Answer prediction method for path-causal reasoning of fair perception cognitive diagnosis task

Abstract

The application discloses a method for answering question prediction of path causal reasoning for fair perception cognitive diagnosis tasks, which comprises the following steps: 1, a cognitive diagnosis task causal graph containing sensitive attributes is established and used for constructing a deep neural network model; 2, a hybrid loss function based on cross-entropy loss is established; 3, the deep neural network model is trained on an educational data set; and 4, the trained network model is used for predicting the answering ability of students. The application can effectively remove the bias caused by the sensitive attributes in the educational data set, so that the answering ability prediction is more fair.

Description

Technical Field

[0001] This invention relates to the field of causal bias correction, specifically a method for predicting answers to a path causal reasoning question in a fairness perception cognitive diagnostic task. Background Technology

[0002] Cognitive diagnostics (CD) is a fundamental component of intelligent education. It requires an intelligent agent to mine students' behavioral data to assess and identify their proficiency in knowledge concepts. It has been applied in various educational scenarios, such as student performance prediction, computer-adaptive testing, and exercise recommendation.

[0003] Researchers have designed numerous neural network-based methods to achieve accurate diagnosis and student modeling. Existing models typically take multiple types of information (e.g., student-practice interaction records, student personal information, exercises) as input and predict students' mastery of each knowledge point. However, while these methods fully utilize student data, they inevitably introduce fairness issues. For example, models often perform better in predicting the performance of students from wealthier families or more developed regions on exercises. If we do not restrict the model's training, it can easily learn associations between student-sensitive information and student performance (e.g., using family wealth to predict a student's proficiency level), which we call spurious correlations. Applying unfair results to real-world scenarios will exacerbate social biases and conflicts, leading to adverse social consequences. Therefore, it is crucial to avoid the model's misuse of student-sensitive attributes while ensuring diagnostic performance. Numerous methods considering fairness have been proposed, such as data reweighting and adversarial learning. However, these strategies still have unavoidable drawbacks. Data resampling methods typically increase / decrease the weights of certain student-practice interactions to achieve fairness goals. However, this strategy violates the principle of cognitive diagnostics, which dictates that the same student should only interact with the same problem once; adversarial learning uses an additional classifier to predict sensitive attributes from user embeddings and directly eliminates the corresponding information. This strategy is too coarse-grained to distinguish between available and sensitive information, leading to a decline in model capability. Summary of the Invention

[0004] To overcome the shortcomings of existing technologies, this invention proposes a path-causal reasoning method for answer prediction in fair perception cognitive diagnostic tasks. This method aims to remove the bias caused by sensitive attributes in educational datasets and ensure the accuracy of predictions, thereby achieving a fairer prediction of answering abilities.

[0005] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0006] The present invention discloses a method for predicting answers using path-causal reasoning in a fairness-perceived cognitive diagnostic task. The fairness-perceived cognitive diagnostic task involves modeling student ability θ given student features u and sensitive attributes A, and then using ability θ to predict the student's answer result y for question V. During the modeling of θ, the labels of the K non-sensitive attributes used for auxiliary modeling are denoted as {Label}. k |k=1,2,...,K};where, Label k The label represents the k-th non-sensitive attribute; the answer prediction method is performed according to the following steps:

[0007] Step 1: Construct a cause-and-effect graph for the cognitive diagnostic task that includes sensitive attributes:

[0008] Let the nodes in the causal graph of the cognitive diagnostic task containing sensitive attributes include: student feature u, sensitive attribute A, student ability θ, and first mediator variable U. f The second mediator variable U d ;

[0009] Let the path in the causal graph of the cognitive diagnostic task containing sensitive attributes include: student feature u through the first mediator variable U f The student-ability path u→U that influences student ability θ f →θ, sensitive attribute A is mediated by the first mediator variable U f The first sensitive ability path A→U that influences student ability θ f →θ, Sensitive attribute A is mediated by the second mediator variable U d The second sensitive ability path A→U that influences student ability θ d →θ;

[0010] Step 2: Construct a deep neural network, including: a representation modeling module, a decoupling prediction module, a bias removal module, and a bias prediction module;

[0011] Step 2.1: Construct a representation modeling module to fuse student features u and sensitive attributes A to obtain student ability θ;

[0012] Interventions were performed on sensitive attribute A and student characteristic u respectively to obtain counterfactual sensitive attribute A. * and counterfactual student characteristics u * The data is then input into the representation modeling module for processing to obtain the counterfactual student ability θ. * ;

[0013] Step 2.2: The decoupling prediction module decouples the sensitive attribute A into the first mediator variable U. f Second mediator variable U d And obtain the predicted value of the kth insensitive attribute. Predicted value of sensitive attribute A And counterfactual sensitivity attribute A * Predicted value

[0014] Step 2.3: The debiasing module uses equation (5) to obtain the unbiased characterization θ. d :

[0015] θ d =σ(θ-βθ) * (5)

[0016] In equation (5), β represents a parameter to be learned;

[0017] Step 2.4: The prediction module processes the student's ability θ and the exercise V to obtain the probability value of the student answering the exercise correctly;

[0018] Step 3: Construct the joint loss function L of the deep neural network;

[0019] Step 4: Train the deep neural network model using the Adam optimizer and calculate the joint loss function L to update the network parameters until the maximum number of iterations is reached or the joint loss function L converges, thereby obtaining the optimal relational reasoning model, which is used to predict students' problem-solving performance.

[0020] The method for predicting answers based on path causal reasoning in a fair perception cognitive diagnostic task, as described in this invention, is characterized in that the representation modeling module includes two perceptron MLPs (·).

[0021] The first multilayer perceptron MLP1(·) processes A to obtain the second mediator variable U. d , convert u and U d After concatenation, the data is input into a second multilayer perceptron (MLP2) for processing to obtain the first mediator variable U. f Finally, use equation (1) to convert U f and U d After fusion, the student's ability θ is obtained;

[0022] θ=σ((1-α)U f +αU d (1)

[0023] In equation (1), α represents a parameter to be learned.

[0024] The decoupling prediction module uses equations (2)-(4) to obtain the predicted value of the k-th insensitive attribute. Predicted value of sensitive attribute A And counterfactual sensitivity attribute A * Predicted value

[0025]

[0026]

[0027]

[0028] In equations (2) to (4), MLP3(·) and MLP4(·) represent the third and fourth multilayer perceptrons, respectively.

[0029] The prediction module uses equations (6)-(9) to calculate the probability of answering the exercise correctly:

[0030]

[0031]

[0032]

[0033]

[0034] In equations (6)-(9), f(·) represents the backbone model; Indicates through θ d The predicted probability value of answering the exercise correctly. Indicates via U f The predicted probability value of answering the exercise correctly. Indicates via U d The predicted probability value of answering the exercise correctly. This represents the probability value of answering the exercise correctly, obtained through θ prediction.

[0035] Step 3 includes:

[0036] Step 3.1: Construct the accuracy loss L using equation (10). ce :

[0037]

[0038] Step 3.2: Construct the insensitive attribute prediction loss L using equations (11) and (12) respectively. cls And the prediction loss of sensitive attributes L rev ;

[0039]

[0040]

[0041] Step 3.3: Group all students according to their sensitivity attribute A, resulting in a disadvantage group (dis), a normal group (gene), and an advantage group (adv), and then analyze the predicted values. and Calculate the predicted mean of the weakest group dis respectively Predicted mean of genes in the normal group The predicted mean of the advantage group's advance. Therefore, the inter-group constraint loss L is calculated using equations (13)-(15). cons :

[0042]

[0043]

[0044]

[0045] In equations (13)-(15), std represents the variance, and E θ This represents the between-group variance of the probability predicted by θ. Indicates via U d The between-group variance of the predicted probabilities;

[0046] Step 3.4: Construct the joint loss function L using equation (16):

[0047] L=w1L ce +w2L cls +w3L rev +w4L cons (16)

[0048] In equation (16), w1, w2, w3, and w4 are four hyperparameters used to balance the losses of each part.

[0049] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the answer prediction method, and the processor is configured to execute the program stored in the memory.

[0050] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program is executed by a processor to perform the steps of the answer prediction method.

[0051] Compared with existing technologies, the beneficial effects of this invention are reflected in:

[0052] 1. Based on sufficient theoretical guidance, this invention constructs a reasonable causal graph for the cognitive diagnostic task of fairness perception, which enables a more accurate description of the relationship between different variables in the data and provides a clear method for calculating the causal effects between different variables in the cognitive diagnostic task.

[0053] 2. This invention proposes a bias-reduction framework adapted to different cognitive diagnostic models. With very few additional parameters, it can significantly improve the bias-reduction performance of cognitive diagnostic models and ensure that the accuracy of exercise prediction does not decrease.

[0054] 3. This invention proposes a joint loss function based on multiple constraints. The deep neural network model trained using the joint loss function can effectively decouple sensitive attributes into a cognitive diagnosis part and a part affecting fairness, making better use of sensitive attributes for prediction while ensuring the fairness of the prediction. Attached Figure Description

[0055] Figure 1 This is a schematic diagram of the framework of the method of the present invention;

[0056] Figure 2 This invention provides a causal graph for a specific path causal reasoning method for fair perception cognitive diagnostic tasks.

[0057] Figure 3 This is a structural diagram of the MLP of the present invention. Detailed Implementation

[0058] In this embodiment, a path-specific causal reasoning method for a fairness-perceived cognitive diagnostic task first models the student features and sensitive attributes in the dataset after constructing a reasonable causal graph for the task. For the student's sensitive attribute A, MLP1 is used to model it to obtain a representation U related to fairness in cognitive diagnostics. d Then, the characteristics u of student i i with U d The data was spliced ​​together, and MLP2 modeling was used to obtain a representation U related to the accuracy of cognitive diagnosis. f Finally, U is fused using a learnable parameter α. d and U f Obtain the student's ability representation θ. To allow U... d and U f This allows for better learning of the components they represent, further constraining them using a decoupled prediction module, and calculating the loss function L. cls L rev and L cons Finally, a bias removal module is used to analyze student features u. i And sensitive attribute A i Intervention to obtain u* and A * The student's ability θ after intervention was calculated. * θ is subtracted from θ by another learnable parameter β. * Obtain the student's ability θ after bias reduction d Remove the unfair parts of student ability θ. Finally, θ d The exercise V is input into the backbone model f(·), and the loss function L is used. ce Constraints are imposed to ensure the accuracy of the model's predictions. These losses are then jointly trained; a schematic diagram of the specific framework is shown below. Figure 1 As shown. Specifically, the method proceeds according to the following steps:

[0059] Step 1: Construct a cause-and-effect graph for the cognitive diagnostic task that includes sensitive attributes, such as... Figure 2 As shown;

[0060] Define the fairness-perceived cognitive diagnostic task as follows: given student features u and sensitive attribute A, model the student's ability θ, and then use the ability θ to predict the exercise V;

[0061] The nodes in the causal graph of a cognitive diagnostic task containing sensitive attributes include: student characteristic u, sensitive attribute A, student ability θ, and mediator variable U. f Mediation variable U d The paths in the cause-effect graph of a cognitive diagnostic task containing sensitive attributes include: path u→U f →θ, path A→U f →θ, path A→U d →θ;Path u→U f →θ represents student characteristics through the mediating variable U f The path that influences student ability θ is completely free of sensitive information; path A→U f →θ represents the sensitive attribute through the mediator variable U f The path that influences student ability θ is the part of the path that is beneficial for prediction due to the sensitive attributes; path A→U d →θ represents the sensitive attribute through the mediator variable U d This path, which influences students' abilities θ, is a factor contributing to prediction unfairness.

[0062] Step 2: Construct a deep neural network model, comprising four modules: a representation modeling module, a decoupling prediction module, a bias removal module, and a prediction module. This module consists of a backbone model f(·) and four perceptron MLPs(·). The structure of each MLP(·) is as follows: Figure 3 As shown;

[0063] Step 2.1: The first multilayer perceptron MLP1(·) processes A to obtain the second mediator variable U.d , convert u and U d After concatenation, the data is input into a second multilayer perceptron (MLP2) for processing to obtain the first mediator variable U. f Finally, use equation (1) to convert U f and U d After fusion, the student's ability θ is obtained;

[0064] U d =σ(MLP1(A)) (1)

[0065] U f =σ(MLP2(concat(u,U) d (2)

[0066] θ=σ((1-α)U f +αU d (3)

[0067] In equations (1)-(3), σ represents the sigmoid function, concat(·,·) represents the concatenation operation, and α is a learnable parameter.

[0068] Step 2.2: The decoupling prediction module uses two constraints on U. f and U d Decouple U f It includes features related to cognitive diagnosis, while U d It contains biased information. First, regarding U... f MLP3 is used to predict K distinct insensitive attributes that are beneficial for predicting student abilities, and then U f and U d Predicting sensitive attributes using a shared MLP4. Constraint U f It can only predict non-sensitive attributes, while U d Sensitive attributes can be predicted;

[0069]

[0070]

[0071]

[0072]

[0073]

[0074] In equations (7) and (8), Label k Let A represent the k-th insensitive attribute, CE(·) represent the cross-entropy loss function, and A *This represents the counterfactual sensitive attribute, which is obtained by averaging all sensitive attributes in the dataset.

[0075] Step 2.3: The bias-reduction module intervenes in student characteristic u and sensitive attribute A, and calculates the counterfactual student ability θ using the same operation as in Step 2.1. * θ * It contains information related to the deviation, subtracted from θ. * Obtain the unbiased characterization θ d Then for θ d Add further constraints L cons To better achieve de-biasing;

[0076]

[0077]

[0078]

[0079] θ d =σ(θ-βθ) * (12)

[0080]

[0081]

[0082]

[0083] In equations (12)-(15), β is a learnable parameter, E represents the variance of the predicted mean in each different student group, and dis, gene, and adv represent the students divided into disadvantaged, average, and advantageous groups based on the sensitivity attribute, respectively.

[0084] Step 3: Construct a joint loss function. Input the obtained student ability θ and exercise V into the backbone model f(·) to obtain the probability value of the student answering the exercise correctly. Then, calculate the cross-entropy loss L using this probability value and the true label. ce To better model U, f U d Loss calculation is performed using θ;

[0085]

[0086]

[0087]

[0088]

[0089]

[0090] In equation (20), y represents the true label. Combining equations (7), (8), (15), and (20), the joint loss function L is constructed:

[0091] L=w1L ce +w2L cls +w3L rev +w4L cons (13)

[0092] In equation (13), w1, w2, w3, and w4 are four hyperparameters used to balance the losses of each part.

[0093] Step 5: Based on all students, exercises, and student labels y in the dataset, train the deep neural network model using the Adam optimizer and calculate the joint loss function L to update the model parameters until the maximum number of iterations is reached or the joint loss function L converges, thereby obtaining the optimal relational reasoning model, which is used to predict students' performance on the exercises.

[0094] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0095] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0096] After training the model, counterfactual samples corresponding to the intervention are generated for each sample during testing. The biased student ability is obtained according to Equation (12), and then the exercises are predicted using Equation (16). The final experimental results are shown in Tables 1 and 2.

[0097] Table 1 shows the experimental results of this invention when the sensitive attribute is "family wealth".

[0098]

[0099] Table 2 shows the experimental results of this invention when the sensitive attribute is "father's education level".

[0100]

[0101] In Tables 1 and 2, the horizontal axis represents experiments conducted on the Australian and Brazilian datasets. IR, AUC, and ACC are metrics related to prediction accuracy, namely recognition rate, area under the ROC curve, and accuracy, respectively. DOA and Equal Opportunity (DOA) are indicators related to predictive fairness, specifically equal opportunity, equal opportunity for disadvantaged groups, and diagnostic accuracy, respectively. In the vertical axis, IRT, MIRT, NCD, and KaNCD represent different backbone models, and Base represents the base model. + The table indicates that sensitive attributes are introduced into the base model, Reg represents a regularization-based method, Adv represents an adversarial method, and PSCRF is the method proposed in this invention. As shown in Tables 1 and 2, the method of this invention achieves good results on different cognitive diagnostic backbone models, improving both the fairness and accuracy of these predictions. These experimental results fully demonstrate the effectiveness of the method of this invention, effectively mitigating the bias caused by sensitive attributes in educational datasets.

Claims

1. A method for predicting answers using path-causal reasoning in a fairness-perceived cognitive diagnostic task, characterized in that, The fairness-perceived cognitive diagnostic task is based on given student characteristics. and sensitive attributes Modeling student abilities Thus utilizing capabilities Answering exercises Answer results To make predictions, in the context of During the modeling process, make the auxiliary modeling The tag with non-sensitive attributes is { | };in, Indicates the first The label contains non-sensitive attributes; the answer prediction method is performed according to the following steps: Step 1: Construct a cause-and-effect graph for the cognitive diagnostic task that includes sensitive attributes: Let the nodes in the causal graph of the cognitive diagnostic task containing sensitive attributes include: student characteristics Sensitive attributes , student ability First mediator variable Second mediator variable ; Let the paths in the causal graph of the cognitive diagnostic task containing sensitive attributes include: student characteristics Through the first mediator variable For students' abilities The student-competency path that has an impact Sensitive attributes Through the first mediator variable For students' abilities The first sensitive-capability path to exert influence Sensitive attributes Through the second mediator variable For students' abilities The second sensitive-capability path that has an impact ; Step 2: Construct a deep neural network, including: a representation modeling module, a decoupled prediction module, a bias removal module, and a prediction module; Step 2.1: Construct a representation modeling module for student features. and sensitive attributes By integrating the data, we can obtain the students' abilities. ; Sensitive attributes and student characteristics Interventions were conducted separately to obtain counterfactual sensitivity attributes. and counterfactual student characteristics The data is then input into the representation modeling module for processing to obtain counterfactual student capabilities. ; The mean value is obtained by averaging all sensitive attributes in the dataset. Step 2.2: The decoupling prediction module will determine the sensitive attributes. Decoupling as the first mediator variable Second mediator variable And using equations (2)-(4) to obtain the first... Predicted values ​​of non-sensitive attributes Sensitive attributes Predicted value and counterfactual sensitivity Predicted value ; (2) (3) (4) In equations (2)-(4), , This refers to the third and fourth multilayer perceptrons; Step 2.3: The bias removal module uses equation (5) to obtain the unbiased characterization. : (5) In equation (5), This represents a parameter to be learned; Represents the sigmoid function; Step 2.4: The prediction module assesses students' abilities. and exercises The data is processed to obtain the probability value of the student answering the exercise correctly; Step 3: Construct the joint loss function for the deep neural network ; Step 3.1: Construct the accuracy loss using equation (10). : (10) In equation (10), Represents the cross-entropy loss function; Indicates passage The predicted probability value of answering the exercise correctly. Indicates passage The predicted probability value of answering the exercise correctly. Indicates passage The predicted probability value of answering the exercise correctly. Indicates passage The predicted probability value of answering the exercise correctly; Step 3.2: Construct the insensitive attribute prediction loss using equations (11) and (12) respectively. and the loss of prediction based on sensitive attributes ; (11) (12) Step 3.3: Arrange all students according to their own sensitivity attributes Grouping was performed to obtain a disadvantage group (dis), a normal group (gene), and a superior group (adv), and the predicted values ​​were used as a basis for further analysis. and Calculate the predicted mean of the weakest group dis respectively. The predicted mean of genes in the normal group The predicted mean of the advantage group's advance. Therefore, equations (13)-(15) are used to calculate the intergroup constraint loss. : (13) (14) (15) In equations (13)-(15), Represents variance. Indicates passage The between-group variance of the predicted probabilities. Indicates passage The between-group variance of the predicted probabilities; Step 3.4: Construct the joint loss function using equation (16). : (16) In equation (16), , , , These are four hyperparameters used to balance the losses of each part; Step 4: Train the deep neural network model using the Adam optimizer and calculate the joint loss function. To update the network parameters until the maximum number of iterations or the joint loss function is reached. The process continues until convergence, thus obtaining the optimal relational reasoning model, which can be used to predict students' performance on questions.

2. The method for predicting answers using path-causal reasoning in a fairness-perceived cognitive diagnostic task according to claim 1, characterized in that, The representation modeling module includes: two perceptrons. ; The first multilayer perceptron right After processing, the second mediator variable is obtained. ,Will and After splicing, the data is input into the second multilayer perceptron. The process is performed to obtain the first mediator variable. Finally, use equation (1) to... and After integration, the students' abilities are obtained. ; (1) In equation (1), This represents a parameter to be learned.

3. The method for predicting answers using path-causal reasoning in a fairness-perceived cognitive diagnostic task according to claim 2, characterized in that, The prediction module uses equations (6)-(9) to calculate the probability of answering the exercise correctly: (6) (7) (8) (9) In equations (6)-(9), This represents the backbone model.

4. An electronic device, comprising a memory and a processor, characterized in that, The memory is used to store programs that support the processor in executing any of the answer prediction methods of claims 1-3, and the processor is configured to execute the programs stored in the memory.

5. A computer-readable storage medium storing a computer program thereon, characterized in that, When the computer program is run by the processor, it performs the steps of any one of the answer prediction methods described in claims 1-3.

Images