on the attention mechanism and the enhanced input representation, a feature-level selection mask is generated; The group-level gating signal is multiplied by the feature-level selection mask to obtain the final feature selection mask for the current decision step.

[0044] During the training of the academic early warning model, group-level constraints are applied to the group-level selection mask. The group-level constraints include regularization terms that control the sparsity of group selection and diversity regularization terms that suppress the repeated selection of the same feature group in adjacent decision steps.

[0045] During the training process of the academic warning model, the loss function, optimizer function, group-level constraints, computational budget constraints, and learnable hyperparameters are adjusted to find the optimal combination of hyperparameters that maximizes the model's performance on the test dataset. The trained academic warning model is then compared with a benchmark model using predetermined evaluation metrics to verify and evaluate its performance. These metrics include Accuracy, Macro-F1, Macro-Recall, Macro-AUC, QWK, and Expected-Cost.

[0046] Step S6 specifically involves: The sample-level adaptive decision step size module calculates the halt probability of a sample in the current decision step based on the feature transformation output and learnable parameters and activation functions. It iteratively updates the remaining computational weights of the sample based on the halt probability and calculates the effective step weights for the current decision step, thereby dynamically allocating differentiated effective decision steps to different samples. Based on the effective step weights, the feature transformation outputs of each decision step are weighted and aggregated to obtain the final decision representation of the sample. The sample is a feature vector representing a single student after data preprocessing.

[0047] Updated remaining computation weights = remaining computation weights of the previous decision step * (1 - halt probability of the current decision step); The remaining computation weight of the previous decision step * the halt probability of the current decision step = the effective step weight of the current decision step; The sum of the effective step weights of each decision step and the final remaining computation weights is 1.

[0048] During the training process of the academic early warning model, a computational budget constraint is applied to the effective step weights. This computational budget constraint is used to encourage the academic early warning model to make decisions at earlier decision steps.

[0049] In step S8, the formula for the cost matrix is: C(y,a)=η(ya)²+α·max(0,ya)+β·max(0,ay); Where C(y,a) represents the cost matrix; y represents the actual academic warning level; a represents the predicted warning action; η, α, and β all represent preset cost coefficients, and α>β, which are used to make the cost of underestimating the high-risk level higher than the cost of overestimating the low-risk level.

[0050] In summary, the advantages of this invention are: 1. By constructing a missing data indication matrix and utilizing a missing data perception fusion module, the missing data pattern itself is transformed into a risk signal, thereby improving the ability to utilize incomplete data and enhancing the accuracy of early warning. By introducing a dual-mask group-level feature selection mechanism based on feature group priors, the reasoning process of experts first locating abnormal behavior domains and then examining specific indicators is simulated, ensuring the stability and interpretability of feature selection, and thus improving the generalization of the model. By designing a cost matrix with asymmetric cost coefficients and constructing a cost-order sensitive decision alignment module, the model minimizes the expected cost when making decisions, thereby explicitly injecting the business logic of "the higher cost of omitting high-risk students" into the model, effectively suppressing the underreporting of high-risk students. Finally, through a sample-level adaptive decision step size module, the number of calculation steps is dynamically allocated to samples with different discrimination difficulties, ensuring the ability to discriminate complex samples while reducing computational redundancy for simple samples, thus optimizing the overall computational efficiency of the model.

[0051] 2. By constructing a missing indicator matrix and designing a "missing information fusion module", we do not simply fill in missing values, but treat the "missing information" itself as an important contextual signal. This module extracts global missing patterns and generates modulation parameters to conditionally enhance numerical features, enabling the model to actively perceive and utilize the distribution patterns of missing data, thereby enhancing the robustness and generalization ability of the model in the case of widespread missing data from multiple sources in the real world.

[0052] 3. Through the dual-mask group-level feature selection module, a two-stage selection strategy of "group-level-feature-level" is implemented. First, the selection is carried out at the group level based on business logic (feature group prior), and then fine-grained feature selection is performed within the selected group. Constraints that promote the sparsity and diversity of group selection are applied. This mechanism not only makes the model's feature selection process more in line with domain knowledge and more interpretable, but also effectively avoids the model from repeatedly relying on a few feature groups in continuous decision steps, thus promoting the breadth and balance of feature utilization.

[0053] 4. An innovative sample-level adaptive decision step size module is introduced. By dynamically calculating the halt probability of each sample at each step, differentiated effective decision steps are allocated to samples of different complexities. This means that for "simple" samples, the model can make decisions earlier to save computing resources; while for "complex" or "difficult" samples, the model can use more computing steps for in-depth analysis. This design can achieve optimal allocation of computing resources, taking into account both efficiency and accuracy.

[0054] 5. The cost-order-level sensitive decision alignment module transforms the model's probability output into the final warning level. The core of this approach is to pre-set an asymmetric cost matrix, explicitly defining that the cost of "underestimating high-risk levels" (missed alerts) is far higher than the cost of "overestimating low-risk levels" (false alarms). This design directly encodes the "better to be strict than lenient" business logic in academic warnings into the model's decision-making process, guiding the model to make more conservative (higher warning level) decisions under uncertainty, thereby directly optimizing the practical application value and risk control capabilities of the technical solution.

[0055] 6. To address the core challenges in academic warning scenarios, such as missing data from multiple sources, complex features, and asymmetric decision costs, a systematic end-to-end deep learning solution is proposed. This solution effectively utilizes missing patterns to enhance model robustness through a missing data perception fusion module, achieves hierarchical filtering that balances business interpretability and selection diversity through dual-mask group-level feature selection, dynamically allocates computational resources through sample-level adaptive decision step size to balance efficiency and accuracy, and finally embeds the key business logic of "underestimating high-risk levels incurs higher costs" into the model output layer through a cost-order sensitive decision alignment module. This ensures that the entire technical solution not only demonstrates advanced predictive performance but also significantly improves its application value in real-world educational environments in terms of practicality, interpretability, and risk control.

[0056] While specific embodiments of the present invention have been described above, those skilled in the art should understand that the specific embodiments described are merely illustrative and not intended to limit the scope of the present invention. Equivalent modifications and variations made by those skilled in the art in accordance with the spirit of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for missing perception fusion and cost-sensitive decision-making for academic early warning, characterized by: Includes the following steps: Step S1: Obtain multi-source behavioral data of college students and construct an initial dataset. Perform data preprocessing on the initial dataset to obtain a training dataset. Step S2: Construct a missing indicator matrix based on the training dataset and perform feature grouping mapping to obtain a joint input containing numerical features, missing indicator information, and feature group priors; Step S3: Input the joint input into the academic early warning model constructed based on the missing information fusion module, the dual-mask group-level feature selection module, the sample-level adaptive decision step size module, the cost-order-level sensitive decision alignment module, the feature transformer, and the classification head; Step S4: Through the missing information fusion module, conditional modulation is performed on the numerical features and missing information to obtain an enhanced input representation; Step S5: In each decision step of the academic early warning model, a feature subset is selected from the enhanced input representation through the dual-mask group-level feature selection module, and the feature subset is input to the feature transformer to obtain the feature transformation output. Step S6: The final decision representation of the sample is generated based on the feature transformation output through the sample-level adaptive decision step size module. Step S7: Input the final decision representation into the classification head to obtain the posterior probability of the category corresponding to each academic warning level; Step S8: Through the cost-order sensitive decision alignment module, based on the preset cost matrix and the category posterior probability, calculate the expected cost of different warning level decisions, output the academic warning level corresponding to the minimum expected cost, and perform academic warning operation based on the academic warning level.

2. The method for missing perception fusion and cost-sensitive decision-making for academic early warning as described in claim 1, characterized in that: Step S1 specifically involves: Acquire multi-source behavioral data from university students, including basic student information, campus card transaction data, exam score data, book borrowing data, and library access control data; Students with valid score records are selected based on the exam score data and used as the modeling subject to create a subject table. The multi-source behavioral data is then left-joined to the subject table at the student granularity to construct the initial dataset. The initial dataset is preprocessed to obtain the training dataset. The preprocessing includes outlier handling, light imputation of missing values, encoding of categorical features, and standardization of continuous features.

3. The method for missing information fusion and cost-sensitive decision-making for academic early warning as described in claim 1, characterized in that: In step S2, the feature grouping mapping divides the features into basic information group, consumption feature group, academic feature group, borrowing feature group, and library entry feature group.

4. The method for missing perception fusion and cost-sensitive decision-making for academic early warning as described in claim 1, characterized in that: In step S3, the academic early warning model is based on a multi-step decision table deep learning structure.

5. The method for missing information fusion and cost-sensitive decision-making for academic early warning as described in claim 1, characterized in that: In step S4, the conditional modulation of the numerical features and the missing information specifically involves: The missing indication information is encoded, and the global missing pattern context is extracted; Based on the global missing pattern context, dimension-wise modulation parameters are generated that act on the numerical features. The numerical features are subjected to an affine transformation using the dimension-wise modulation parameters and then added to the embedded representation of the missing indication information to obtain an enhanced input representation.

6. The method for missing perception fusion and cost-sensitive decision-making for academic early warning as described in claim 1, characterized in that: Step S5 specifically involves: In each decision step of the academic early warning model, the dual-mask group-level feature selection module first performs group-level feature selection on the enhanced input representation based on the feature group prior, then performs feature-level selection within the selected group, generates the feature selection mask for the current decision step, and uses the feature selection mask to filter out a feature subset from the enhanced input representation, inputs the feature subset to the feature transformer, and obtains the feature transformation output for the current decision step. The specific steps for performing group-level feature selection on the enhanced input representation based on the prior of the feature group, followed by feature-level selection within the selected group to generate the feature selection mask for the current decision step are as follows: Based on the feature group prior, a feature group attribution matrix is ​​constructed. The enhanced input representation is then mapped to the group space through the feature group attribution matrix to obtain a group-level representation. Based on the attention mechanism and the group-level representation, a group-level selection mask is generated; The group-level selection mask is mapped back to the original feature space to obtain the group-level gating signal; Based on the attention mechanism and the enhanced input representation, a feature-level selection mask is generated; The group-level gating signal is multiplied by the feature-level selection mask to obtain the final feature selection mask for the current decision step.

7. The method for missing perception fusion and cost-sensitive decision-making for academic early warning as described in claim 6, characterized in that: During the training of the academic early warning model, group-level constraints are applied to the group-level selection mask. The group-level constraints include regularization terms that control the sparsity of group selection and diversity regularization terms that suppress the repeated selection of the same feature group in adjacent decision steps.

8. The method for missing perception fusion and cost-sensitive decision-making for academic early warning as described in claim 1, characterized in that: Step S6 specifically involves: The sample-level adaptive decision step size module calculates the halt probability of a sample in the current decision step based on the feature transformation output and learnable parameters and activation functions. It iteratively updates the remaining computational weights of the sample based on the halt probability and calculates the effective step weights for the current decision step, thereby dynamically allocating differentiated effective decision steps to different samples. Based on the effective step weights, the feature transformation outputs of each decision step are weighted and aggregated to obtain the final decision representation of the sample. The sample is a feature vector representing a single student after data preprocessing.

9. The method for missing information fusion and cost-sensitive decision-making for academic early warning as described in claim 8, characterized in that: During the training process of the academic early warning model, a computational budget constraint is applied to the effective step weights. This computational budget constraint is used to encourage the academic early warning model to make decisions at earlier decision steps.

10. The method for missing information fusion and cost-sensitive decision-making for academic early warning as described in claim 1, characterized in that: In step S8, the formula for the cost matrix is: C(y,a)=η(ya)²+α·max(0,ya)+β·max(0,ay); Where C(y,a) represents the cost matrix; y represents the actual academic warning level; a represents the predicted warning action; η, α, and β all represent preset cost coefficients, and α>β, which are used to make the cost of underestimating the high-risk level higher than the cost of overestimating the low-risk level.

Images