Deep knowledge tracing method based on learning behavior collaborative perception and heterogeneous relationship graph

Abstract

A deep knowledge tracing method based on learning behavior collaborative perception and heterogeneous relationship graph is composed of the steps of data set preprocessing, constructing a multi-factor interaction network, training the multi-factor interaction network, verifying the multi-factor interaction network and testing the multi-factor interaction network. The multi-factor interaction network is adopted in the method, the personalized learning behavior characteristics and exercise characteristics of learners are embedded, the attention enhanced time state updating module is used, the complex interaction relationship between different learning behavior characteristics and exercise characteristics is deeply analyzed, the key features influencing the knowledge state are extracted, and the knowledge state diagnosis of the learning behavior and exercise characteristics enhancement is realized. Compared with the existing knowledge tracing model, the method has good learning effect evaluation performance, high exercise answer prediction accuracy and strong generalization ability, and can be used for knowledge state diagnosis and personalized teaching intervention in an intelligent education system.

Description

Technical Field

[0001] This invention belongs to the field of intelligent education and artificial intelligence technology, specifically involving a knowledge tracking method in the context of online education platforms. Background Technology

[0002] Knowledge tracking, as a core technology of intelligent education systems, aims to dynamically diagnose learners' knowledge mastery status and accurately predict future learning performance by modeling the historical interaction sequences between learners and the system. Its core objective is to provide a scientific and objective basis for educational decision-making through the system's collection and analysis of evidence, thereby ensuring and promoting learners' holistic development.

[0003] With the rapid development of deep learning technology, knowledge tracking models have undergone a transformation from traditional probabilistic models to deep neural networks. Current research mainly focuses on methods based on learner behavioral features and methods based on practice embedding features.

[0004] Learner behavior feature modeling methods focus on the behavioral sequences generated during the learning process, such as answer results, response time, and number of attempts. They capture the dynamic evolution of knowledge states through temporal modeling methods such as recurrent neural networks or Transformers. These methods can directly utilize raw interaction data, but they are insufficient in characterizing the complex synergistic effects between behaviors and are easily affected by noise in the behavioral data.

[0005] Modeling methods based on practice embeddings focus on constructing high-quality practice representations, optimizing practice embeddings using graph neural networks or pre-training techniques to alleviate data sparsity and enhance model interpretability. While these methods can more robustly represent learned content by introducing structured prior knowledge, they still fall short in deeply exploring implicit connections between practices and cognitive dependencies between concepts.

[0006] While the aforementioned methods have made progress in knowledge tracing, in learning behavior modeling, most models simply concatenate multidimensional learning behaviors (such as process behaviors, outcome behaviors, and temporal behaviors) or use them as additional feature inputs, failing to effectively quantify the dynamic influence and inherent synergistic mechanisms between different types of behaviors. Learners' cognitive processes are the result of the combined effects of multiple behaviors; ignoring the intrinsic connections between these behaviors leads to superficial diagnoses of knowledge states. In practice representation, the focus is primarily on explicit associations between practices and concepts, neglecting implicit associations between practices derived from shared relationships and cognitive dependencies between concepts. This simplified modeling approach to knowledge structure networks limits the discriminative power and generalization ability of practice embeddings, thus affecting the accurate updating of learners' knowledge states.

[0007] The synergistic effects of learning behaviors and the complex heterogeneous relationships between practice and concepts are key to accurately characterizing the evolution of knowledge states. Therefore, constructing a knowledge tracing model that can deeply integrate behavioral synergistic mechanisms and knowledge structure networks is of great significance for improving the accuracy and interpretability of knowledge tracing. Summary of the Invention

[0008] The technical problem to be solved by the present invention is to overcome the shortcomings of the prior art and provide a deep knowledge tracking method based on learning behavior collaborative perception and heterogeneous relationship graphs that is highly accurate, interpretable, and robust.

[0009] The technical solution adopted to solve the above technical problems consists of the following steps:

[0010] (1) Dataset preprocessing

[0011] We selected the publicly available educational datasets ASSISTments2009, ASSISTments2012, and ASSISTChall, extracted students' practice interaction sequences, multidimensional learning behavior data, and practice-concept association information, and removed student samples with fewer than 50 interaction records or missing key data as the dataset. The dataset was then divided into training set, validation set, and test set in a 7:1:2 ratio.

[0012] (2) Extracting key features

[0013] Key features are extracted using a relational graph convolutional neural network.

[0014] (3) Constructing a multi-factor interaction network

[0015] The multi-factor interaction network consists of a multi-dimensional learning behavior representation module, a practice embedding enhancement module, a temporal state update module, and a question prediction module. The outputs of the multi-dimensional learning behavior representation module and the practice embedding enhancement module are connected to the temporal state update module, and the temporal state update module is connected in series with the question prediction module.

[0016] (4) Training a multi-factor interaction network

[0017] 1) Construct the cross-entropy loss function

[0018] Construct the cross-entropy loss function as follows :

[0019]

[0020] in, For sample size, The value can be a finite number of positive integers. For the t-th data point, the actual answer result is... Predict the probability for the t-th data point. Here, is the regularization hyperparameter, and n is the total number of weight parameters. Let i be the i-th parameter in the model.

[0021] 2) Training a multi-factor interaction network

[0022] The training set was fed into a multi-factor interactive network for training. The server graphics card used for training was an NVIDIA GeForce RTX 5090. The initial learning rate was 0.001, the number of training epochs was 80, the batch size was 128, and the Adam optimizer was used. Training continued until the cross-entropy loss function was reached. convergence.

[0023] (5) Validate the multi-factor interaction network

[0024] The validation set is input into the trained multi-factor interactive network for validation, predicting the probability of students answering questions correctly, and outputting the area under the receiver operating characteristic curve, accuracy, root mean square error, and coefficient of determination to evaluate the network performance.

[0025] (6) Testing multi-factor interaction networks

[0026] The test set is input into the trained multi-factor interaction network for testing, predicting the probability of students answering the exercises correctly, and outputting deep knowledge tracking results.

[0027] In step (3) of the present invention, the multi-factor interaction network is constructed by connecting the behavior feature encoding layer, the fully connected layer 1, the fully connected layer 2, the multi-head attention layer 1, and the fully connected layer 3 in series.

[0028] In step (3) of the present invention, the construction of a multi-factor interaction network is performed by the time-series state update module, which is composed of a fully connected layer 4, a multi-head attention layer 2, a fully connected layer 5, a long short-term memory network layer, a fully connected layer 6, and a normalization layer connected in series.

[0029] In step (3) of the present invention, the multi-factor interaction network is constructed, and the answer prediction module is composed of a knowledge state focusing layer, a fully connected layer 7, and a fully connected layer 8 connected in series.

[0030] The method for constructing the multi-head attention layer 1 of the present invention is shown in equation (1):

[0031] (1)

[0032]

[0033]

[0034]

[0035]

[0036]

[0037]

[0038]

[0039]

[0040]

[0041]

[0042] in, This represents multi-head attention layer 1. This represents the activation function. Indicates the flattening operation. The output matrix of multi-head attention. For embedded dimensions, , This indicates a horizontal splicing operation. This is the attention output matrix for the corresponding head. Indicates which head, h represents the number of heads that receive attention. , For behavioral collaborative attention matrix, This represents the activation function. This indicates the scaling factor of the header. , , These are the projection matrices for the corresponding heads. , , , For value matrices, For querying the matrix, The key matrix, , , , , , , , For a trainable parameter matrix, For bias terms, For the row matrix, To learn behavior vectors, For the learning result vector, To learn the time vector, This indicates a vertical splicing operation.

[0043] The method for constructing the behavior feature encoding layer of the present invention includes learning behavior vectors. Learning result vector Learning time vector The construction methods for the three vectors are shown in equation (2):

[0044] Determine the learning behavior vector using the following formula :

[0045] (2)

[0046]

[0047] in, This represents the activation function. For a trainable parameter matrix, For the bias term, Let the number of attempts be a vector. To indicate the number of requests, Indicates an embedded function. Number of attempts This indicates the number of requests.

[0048] The learning result vector is determined by the following formula. :

[0049]

[0050]

[0051] in, For the learning result vector, For a trainable parameter matrix, As the vector for identifying exercises, For bias terms, For the learning outcome data, Data is used to identify the exercises.

[0052] Determine the learning time vector using the following formula :

[0053]

[0054]

[0055] in, Let the answer time vector be... For a trainable parameter matrix, , For the concept repetition frequency vector, For bias terms, For answer time data, The number of times the concept is repeated.

[0056] The method for constructing the multi-head attention layer 2 of the present invention is shown in equation (3):

[0057] (3)

[0058]

[0059]

[0060]

[0061]

[0062]

[0063]

[0064]

[0065]

[0066]

[0067]

[0068] in, This indicates multi-head attention layer 2. This represents the activation function. Indicates the flattening operation. The output matrix of multi-head attention. , This indicates a horizontal splicing operation. This is the attention output matrix for the corresponding head. , Indicates which head, For behavioral collaborative attention matrix, , This indicates the scaling factor of the header. , , These are the projection matrices for the corresponding heads. , , , For querying the matrix, , The key matrix, , For value matrices, , , , , , , , , These are the trainable parameter matrices, , , , , , , , , For bias terms, Indicates fusion features, To learn behavioral feature vectors, This indicates a vertical splicing operation. To practice embedding augmentation vectors.

[0069] The Long Short-Term Memory Network Layer of the Invention The construction method is shown in equation (4):

[0070] (4)

[0071]

[0072]

[0073]

[0074]

[0075]

[0076]

[0077] in, This indicates the implicit state at the current moment. These represent the input gate, forget gate, and output gate, respectively. This indicates the current state of the cell. This indicates the cell state at the previous moment. This represents an intermediate variable that temporarily stores new information. This represents the implicit state at the previous moment. This indicates cognitive logic enhancement features. This represents the activation function. , , , , These are the trainable parameter matrices, , , , , , , , , , As a bias term, k represents the number of knowledge points. .

[0078] The knowledge state focusing layer of the present invention The construction method is shown in equation (5):

[0079] (5)

[0080] in, For practice The corresponding concept vector, This is the knowledge state matrix at the current moment.

[0081] Because this invention employs a multi-factor interactive network, deeply integrating multi-dimensional behavioral features of the learning process, results, and time, and extracts key features through a heterogeneous relational graph convolutional network, it solves the problems of shallow utilization of behavioral features and fragmented modeling of practice relationships in existing knowledge tracking models. This accurately captures the dynamic evolution of learners' knowledge states and cognitive logic. Compared with existing knowledge tracking methods, this invention outperforms existing methods in key evaluation metrics and has advantages such as high accuracy in knowledge state diagnosis, strong generalization ability, and good interpretability. It can be used for personalized learning path recommendation, adaptive teaching intervention, and dynamic evaluation of learning outcomes in intelligent education systems. Attached Figure Description

[0082] Figure 1 This is a flowchart of Embodiment 1 of the present invention.

[0083] Figure 2 This is a schematic diagram of the structure of a multi-factor interaction network.

[0084] Figure 3 yes Figure 2 A schematic diagram of the structure of the multidimensional learning behavior representation module.

[0085] Figure 4 yes Figure 2 A schematic diagram of the structure of the timing state update module.

[0086] Figure 5 yes Figure 2 A schematic diagram of the answer prediction module. Specific implementation methods

[0087] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments, but the present invention is not limited to the following embodiments.

[0088] Example 1

[0089] like Figure 1 As shown, the deep knowledge tracing method based on collaborative perception of learning behavior and heterogeneous relationship graph in this embodiment consists of the following steps.

[0090] (1) Dataset preprocessing

[0091] We selected the publicly available educational datasets ASSISTments2009, ASSISTments2012, and ASSISTChall, extracted students' practice interaction sequences, multidimensional learning behavior data, and practice-concept association information, and removed student samples with fewer than 50 interaction records or missing key data as the dataset. The dataset was then divided into training set, validation set, and test set in a 7:1:2 ratio.

[0092] (2) Extracting key features

[0093] Key features are extracted using a relational graph convolutional neural network.

[0094] (3) Constructing a multi-factor interaction network

[0095] Figure 2 A schematic diagram of the multi-factor interaction network in this embodiment is provided. Figure 2 In this embodiment, the multi-factor interaction network is composed of a multi-dimensional learning behavior representation module, a practice embedding enhancement module, a temporal state update module, and a question prediction module connected together. The outputs of the multi-dimensional learning behavior representation module and the practice embedding enhancement module are connected to the temporal state update module, and the temporal state update module is connected in series with the question prediction module.

[0096] Figure 3 shows Figure 2 A schematic diagram of the structure of the multidimensional learning behavior representation module. Figure 3 In this embodiment, the multidimensional learning behavior representation module is composed of a behavior feature encoding layer and fully connected layer 1, fully connected layer 2, multi-head attention layer 1, and fully connected layer 3 connected in series. The fully connected layer 1, fully connected layer 2, and fully connected layer 3 have the same structure.

[0097] The method for constructing the multi-head attention layer 1 in this embodiment is shown in equation (1):

[0098] (1)

[0099]

[0100]

[0101]

[0102]

[0103]

[0104]

[0105]

[0106]

[0107]

[0108]

[0109] in, This represents multi-head attention layer 1. This represents the activation function. Indicates the flattening operation. The output matrix of multi-head attention. For embedded dimensions, In this embodiment, d takes the value 2. 7 , This indicates a horizontal splicing operation. This is the attention output matrix for the corresponding head. Indicates which head, h represents the number of heads that receive attention. In this embodiment, the value of h is 2. 2 , For behavioral collaborative attention matrix, This represents the activation function. This indicates the scaling factor of the header. , , These are the projection matrices for the corresponding heads. , , , For value matrices, For querying the matrix, The key matrix, , , , , , , , For a trainable parameter matrix, For bias terms, For the row matrix, To learn behavior vectors, For the learning result vector, To learn the time vector, This indicates a vertical splicing operation.

[0110] The method for constructing the behavior feature encoding layer in this embodiment includes learning behavior vectors. Learning result vector Learning time vector The construction methods for the three vectors are as follows:

[0111] Determine the learning behavior vector according to formula (2). :

[0112] (2)

[0113]

[0114] in, This represents the activation function. For a trainable parameter matrix, For the bias term, Let the number of attempts be a vector. To indicate the number of requests, Indicates an embedded function. Number of attempts This indicates the number of requests.

[0115] The learning result vector is determined by the following formula. :

[0116]

[0117]

[0118] in, For the learning result vector, For a trainable parameter matrix, As the vector for identifying exercises, For bias terms, For the learning outcome data, Data is used to identify the exercises.

[0119] Determine the learning time vector using the following formula :

[0120]

[0121]

[0122] in, Let the answer time vector be... For a trainable parameter matrix, , For the concept repetition frequency vector, For bias terms, For answer time data, The number of times the concept is repeated.

[0123] Figure 4 Given Figure 2 A schematic diagram of the structure of the time-series state update module. Figure 4 In this embodiment, the temporal state update module is composed of a fully connected layer 4, a multi-head attention layer 2, a fully connected layer 5, a long short-term memory network layer, a fully connected layer 6, and a normalization layer connected in series. The structures of the fully connected layers 4, 5, and 6 are the same as those of the fully connected layer 1.

[0124] The method for constructing the multi-head attention layer 2 in this embodiment is shown in equation (3):

[0125] (3)

[0126]

[0127]

[0128]

[0129]

[0130]

[0131]

[0132]

[0133]

[0134]

[0135]

[0136] in, This indicates multi-head attention layer 2. This represents the activation function. Indicates the flattening operation. The output matrix of multi-head attention. , This indicates a horizontal splicing operation. This is the attention output matrix for the corresponding head. , Indicates which head, For behavioral collaborative attention matrix, , This indicates the scaling factor of the header. , , These are the projection matrices for the corresponding heads. , , , For querying the matrix, , The key matrix, , For value matrices, , , , , , , , , These are the trainable parameter matrices, , , , , , , , , For bias terms, Indicates fusion features, To learn behavioral feature vectors, This indicates a vertical splicing operation. To practice embedding augmentation vectors.

[0137] The Long Short-Term Memory Network Layer in this embodiment The construction method is shown in equation (4):

[0138] (4)

[0139]

[0140]

[0141]

[0142]

[0143]

[0144]

[0145] in, This indicates the implicit state at the current moment. These represent the input gate, forget gate, and output gate, respectively. This indicates the current state of the cell. This indicates the cell state at the previous moment. This represents an intermediate variable that temporarily stores new information. This represents the implicit state at the previous moment. This indicates cognitive logic enhancement features. This represents the activation function. , , , , These are the trainable parameter matrices, , , , , , , , , , As a bias term, k represents the number of knowledge points. In this embodiment, k is set to 250.

[0146] Figure 5 Given Figure 2 A schematic diagram of the answer prediction module. Figure 5 In this embodiment, the question prediction module is composed of a knowledge state focusing layer and fully connected layers 7 and 8 connected in series. The structures of fully connected layers 7 and 8 are the same as those of fully connected layer 1.

[0147] This embodiment features a knowledge state focusing layer. The construction method is shown in equation (5):

[0148] (5)

[0149] in, For practice The corresponding concept vector, This is the knowledge state matrix at the current moment.

[0150] Because this invention employs a multi-factor interactive network, embedding learners' personalized learning behavior characteristics and practice characteristics, and using an attention-enhanced temporal state update module, it deeply analyzes the complex interactive relationships between different learning behavior characteristics and practice characteristics, extracts key features that affect knowledge state, and achieves knowledge state diagnosis enhanced by learning behavior and practice characteristics.

[0151] (4) Training a multi-factor interaction network

[0152] 1) Construct the cross-entropy loss function

[0153] Construct the cross-entropy loss function as follows :

[0154]

[0155] in, For sample size, The value range is 3000 to 50000, in this embodiment... The value range is 25000. For the t-th data point, the actual answer result is... Predict the probability for the t-th data point. As a regularization hyperparameter, this embodiment The value is 0.0001, and n is the total number of weight parameters. This refers to the i-th parameter in the model;

[0156] 2) Training a multi-factor interaction network

[0157] The training set was fed into a multi-factor interactive network for training. The server graphics card used for training was an NVIDIA GeForce RTX 5090. The initial learning rate was 0.001, the number of training epochs was 80, the batch size was 128, and the Adam optimizer was used. Training continued until the cross-entropy loss function was reached. convergence;

[0158] (5) Validate the multi-factor interaction network

[0159] The validation set is input into the trained multi-factor interactive network for validation, predicting the probability of students answering questions correctly, and outputting the area under the receiver operating characteristic curve, accuracy, root mean square error, and coefficient of determination index to evaluate the network performance.

[0160] (6) Testing multi-factor interaction networks

[0161] The test set is input into the trained multi-factor interaction network for testing, predicting the probability of students answering the exercises correctly, and outputting deep knowledge tracking results.

[0162] Complete a deep knowledge tracing method based on collaborative perception of learning behavior and heterogeneous relationship graphs.

[0163] Example 2

[0164] The deep knowledge tracing method based on learning behavior collaborative perception and heterogeneous relationship graph in this embodiment consists of the following steps.

[0165] (1) Dataset preprocessing

[0166] The steps are the same as in Example 1.

[0167] (2) Extracting key features

[0168] The steps are the same as in Example 1.

[0169] (3) Constructing a multi-factor interaction network

[0170] The multi-factor interaction network structure is the same as in Example 1.

[0171] The method for constructing the multi-head attention layer 1 in this embodiment is as shown in equation (1):

[0172] The expression of equation (1) is the same as that in Example 1.

[0173] In equation (1), For embedded dimensions, In this embodiment, d takes the value 2. 5 h represents the number of heads that receive attention. In this embodiment, the value of h is 2. 0 The meanings and value ranges of other parameters and variables are the same as in Example 1.

[0174] The method for constructing the behavior feature encoding layer in this embodiment includes learning behavior vectors. Learning result vector Learning time vector The construction methods for the three vectors are as follows:

[0175] Determine the learning behavior vector according to formula (2). :

[0176] The expression of equation (2) is the same as that in Example 1.

[0177] In equation (2), the meanings and ranges of the parameters and variables are the same as in Example 1.

[0178] The method for constructing the multi-head attention layer 2 in this embodiment is shown in equation (3):

[0179] In equation (3), the meanings and ranges of the parameters and variables are the same as in Example 1.

[0180] The Long Short-Term Memory Network Layer in this embodiment The construction method is shown in equation (4):

[0181] In equation (4), k represents the number of knowledge points. In this embodiment, k is set to 100, and the meanings and value ranges of other parameters and variables are the same as in embodiment 1.

[0182] This embodiment features a knowledge state focusing layer. The construction method is shown in equation (5):

[0183] In equation (5), the meanings and ranges of the parameters and variables are the same as in Example 1.

[0184] (4) Training a multi-factor interaction network

[0185] 1) Construct the cross-entropy loss function

[0186] Construct the cross-entropy loss function as follows :

[0187]

[0188] in, For sample size, The value range is 3000 to 50000, in this embodiment... The value range is 3000, and the meanings and value ranges of other parameters and variables are the same as in Example 1.

[0189] The other steps in this procedure are the same as in Example 1.

[0190] The other steps are the same as in Example 1.

[0191] Complete a deep knowledge tracing method based on collaborative perception of learning behavior and heterogeneous relationship graphs.

[0192] Example 3

[0193] The deep knowledge tracing method based on learning behavior collaborative perception and heterogeneous relationship graph in this embodiment consists of the following steps.

[0194] (1) Dataset preprocessing

[0195] The steps are the same as in Example 1.

[0196] (2) Extracting key features

[0197] The steps are the same as in Example 1.

[0198] (3) Constructing a multi-factor interaction network

[0199] The multi-factor interaction network structure is the same as in Example 1.

[0200] The method for constructing the multi-head attention layer 1 in this embodiment is as shown in equation (1):

[0201] The expression of equation (1) is the same as that in Example 1.

[0202] In equation (1), For embedded dimensions, In this embodiment, d takes the value 2. 8 h represents the number of heads that receive attention. In this embodiment, the value of h is 2. 4 The meanings and value ranges of other parameters and variables are the same as in Example 1.

[0203] The method for constructing the behavior feature encoding layer in this embodiment includes learning behavior vectors. Learning result vector Learning time vector The construction methods for the three vectors are as follows:

[0204] Determine the learning behavior vector according to formula (2). :

[0205] The expression of equation (2) is the same as that in Example 1.

[0206] In equation (2), the meanings and ranges of the parameters and variables are the same as in Example 1.

[0207] The method for constructing the multi-head attention layer 2 in this embodiment is shown in equation (3):

[0208] In equation (3), the meanings and ranges of the parameters and variables are the same as in Example 1.

[0209] The Long Short-Term Memory Network Layer in this embodiment The construction method is shown in equation (4):

[0210] In equation (4), k represents the number of knowledge points. In this embodiment, k is 400, and the meanings and value ranges of other parameters and variables are the same as in embodiment 1.

[0211] This embodiment features a knowledge state focusing layer. The construction method is shown in equation (5):

[0212] In equation (5), the meanings and ranges of the parameters and variables are the same as in Example 1.

[0213] (4) Training a multi-factor interaction network

[0214] 1) Construct the cross-entropy loss function

[0215] Construct the cross-entropy loss function as follows :

[0216]

[0217] in, For sample size, The value range is 3000 to 50000, in this embodiment... The value range is 50000, and the meanings and value ranges of other parameters and variables are the same as in Example 1.

[0218] The other steps in this procedure are the same as in Example 1.

[0219] The other steps are the same as in Example 1.

[0220] Complete a deep knowledge tracing method based on collaborative perception of learning behavior and heterogeneous relationship graphs.

[0221] To verify the beneficial effects of the present invention, the inventors conducted comparative experiments using the deep knowledge tracing method based on learning behavior collaborative perception and heterogeneous relationship graphs from Embodiment 1 of the present invention, and the following methods: deep knowledge tracing method based on recurrent neural networks (hereinafter referred to as DKT), knowledge tracing method based on dynamic key-value memory networks (hereinafter referred to as DKVMN), knowledge tracing method based on Transformer architecture (hereinafter referred to as SAKT), knowledge tracing method based on attention mechanism (hereinafter referred to as AKT), knowledge tracing method based on learning and forgetting rules (hereinafter referred to as LPKT), knowledge tracing method based on learning and forgetting rules extension (hereinafter referred to as LPKT-S), and knowledge tracing method based on graph convolutional networks (hereinafter referred to as GIKT).

[0222] Evaluation metrics were determined as follows: To conduct a scientific and comprehensive performance evaluation and comparison, four evaluation metrics widely used in the knowledge tracing field were selected: Area Under the ROC Curve (AUC), Accuracy (ACC), Root Mean Square Error (RMSE), and Coefficient of Determination (R²). 2 These metrics comprehensively evaluate model performance based on model discrimination, classification accuracy, prediction precision, and goodness of fit.

[0223] AUC: Measures the model's overall ability to distinguish between positive and negative samples (correct / incorrect answers). A value closer to 1 indicates better model performance. The calculation formula is:

[0224]

[0225] Where TP, TN, FP, and FN represent the number of true examples, true negative examples, false positive examples, and false negative examples, respectively.

[0226] ACC: Calculates the proportion of samples correctly predicted by the model out of the total samples, reflecting the overall classification accuracy.

[0227]

[0228] Where TP, TN, FP, and FN represent the number of true examples, true negative examples, false positive examples, and false negative examples, respectively.

[0229] RMSE: Measures the root mean square error between the predicted probability and the true label. The smaller the value, the more accurate the prediction.

[0230]

[0231] Where N is the total number of samples, For real labels, To predict probabilities.

[0232] R 2 This reflects the model's ability to explain data variation; the closer the value is to 1, the better the model fits the data.

[0233]

[0234] in, For real labels, To predict probabilities, It is the average of the actual labels.

[0235] The results of the comparative experiment are shown in Table 1.

[0236] Table 1. Experimental results of the method in Example 1 and the comparative experimental method.

[0237]

[0238] As shown in Table 1, on the ASSISTments2009, ASSISTments2012, and ASSISTChall datasets, the AUC of the method in Example 1 is 0.841, 0.762, and 0.883, respectively, which is higher than that of the DKT, DKVMN, AKT, SAKT, LPKT, LPKT-S, and GIKT methods. Similarly, on the three datasets, the ACC of the method in Example 1 is 0.773, 0.762, and 0.799, respectively, which is higher than that of the DKT, DKVMN, AKT, SAKT, LPKT, LPKT-S, and GIKT methods. On the ASSISTments2012 and ASSISTChall datasets, the RMSE of the method in Example 1 is 0.399 and 0.364, respectively, which is lower than that of the DKT, DKVMN, AKT, SAKT, LPKT, and LPKT-S methods. On the ASSISTments2012 and ASSISTChall datasets, Example 1 method R 2 The scores are 0.223 and 0.397, which are higher than those of the DKT, DKVMN, AKT, SAKT, LPKT, and LPKT-S methods. Experiments show that the method of this invention has better accuracy in knowledge state assessment.

Claims

1. A deep knowledge tracing method based on collaborative perception of learning behavior and heterogeneous relationship graphs, characterized in that... It consists of the following steps: (1) Dataset preprocessing We selected the publicly available educational datasets ASSISTments2009, ASSISTments2012, and ASSISTChall, extracted students' practice interaction sequences, multidimensional learning behavior data, and practice-concept association information, and removed student samples with fewer than 50 interaction records or missing key data as the dataset. The dataset was then divided into training set, validation set, and test set in a ratio of 7:1:

2. (2) Extracting key features Key features are extracted using a relational graph convolutional neural network; (3) Constructing a multi-factor interaction network The multi-factor interaction network consists of a multi-dimensional learning behavior representation module, a practice embedding enhancement module, a temporal state update module, and a question prediction module. The outputs of the multi-dimensional learning behavior representation module and the practice embedding enhancement module are connected to the temporal state update module, and the temporal state update module is connected in series with the question prediction module. (4) Training a multi-factor interaction network 1) Construct the cross-entropy loss function Construct the cross-entropy loss function as follows : in, For sample size, The value can be a finite number of positive integers. For the first t These are the actual answers to the questions. For the first t Predict the probability of each data point. For regularization hyperparameters, n The total number of weight parameters. For the model i One parameter; 2) Training a multi-factor interaction network The training set was fed into a multi-factor interactive network for training. The server graphics card used for training was an NVIDIA GeForce RTX 5090. The initial learning rate was 0.001, the number of training epochs was 80, the batch size was 128, and the Adam optimizer was used. Training continued until the cross-entropy loss function was reached. convergence; (5) Validate the multi-factor interaction network The validation set is input into the trained multi-factor interactive network for validation, predicting the probability of students answering questions correctly, and outputting the area under the receiver operating characteristic curve, accuracy, root mean square error, and coefficient of determination index to evaluate the network performance. (6) Testing multi-factor interaction networks The test set is input into the trained multi-factor interaction network for testing, predicting the probability of students answering the exercises correctly, and outputting deep knowledge tracking results.

2. The deep knowledge tracing method based on collaborative perception of learning behavior and heterogeneous relationship graphs according to claim 1, characterized in that: In step (3), the multi-factor interaction network is constructed. The multi-dimensional learning behavior representation module is composed of a behavior feature encoding layer and a fully connected layer 1, a fully connected layer 2, a multi-head attention layer 1, and a fully connected layer 3 connected in series.

3. The deep knowledge tracing method based on collaborative perception of learning behavior and heterogeneous relationship graphs according to claim 1, characterized in that: In step (3), the multi-factor interaction network is constructed. The time-series state update module is composed of a fully connected layer 4, a multi-head attention layer 2, a fully connected layer 5, a long short-term memory network layer, a fully connected layer 6, and a normalization layer connected in series.

4. The deep knowledge tracing method based on collaborative perception of learning behavior and heterogeneous relationship graphs according to claim 1, characterized in that: In step (3), the multi-factor interaction network is constructed. The answer prediction module is composed of a knowledge state focusing layer, a fully connected layer 7, and a fully connected layer 8 connected in series.

5. The deep knowledge tracing method based on collaborative perception of learning behavior and heterogeneous relationship graphs according to claim 2, characterized in that, The construction method of the multi-head attention layer 1 is shown in equation (1): （1） in, This represents multi-head attention layer 1. This represents the activation function. Indicates the flattening operation. The output matrix of multi-head attention. For embedded dimensions, , This indicates a horizontal splicing operation. This is the attention output matrix for the corresponding head. Indicates which head, h represents the number of heads that receive attention. , For behavioral collaborative attention matrix, This represents the activation function. This indicates the scaling factor of the header. , , These are the projection matrices for the corresponding heads. , , , For value matrices, For querying the matrix, The key matrix, , , , , , , , For a trainable parameter matrix, For bias terms, For the row matrix, To learn behavior vectors, For the learning result vector, To learn the time vector, This indicates a vertical splicing operation.

6. The deep knowledge tracing method based on collaborative perception of learning behavior and heterogeneous relationship graphs according to claim 2, characterized in that, The method for constructing the behavior feature encoding layer includes learning behavior vectors. Learning result vector Learning time vector The construction methods for the three vectors are shown in equation (2): Determine the learning behavior vector using the following formula : （2） in, This represents the activation function. For a trainable parameter matrix, For the bias term, Let the number of attempts be a vector. To indicate the number of requests, Indicates an embedded function. Number of attempts To indicate the number of requests; The learning result vector is determined by the following formula. : in, For the learning result vector, For a trainable parameter matrix, As the vector for identifying exercises, For bias terms, For the learning outcome data, Identify data for the exercises; Determine the learning time vector using the following formula : in, Let the answer time vector be... For a trainable parameter matrix, , For the concept repetition frequency vector, For bias terms, For answer time data, The number of times the concept is repeated.

7. The deep knowledge tracing method based on collaborative perception of learning behavior and heterogeneous relationship graphs according to claim 3, characterized in that, The method for constructing the multi-head attention layer 2 is shown in equation (3): （3） in, This indicates multi-head attention layer 2. This represents the activation function. Indicates the flattening operation. The output matrix of multi-head attention. , This indicates a horizontal splicing operation. This is the attention output matrix for the corresponding head. , Indicates which head, For behavioral collaborative attention matrix, , This indicates the scaling factor of the header. , , These are the projection matrices for the corresponding heads. , , , For querying the matrix, , The key matrix, , For value matrices, , , , , , , , , These are the trainable parameter matrices, , , , , , , , , For bias terms, Indicates fusion features, To learn behavioral feature vectors, This indicates a vertical splicing operation. To practice embedding augmentation vectors.

8. The deep knowledge tracing method based on collaborative perception of learning behavior and heterogeneous relationship graphs according to claim 3, characterized in that, The Long Short-Term Memory Network Layer The construction method is shown in equation (4): （4） in, This indicates the implicit state at the current moment. These represent the input gate, forget gate, and output gate, respectively. This indicates the current state of the cell. This indicates the cell state at the previous moment. This represents an intermediate variable that temporarily stores new information. This represents the implicit state at the previous moment. This indicates cognitive logic enhancement features. This represents the activation function. , , , , These are the trainable parameter matrices, , , , , , , , , , As a bias term, k represents the number of knowledge points. .

9. The deep knowledge tracing method based on collaborative perception of learning behavior and heterogeneous relationship graphs according to claim 4, characterized in that, The knowledge state focusing layer The construction method is shown in equation (5): （5） in, For practice The corresponding concept vector, This is the knowledge state matrix at the current moment.

Images