Knowledge tracking method based on online learning dynamic interaction graph and related equipment

Abstract

The invention belongs to the technical field of wisdom education, and provides a knowledge tracking method based on an online learning dynamic interaction graph and related equipment. The method comprises the following steps: step A, obtaining historical learning parameters of a target student and performing effect quantification to obtain learning performance of the target student; step B, constructing a learner-exercise interaction graph and an exercise-knowledge interaction graph based on the learning performance; c, performing embedded learning on the learner-exercise interaction graph and the exercise-knowledge interaction graph by using a graph convolutional network to obtain an exercise embedded feature representation; and step D, arranging the exercise embedding feature representations into an exercise embedding feature sequence according to the answering sequence, and inputting the exercise embedding feature sequence into a preset LSTM model to obtain the probability that the target student correctly answers the target exercise. The method can improve the prediction performance of knowledge tracking.

Description

Technical Field

[0001] The present invention belongs to the field of smart education technology, and in particular relates to a knowledge tracking method based on an online learning dynamic interactive graph and related equipment. Background Art

[0002] Today's online education platforms (such as MOOCs, Udemy, Lynda, Udacity, Coursera, and edX) have attracted many scholars to online learning. The advantage of this learning model is that it can fully record learners' learning trajectories, thereby evaluating their knowledge mastery and learning performance. Knowledge tracking technology, as a key technology and research hotspot for online education platforms, aims to assess learners' knowledge mastery based on their learning records and predict their future learning performance. Using knowledge tracking technology, learners can focus their time and energy on knowledge points that have not yet been mastered, thereby improving their learning efficiency. At the same time, for educators, knowledge tracking technology can also provide more accurate teaching feedback and strategic recommendations, promoting the overall improvement of teaching quality.

[0003] Early knowledge tracing technologies primarily relied on statistical methods, such as Bayesian Knowledge Tracing (BKT) and Dynamic Bayesian Knowledge Tracing (DBKT). While these methods can track learners' knowledge states, their modeling of knowledge states is overly simplistic, making them inadequate for the complex knowledge systems found in existing courses. With the rise of deep learning technology, the Deep Knowledge Tracing (DKT) method models learning states by treating learners' answer sequences as time series data, demonstrating superior ability to represent dynamic knowledge states. However, knowledge tracing methods based on deep learning models primarily focus on learning exercises and knowledge features, neglecting to fully understand learners' learning behaviors. In fact, the effectiveness of knowledge acquisition can be more profoundly reflected through interactive details during the learning process, such as learning ability (answer accuracy), number of attempts, and number of prompts. In recent years, with the continued maturity of graph neural networks (GNNs), graph-based knowledge tracing methods, such as GKT (Graph-based Knowledge Tracing), GIKT (Graph-based Interaction Knowledge Tracing), and SKT (Structure-based Knowledge Tracing), have gradually become a research hotspot. Graph-based knowledge tracing models transform the knowledge tracing task into a time series node-level classification problem within GNNs, demonstrating strong performance in learning feature extraction. However, current graph-based knowledge tracing models focus on modeling the dynamic interaction between exercises and knowledge, neglecting the detailed interactions between learners during the exercise-answering process. In fact, the detailed interactions between learners and exercises are more important for effectively predicting learners' learning performance. Summary of the Invention

[0004] To address the problem that existing models do not adequately model the detailed information of learners' answers, this paper proposes a knowledge tracking method based on an online learning dynamic interaction graph. By capturing the complex dependencies between learners, exercises, and knowledge, it improves the predictive performance of knowledge tracking and enriches the detailed information of learners' answers in the model.

[0005] In a first aspect, the present invention provides a knowledge tracking method based on an online learning dynamic interaction graph, comprising:

[0006] Step A: Obtain students’ historical learning parameters and quantify the effects to obtain the learning performance of the target students;

[0007] Step B: constructing a learner-exercise interaction graph and an exercise-knowledge interaction graph based on the learning performance;

[0008] Step C: Using a graph convolutional network to perform embedding learning on the learner-exercise interaction graph and the exercise-knowledge interaction graph to obtain an exercise embedding feature representation;

[0009] Step D: Arrange the exercise embedding feature representations into an exercise embedding feature sequence according to the target student's answering order, input the exercise embedding feature sequence into a preset LSTM model, and obtain the probability that the target student correctly answers the target exercise.

[0010] Furthermore, the learning parameters include answer accuracy, number of attempts and number of prompts.

[0011] Furthermore, the step A specifically includes:

[0012] Step A1: Quantify the accuracy of answering questions. Quantify the students' accuracy of answering questions to represent their learning ability, as shown below:

[0013]

[0014] Among them, a k represents the student's answer to the kth exercise, K represents the number of exercises, c i Indicates learning ability;

[0015] Step A2: Quantify the number of attempts based on the student's performance on a specific exercise. j Calculate the attempt factor based on the number of attempts:

[0016]

[0017] in, represents the average number of attempts, p ij Represents a specific exercise q j Number of attempts, PC ij Represents a specific exercise q j The attempt factor;

[0018] Calculate the question answering knowledge benefit factor based on the attempt factor

[0019]

[0020] Among them, α, β and η represent adjustable parameters;

[0021] Step A3: Quantify the number of prompts based on the student's progress on a specific exercise. j Calculate the prompt factor based on the number of prompts:

[0022]

[0023] in, represents the average number of prompts, n ij Represents a specific exercise q j Number of prompts, NC ij Represents a specific exercise q j The prompt factor;

[0024] Calculate the prompt knowledge benefit factor based on the prompt factor

[0025]

[0026] Among them, α, β and η represent adjustable parameters;

[0027] Step A4: Comprehensively represent learning performance by linearly combining the learning ability, answering knowledge benefit factor, and prompting knowledge benefit factor to obtain the student's performance on a specific exercise q. j Learning performance on:

[0028]

[0029] Among them, w c , w p and w n represents a learnable parameter to adjust the relative importance of each knowledge acquisition vector;

[0030] Step A5: Repeat steps A1 to A4 until each student's learning performance on all exercises is obtained.

[0031] Furthermore, the student-exercise interaction diagram is specifically represented as Where U represents the set of learners, Q represents the set of exercises, represents the relationship between the learner and the exercise, determined by the learning performance;

[0032] The exercise-knowledge interaction graph is specifically expressed as G={Q, S, A}, where S represents the knowledge set and A represents the relationship between exercises and knowledge.

[0033] Furthermore, step C specifically includes:

[0034] Step C1: Input the learner-exercise interaction graph into the graph convolutional network to obtain the first deep embedding representation of the exercise, which is expressed by the following formula:

[0035]

[0036] in, and represents the learnable weights and biases, l and l-1 represent the number of embedding layers of the graph convolutional network, Indicates that from the student node Pass to exercise q j learning performance information, Indicates the number of student nodes connected to the exercises, Represents the information conveyed by the exercise itself, and ReLU represents the nonlinear activation function;

[0037] Step C2: Input the exercise-knowledge interaction graph into the graph convolutional network to obtain a second deep embedding representation of the exercise, which is expressed by the following formula:

[0038]

[0039]

[0040] in, and denotes the learnable weights and biases, Indicates that from the knowledge node Pass to exercise q j Information, Indicates the number of knowledge nodes connected by the exercises, Representation Exercise q j The information it conveys;

[0041] Step C3: Fuse the first deep embedding representation and the second deep embedding representation to obtain the third exercise embedding feature

[0042]

[0043] in, and represents the weight, represents the first deep embedding representation of all exercises, and q represents the second deep embedding representation of all exercises.

[0044] Furthermore, an attention mechanism is introduced in step D.

[0045] Furthermore, step D specifically includes:

[0046] Step D1: For each exercise, concatenate the exercise and the corresponding answer to form the final exercise representation:

[0047]

[0048] Among them, e t represents the tth final exercise, ReLU represents the ReLU function, w t and bt denote weight and bias respectively, represents the embedding feature representation of the t-th exercise, a t Represents The corresponding answers indicate;

[0049] Step D2: Arrange the final exercise representations into an exercise embedding feature sequence according to the target student's answering order, input the exercise embedding feature sequence into the preset LSTM model, use the answering order as the time step, and use the LSTM input gate, forget gate, output gate, and cell state update mechanism to learn the target student's knowledge state from the input representation. The formula is as follows:

[0050] i t =σ(w ei e t +w hi h t-1 +b i )

[0051] g t =σ(w ef e t +w hf h t-1 +b f )

[0052] o t =σ(w eo e t +w ho h t-1 +b o )

[0053] c t =g t c t-1 +i t tanh(w ec e t +w hc h t-1 +b c ) h t =o t tanh(c t )

[0054] Among them, h t-1 represents the output of the previous time step, w * and b * represents the network parameters, i t 、g t and o t Represent the input gate, forget gate and output gate respectively, c tRepresents the unit storage vector. σ(·) represents the Sigmoid function, h t Represents the current state of knowledge;

[0055] Step D3: Use the attention mechanism to capture the relationship between the target exercise and the historical state, and select the top k states with the highest attention scores as the relevant historical knowledge state set I k :

[0056] r y′ ={f(q t ′,q t ),t′∈[tl,…,t-1]}

[0057]

[0058] Where f(·) represents the attention ranking function, which is used to calculate the relevance between the top k history exercises and the target exercise, and the topk(·) function selects the knowledge state of the top k most relevant history exercises. k represents a set of k historical knowledge states;

[0059] Step D4: Predicting the probability of correctly answering the target question based on the relevant historical knowledge state set, which is expressed by the following formula:

[0060]

[0061] Among them, f i represents the interaction term between the current knowledge state and the historical knowledge state, f j represents the interaction term between the target exercise embedding and the corresponding knowledge embedding, S q Representation Exercises The knowledge involved is embedded in the representation, P t represents the probability that the target student correctly answers the target exercise, represents the attention mechanism, and g(·) represents the inner product.

[0062] In a second aspect, the present invention provides a knowledge tracking device based on an online learning dynamic interaction graph, comprising:

[0063] The learning performance module is used to obtain students' historical learning parameters and quantify the effects to obtain the learning performance of the target students;

[0064] An interaction graph construction module constructs a learner-exercise interaction graph and an exercise-knowledge interaction graph based on the learning performance;

[0065] An embedding feature learning module uses a graph convolutional network to perform embedding learning on the learner-exercise interaction graph and the exercise-knowledge interaction graph to obtain an exercise embedding feature representation;

[0066] The exercise answer prediction module is used to arrange the exercise embedding feature representations into an exercise embedding feature sequence according to the target student's answer order, input the exercise embedding feature sequence into a preset LSTM model, and obtain the probability that the target student correctly answers the target exercise.

[0067] In a third aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the method described above when executing the program.

[0068] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program executes the method described above when executed by a processor.

[0069] The beneficial effects of the present invention are:

[0070] The present invention takes learners' online answering records as the research object, and quantifies the impact of answering accuracy, number of attempts and number of prompts on learners' knowledge gains, providing a new way to evaluate learners' knowledge status in a fine-grained and refined manner, and improving the accuracy of answer prediction.

[0071] The present invention uses a graph convolutional network (GCN) model to perform embedding learning on the learner-exercise interaction graph structure and the exercise-knowledge interaction graph structure, which can effectively obtain the complex dependency relationship between learners, exercises, and knowledge, and realize the comprehensive feature representation of learners, exercises, and knowledge; in addition, the present invention uses a graph embedding feature fusion mechanism to fuse the exercise representations in the two interaction graphs to obtain an enhanced embedded feature representation that includes the learner's answering performance and knowledge mastery, thereby effectively evaluating the learner's knowledge status and further improving the accuracy of answer prediction.

[0072] The proposed method has been extensively tested for performance analysis and effectiveness on three publicly available educational datasets: ASSISTments2009, ASSISTments2012, and Junyi. The results demonstrate that the knowledge tracing method based on the online learning dynamic interaction graph significantly outperforms all baseline models in predicting learners' future performance, achieving an ACC (Accuracy) exceeding 85% and an AUC (Area Under the Curve) exceeding 82%. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 A flowchart of a knowledge tracking method based on an online learning dynamic interaction graph provided by an embodiment of the present invention;

[0074] Figure 2A schematic diagram of a framework of a knowledge tracking method based on an online learning dynamic interaction graph provided by an embodiment of the present invention;

[0075] Figure 3 A schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0076] To make the objectives, technical solutions, and advantages of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0077] like Figure 1 and Figure 2 As shown, an embodiment of the present invention provides a knowledge tracking method based on an online learning dynamic interaction graph, including:

[0078] Step A: Obtain students’ historical learning parameters and quantify the effects to obtain the learning performance of the target students.

[0079] The learning parameters include answer accuracy, number of attempts, and number of prompts. In this embodiment of the present invention, answer accuracy, number of attempts, and number of prompts are selected as key features to quantitatively evaluate the learner's knowledge acquisition dynamics.

[0080] Step B: Construct learner-exercise interaction diagram and exercise-knowledge interaction diagram based on learning performance.

[0081] Among them, the student-exercise interaction diagram is specifically expressed as Where U represents the set of learners, Q represents the set of exercises, It represents the relationship between learners and exercises, which is determined by learning performance; the exercise-knowledge interaction graph is specifically expressed as G = {Q, S, A}, where Q represents the exercise set, S represents the knowledge set, and A represents the relationship between exercises and knowledge.

[0082] Specifically, the learner-exercise interaction diagram can reflect the dynamic interaction between the learner and the exercise, and the exercise-knowledge interaction diagram can reflect the complex relationship between the exercise and the knowledge. The association weight between the learner and the exercise is determined according to the learner's learning performance on the exercise. If the learner answers the exercise, then A uq =g ij , otherwise A uq = 0. The association weight between the exercise and the knowledge is determined by whether the exercise involves the knowledge. If so, A qs =1, otherwise A qs =0.

[0083] Step C: Use the graph convolutional network to embed the learner-exercise interaction graph and the exercise-knowledge interaction graph to obtain the exercise embedding feature representation.

[0084] By using graph convolutional networks to embed two interaction graphs, we can obtain the implicit dynamic embedding representations in the two graph structures. By fusing them, we can obtain the embedded feature representation of exercises that includes learners' answering performance and knowledge mastery.

[0085] Step D: Arrange the exercise embedding feature representations into an exercise embedding feature sequence according to the target student's answer order, input the exercise embedding feature sequence into the preset LSTM model, and obtain the probability that the target student correctly answers the target exercise.

[0086] The method provided by the embodiment of the present invention first quantifies the online learning process of online learners and models the impact of learning dynamics (learners' learning ability (students' answering accuracy), number of answering attempts, and number of answering prompts) on learners' knowledge gains; then analyzes learners' answering performance, models the dynamic interaction process among learners, exercises, and knowledge, constructs learner-exercise interaction graphs and exercise-knowledge interaction graphs, and expresses the dynamic interaction process between learners and exercises, and exercises and knowledge. Graph convolutional networks (GCNs) are then used to learn the dynamic interaction features between learners and exercises, and between exercises and knowledge, and perform feature fusion on them to comprehensively express learners' learning performance and knowledge mastery. Finally, the method inputs the exercise embedding feature representation output in the previous step into the LSTM structure to capture the learners' knowledge state changes and predict their future performance in answering questions.

[0087] Based on the above embodiment, this embodiment provides a learning parameter quantization method, which specifically includes:

[0088] Step A1: Quantify the accuracy of answering questions. Quantify the students' accuracy of answering questions to represent their learning ability, as shown below:

[0089]

[0090] Among them, a k represents the student's answer to the kth exercise, K represents the number of exercises, c i Indicates learning ability.

[0091] Step A2: Quantify the number of attempts based on the student's performance on a specific exercise. j Calculate the attempt factor based on the number of attempts:

[0092]

[0093] in, represents the average number of attempts, p ij Represents a specific exercise q j Number of attempts, PC ij Represents a specific exercise q j The attempt factor. After research, the number of attempts p ij The probability that a learner's actual number of attempts on a specific exercise is higher than the average number of attempts is approximately Poisson distribution. The higher the attempt factor, the more difficult it is for the learner to solve the exercise.

[0094] Calculating the knowledge benefit factor of answering questions based on the attempt factor

[0095]

[0096] Among them, α, β and η represent adjustable parameters;

[0097] Step A3: Quantify the number of prompts based on the student's progress on a specific exercise. j Calculate the prompt factor based on the number of prompts:

[0098]

[0099] in, represents the average number of prompts, n ij Represents a specific exercise q j Number of prompts, NC ij Represents a specific exercise q j The hint factor. Similar to the number of attempts, the number of hints also follows a Poisson distribution. The hint factor reflects the probability that a learner's actual number of hints on a particular problem is higher than the average number of hints. Higher hint factors indicate greater difficulty for learners solving the problem.

[0100] Calculate the prompt knowledge benefit factor based on the prompt factor

[0101]

[0102] Here, α, β, and η represent adjustable parameters.

[0103] Step A4: Comprehensively represent learning performance by linearly combining learning ability, answering knowledge benefit factor and prompting knowledge benefit factor to obtain the student's performance on a specific exercise q. j Learning performance on:

[0104]

[0105] Among them, w c , w p and w nrepresents a learnable parameter to adjust the relative importance of each knowledge acquisition vector.

[0106] Step A5: Repeat steps A1 to A4 until each student's learning performance on all exercises is obtained.

[0107] Based on the above embodiment, this embodiment provides the specific steps of step C:

[0108] Step C1: Input the learner-exercise interaction graph into the graph convolutional network to obtain the first deep embedding representation of the exercise, which is expressed by the following formula:

[0109]

[0110] in, and represents the learnable weights and biases, l and l-1 represent the number of embedding layers of the graph convolutional network, Indicates that from the student node Pass to exercise q j learning performance information, Indicates the number of student nodes connected to the exercises, Represents the information conveyed by the exercise itself, ReLU represents the nonlinear activation function, Representation Exercise q j The first deep embedding representation of .

[0111] Step C2: Input the exercise-knowledge interaction graph into the graph convolutional network to obtain the second deep embedding representation of the exercise, which is expressed by the following formula:

[0112]

[0113] in, and denotes the learnable weights and biases, Indicates that from the knowledge node Pass to exercise q j Information, Indicates the number of knowledge nodes connected by the exercises, Representation Exercise q j The information conveyed by itself, q j Representation Exercise q j The second deep embedding representation of .

[0114] In addition, for any knowledge node s r ,,This method also uses the same feature aggregation strategy to obtain knowledge representation s r .

[0115] Based on the above steps, all exercises can be embedded with feature representations, which are recorded as and q, as well as the knowledge embedding feature representation, denoted as s.

[0116] Step C3: Fuse the first deep embedding representation and the second deep embedding representation to obtain the third exercise embedding feature

[0117]

[0118] in, and represents the weight, represents the first deep embedding representation of all exercises, and q represents the second deep embedding representation of all exercises.

[0119] Based on the above embodiment, in order to further solve the problem of using the LSTM model to simulate the learner's answering process and generate the learner's knowledge hidden state at each moment, an attention mechanism is introduced in step D to select the knowledge state most closely related to the current exercise as an auxiliary feature fusion prediction process. Accordingly, step D specifically includes:

[0120] Step D1: For each exercise, concatenate the exercise and the corresponding answer to form the final exercise representation:

[0121]

[0122] Among them, e t represents the tth final exercise, ReLU represents the ReLU function, w t and b t denote weight and bias respectively, represents the embedding feature representation of the t-th exercise, a t Represents The corresponding answers indicate;

[0123] Step D2: Arrange the final exercise representations into an exercise embedding feature sequence according to the target student's answer order. Input the exercise embedding feature sequence into the preset LSTM model, use the answer order as the time step, and use the LSTM input gate, forget gate, output gate, and cell state update mechanism to learn the target student's knowledge state from the input representation. The formula is as follows:

[0124] i t =σ(w ei e t +w hi h t-1 +b i )

[0125] g t =σ(w ef e t +w hf ht-1 +b f )

[0126] o t =σ(w eo e t +w ho h t-1 +b o )

[0127] c t =g t c t-1 +i t tanh(w ec e t +w hc h t-1 +b c ) h t =o t tanh(c t )

[0128] Among them, h t-1 represents the output of the previous time step, w * and b * represents the network parameters, i t 、g t and o t Represent the input gate, forget gate and output gate respectively, c t Represents the unit storage vector. σ(·) represents the Sigmoid function, h t Represents the current state of knowledge;

[0129] Step D3: Use the attention mechanism to capture the relationship between the target exercise and the historical state, and select the top k states with the highest attention scores as the relevant historical knowledge state set I k :

[0130] r t′ ={f(q t′ ,q t ),t′∈[tl,…,t-1]}

[0131]

[0132] Where f(·) represents the attention ranking function, which is used to calculate the relevance between the top k history exercises and the target exercise, and the topk(·) function selects the knowledge state of the top k most relevant history exercises. k represents a set of k historical knowledge states;

[0133] Step D4: Predict the probability of correctly answering the target question based on the relevant historical knowledge state set, which is expressed by the following formula:

[0134]

[0135] Among them, f represents the interaction term between the current knowledge state and the historical knowledge state, f j represents the interaction term between the target exercise embedding and the corresponding knowledge embedding, S q Representation Exercises The knowledge involved is embedded in the representation, P t represents the probability that the target student correctly answers the target exercise, represents the attention mechanism, and g(·) represents the inner product.

[0136] An embodiment of the present invention further provides a knowledge tracking device based on an online learning dynamic interaction graph, comprising:

[0137] The learning performance module is used to obtain students' historical learning parameters and quantify the effects to obtain the learning performance of the target students;

[0138] Interaction graph construction module, which constructs learner-exercise interaction graph and exercise-knowledge interaction graph based on learning performance;

[0139] The embedding feature learning module uses a graph convolutional network to embed the learner-exercise interaction graph and the exercise-knowledge interaction graph to obtain the exercise embedding feature representation;

[0140] The exercise answer prediction module is used to arrange the exercise embedding feature representations into an exercise embedding feature sequence according to the target student's answer order, input the exercise embedding feature sequence into the preset LSTM model, and obtain the probability that the target student correctly answers the target exercise.

[0141] like Figure 3 As shown, an embodiment of the present invention further provides an electronic device, which may include: a processor (processor) 301, a communication interface (Communications Interface) 302, a memory (memory) 303 and a communication bus 304, wherein the processor 301, the communication interface 302, and the memory 303 communicate with each other via the communication bus 304. The processor 301 may call the logic instructions in the memory 303 to execute the method provided in the above embodiment, for example, including:

[0142] Step A: Obtain the student's historical learning parameters and quantify the effects to obtain the target student's learning performance; Step B: Construct a learner-exercise interaction graph and an exercise-knowledge interaction graph based on the learning performance; Step C: Use a graph convolutional network to embed the learner-exercise interaction graph and the exercise-knowledge interaction graph to obtain the exercise embedding feature representation; Step D: Arrange the exercise embedding feature representation into an exercise embedding feature sequence according to the target student's answer order, input the exercise embedding feature sequence into a preset LSTM model, and obtain the probability of the target student correctly answering the target exercise.

[0143] In addition, when the logic instructions in the above-mentioned memory 303 are implemented in the form of a software functional unit and sold or used as an independent product, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0144] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the method provided in the above embodiment is implemented, for example, including:

[0145] Step A: Obtain the student's historical learning parameters and quantify the effects to obtain the target student's learning performance; Step B: Construct a learner-exercise interaction graph and an exercise-knowledge interaction graph based on the learning performance; Step C: Use a graph convolutional network to embed the learner-exercise interaction graph and the exercise-knowledge interaction graph to obtain the exercise embedding feature representation; Step D: Arrange the exercise embedding feature representation into an exercise embedding feature sequence according to the target student's answer order, input the exercise embedding feature sequence into a preset LSTM model, and obtain the probability of the target student correctly answering the target exercise.

[0146] To verify the effectiveness of the proposed method, the present invention conducted extensive performance analysis and effectiveness evaluation experiments on three public educational datasets: Assistments2009, Assistments2012, and Junyi. The results are shown in Table 1.

[0147] Table 1 Experimental results of different models and different data sets

[0148]

[0149]

[0150] As can be seen from Table 1, the knowledge tracking method based on the online learning dynamic interaction graph provided by the present invention significantly outperforms all baseline models in predicting learners' future performance in problem-solving. Its performance evaluation index ACC (Accuracy) exceeds 85%, and its prediction accuracy AUC (Area Under the Curve) exceeds 82%.

[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A knowledge tracking method based on online learning dynamic interaction graph, characterized in that: include: Step A: Obtain students’ historical learning parameters and quantify the effects to obtain the learning performance of the target students; Step B: constructing a learner-exercise interaction graph and an exercise-knowledge interaction graph based on the learning performance; Step C: Using a graph convolutional network to perform embedding learning on the learner-exercise interaction graph and the exercise-knowledge interaction graph to obtain an exercise embedding feature representation; Step D: Arrange the exercise embedding feature representations into an exercise embedding feature sequence according to the target student's answering order, input the exercise embedding feature sequence into a preset LSTM model, and obtain the probability that the target student correctly answers the target exercise.

2. The knowledge tracking method based on online learning dynamic interaction graph according to claim 1 is characterized in that: The learning parameters include answer accuracy, number of attempts and number of prompts.

3. The knowledge tracking method based on online learning dynamic interaction graph according to claim 2 is characterized in that: The step A specifically includes: Step A1: Quantify the accuracy of answering questions. Quantify the students' accuracy of answering questions to represent their learning ability, as shown below: Among them, a k represents the student's answer to the kth exercise, K represents the number of exercises, c i Indicates learning ability; Step A2: Quantify the number of attempts based on the student's performance on a specific exercise. j Calculate the attempt factor based on the number of attempts: in, represents the average number of attempts, p ij Represents a specific exercise q j Number of attempts, PC ij Represents a specific exercise q j The attempt factor; Calculate the question answering knowledge benefit factor based on the attempt factor Among them, α, β and η represent adjustable parameters; Step A3: Quantify the number of prompts based on the student's progress on a specific exercise. j Calculate the prompt factor based on the number of prompts: in, represents the average number of prompts, n ij Represents a specific exercise q j Number of prompts, NC ij Represents a specific exercise q j The prompt factor; Calculate the prompt knowledge benefit factor based on the prompt factor Among them, α, β and η represent adjustable parameters; Step A4: Comprehensively represent learning performance by linearly combining the learning ability, answering knowledge benefit factor, and prompting knowledge benefit factor to obtain the student's performance on a specific exercise q. j Learning performance on: Among them, w c , w p and w n represents a learnable parameter to adjust the relative importance of each knowledge acquisition vector; Step A5: Repeat steps A1 to A4 until each student's learning performance on all exercises is obtained.

4. The knowledge tracking method based on online learning dynamic interaction graph according to claim 1 is characterized in that: The student-exercise interaction diagram is specifically represented as follows: Where U represents the set of learners, Q represents the set of exercises, represents the relationship between the learner and the exercise, determined by the learning performance; The exercise-knowledge interaction graph is specifically expressed as G={Q, S, A}, where S represents the knowledge set and A represents the relationship between exercises and knowledge.

5. The knowledge tracking method based on online learning dynamic interaction graph according to claim 1 is characterized in that: Step C specifically includes: Step C1: Input the learner-exercise interaction graph into the graph convolutional network to obtain the first deep embedding representation of the exercise, which is expressed by the following formula: in, and represents the learnable weights and biases, l and l-1 represent the number of embedding layers of the graph convolutional network, Indicates that from the student node Pass to exercise q j learning performance information, Indicates the number of student nodes connected to the exercises, Represents the information conveyed by the exercise itself, and ReLU represents the nonlinear activation function; Step C2: Input the exercise-knowledge interaction graph into the graph convolutional network to obtain a second deep embedding representation of the exercise, which is expressed by the following formula: in, and denotes the learnable weights and biases, Indicates that from the knowledge node Pass to exercise q j Information, Indicates the number of knowledge nodes connected by the exercises, Representation Exercise q j The information it conveys; Step C3: Fuse the first deep embedding representation and the second deep embedding representation to obtain the third exercise embedding feature in, and represents the weight, represents the first deep embedding representation of all exercises, and q represents the second deep embedding representation of all exercises.

6. The knowledge tracking method based on online learning dynamic interaction graph according to claim 1 is characterized in that: The attention mechanism is introduced in step D.

7. The knowledge tracking method based on online learning dynamic interaction graph according to claim 6 is characterized in that: Step D specifically includes: Step D1: For each exercise, concatenate the exercise and the corresponding answer to form the final exercise representation: Among them, e t represents the tth final exercise, ReLU represents the nonlinear activation function, w t and b t denote weight and bias respectively, represents the embedding feature representation of the t-th exercise, a t Represents The corresponding answers indicate; Step D2: Arrange the final exercise representations into an exercise embedding feature sequence according to the target student's answering order, input the exercise embedding feature sequence into the preset LSTM model, use the answering order as the time step, and use the LSTM input gate, forget gate, output gate, and cell state update mechanism to learn the target student's knowledge state from the input representation. The formula is as follows: i t =σ(w ei e t +w hi h t-1 +b i ) g t =σ(w ef e t +w hf h t-1 +b f ) you are t σ(w eo e t +w ho h t-1 +b o ) c t =g t c t-1 +i t tanh(w ec e t +w hc h t-1 +b c ) h t =o t fishy t ) Among them, h t-1 represents the output of the previous time step, w * and b * represents the network parameters, i t 、g t and o t Represent the input gate, forget gate and output gate respectively, c t Represents the unit storage vector. σ(·) represents the Sigmoid function, h t Represents the current state of knowledge; Step D3: Use the attention mechanism to capture the relationship between the target exercise and the historical state, and select the top k states with the highest attention scores as the relevant historical knowledge state set I k : r t′ ={f(q t′ ,q t ),t′∈[t-l,…,t-1]} Where f(·) represents the attention ranking function, which is used to calculate the relevance between the top k history exercises and the target exercise, and the topk(·) function selects the knowledge state of the top k most relevant history exercises. k represents a set of k historical knowledge states; Step D4: Predicting the probability of correctly answering the target question based on the relevant historical knowledge state set, which is expressed by the following formula: Among them, fi represents the interaction term between the current knowledge state and the historical knowledge state, f j represents the interaction term between the target exercise embedding and the corresponding knowledge embedding, S q Representation Exercises The knowledge involved is embedded in the representation, P t represents the probability that the target student correctly answers the target exercise, represents the attention mechanism, and g(·) represents the inner product.

8. A knowledge tracking device based on online learning dynamic interaction graph, characterized in that: include: The learning performance module is used to obtain students' historical learning parameters and quantify the effects to obtain the learning performance of the target students; An interaction graph construction module constructs a learner-exercise interaction graph and an exercise-knowledge interaction graph based on the learning performance; An embedding feature learning module uses a graph convolutional network to perform embedding learning on the learner-exercise interaction graph and the exercise-knowledge interaction graph to obtain an exercise embedding feature representation; The exercise answer prediction module is used to arrange the exercise embedding feature representations into an exercise embedding feature sequence according to the target student's answer order, input the exercise embedding feature sequence into a preset LSTM model, and obtain the probability that the target student correctly answers the target exercise.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 7 is executed.

Images