Interdisciplinary student score prediction method based on personalized federal learning
Through personalized federated learning and deep learning technology, an interdisciplinary science achievement prediction model is built, which solves the problems of inefficiency, lack of privacy protection and comprehensiveness in the existing technology, and realizes intelligent, accurate and objective prediction of students' course scores.
Patent Information
- Application Number
- CN202510307840.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-03-17
- Publication Date
- 2025-06-13
AI Technical Summary
The existing technology of middle school student achievement prediction methods are inefficient, lack of privacy protection and comprehensiveness.
The interdisciplinary science-generating achievement prediction method based on personalized federated learning is adopted. By collecting students' historical achievement data and personal information data from multiple subjects, using reinforcement learning to screen the optimal predictor, a student achievement prediction model based on deep learning is constructed, and training is combined with personalized federated learning strategies.
It realizes intelligent prediction of students' course scores, improves the accuracy and objectivity of predictions, enhances privacy protection, and is suitable for diverse interdisciplinary courses.
Smart Images
Figure CN120146307A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of educational process mining and student performance prediction, and particularly provides an interdisciplinary student performance prediction method based on personalized federated learning. Background Art
[0002] Educational process mining is a discipline dedicated to the in-depth analysis of educational data. In the wide application of educational process mining, student performance prediction is a key research area, which estimates the performance of courses by considering multiple factors. With the development of advanced technologies such as artificial intelligence, student performance prediction is attracting more and more research interest in the academic and industrial fields. In higher education, accurately predicting the learning performance of each course is crucial, which can support the personalized training of learners and also play an important role in educational and teaching improvement. The course types in higher education institutions are very diverse, and many courses are interdisciplinary. For these courses, learners from different colleges or institutions bring unique prior knowledge and abilities to the same learning process. However, most of the existing prediction methods adopt a centralized modeling strategy, ignoring the inherent differences among learners. Although there are some methods that use a distributed modeling strategy, they focus more on privacy protection rather than the personality information of learners from different disciplines.
[0003] With the application of big data and artificial intelligence technologies, machine learning, especially deep learning technologies, has demonstrated excellent performance in multiple fields and provided new solutions. Therefore, how to use these technologies to predict student performance, greatly reduce the labor cost of educational evaluation, and achieve fast, objective and accurate prediction has become an urgent problem to be solved. Summary of the Invention
[0004] In view of this, the purpose of the present invention is to provide an interdisciplinary student performance prediction method based on personalized federated learning to solve the problems of low efficiency, lack of privacy protection and comprehensiveness in the existing student performance prediction methods.
[0005] The technical solution provided by the present invention is: an interdisciplinary student performance prediction method based on personalized federated learning, including the following steps:
[0006] S1: Collect historical performance data and personal information data of a large number of students in multiple disciplines and process them to obtain a sample data set, wherein the label of the sample data set is the grading level corresponding to the score of the course to be predicted;
[0007] S2: Use reinforcement learning to screen the historical performance in the sample data set and select the optimal set of prediction factors;
[0008] S3: Construct a student performance prediction model based on deep learning, and combine it with a personalized federated learning strategy. Use the sample data set to train and test the student performance prediction model to obtain a trained student performance prediction model. Among them, the student performance prediction model is a residual convolutional neural network model including a convolutional layer, a residual connection block, a pooling layer, and a fully connected layer. The personalized federated learning strategy is implemented by selecting a part of the network layers as personalized layers;
[0009] S4: Use the trained student performance prediction model to classify the courses that the student needs to predict to obtain the corresponding prediction results.
[0010] Preferably, in S1, the processing of historical performance data includes removing unreasonable data, numericalization, and normalization.
[0011] Further preferably, in S1, the personal information data includes gender and major information. The processing of personal information data is to convert the personal information data from text into numerical data that can be processed by a machine learning model.
[0012] Further preferably, the steps of selecting the optimal set of prediction factors in S2 are as follows:
[0013] S21: Use the EXCEL tool to process the data to obtain aligned historical performance;
[0014] S22: Define the feature selection problem as a Markov decision process. Among them, the state space is the set of all possible subsets of the features of the previous courses, and the action is the process of selecting a feature from the features of the previous courses that have not been included in the model and adding it to the model. The reward function is determined by the improvement of the model accuracy by the selected subset of the features of the previous courses, that is, the difference in accuracy between the current state and the next state;
[0015] S23: Adopt an ε-greedy strategy, randomly select features with a probability of ε for exploration, select the current optimal features of the previous courses with a probability of 1 - ε for exploitation, evaluate the state value through the temporal difference algorithm, and update the value function according to the obtained reward and the state value when visiting the previous state;
[0016] S24: Introduce the concept of average reward to evaluate the importance of features. Feature selection is divided into a random stage and an ε-greedy stage. In the random stage, the agent randomly selects features to explore the environment; in the ε-greedy stage, the agent selects features according to the AoR value to maximize the prediction accuracy. AoR is the average value of the rewards obtained by the features when they are selected, as shown in the following formula:
[0017] AoR A = Average{V(F t ) - V(F t+1 )};
[0018] Among them, V(F t ) is the evaluation value of the value function under the current previous lesson feature set Ft, and V(F t+1 ) is the evaluation value of the value function under the new previous lesson feature set F t+1 after the execution of action A;
[0019] S25: Repeat the above process until the preset number of iterations or policy convergence is reached. Finally, by comparing the accuracies of different feature subsets, select the best previous lesson feature set, and determine the features included therein as the most valuable predictors for performance prediction.
[0020] Further preferably, in S3, the residual convolutional network includes two convolutional layers, one residual block, one pooling layer, and three fully connected layers. The two convolutional layers are connected in sequence. The first convolutional layer is used to expand the number of input channels to 16, and the second convolutional layer is used to expand the number of input channels to 32. A two-dimensional batch normalization layer and an activation function are connected after each convolutional layer. The residual block is located after the convolutional layer. The residual block includes two convolutional layers and two batch normalization layers, which are used to directly add the input to the convolutional output to achieve residual connection. The pooling layer is located after the residual block, and the three fully connected layers are connected in sequence after the pooling layer.
[0021] Further preferably, the input channel number of the residual block is 32, the output channel number is 64, and the stride is 2; the pooling window size of the pooling layer is 2×2, and the stride is 2; the output feature number of the first fully connected layer is 1024, the output feature number of the second fully connected layer is 512, and the output feature number of the third fully connected layer is the classification result.
[0022] Further preferably, in S3, the process of training the student performance prediction model in combination with the personalized federated learning strategy is as follows:
[0023] S31: Treat each subject as a different client, and set the same residual convolutional neural network model for each client. Among them, the first three layers of the network are used as non-personalized layers, and the last three layers are used as personalized layers;
[0024] S32: The server initializes the model parameters of the non-personalized layer and distributes them to all clients to complete model initialization;
[0025] S33: The client uses its local dataset to train the network;
[0026] S34: After the training is completed, the institution uploads the updated non-personalized layer parameters to the server side;
[0027] S35: The server receives the updates of the non-personalized layer parameters from each client and generates a new global model by calculating the globally aggregated model parameters;
[0028] S36: The server sends the non-personalized layer parameters of the aggregated global model back to all clients for the next round of training;
[0029] S37: Repeat S33 to S36 until the stopping condition is met and the training is completed.
[0030] Further preferably, in S3, during the model training process, the loss function adopts the cross-entropy loss function, as shown in the following formula:
[0031]
[0032] In the formula, y oc is the true label, p oc is the probability that the model predicts that the sample o belongs to the category c, M is the total number of categories, and y oc is a binary indicator, which is 1 if the category c is the correct classification of the sample o, and 0 otherwise.
[0033] The interdisciplinary student performance prediction method based on personalized federated learning provided by the present invention adopts personalized federated learning and a residual convolutional neural network, and can intelligently predict the final grade of a certain course for students. This method can make predictions through the historical grades of previous courses and personal information before the course, and can also make predictions during the course learning process through the historical grades of previous courses, personal information, and process data of this course, such as mid-term grades, etc. Thus, more accurate and objective student performance predictions are provided.
[0034] The interdisciplinary student performance prediction method based on personalized federated learning provided by the present invention applies the personalized federated learning technology to the education evaluation process, improves the teaching quality and learning efficiency, and also promotes the development of the fields of educational technology and artificial intelligence. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] The following further elaborates the present invention in detail in conjunction with the drawings and embodiments:
[0036] Figure 1 is the flowchart of the interdisciplinary student performance prediction method based on personalized federated learning provided by the present invention;
[0037] Figure 2 is the structural schematic diagram of the residual convolutional neural network adopted by the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS
[0038] The following further elaborates the technical solution of the present invention in detail.
[0039] To solve the problems existing in the prior art, such as Figure 1 As shown, the present invention provides an interdisciplinary student performance prediction method based on personalized federated learning, comprising the following steps:
[0040] S1: Collect historical performance data and personal information data of a large number of students in multiple disciplines and process them to obtain a sample data set, wherein the label of the sample data set is the grading level corresponding to the score of the course to be predicted;
[0041] Among them, each sample in the sample data set is labeled according to a unified grading standard, so as to facilitate the learning of the prediction logic by the student performance prediction model;
[0042] Processing the data includes removing unreasonable data, numericalization, normalization, etc.;
[0043] Preferably, the labels are divided into 4 categories, corresponding to the grading results (1-4) respectively, wherein the grading results are manually carried out according to a unified standard, and the grading standard is as follows:
[0044] 1) 1 point: The predicted course score is in the range of 80-100;
[0045] 2) 2 points: The predicted course score is in the range of 60-79;
[0046] 3) 3 points: The predicted course score is in the range of 40-59;
[0047] 4) 4 points: The predicted course score is in the range of 0-39;
[0048] Record each score label of the data set according to this standard, and then convert each grading label into a sequence of length 4 through one-hot encoding, wherein the sequence corresponding to 4 points is [0,0,0,1], and the other scores can be deduced by analogy.
[0049] S2: Use reinforcement learning to screen the historical scores in the sample data set and select the optimal set of prediction factors;
[0050] Specifically, it includes the following steps:
[0051] S21: Use the EXCEL tool to process the data to obtain aligned historical scores;
[0052] S22: Define the feature selection problem as a Markov decision process, where the state space is the set of all possible subsets of previous-course features, the action is the process of selecting a feature from the previous-course features not yet included in the model and adding it to the model, and the reward function is determined by the improvement in the model accuracy by the selected subset of previous-course features, i.e., the difference in accuracy between the current state and the next state (after including an additional feature);
[0053] S23: Adopt an ε-greedy strategy, randomly select features for exploration with a probability of ε, and select the current optimal previous-course feature for exploitation with a probability of 1 - ε. Evaluate the state value through the temporal difference algorithm, and update the value function according to the obtained reward and the state value when visiting the previous state;
[0054] S24: Introduce the concept of average reward to evaluate the importance of features. Feature selection is divided into a random stage and an ε-greedy stage. In the random stage, the agent randomly selects features to explore the environment; in the ε-greedy stage, the agent selects features according to the AoR value to maximize the prediction accuracy. AoR is the average of the rewards obtained when the feature is selected, as shown in the following formula:
[0055] AoR A = Average{V(F t ) - V(F t+1 )};
[0056] where V(F t ) is the evaluation value of the value function under the current set of previous-course features Ft, and V(F t+1 ) is the evaluation value of the value function under the new set of previous-course features F t+1 after performing the action A (selecting the feature f);
[0057] S25: Repeat the above process until the preset number of iterations is reached or the strategy converges. Finally, by comparing the accuracies of different feature subsets, select the best set of previous-course features, and determine the features included therein as the most valuable predictors for performance prediction.
[0058] S3: Construct a student performance prediction model based on deep learning, and combine it with a personalized federated learning strategy. Use the sample data set to train and test the student performance prediction model to obtain a trained student performance prediction model. Among them, the student performance prediction model is a residual convolutional neural network model including a convolutional layer, a residual connection block, a pooling layer, and a fully connected layer. The structure of the residual convolutional network is as Figure 2As shown, specifically, the residual convolutional network includes two convolutional layers (Conv1 and Conv2), a residual block (Res), a pooling layer (MP), and three fully connected layers (Fc1, Fc2, Fc3). The two convolutional layers are connected in sequence. The first convolutional layer (Conv1) is used to expand the number of input channels to 16, and the second convolutional layer (Conv2) is used to expand the number of input channels to 32. A two-dimensional batch normalization layer (Bn1, Bn2) and an activation function are connected behind each convolutional layer. The activation function adopts the ReLU activation function. The residual block (Res) is located behind the convolutional layer. The number of input channels of the residual block is 32, the number of output channels is 64, and the stride is 2. The residual block includes two convolutional layers (RC1, RC2) and two batch normalization layers (RB1, RB2), which are used to directly add the input to the convolutional output to achieve residual connection. The pooling layer (MP) is located behind the residual block (Res). The pooling layer uses a max pooling layer to downsample the feature map and reduce the size of the feature map. The size of the pooling window is 2×2, and the stride is 2. The three fully connected layers (Fc1, Fc2, Fc3) are located behind the pooling layer. The number of input features of the first fully connected layer (Fc1) is dynamically calculated according to the size of the feature map output by the convolutional layer, and the number of output features is 1024. This layer is followed by a ReLU activation function and a Dropout layer. The number of input features of the second fully connected layer (Fc2) is 1024, and the number of output features is 512. This layer is followed by a ReLU activation function and a Dropout layer. The number of input features of the third fully connected layer (Fc3) is 512, and the number of output features is the classification result.
[0059] The processing process of the residual convolutional network for the input data is as follows:
[0060] First, the input data passes through two layers of convolution (Conv1 and Conv2) to extract low-level features, and the output is C out (x). Among them, in order to improve the stability and speed of network training, two-dimensional batch normalization is applied behind each convolutional layer;
[0061] Then, C out (x) is input into the residual block (Res) to further extract high-level features and enhance the feature expression through residual connection. The formula is as follows: Res(x) = RB 2 (RC 2 (RB 1 (RC 1 (C out (x)))))+S(x);
[0062] Next, the maximum pooling layer MP is used to reduce the spatial dimension of the feature map. After downsampling by the maximum pooling layer, the size of the feature map is reduced, and the output is Pool(x). The formula is as follows: Pool(x) = MP(Res(x); k; S Pool ) In the formula, k and S Pool represent the size and stride of the pooling kernel;
[0063] Finally, the output of the convolutional layer is flattened into a one-dimensional vector and input into the fully connected layer. After passing through two fully connected layers (Fc1, Fc2) and the ReLU activation function to extract features, the prediction result Fc(x) is obtained through the last fully connected layer (Fc3). The formula is as follows: Fc(x) = W 3 ·σ(W 2 ·σ(W 1 ·x + b 1 ) + b 2 ) + b 3 .
[0064] The personalized federated learning strategy is implemented by selecting a part of the network layers as personalized layers.
[0065] Among them, the process of training the student performance prediction model in combination with the personalized federated learning strategy is as follows:
[0066] S31: Each subject is regarded as a different client, and each client sets the same residual convolutional neural network model. Among them, the first three layers of the network (convolutional layer 1 (Conv1), convolutional layer 2 (Conv2), residual block (Res)) are used as non-personalized layers, and the last three layers (fully connected layer 1 (Fc1), fully connected layer 2 (Fc2), fully connected layer 3 (Fc3)) are used as personalized layers;
[0067] S32: The server initializes the non-personalized layer model parameters and distributes them to all clients to complete model initialization;
[0068] S33: The client uses its local dataset to train the network;
[0069] S34: After training is completed, the institution uploads the updated non-personalized layer parameters to the server side;
[0070] S35: The server receives the non-personalized layer parameter updates from each client and generates a new global model by calculating the globally aggregated model parameters;
[0071] S36: The server sends the non-personalized layer parameters of the aggregated global model back to all clients for the next round of training;
[0072] S37: Repeat S33 to S36 until the stop condition is met and the training is completed.
[0073] During the model training process, the cross - entropy loss function is used as the loss function, as shown in the following formula:
[0074]
[0075] In the formula, y oc is the true label, p oc is the probability that the model predicts that the sample o belongs to the category c, M is the total number of categories, and y oc is a binary indicator, which is 1 if the category c is the correct classification of the sample o, and 0 otherwise.
[0076] The training of the model can be decomposed into two main iterative steps: First, the model is trained on the training set in the training mode, and second, it is verified on the test set in the evaluation mode. After the model is trained, the parameter information such as the weights of the model is saved to facilitate re - loading the model after training is completed, or for further testing or deployment.
[0077] S4: Use the trained student performance prediction model to classify the courses that the student needs to predict to obtain the corresponding prediction results;
[0078] Specifically, it includes the following steps:
[0079] Data upload: Through the submission interface, upload the student's historical performance data and personal information data (major and gender, etc.);
[0080] Data check: Conduct a preliminary format and security check on the uploaded data;
[0081] Data pre - processing: Normalize the uploaded data and convert personal information data, etc. from text to numerical data that can be processed by the machine learning model;
[0082] Model scoring: Classify the processed performance data based on the student performance prediction model to obtain the corresponding prediction results.
[0083] Among them, the prediction results can achieve a quantitative evaluation of the course grades that the student needs to predict.
[0084] The above has described the embodiments of the present invention in detail with reference to the accompanying drawings. However, the present invention is not limited to the above - described embodiments. Within the scope of knowledge possessed by those of ordinary skill in the art, various changes can be made without departing from the gist of the present invention.
Claims
1. An interdisciplinary student performance prediction method based on personalized federated learning, characterized by: The steps include: S1: Collect and process a large amount of historical performance data and personal information data of students in multiple subjects to obtain a sample data set, wherein the label of the sample data set is the rating level corresponding to the course score to be predicted; S2: Use reinforcement learning to filter the historical scores in the sample data set and select the optimal set of predictive factors; S3: Construct a student score prediction model based on deep learning, and combine it with a personalized federated learning strategy, use a sample data set to train and test the student score prediction model, and obtain a trained student score prediction model, wherein the student score prediction model is a residual convolutional neural network model including a convolutional layer, a residual connection block, a pooling layer, and a fully connected layer, and the personalized federated learning strategy is implemented by selecting a part of the network layer as a personalized layer; S4: Use the trained student performance prediction model to classify the courses that students need to predict and obtain the corresponding prediction results.
2. The method for predicting interdisciplinary student performance based on federated learning according to claim 1, characterized in that: In S1, the processing of historical performance data includes removing unreasonable data, digitizing and normalizing.
3. The method for predicting interdisciplinary student performance based on federated learning according to claim 1, characterized in that: In S1, the personal information data includes gender and professional information, and the processing of the personal information data is to convert the personal information data from text into numerical data that can be processed by the machine learning model.
4. The method for predicting interdisciplinary student performance based on federated learning according to claim 1, characterized in that: The steps to select the optimal set of predictors in S2 are as follows: S21: Use EXCEL tools to process the data and obtain aligned historical scores; S22: The feature selection problem is defined as a Markov decision process, where the state space is the set of all possible feature subsets of previous courses, the action is the process of selecting a feature from the features of previous courses that are not yet included in the model to add to the model, and the reward function is determined by the improvement of the model accuracy caused by the selected feature subset of previous courses, that is, the difference between the accuracy of the current state and the next state; S23: Adopt the ε-greedy strategy, randomly select features for exploration with a probability of ε, select the current optimal previous course features for utilization with a probability of 1-ε, evaluate the state value through the temporal difference algorithm, and update the value function according to the reward obtained and the state value when the state was previously visited; S24: The concept of average reward is introduced to evaluate the importance of features. Feature selection is divided into a random stage and an ε-greedy stage. In the random stage, the agent randomly selects features to explore the environment; in the ε-greedy stage, the agent selects features according to the AoR value to maximize the prediction accuracy. AoR is the average value of the reward obtained when the feature is selected, as shown in the following formula: AoR A =Average{V(F t )-V(F t+1 )}; Among them, V(F t ) is the value function evaluation value under the current previous lesson feature set Ft, V(F t+1 ) is the new previous feature set F after executing action A t+1 The value function evaluation value under ; S25: Repeat the above process until the preset number of iterations is reached or the strategy converges. Finally, by comparing the accuracy of different feature subsets, the best feature set of the previous course is selected, and the features contained therein are determined as the most valuable predictors for grade prediction.
5. The method for predicting interdisciplinary student performance based on federated learning according to claim 1, characterized in that: In S3, the residual convolution network includes two convolution layers, a residual block, a pooling layer and three fully connected layers. The two convolution layers are connected in sequence. The first convolution layer is used to expand the number of input channels to 16, and the second convolution layer is used to expand the number of input channels to 32. Each convolution layer is connected to a two-dimensional batch normalization layer and an activation function. The residual block is located after the convolution layer. The residual block includes two convolution layers and two batch normalization layers, which are used to add the input directly to the convolution output to realize residual connection. The pooling layer is located after the residual block, and the three fully connected layers are connected to the pooling layer in sequence.
6. The method for predicting interdisciplinary student performance based on federated learning according to claim 5, characterized in that: The number of input channels of the residual block is 32, the number of output channels is 64, and the step size is 2; the pooling window size of the pooling layer is 2×2, and the step size is 2; the number of output features of the first fully connected layer is 1024, the number of output features of the second fully connected layer is 512, and the number of output features of the third fully connected layer is the classification result.
7. The method for predicting interdisciplinary student performance based on federated learning according to claim 1, characterized in that: In S3, the process of training the student performance prediction model in combination with the personalized federated learning strategy is as follows: S31: Each subject is regarded as a different client, and the same residual convolutional neural network model is set for each client, where the first three layers of the network are used as non-personalized layers and the last three layers are used as personalized layers; S32: The server initializes the non-personalized layer model parameters and sends them to all clients to complete the model initialization; S33: The client trains the network using its local dataset; S34: After the training is completed, the organization uploads the updated non-personalized layer parameters to the server; S35: the server receives the non-personalized layer parameter updates from each client, and generates a new global model by calculating the globally aggregated model parameters; S36: The server sends the aggregated non-personalized layer parameters of the global model back to all clients for the next round of training; S37: Repeat S33 to S36 until the stop condition is met and the training is completed.
8. The method for predicting interdisciplinary student performance based on federated learning according to claim 1, characterized in that: In S3, during the model training process, the loss function uses the cross entropy loss function, as shown in the following formula: In the formula, y oc is the true label, p oc is the probability that the model predicts that sample o belongs to category c, M is the total number of categories, and y oc is a binary indicator that is 1 if class c is the correct classification for sample o and 0 otherwise.