A knowledge tracking method based on diverse learning behaviors

Abstract

The present invention belongs to the field of knowledge tracking technology, and specifically discloses a knowledge tracking method based on diverse learning behaviors. The knowledge tracking method based on diverse learning behaviors in the present invention includes three parts: data preprocessing and vector representation encoding, a multi-head self-attention mechanism that integrates forgetting behaviors and transformation relationships between behaviors, and model training and prediction. Based on a variety of learning behavior data, the present invention models knowledge tracking problems from the perspective of the learner's full-process learning scenario, so that the constructed diverse learning behavior knowledge tracking model can model the transformation relationship of different types of learning behaviors and the learner's personalized learning mode. At the same time, it introduces learner forgetting behaviors based on learning time and number of repetitions to better model the changes in the learner's knowledge state during the entire learning process, and effectively obtain the learner's knowledge state information from diverse learning behaviors, thereby completing the knowledge tracking task.

Description

Technical Field

[0001] The present invention belongs to the technical field of knowledge tracing, and relates to a knowledge tracing method based on diverse learning behaviors. Background Art

[0002] Most of the existing knowledge tracking methods focus on students' interactive behaviors in exercises during course learning.

[0003] However, in real learning scenarios, the interaction between students and knowledge content is not limited to course exercises, but also includes a variety of learning behaviors such as learning course video materials, learning text materials, and course discussions.

[0004] Introducing diverse learning behavior data into the knowledge tracking model is conducive to the model's evaluation of students' knowledge status during the course learning process, and it also helps the model to mine students' learning behavior patterns from the data.

[0005] There are two challenges to be addressed when mining learning behavior interaction data that contains multiple types of behavior patterns:

[0006] 1) Different types of interactive behaviors between learners and learning content have complex dependencies; the dependencies between learning behaviors mean that for a piece of knowledge content, different learning behaviors are not independent, but are connected. For example, some students will participate in classroom discussions, answer exercises, and other activities after completing the course video learning. At the same time, students' participation in the previous links will also affect their performance in classroom discussions, teacher Q&A, and answering exercises.

[0007] 2) Due to the individual characteristics of learners, learners' learning behavior patterns are diverse. The diversity of learning patterns means that each student's learning habits are personalized. For example, some students prefer to learn new concepts through rigorous text materials, while others prefer to understand new concepts through vivid video materials.

[0008] Therefore, how to model the dependency between learners and learning behaviors, and how to mine learners' personalized learning patterns from learning behavior interaction data are urgent issues to be solved when introducing diverse learning behavior data into knowledge tracking methods. Summary of the invention

[0009] The purpose of the present invention is to propose a knowledge tracking method based on diverse learning behaviors. The method is based on a variety of learning behavior data and models the knowledge tracking problem from the perspective of the learner's full learning scenario, so that the constructed diverse learning behavior knowledge tracking model can model the conversion relationship between different types of learning behaviors and the learner's personalized learning mode. At the same time, the learner's forgetting behavior based on learning time and number of repetitions is introduced to better model the changes in the learner's knowledge state during the entire learning process, and effectively obtain the learner's knowledge state information from diverse learning behaviors, thereby completing the knowledge tracking task.

[0010] In order to achieve the above object, the present invention adopts the following technical scheme:

[0011] A knowledge tracking method based on diverse learning behaviors includes the following steps:

[0012] Step 1. Extract the interaction sequence data of the learning process from the learners’ logs on the online learning platform;

[0013] Interaction sequence data includes four types of feature information, namely, the knowledge content X learned by the learner, the learning behavior taken by the learner, , the exercise Q of the learner's exercise answering behavior interaction, and the time when the learner's learning behavior occurs ;

[0014] When designing different types of learning behaviors, a diverse set of learning behaviors was constructed. ; Among them, the learning behavior set It includes four types of learning behaviors: video learning, text material reading, correct answers to exercises, and incorrect answers to exercises;

[0015] Use label encoding method to obtain the knowledge content X in the interactive sequence data and the learning behavior taken by the learner , and the exercise Q of the learner's exercise answering behavior interaction is pre-coded;

[0016] Step 2. Build a diverse learning behavior knowledge tracking model that integrates the transformation relationship between behaviors and forgetting features, which includes a representation vector encoding module, a multi-head self-attention module that integrates the transformation relationship between behaviors and forgetting features, and a feature fusion and prediction module;

[0017] The representation vector encoding module is a set of knowledge content X and interactive learning behaviors. , the three MAP matrices corresponding to the interactive exercise Q are randomly initialized, and the scalar code in the pre-coding stage is converted into a high-dimensional vector code;

[0018] In order to enable the model to capture the diverse feature information of each interaction, the knowledge content representation vector and the learning behavior representation vector of each interaction are concatenated, and the concatenated representation vector is mapped with the help of a layer of fully connected neural network;

[0019] For the exercise answering behavior, the exercise information of learners' interaction is further introduced for encoding;

[0020] The multi-head self-attention module integrates the transformation relationship between behaviors and the forgetting feature, and represents the interactive item vector with behavioral features, knowledge content features, and exercise features. , input multi-head self-attention mechanism;

[0021] When calculating the attention score, the cosine similarity method is used to model the transformation relationship between different learning behaviors;

[0022] By calculating the time difference between the current interaction behavior and other interactions in the sequence and the frequency of repeated knowledge content, the knowledge forgetting process is modeled, so that the model pays more attention to the knowledge content that the learner has learned recently and has repeatedly learned;

[0023] The attention score is calculated based on the conversion relationship between different learning behaviors and the forgetting characteristics of the learner's knowledge state, and the learner's individual learning pattern characteristics are mined from each learner's own interaction sequence data;

[0024] The feature fusion and prediction module first calculates the weighted value matrix of the attention score for the attention head; then it uses the splicing method to aggregate the information of multiple attention heads to obtain the knowledge state representation vector of the student at the current moment;

[0025] On this basis, the model input representation vector of the next moment and the student's knowledge state representation vector of the current moment are concatenated, and feature extraction is performed through the fully connected layer, and then the sigmoid function is used to map it to the probability value of the learner's correct answer;

[0026] Step 3. Use the cloze task to train the constructed diverse learning behavior knowledge tracking model. After the model is trained, use the trained knowledge tracking model to predict the learners' exercise answering performance results.

[0027] The present invention has the following advantages:

[0028] As described above, the present invention designs a knowledge tracking method based on diverse learning behaviors for knowledge tracking tasks. The method of the present invention enables the model to have the ability to process diverse learning behavior feature information. In terms of behavioral feature processing, the present invention has the ability to model the heterogeneity between diverse learning behaviors and can also model the transformation relationship between learning behaviors. In terms of knowledge state modeling, the present invention further models the forgetting behavior of learners in the learning process and the transformation relationship between different learning behaviors on the basis of modeling the dependency relationship between sequence data using the self-attention mechanism. The above characteristics enable the model of the present invention to effectively obtain the learner's knowledge state information from diverse learning behaviors, thereby effectively completing the knowledge tracking task. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Schematic diagram of the flow of a knowledge tracking method based on diverse learning behaviors in an embodiment of the present invention.

[0030] Figure 2 This is a processing flow chart of the diverse learning behavior knowledge tracking model constructed in an embodiment of the present invention.

[0031] Figure 3 Schematic diagram of the multi-head self-attention mechanism that integrates the association information of forgetting behavior and diverse learning behaviors. DETAILED DESCRIPTION

[0032] The present invention is further described in detail below with reference to the accompanying drawings and specific embodiments:

[0033] Example 1

[0034] The present invention discloses a knowledge tracking method based on diverse learning behaviors, which models the transformation relationship between different learning behaviors through cosine similarity. In terms of personalized learning model modeling, a multi-head self-attention mechanism that integrates the transformation relationship of different learning behaviors is used to mine personalized learning models from the interactive sequence data of each learner.

[0035] like Figure 1 As shown, the knowledge tracking method based on diverse learning behaviors in this embodiment includes the following steps:

[0036] Step 1. Extract interaction sequence data during the learning process from the logs of the online learning platform according to the learners.

[0037] Interaction sequence data includes four types of feature information, namely, the knowledge content X learned by the learner, the learning behavior taken by the learner, , the exercise Q of the learner's exercise answering behavior interaction, and the time when the learner's learning behavior occurs .

[0038] When designing different types of learning behaviors, a diverse set of learning behaviors was constructed. ; Learning behavior set It includes four learning behaviors: video learning, text material reading, answering exercises correctly, and answering exercises incorrectly.

[0039] Answering exercises correctly and answering exercises incorrectly are considered two different learning behaviors to facilitate subsequent model training.

[0040] Specifically define the knowledge tracking tasks in various learning behavior scenarios:

[0041] set up , , and They represent the knowledge content of learners' interaction, the exercises they answer, the specific learning behaviors they take, and the specific practices of the interaction methods.

[0042] in , , and They represent the amount of knowledge content, the number of exercises, the number of types of learning behaviors that can be taken, and the total amount of time recorded in the entire course. Represents knowledge content, Indicates exercises, Indicates interactive behavior. Specific time.

[0043] When constructing interactive behavior data, the learner's correct answer and answer errors Divided into two different interactive behaviors.

[0044] For a learner, a set of interaction sequences will be generated during the online learning process. ,Each interaction item in the interaction sequence is composed of multiple ,interaction features (knowledge content, interaction behavior).

[0045] In particular, for exercise interaction, the interaction item is composed of three features: (knowledge content, exercise, interaction behavior). .

[0046] in, Indicates the maximum length of the interaction sequence, Indicates the first time, The value range is 1- .

[0047] The knowledge tracking task in the scenario of diverse learning behaviors aims to predict the first The learner's performance on the knowledge content related exercises at the moment. Based on the above definition, the problem is formalized as:

[0048] ;

[0049] in Indicates that the interaction sequence data contains Data for all times other than the time.

[0050] The above problem is formally described as follows: by using the contextual data in the interaction sequence, including the knowledge content learned by the learner at other times, the learning method adopted, and the exercises of other answering behaviors, the model output is maximized to determine the learner’s The probability of learners’ actual performance on content-related exercises at that moment.

[0051] Different from the traditional knowledge tracking model which mainly focuses on students' answering behavior of exercises, the present invention focuses on the learning behavior of learners in the whole process of online course learning, including answering behavior of exercises, video learning behavior and text material learning behavior.

[0052] In addition, the present invention is inspired by the traditional knowledge tracking model and introduces the learning forgetting law in the process of modeling the learner's knowledge state change law, so that the model can better fit the learner's knowledge state change during the learning process.

[0053] The model cannot learn directly based on text content. The present invention uses a label encoding method to encode the knowledge content X in the interactive sequence data and the learning behavior taken by the learner. , and pre-encode the exercises Q that interact with learners’ exercise answering behaviors.

[0054] In order to facilitate the model to process sequence data, the present invention selects a fixed length of 200 to segment overlong interaction sequence data; for interaction sequence data with a length of less than 200, the empty positions are filled with 0 values. Here, the fixed length of 200 refers to 200 interactions. There are differences in learning progress between different learners, and the lengths of interaction sequences recorded in the logs are different. By truncating overlong sequences and padding overshort sequences, the sequence data input to the model is kept at a fixed length, which is convenient for model processing.

[0055] Step 2. Build a diverse learning behavior knowledge tracking model, which includes a representation vector encoding module, a multi-head self-attention module, and a feature fusion and prediction module, such as Figure 2 shown.

[0056] The representation vector encoding module represents the knowledge content X and the interactive learning behavior set , the three MAP matrices corresponding to the interactive exercise Q are randomly initialized, and the scalar code in the precoding stage is converted into a high-dimensional vector code.

[0057] In order to enable the model to capture the diverse feature information of each interaction, the knowledge content representation vector and the learning behavior representation vector of each interaction are spliced, and the spliced ​​representation vector is mapped with the help of a layer of fully connected neural network.

[0058] For exercise answering behavior, the exercise information of learners' interaction is further introduced for encoding.

[0059] The processing of the representation vector encoding module is as follows:

[0060] Through the lecun_normal method, the learning behavior set of knowledge content X and interaction , the three map matrices corresponding to the interactive exercise Q are randomly initialized, and the label encoding in the pre-coding stage is converted into a high-dimensional vector encoding. The formula is expressed as follows:

[0061] ;

[0062] in, Indicates The knowledge content that learners learn at all times, Indicates Exercises that students answer at all times, No. The learning behaviors taken by learners at all times, , , They respectively represent the corresponding representation vectors, i.e., high-dimensional vector encodings.

[0063] In order to enable the model to better capture the diverse feature information of each interaction, the knowledge content representation vector and the learning behavior representation vector of each interaction are spliced, and the exercise information of the learner interaction is further introduced to encode the exercise answering behavior.

[0064] In order to maintain the uniformity of dimension, the concatenated representation vector is mapped with the help of a layer of fully connected neural network.

[0065] The above process is expressed in the formula:

[0066] ;

[0067] in, Indicates The learning behaviors taken by learners at all times, represents the weight coefficient, represents a fully connected neural network layer, Indicates The representation vector of the model input at this moment, Represents a subset of learning behaviors, including both correct and incorrect answers to exercises.

[0068] Different from the traditional knowledge tracking method, this paper proposes a unique construction method for constructing model input representation vectors in various learning behavior scenarios, that is, by splicing the interactive knowledge content representation vector and the learning behavior representation vector, the joint features of the interactive behavior and the knowledge content at the same time are obtained. In addition, for the exercise answering behavior, exercise information is further introduced to improve the accuracy of the model in predicting the learner's exercise answering performance.

[0069] like Figure 3 As shown in the figure, the multi-head self-attention module integrates various learning behaviors and forgetting features, that is, the interactive project representation vector with behavioral features, knowledge content features and exercise features , input the multi-head self-attention mechanism, the formula is expressed as follows:

[0070] ;

[0071] in, Depend on arrive It is a tensor obtained by concatenating the entire interaction sequence data through the representation vector encoding module; Indicates The query tensor corresponding to the attention head, Indicates The key tensor corresponding to the attention head, Indicates The value tensor corresponding to each attention head; , , Represent the corresponding projection matrices respectively.

[0072] Since different learning behaviors have different effects on the results of learners' exercise answers, and there are also effects between non-exercise answering types of learning behaviors, it is necessary to consider the conversion relationship between different learning behaviors when calculating the attention score. The present invention uses the cosine similarity method to model the conversion relationship between different learning behaviors, and the formula is expressed as follows:

[0073] ;

[0074] in, and Respectively represent Moment and The representation vector corresponding to the learning behavior taken by the learner at each moment, Indicates the calculation of the inner product value of two representation vectors. Represents the transformation relationship between different learning behaviors.

[0075] The present invention starts from the conversion relationship between specific learning behaviors, based on the representation vector of the learning behavior, and uses the cosine similarity method to directly model the conversion relationship between different learning behaviors.

[0076] By calculating the time difference between the current interaction behavior and other interaction behaviors in the sequence and the frequency of repeated occurrence of knowledge content, the knowledge forgetting process is modeled, so that the model pays more attention to the knowledge content that learners have learned recently and repeatedly learned.

[0077] In addition to the conversion relationship between different learning behaviors, the modeling of the changes in the learner's knowledge status during the entire learning process also needs to consider the forgetting rules of the knowledge content that the learner has learned.

[0078] The present invention models the knowledge forgetting process by calculating the time difference between the current interactive behavior and other interactive behaviors in the same sequence and the frequency of repeated occurrence of knowledge content. Specifically, as the time interval increases, the knowledge content that the learner has learned in the past will gradually be forgotten. At the same time, for knowledge content that has been repeatedly learned many times, the learner's forgetting speed for the corresponding knowledge content will decrease. The present invention calculates the time difference between two learning behaviors to represent the time interval after the learning behavior occurs. At the same time, considering the number of times the corresponding knowledge content is repeatedly learned at the historical moment, the degree of reduction in the forgetting speed under repeated learning is expressed in inverse form. The specific formula is as follows:

[0079] ;

[0080] in, Indicated in Knowledge content that appears all the time The frequency of occurrence, and Respectively represent Moment and The timestamp value of the moment; Represents the process of knowledge forgetting.

[0081] The attention score is calculated based on the conversion relationship between different learning behaviors and the forgetting characteristics of the learner's knowledge state, and the learner's individual learning pattern characteristics are mined from each learner's own interaction sequence data. The formula is as follows:

[0082] ;

[0083] in represent The attention heads correspond to the attention score matrix Position attention score; represents the natural number base, represents the vector dimension, Representative The query tensor corresponding to the attention head corresponds to the The value of the moment, Representative The key tensor corresponding to the attention head corresponds to the The value of the moment.

[0084] The attention mechanism proposed in the present invention integrates the transformation relationship between learning behaviors, the forgetting characteristics of knowledge states, and the correlation relationship between knowledge contents. This feature enables the model to consider the correlation relationship between knowledge contents and the transformation relationship between learning behaviors when paying attention to forgetting characteristics. When predicting the knowledge state of learners, the model can pay more attention to the relevant knowledge content learned recently or multiple times, and at the same time pay attention to the learning behaviors that have a higher impact on the performance results of answering exercises.

[0085] The feature fusion and prediction module first calculates the weighted value matrix of the attention score for the attention head; then it uses the splicing method to aggregate the information of multiple attention heads to obtain the knowledge status representation vector of the student at the current moment.

[0086] On this basis, the model input representation vector of the next moment and the student’s current knowledge state representation vector are concatenated, and feature extraction is performed through the fully connected layer, and then the sigmoid function is used to map it to the probability value of the learner’s correct answer.

[0087] Specifically, the processing flow of the feature fusion and prediction module is as follows:

[0088] First, for the attention head Calculate the weighted value matrix of attention scores ; Then, the splicing method is used to aggregate the information of multiple attention heads to predict the student’s knowledge state representation vector at the next moment ;

[0089] On this basis, the model input representation vector of the next moment is concatenated with the student's knowledge state representation vector at the current moment, and features are extracted through the fully connected layer, and then the sigmoid function is used to map the probability value of the learner's correct answer. ;

[0090] ;

[0091] ;

[0092] ;

[0093] in, Indicates The value tensor corresponding to the attention heads; Concat() means concatenation, represents a weight matrix used to aggregate subspace information from different attention heads. Indicates The learner’s knowledge status at all times; Represents the model prediction The probability that a learner answers an exercise correctly at any given moment.

[0094] Step 3. Use the cloze task to train the constructed diverse learning behavior knowledge tracking model. After the model is trained, use the trained knowledge tracking model to predict the learner's exercise answering performance, that is, the probability value of the correct answer.

[0095] First, based on the fixed-length sequence data obtained in step 1, 15% of the knowledge content corresponding to the exercise answering behavior is randomly selected from the interactive sequence data according to the learning behavior type for masking. Specifically, a special knowledge content [mask] is introduced to represent the masked items, and then the masked sequence data is label encoded. The encoded interactive sequence data is passed through the model's representation vector encoding module to fuse the three different sets of feature information that constitute the interactive items and convert them into the model's input representation vector.

[0096] Then, based on the multi-head self-attention module constructed in step 2 that integrates the transformation relationship between behaviors and the forgetting features, the attention score of the contextual interaction data of the masked position is calculated. Through the feature fusion and prediction module in step 2, features are extracted based on the attention score to obtain the learner's knowledge state representation vector predicted by the model at the masked moment. Based on the knowledge representation vector, the probability value of the learner answering the corresponding exercise correctly at the masked moment is estimated.

[0097] Finally, the binary cross entropy loss function is used to calculate the loss between the estimated value and the actual answer result, and the back propagation method is used to optimize the model based on the loss value. The formula is as follows:

[0098] ;

[0099] in, Representative Time to answer the exercises, Representative The answer result predicted by the model at that moment, The subset of exercise answering behaviors representing the learning behavior set includes both correct exercise answering and incorrect exercise answering.

[0100] The present invention combines the characteristics of the self-attention mechanism and proposes a cloze task for the field of knowledge tracking, so that the model no longer focuses only on historical interaction data, but predicts the learner's knowledge status based on contextual interaction information. This feature helps the model to explore the changing rules of the learner's knowledge status from the interaction sequence.

[0101] After the model training is completed, the trained knowledge tracking model is used to predict the learner's exercise answering performance results. That is, for a given learner's historical interaction sequence log data, it is preprocessed according to step 1 to obtain input data that can be processed by the model, and then input into the trained knowledge tracking model to predict the learner's exercise answering performance results.

[0102] The knowledge tracking method based on diverse learning behaviors proposed in the present invention is based on a multi-head self-attention mechanism that integrates the conversion relationship between forgetting behavior and learning behavior, mines the changing rules of the learner's knowledge state in the learning process from the interactive sequence data containing the learner's diverse learning behaviors, models the learner's forgetting behavior in the learning process through the time interval of knowledge content learning and the number of repetitions of related learning behaviors, enhances the influence of recently learned knowledge content and repeatedly learned knowledge content on the learner's current knowledge state, models the conversion relationship between different learning behavior representations through cosine similarity, and adjusts the calculation result of the attention score, so that the model focuses on the related learning behavior interactions that have an obvious correlation with the exercise answering behavior, thereby being able to effectively obtain the learner's knowledge state information from diverse learning behaviors, thereby effectively completing the knowledge tracking task.

[0103] Of course, the above description is only a preferred embodiment of the present invention, and the present invention is not limited to the above embodiments. It should be noted that all equivalent substitutions and obvious deformation forms made by any technician familiar with the field under the guidance of this specification fall within the essential scope of this specification and should be protected by the present invention.

Claims

1. A knowledge tracking method based on diverse learning behaviors, characterized in that: The steps include: Step 1. Extract the interaction sequence data of the learning process from the learners’ logs on the online learning platform; Interaction sequence data includes four types of feature information, namely, the knowledge content X learned by the learner, the learning behavior taken by the learner, , the exercise Q of the learner's exercise answering behavior interaction, and the time when the learner's learning behavior occurs ; Step 2. Build a diverse learning behavior knowledge tracking model that integrates the transformation relationship between behaviors and forgetting features, which includes a representation vector encoding module, a multi-head self-attention module that integrates the transformation relationship between behaviors and forgetting features, and a feature fusion and prediction module; The representation vector encoding module concatenates the knowledge content representation vector and the learning behavior representation vector of each interaction, and maps the concatenated representation vector with the help of a fully connected neural network. For the exercise answering behavior, the exercise information of learners' interaction is further introduced for encoding; The multi-head self-attention module integrates the transformation relationship between behaviors and the forgetting characteristics, and inputs the interactive item representation vector with behavior characteristics, knowledge content characteristics and exercise characteristics into the multi-head self-attention mechanism; When calculating the attention score, the cosine similarity method is used to model the transformation relationship between different learning behaviors; The attention score is calculated based on the conversion relationship between different learning behaviors and the forgetting characteristics of the learner's knowledge state, and the learner's individual learning pattern characteristics are mined from each learner's own interaction sequence data; Step 3. Use the cloze task to train the constructed diverse learning behavior knowledge tracking model. After the model is trained, use the trained knowledge tracking model to predict the learners' exercise answering performance results.

2. The knowledge tracking method based on diverse learning behaviors according to claim 1 is characterized in that: In step 2, the processing process of the characterization vector encoding module is as follows: Through the lecun_normal method, the learning behavior set of knowledge content X and interaction , the three map matrices corresponding to the interactive exercise Q are randomly initialized, and the label encoding in the pre-coding stage is converted into a high-dimensional vector encoding. The formula is expressed as follows: ； in, Indicates The knowledge content that learners learn at all times, Indicates Exercises that students answer at all times, No. The learning behaviors taken by learners at all times, , , They respectively represent the corresponding representation vectors, i.e., high-dimensional vector encodings; The knowledge content representation vector and learning behavior representation vector of each interaction are concatenated; the concatenated representation vector is mapped with the help of a layer of fully connected neural network. For the exercise answering behavior, the exercise information of the learner interaction is further introduced for encoding. The specific formula is expressed as follows: ； in, Indicates The learning behaviors taken by learners at all times, represents the weight coefficient, represents a fully connected neural network layer, Indicates The representation vector of the model input at that moment; Represents a subset of learning behaviors, including both correct and incorrect answers to exercises.

3. The knowledge tracking method based on diverse learning behaviors according to claim 2 is characterized in that: In step 2, the interactive item representation vector with behavior features, knowledge content features and exercise features is Enter the multi-head self-attention mechanism, the formula is expressed as follows: ； in, Represents the tensor obtained by the representation vector encoding module for the entire interaction sequence data; Representation vector by interaction item arrive Spliced ​​together, Indicates the maximum length of the interaction sequence; Indicates The query tensor corresponding to the attention head, Indicates The key tensor corresponding to the attention head, Indicates The value tensor corresponding to each attention head; , , Represent the corresponding projection matrices respectively.

4. The knowledge tracking method based on diverse learning behaviors according to claim 3 is characterized in that: In step 2, when calculating the attention score, the conversion relationship between different learning behaviors is considered, and the conversion relationship between different learning behaviors is modeled by the cosine similarity method. The formula is expressed as follows: ； in, and Respectively represent Moment and The representation vector corresponding to the learning behavior taken by the learner at each moment, Indicates the calculation of the inner product value of two representation vectors. Represents the transformation relationship between different learning behaviors.

5. The knowledge tracking method based on diverse learning behaviors according to claim 4 is characterized in that: In step 2, by calculating the time difference between the current interaction behavior and other interaction behaviors in the sequence and the frequency of repeated occurrence of knowledge content, the knowledge forgetting process is modeled so that the model pays more attention to the knowledge content that the learner has learned recently and has repeatedly learned. The specific formula is as follows: ； in, Indicated in Knowledge content that appears all the time The frequency of occurrence, and Respectively represent Moment and The timestamp value of the moment; Represents the process of knowledge forgetting.

6. The knowledge tracking method based on diverse learning behaviors according to claim 5 is characterized in that: In step 2, the attention score is calculated based on the conversion relationship between different learning behaviors and the forgetting characteristics of the learner's knowledge state, and the learner's individual learning pattern characteristics are mined from each learner's own interaction sequence data. The formula is as follows: ； in represent The attention heads correspond to the attention score matrix Position attention score; represents the natural number base, represents the vector dimension, Representative The query tensor corresponding to the attention head corresponds to the The value of the moment, Representative The key tensor corresponding to the attention head corresponds to the The value of the moment.

7. The knowledge tracking method based on diverse learning behaviors according to claim 6 is characterized in that: In step 2, the feature fusion and prediction module first calculates the value matrix of the weighted attention score for the attention head; then uses the splicing method to aggregate the information of multiple attention heads to obtain the knowledge state representation vector of the student at the current moment; On this basis, the model input representation vector of the next moment and the student's knowledge state representation vector of the current moment are concatenated, and feature extraction is performed through the fully connected layer, and then the sigmoid function is used to map it to the probability value of the learner's correct answer; The processing flow of the feature fusion and prediction module is as follows: First, for the attention head Calculate the weighted value matrix of attention scores ; Then, the splicing method is used to aggregate the information of multiple attention heads to predict the student’s knowledge state representation vector at the next moment ; On this basis, the model input representation vector of the next moment is concatenated with the student's knowledge state representation vector at the current moment, and features are extracted through the fully connected layer, and then the sigmoid function is used to map the probability value of the learner's correct answer. ; ； ； ； in, Indicates The value tensor corresponding to the attention heads; Concat() means concatenation, represents a weight matrix used to aggregate subspace information from different attention heads. Indicates The learner’s knowledge status at all times; Represents the model prediction The probability that a learner answers an exercise correctly at any given moment.

8. The knowledge tracking method based on diverse learning behaviors according to claim 1, characterized in that: In step 1, when designing different learning behavior types, a diverse learning behavior set is constructed. ; Among them, the learning behavior set It includes four learning behaviors: video learning, text material reading, correct answers to exercises, and incorrect answers to exercises; Use label encoding method to obtain the knowledge content X in the interactive sequence data and the learning behavior taken by the learner , and the exercise Q of the learner's exercise answering behavior interaction is pre-coded; In order to facilitate the model to process sequence data, a fixed length of 200 is selected to split the interaction sequence data that is too long; for the interaction sequence data with a length of less than 200, vacant positions are filled with 0 values.

9. The knowledge tracking method based on diverse learning behaviors according to claim 1, characterized in that: In step 3, the cloze task is used as the model training target, and the model training process is as follows: First, based on the fixed-length sequence data obtained in step 1, 15% of the knowledge content corresponding to the exercise answering behavior is randomly selected from the interactive sequence data according to the learning behavior type for masking; specifically, a special knowledge content [mask] is introduced to represent the masked items, and then the masked interactive sequence data is label encoded; the encoded sequence data is passed through the model's representation vector encoding module to fuse the three different sets of feature information constituting the interactive items and transform them into the model's input representation vector; Then, based on the multi-head self-attention module constructed in step 2 that integrates the transformation relationship between behaviors and the forgetting features, the attention score of the contextual interaction data of the masked position is calculated; through the feature fusion and prediction module in step 2, features are extracted based on the attention score to obtain the learner's knowledge state representation vector predicted by the model at the masked moment; Based on the knowledge state representation vector, estimate the probability value of the learner answering the corresponding exercise correctly at the masked moment; Finally, the binary cross entropy loss function is used to calculate the loss between the estimated value and the actual answer result, and the back propagation method is used to optimize the model based on the loss value. The formula is as follows: ； in, Representative Time to answer the exercises, Representative The answer result predicted by the model at that moment, The subset of exercise answering behaviors representing the learning behavior set includes both correct exercise answering and incorrect exercise answering.

10. The knowledge tracking method based on diverse learning behaviors according to claim 1, characterized in that: In step 3, after the model training is completed, the given learner historical interaction sequence log data is preprocessed in the manner of step 1 to obtain input data that can be processed by the model, and input into the trained knowledge tracking model to predict the learner's exercise answering performance results.

Images