Multi-target exercise recommendation method based on learner knowledge state prediction

Abstract

The invention relates to the technical field of intelligent teaching, and particularly provides a multi-target exercise recommendation method based on learner knowledge state prediction, which specifically comprises the following steps: 1) predicting the knowledge concept mastering degree of a learner by using a knowledge tracking model; 2) predicting the learning progress of the learner by using the time sequence prediction model; 3) screening out a candidate exercise set from the exercise library through the predicted values of the first two models; 4) further optimizing and screening the candidate exercise set through an optimization algorithm to generate an exercise recommendation list; and 5) arranging the exercises in the exercise recommendation list according to the difficulty from small to large through an exercise difficulty function so as to meet the requirement of the learner for learning step by step. According to the invention, accurate prediction of the learning state of the learner is realized, the accuracy, diversity and novelty of the exercise recommendation system are improved, and personalized exercises are recommended to students from the multi-objective perspective.

Description

Technical Field

[0001] The present invention relates to the field of smart teaching technology, and more specifically, to a multi-objective exercise recommendation method based on learner knowledge state prediction. Background Art

[0002] With the rapid development of the "Internet Plus" and the information-based education era, learning information resources are growing exponentially. Learners need to quickly and efficiently find appropriate learning resources from this vast and complex information pool. This has led to the emergence of personalized learning operations. Personalized exercise recommendation plays a crucial role in personalized learning. It aims to automatically generate personalized exercises based on learners' learning profiles, knowledge levels, and learning needs, helping them effectively consolidate their knowledge and improve learning outcomes. In recent years, a variety of recommendation algorithms have been applied to exercise recommendation, addressing a wide range of challenges. Much of this research has focused on educational psychology and data mining, designing diverse model structures driven by extensive educational data to recommend appropriate exercises for learners. Examples include collaborative filtering and cognitive diagnostics. Among these, collaborative filtering-based exercise recommendation methods analyze learners' feedback and preferences based on behavioral data across learners to recommend exercises that meet their learning needs. This approach assumes that learners with similar learning behaviors also share similar exercise preferences. By constructing a similarity matrix, it predicts learners' interest in unlearned exercises, thereby providing personalized exercise recommendations. For example, Khairil et al. proposed an exercise recommendation method based on the similarity of attributes between exercises and learning objectives. Esteban et al. focused on student relationships and learning resources and proposed a hybrid filtering recommendation method that uses multiple criteria related to student and course information to recommend the most appropriate courses. However, these methods ignore the learner's mastery of various knowledge concepts and have certain limitations in recommendation accuracy. Exercise recommendation methods based on cognitive diagnosis use the subject knowledge structure and learner learning history to create a knowledge model to recommend the most appropriate exercises for learners. Early methods used knowledge graphs to display the subject knowledge structure and form learning paths over the learning process. Current mainstream methods mainly combine knowledge tracking and personalized recommendation technologies to provide more intelligent personalized learning support. However, these methods only consider the level of knowledge mastery and ignore the learner's learning progress, making it difficult to achieve truly personalized development for learners.

[0003] Therefore, to recommend more personalized exercises to learners, it is necessary to consider both the learner's learning progress and their mastery of various concepts. Learning progress, as a temporal representation of knowledge construction validity, is essentially the rate of accumulation of effective knowledge units—that is, the number of knowledge concepts mastered by a learner over a period of time. This rate is generally positively correlated with learning time. Traditional research, based on the assumption of time-knowledge linearity, uses temporal models such as LSTM to infer progress from practice frequency. However, mastery learning theory suggests that learning progress is essentially driven by the rate of correct knowledge mastery. Its core mechanism requires crossing a threshold of accuracy to trigger effective progress accumulation, fundamentally negating the assumption of linear progress that relies solely on practice duration. A limitation of existing methods is that they directly equate practice records with learning progress, ignoring the moderating role of answer quality. This can lead to misjudgments of learning progress. For example, if learner A repeatedly makes mistakes in a trigonometric function conversion exercise, their progress value will be overestimated; whereas learner B achieves true knowledge construction through correct deductions. Furthermore, when predicting a learner's mastery of knowledge concepts, most methods employ knowledge tracking models. However, current knowledge tracking models are generally limited by the Markov assumption of temporal modeling. For example, the DKT and AKT both assume that a learner's knowledge state is only related to the most recent k-step interactions. This strong assumption can lead to flaws when faced with knowledge concept switching: According to cognitive load theory, during learners' practice, when non-continuous knowledge concept switching occurs, the parallel activation of new and old concepts can cause the external cognitive load to exceed the threshold, resulting in working memory overload, which can increase the misjudgment rate of the learner's knowledge state. For example, when the interval between concept switching steps is too long, the limited window modeling mechanism of traditional models, based on the Markov assumption of recent interactions, can induce systematic biases, ignoring the state before the knowledge concept switching, forcibly truncating the early, unmastered knowledge state, and causing the mastery prediction of the switched knowledge concept to deviate from the true cognitive trajectory. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies and provide a multi-objective exercise recommendation method based on the prediction of learner knowledge status. This method can accurately predict the learner's learning status and effectively improve the accuracy, diversity and novelty of exercise recommendations.

[0005] The technical solution for achieving the purpose of the present invention is:

[0006] A multi-objective exercise recommendation method based on the prediction of learners' knowledge status is mainly divided into three modules, namely the learning progress prediction module, the knowledge concept mastery degree prediction module and the exercise filtering module. The model takes the learner's historical answer record as input. First, in the learning progress prediction module, we use the advanced time series prediction model to predict the student's learning progress. Secondly, in the knowledge concept mastery degree prediction module, we use the advanced knowledge tracking model to predict the student's knowledge concept mastery degree. Finally, the exercise filtering module first uses the learning progress and knowledge concept mastery degree obtained from the first two modules to filter the candidate exercise sets from the exercise library, and then uses the Sine Hippo optimization algorithm in this module to further optimize and screen the candidate sets to maximize the diversity of the recommendation list and generate the optimal exercise recommendation list. Finally, through the exercise difficulty function, the exercises in the recommendation list are arranged in order of difficulty from small to large. The method includes the following steps:

[0007] 1) Learning Progress Prediction Module: This module aims to predict students' learning progress, specifically, the coverage of knowledge concepts. We use the knowledge concept sequence from time 0 to time t to predict the probability of each knowledge concept occurring at time t+1. To this end, we constructed a time series prediction model. First, we embed the student's exercises, the knowledge concepts contained in the exercises, and the answers as input into the model. Because normalization on the sequence and feature mixer can slow training and cause information interference, we perform normalization before information mixing:

[0008] x n =Layernorm(x)

[0009] Then we extract information from three branches: sequence, feature, and channel. The specific calculation formula is as follows:

[0010] y s =W2g l (W1x s )

[0011] y f =W4g l (W3x f )

[0012] y c =W6g l (W5x c )

[0013] Then we obtain the hidden state sequence y containing sequential, cross-channel and cross-feature dependencies:

[0014] y=Trans(y s )+Trans(y f)+Trans(y c )

[0015] Where Trans is the transposition operation. By doing so, the model is ultimately able to capture the cross-dimensional information of the recommendation. When predicting the next item, given the hidden state related to the item to be predicted, the similarity is obtained by calculating the dot product between the hidden state and the embedding of all candidate items, and then the prediction results are sorted according to the similarity. The final output of the model is a vector whose length is equal to the number of knowledge concepts in the exercise, denoted as: M(C) = [m(c1), m(c2), ..., m(c t )], where each element is the probability of occurrence of the corresponding knowledge concept at time t+1, which is the learner's learning progress.

[0016] 2) Knowledge Concept Mastery Prediction Module: The probability of correctly answering a knowledge concept reflects the student's mastery of the corresponding knowledge concept. In this module, we predict the student's mastery of each knowledge concept based on their historical learning history. Specifically, we designed a knowledge tracking model consisting of a weight module and a memory module. The memory module models the learner's long-term memory knowledge, while the self-attention-based neural network in the weight module models the learner's recent knowledge, automatically balancing the trade-offs between these two types of knowledge.

[0017] In the model embedding module, we take the exercises, knowledge points and students’ answers as input, and then extract the students’ embedding vector h from its matrix t Then obtain the key storage matrix M in the memory module k , to calculate the weight vector w associated with the exercise and all potential concepts t , and its specific calculation formula is as follows:

[0018]

[0019] Then use the obtained exercise-potential concept related weight vector w t To calculate the current relevant representation vector of the relevant exercise The specific calculation form is as follows:

[0020]

[0021] At the same time, the self-attention neural network in the weight module also obtains the learner's recent knowledge state

[0022]

[0023] Then we represent the vector and the learner's recent knowledge state The vertical links are then represented by a fully linked layer, and finally the Sigmoid activation function is used to predict the probability of the correct answer for each concept. The specific calculation is as follows:

[0024]

[0025] Where y represents the degree of mastery of each concept, i.e., Z(C) = y. However, the probability of a student correctly answering an exercise can usually be used to indicate the difficulty of the exercise for that student. Since an exercise contains one or more knowledge concepts, we can calculate the probability of a student correctly answering a particular exercise based on the student's mastery of the knowledge concepts:

[0026]

[0027] Then we use To express the problem e i (C) Difficulty for students:

[0028] 3) Exercise filtering module: We obtain the student’s learning progress F(C) and knowledge concept mastery level Z(C) through the above two modules, and then input these two values ​​into the first part of the exercise filtering module to filter the exercises, thereby obtaining our candidate set CE.

[0029] We use cosine similarity to calculate the similarity between the knowledge concept e(C) contained in a certain exercise and the student's learning progress F(C), and then calculate the expected difficulty δ and D(e i (C) The distance between them, and then get the Euclidean norm Ω of the two values e(C) , which is expressed as follows:

[0030]

[0031] We will sort the exercises in the exercise bank EB according to Ω e(C) Arranged from smallest to largest, the exercises in the first section form our candidate set (CE). To expand the scope and diversity of students' potential interests, we designed the second part of the exercise filtering module, the exercise list optimization module, to further screen the CE. We introduced and improved the Hippo optimization algorithm, developing a new Hippo search algorithm (S-OH) based on the Sine transform. The Sine Hippo search algorithm generates a final recommendation list based on the candidate subset after adaptively updating the positions of multiple groups of hippos. The basic formula of the Sin chaotic map is as follows:

[0032] x n+1 =αsin(π / x n )

[0033] In the Hippo Optimization Algorithm, a hippo population typically consists of several female hippos, several juvenile hippos, several male hippos, and a dominant male hippo. Depending on the situation, the hippos in the population exhibit different behavioral patterns. In the Hippo Optimization Algorithm, these behavioral patterns are categorized into three main types: updating hippos' positions in rivers or ponds, defending against predators, and escaping predators to safe areas. Through these three behavioral patterns, hippos continuously adjust and optimize the population structure, ultimately finding the optimal group distribution. Therefore, we need to design a fitness function to assess the safety of hippo positions and determine the optimal location.

[0034] The fitness function is the objective function

[0035] Here, we choose Euclidean distance as the fitness function to evaluate the sum of distances between all exercises in the recommendation list. The distance between two exercises can be calculated by the equation:

[0036]

[0037] In the exercise recommendation task, the diversity of exercises is related to the value of the fitness. The larger the fitness, the smaller the similarity between two exercises and the greater the diversity of exercises. In other words, the diversity of exercises is proportional to the size of the fitness. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 Schematic diagram of the system architecture in the embodiment;

[0039] Figure 2 Schematic diagram of the knowledge tracking model in the embodiment;

[0040] Figure 3 Schematic diagram of the Hippo algorithm in the embodiment;

[0041] Figure 4 This is a schematic diagram of the distribution of exercises in the recommended list in the embodiment;

[0042] Figure 5 This is a schematic diagram of the comparison of knowledge tracking AUC in the embodiment;

[0043] Figure 6 This is a flow chart of the recommended method for practicing the present invention. DETAILED DESCRIPTION

[0044] The present invention will be further described below with reference to the accompanying drawings and embodiments, but the present invention is not limited thereto.

[0045] Example:

[0046] A multi-objective exercise recommendation method based on learner knowledge status prediction, such as Figure 1As shown: It is mainly divided into three modules, namely the learning progress prediction module, the knowledge concept mastery prediction module and the exercise filtering module. The model takes the learner's historical answer records as input. First, in the learning progress prediction module, we use the advanced time series prediction model to predict the student's learning progress. Secondly, in the knowledge concept mastery prediction module, we use the advanced knowledge tracking model to predict the student's knowledge concept mastery. Finally, the exercise filtering module first uses the learning progress and knowledge concept mastery obtained from the first two modules to filter the candidate exercise sets from the exercise library, and then uses the Sine Hippo optimization algorithm in this module to further optimize and screen the candidate sets, so as to maximize the diversity of the recommendation list and generate the optimal exercise recommendation list. Finally, through the exercise difficulty function, the exercises in the recommendation list are arranged in order of difficulty from small to large. The method includes the following steps:

[0047] 1) Learning Progress Prediction Module: This module aims to predict students' learning progress, specifically, the coverage of knowledge concepts. We use the knowledge concept sequence from time 0 to time t to predict the probability of each knowledge concept occurring at time t+1. To this end, we construct a time series prediction model. First, we embed the student's exercises, the knowledge concepts contained in the exercises, and the answers as input into the model. Because normalization on the sequence and feature mixers can slow training and cause information interference, we perform normalization before information mixing. We then extract information from three branches: sequence, feature, and channel. This yields a hidden state sequence y containing sequential, cross-channel, and cross-feature dependencies. When predicting the next item, given the hidden state associated with the item to be predicted, we calculate the dot product between this hidden state and the embeddings of all candidate items to obtain similarity. The prediction results are then ranked based on the similarity. The final output of the model is a vector whose length equals the number of knowledge concepts in the exercise, denoted as: M(C) = [m(c1), m(c2), ..., m(c t )], where each element is the probability of occurrence of the corresponding knowledge concept at time t+1, which is the learner's learning progress.

[0048] 2) Knowledge Concept Mastery Prediction Module: The probability of correctly answering a knowledge concept reflects the student's mastery of the corresponding knowledge concept. In this module, we predict the student's mastery of each knowledge concept based on their historical learning history. Specifically, we designed a knowledge tracking model consisting of a weight module and a memory module. The memory module models the learner's long-term memory knowledge, while the self-attention-based neural network in the weight module models the learner's recent knowledge, automatically balancing the trade-offs between these two types of knowledge.

[0049] In the model embedding module, we take the exercises, knowledge points and students’ answers as input, and then extract the students’ embedding vector h from its matrix t Then obtain the key storage matrix M in the memory module k , to calculate the weight vector w associated with the exercise and all potential concepts t Then use the obtained exercise-potential concept related weight vector w t To calculate the current relevant representation vector of the relevant exercise At the same time, the self-attention neural network in the weight module also obtains the learner's recent knowledge state Then we represent the vector and the learner's recent knowledge state The vertical link is then represented by a fully linked layer, and finally the Sigmoid activation function is used to predict the probability y of the correct answer for each concept. y represents the degree of mastery of each concept, that is, Z(C) = y. However, the probability of a student correctly answering an exercise can usually be used to indicate the difficulty of the exercise for him. Since an exercise contains one or more knowledge concepts, we can calculate the probability of a student correctly answering a certain exercise by the degree of mastery of the knowledge concepts. Then we use To express the problem e i (C) Difficulty for students:

[0050] 3) Exercise filtering module: We obtain the student’s learning progress F(C) and knowledge concept mastery level Z(C) through the above two modules, and then input these two values ​​into the first part of the exercise filtering module to filter the exercises, thereby obtaining our candidate set CE.

[0051] We use cosine similarity to calculate the similarity between the knowledge concept e(C) contained in a certain exercise and the student's learning progress F(C), and then calculate the expected difficulty δ and D(e i (C) The distance between them, and then get the Euclidean norm Ω of the two values e(C) , we will exercise the EB exercise according to Ω e(C) Arrange the exercises from smallest to largest, and select the exercises from the front to form our candidate set (CE). To expand the scope and diversity of students' potential interests, we designed the second part of the exercise filtering module, the exercise list optimization module, to further filter the CE. We introduced and improved the Hippo optimization algorithm, developing a new Hippo optimization algorithm based on the Sine transform. The Sine Hippo optimization algorithm adaptively updates the positions of multiple Hippo groups to generate the final recommendation list based on the candidate subset.

[0052] In the Hippo Optimization Algorithm, a hippo population typically consists of several female hippos, several juvenile hippos, several male hippos, and a dominant male hippo. Depending on the situation, hippos in the population exhibit different behavioral patterns. In the Hippo Optimization Algorithm, these behavioral patterns are categorized into three main types: updating hippos' positions in rivers or ponds, defending against predators, and escaping predators to safe areas. Through these three behavioral patterns, hippos are able to continuously adjust and optimize the population structure, thereby finding the optimal group distribution. Therefore, we need to design a fitness function to assess the safety of hippo positions and thus determine the optimal location. This fitness function is the objective function. Here, we choose Euclidean distance as the fitness function, which evaluates the sum of the distances between all exercises in the recommended list. In the exercise recommendation task, the diversity of exercises is related to the value of this fitness. A larger fitness value indicates less similarity between two exercises and greater diversity of exercises. In other words, the diversity of exercises is directly proportional to the fitness value.

[0053] This technical solution uses learning status prediction and optimization algorithm selection to recommend exercises. Using a knowledge tracking model and a time series prediction model, the learner's learning status (level of mastery of knowledge concepts and learning progress) is accurately predicted using their practice interaction data. These two predicted values ​​are then used to select exercise sets from the exercise database. The Sine Hippo optimization algorithm is then used to further optimize and select these sets, increasing the diversity of the exercises. Finally, an exercise difficulty function is used to sort the recommended exercises from easy to difficult.

[0054] During the experiment, we compared the exercises recommended by our model with those recommended by other baseline models. First, in terms of three performance indicators, diversity, accuracy, and novelty, our model outperformed most of the baseline models. Figure 4 As shown in Figure 1, a detailed comparison of the quality of exercise recommendation lists of different methods on three datasets is shown. The upper right corner of the figure is the ideal area, and the closer to the upper right corner, the better the performance. Compared with other models, the exercise recommendation list generated by our model is closer to the ideal area in the upper right corner, which shows that our model can better achieve a good balance between accuracy and novelty. In addition, in order to verify the accuracy of the knowledge tracking model, as shown in Figure 1, Figure 5 As shown in Figure 3, we compare its AUC indicator with other baseline models, and the results show that its performance is higher than other models and the prediction is more accurate.

[0055] The important contributions of this embodiment are as follows:

[0056] a. A time series prediction model was constructed. Based on the working memory mechanism, the model introduced the learner's response status to adjust the learning progress, and through a three-way mixed input method, the efficiency of information extraction was improved, thereby better predicting the learner's learning progress.

[0057] b. A knowledge tracking model was designed, consisting of a memory module and a weighting module. The memory module stores and updates the learner's long-term memory during the learning process. The weighting module automatically measures the relevance between the current knowledge concept and the knowledge concepts in the memory module. It also considers the learning status before and after the knowledge concept is switched, and then predicts the learner's mastery of the knowledge concept.

[0058] c. A multi-module exercise screening framework is proposed. Exercises are screened using the parameters obtained from the first two models, and the Hippo optimization algorithm is introduced to optimize the exercise list, thereby recommending more personalized exercises to learners.

[0059] d. Experiments are conducted on three public datasets and the results are compared with the baselines, revealing the advantages of our method in performance.

[0060] The above is a schematic description of the present invention and its embodiments, which is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by this and, without departing from the purpose of the present invention, designs a structure and embodiment similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.

Claims

1. A multi-objective exercise recommendation method based on learner knowledge state prediction, characterized by: The following steps are involved: Step 1: Process the interaction data between users and exercises in the three datasets of ASSISTments2009, ASSISTments2017, and Algebra2005, and input them into the knowledge tracking model to predict the learner's mastery of knowledge concepts; Step 2: Input the interaction data into the time series prediction model to predict the learner's learning progress; Step 3: Based on the prediction results, use the two prediction values ​​to preliminarily screen the question bank and generate candidate question sets; Step 4: Use the improved Hippo optimization algorithm to further optimize and screen the candidate exercise sets and generate a recommended exercise list; Step 5: Use the exercise difficulty function to calculate the difficulty of each exercise in the exercise recommendation list, arrange them in ascending order of difficulty, and generate the final exercise recommendation list.

2. The multi-objective exercise recommendation method based on learner knowledge status prediction according to claim 1, characterized in that: In step 1, in the exercise recommendation system, predicting the learner's mastery of knowledge concepts in their learning state requires a knowledge tracking approach. Therefore, we designed a knowledge tracking model consisting of a memory module and a weighting module. The memory module stores and updates the learner's long-term memory during the learning process. The weighting module automatically measures the relevance between the current knowledge concept and the knowledge concepts in the memory module by weighting them. It also considers the learning progress before and after switching knowledge concepts to predict the learner's mastery of the knowledge concept.

3. The multi-objective exercise recommendation method based on learner knowledge status prediction according to claim 2, characterized in that: In step 2, we predict the learner's learning progress. Since the learner's practice interaction sequence is a time-ordered sequence, we designed a time series prediction model. This model uses the learner's practice interactions from time 0 to time t to predict the probability of each knowledge concept appearing at time t+1, representing the learner's learning progress. Furthermore, based on working memory mechanisms, the model incorporates the learner's response status to adjust the learning progress. This three-way mixed input method improves information extraction efficiency, thereby better predicting the learner's learning progress.

4. The multi-objective exercise recommendation method based on learner knowledge status prediction according to claim 3, characterized in that: In step three, we use the Euclidean norm of the two predicted values ​​to perform a preliminary screening of the problem set, generating a candidate problem set. This process involves calculating the Euclidean norm for each problem, sorting them from smallest to largest, and selecting the problems at the beginning to form the candidate problem set.

5. The multi-objective exercise recommendation method based on learner knowledge status prediction according to claim 4, characterized in that: In step 4, since some redundant exercises exist in the candidate problem set, an optimization algorithm is needed to further refine and filter them. Here, we improve the Hippo optimization algorithm by using a Sine chaotic map to generate its population, making the traversal more uniform. Using the Sine Hippo optimization algorithm to optimize and filter the exercises effectively removes redundant items and increases the diversity of the recommended problem list.

6. The multi-objective exercise recommendation method based on learner knowledge status prediction according to claim 5, characterized in that: In step five, we consider that learners' knowledge acquisition follows a gradual progression, meaning that they need to learn addition and subtraction before learning multiplication and division. In other words, the difficulty of successive exercises should increase from simple to complex. To address this, we used a squared loss function as the exercise difficulty function. By calculating the difficulty of each exercise in the exercise list, we sorted them from least difficult to most difficult.

Images