Multi-source heterogeneous learning path planning method based on personalized constraints

Abstract

The invention discloses a multi-source heterogeneous learning path planning method based on personalized constraints, which solves the problems of single path, data staticization and the like in the prior art, and comprises the following steps: acquiring multi-source behavior data of a learner, carrying out feature modeling, and constructing a high-dimensional behavior portrait; according to the high-dimensional behavior portrait, utilizing a neural collaborative filtering algorithm to predict the interest and mastering probability of a learner to any knowledge point, and generating a preliminary learning path; carrying out dominant and implicit evaluation on the knowledge mastering state of the learner by adopting a cognitive diagnosis model, and carrying out dynamic correction on the preliminary learning path to obtain a corrected learning path; performing dimension reduction on the multi-source behavior data by adopting a principal component analysis algorithm to extract a core factor, and forming an evaluation basis for path optimization; and constructing a multi-objective path function model, and obtaining an optimal dynamic learning path by adopting a multi-objective optimization path algorithm in combination with the corrected learning path and the core factor. The method has high practical value and popularization value in the technical field of learning path planning.

Description

Technical Field

[0001] The present invention relates to the technical field of learning path planning, and in particular to a multi-source heterogeneous learning path planning method based on personalized constraints. Background Art

[0002] Learning path planning is the process of systematically designing and arranging learning content, methods, resources, and time based on factors such as individual learning goals, current knowledge level, learning ability, and time resources. Currently, existing educational learning platforms suffer from the following problems: learning resources are fragmented and unsystematic, student ability assessment relies on single assessments and is highly haphazard, recommended paths cannot dynamically adapt to learners' development, learning effect evaluation is ambiguous, and multi-objective optimization capabilities are lacking.

[0003] The learning path planning method in the existing technology, such as the public technology of "Publication No. CN112734142A, named Resource Learning Path Planning Method and Device Based on Deep Learning", first collects the learning data click records of virtual objects in real time and stores them in the virtual object database; then obtains the target data from the virtual object database according to the screening conditions to form a sample set for model training to obtain a prediction model; finally, if the initial input features of the current course sub-trajectory information are received, they are input into the prediction model for calculation to obtain the output result.

[0004] Another example is the publicly available technology, "System and Method for Intelligent English Vocabulary Learning Path Planning Based on Reinforcement Learning," with publication number CN119904005A. This includes a preliminary learning path planning module, a learning device feedback module, a learning effect analysis module, and a learning path update module. This technology provides a reinforcement learning-based intelligent English vocabulary learning path planning system and method. This system obtains learners' basic vocabulary test results, planned learning time, and goals, and combines dynamic vocabulary matching with learning starting point settings to achieve personalized learning path planning.

[0005] The above technologies have the following problems: First, the path generation method is single and has poor adaptability: Most existing path planning systems are based on collaborative filtering or static rules, ignoring the learner's dynamic state and cognitive evolution process. Second, ability assessment is static and highly generalized: Most systems use initial baseline tests that cannot reflect students' true real-time cognitive level and path deviation trends. Third, data structure utilization is insufficient: They fail to effectively process structured, semi-structured, and unstructured learning data, and are unable to extract deep information features from multi-dimensional behaviors. Fourth, they lack multi-objective optimization capabilities: Traditional methods struggle to achieve a trade-off between multiple objectives such as "learning efficiency, interest, and mastery."

[0006] Therefore, there is an urgent need to propose a logically simple and dynamic multi-source heterogeneous learning path planning method based on personalized constraints. Summary of the Invention

[0007] In response to the above problems, the present invention aims to provide a multi-source heterogeneous learning path planning method based on personalized constraints. The technical solutions adopted by the present invention are as follows: The multi-source heterogeneous learning path planning method based on personalized constraints includes the following steps: Obtain learners' multi-source behavioral data, perform feature modeling, and construct high-dimensional behavioral profiles; Based on high-dimensional behavioral portraits and using neural collaborative filtering algorithms, we predict learners' interest in and probability of mastering any knowledge point, and generate a preliminary learning path. Use cognitive diagnostic models to conduct explicit and implicit assessments of learners' knowledge mastery status, and dynamically revise the initial learning path to obtain the revised learning path; The principal component analysis algorithm is used to reduce the dimension of multi-source behavioral data to extract core factors and form the evaluation basis for path optimization; A multi-objective path function model is constructed, and combined with the revised learning path and core factors, a multi-objective optimization path algorithm is used to obtain the optimal dynamic learning path.

[0008] Compared with the prior art, the present invention has the following beneficial effects: The present invention performs feature modeling based on multi-source behavioral data, which serves as the input of the neural collaborative filtering model, taking into account both static ability differences and dynamic behavioral characteristics, and providing a high-dimensional data foundation for subsequent interest prediction and path generation.

[0009] This paper uses a neural collaborative filtering algorithm to predict a learner's interest and mastery probability for any given knowledge point, generating preliminary learning paths that enable highly personalized recommendations and leverage group collaboration signals. This paper utilizes neural collaborative filtering, combining collaborative filtering with a cognitive diagnostic framework. This approach leverages effective feedback from similar learners to enhance recommendation quality in cold-start and sparse data environments. Furthermore, the proposed model performs better in predicting interest and mastery statistics, facilitating path construction based on individual differences.

[0010] This invention utilizes a cognitive diagnostic model to perform both explicit and implicit assessments of a learner's knowledge mastery status and dynamically corrects the initial learning path to produce a revised learning path. This approach offers the advantages of more accurate characterization of mastery status and automatic dynamic adaptation. The present invention's explicit diagnosis (determining mastery based on test scores) provides a clear feedback mechanism, while the implicit assessment's "efficiency-difficulty" metric helps identify potential weaknesses and improves the synchronization of the path with the actual learning state. This combination of explicit and implicit methods allows for more precise path adjustment, avoids misjudgments, and enhances the efficiency of personalized dynamic adaptation.

[0011] The present invention uses the principal component analysis algorithm to reduce the dimension of multi-source behavior data, extract core factors, and form an evaluation basis for path optimization. The advantages are as follows: dimension reduction and denoising, and improving the model efficiency. The present invention uses the principal component analysis algorithm (PCA) to efficiently compress highly correlated behavior data, remove noise and redundant dimensions, and reduce the model complexity and overfitting risk. In addition, the present invention considers the principal factors in path optimization, can associate the path with the key behavior preferences of students, and improve the path quality and interpretability.

[0012] The present invention obtains the optimal dynamic learning path by constructing a multi-objective path function model, combining the corrected learning path and core factors, and using a multi-objective optimization path algorithm. The advantages are as follows: multi-objective balance and ensuring personalization and global optimality. Among them, the present invention realizes a more comprehensive path planning by simultaneously optimizing dimensions such as "learning time", "interest maintenance", "smoothness", and "coverage". In addition, the present invention adopts the Pareto optimal strategy to avoid single-objective bias and ensure the optimal compromise among multiple objectives. The final path of the present invention has stronger generalization and application capabilities and performs stably in a real teaching environment.

[0013] In summary, the present invention has the advantages of simple logic, dynamic reliability, etc., and has high practical value and promotion value in the technical field of learning path planning. Brief Description of the Drawings

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required in the embodiments. It should be understood that the following drawings only show some embodiments of the present invention, and therefore should not be regarded as limiting the protection scope. For those skilled in the art, without creative efforts, other related drawings can also be obtained based on these drawings.

[0015] Figure 1 It is the logic flowchart of the present invention. Detailed Embodiments

[0016] To make the objectives, technical solutions, and advantages of the present application clearer, the following further illustrates the present invention with reference to the drawings and embodiments. The embodiments of the present invention include but are not limited to the following embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the protection scope of the present application.

[0017] As Figure 1 shown, this embodiment provides a multi-source heterogeneous learning path planning method based on personalized constraints, which integrates a personalized learning path planning system of neural collaborative filtering, cognitive diagnosis modeling, principal component analysis, and multi-objective optimization to realize the closed-loop optimization of path generation-evaluation-update.

[0018] Specifically, the multi-source heterogeneous learning path planning method based on personalized constraints includes the following steps: In the first step, feature modeling is performed on the multi-source behavior data of learners to form a high-dimensional behavior portrait. Here, the multi-source behavior data can be obtained by integrating students' background academic records, online learning behaviors, extracurricular activity participation, etc. The high-dimensional behavior portrait is a method of comprehensively depicting user behavior characteristics through multi-dimensional data. It usually involves multiple data sources and complex feature vectors, and can more accurately describe users' behavior patterns, preferences, and potential needs.

[0019] In the second step, the neural collaborative filtering algorithm is used to predict the probability of learners' interest and mastery of each knowledge point, and a preliminary learning path is generated. Here, the multi-source static and dynamic behavior data of learners (that is, the multi-source behavior data is divided into multi-source static data and multi-source dynamic behavior data) are subjected to feature modeling, and the neural collaborative filtering algorithm is used to predict their interest and mastery probability of each knowledge point. Here, its output not only provides basic data for subsequent cognitive diagnosis, but also provides a basis for path ranking and recommendation. Specifically: Here, the high-dimensional behavior portrait contains the states of several learners, and the state of the learner is composed of the splicing of three types of vectors, and its expression is:

[0020] Among them, represents the transposed matrix of the static basic information vector of the th learner, which is used to store the relatively fixed and static basic attribute information of the learner, such as feature data that does not change dynamically during the learning process, such as gender, age, educational background, etc., and organizes this information in matrix form; represents the transposed matrix of the behavior process feature vector of the th learner, which is used to record the behavior-related characteristics of the learner during the learning process, such as dynamic behavior data such as learning duration distribution, answering interaction frequency, learning path selection, etc., and presents the characteristics of these behavior process dimensions in matrix form; represents the transposed matrix of the cognitive result feature vector of the th learner, which reflects the achievement characteristics at the cognitive level after the learner's learning, such as data reflecting learning output and cognitive level such as test scores, knowledge mastery degree evaluation results, skill certification status, etc., and stores them in matrix form as well.

[0021] Here, the above feature vectors are used as the input of the neural collaborative filtering model, taking into account both static ability differences and dynamic behavior characteristics, and providing a high-dimensional data basis for subsequent interest prediction and path generation.

[0022] In this embodiment, the neural collaborative filtering algorithm is used to predict the interest and mastery probability of the th learner in the th knowledge point, and its expression is: where represents the th dimension of the behavior embedding vector of the th learner, which is used to characterize the attributes of the learner in different latent dimensions, such as the performance of learning style, knowledge base preference, etc. on the corresponding latent features; represents the th dimension of the embedding vector of the th knowledge point, which is used to describe the characteristics of the knowledge point in each latent feature dimension, such as the difficulty of the knowledge point, the manifestation of the category of the knowledge domain to which it belongs on the corresponding latent feature, etc.; b represents the system bias term; E represents the identity matrix; N represents the number of dimensions of the embedding vector, that is, the total number of latent dimensions used to characterize learners and knowledge points, which is a parameter set by the model and determines the scale of the latent feature decomposition of users and knowledge.

[0023] The interest and mastery probability of the th learner in the th knowledge point is sorted from large to small to generate a preliminary learning path. Here, the higher the score of the interest and mastery probability of the th learner in the th knowledge point , the more suitable it is for the current path node. This scoring result will directly affect the priority of subsequent path recommendations.

[0024] Here, the state fusion embedding representation and score correction of several learners are carried out, and its expression is:

[0025] where M represents the total number of other learners considered in the fusion process; represents the total number of knowledge points; represents the preference weight of the th learner in the th knowledge point; represents the non-linear activation function, such as Sigmoid, ReLU, etc. commonly used, and its role is to perform non-linear transformation on the operation result in the parentheses, introduce non-linear factors, and enable the model to fit more complex relationships; represents the transposed matrix of the th dimension of the fusion feedback vector of the th knowledge point, which is used to describe the characteristics of knowledge points, teaching strategies, etc. in each latent feature dimension; represents the embedding vectors of learners; Indicates the learners and Similarity of learners; Indicates the measurement of learners and The weight coefficient of the relevant features of each knowledge point; Indicates the The original features of the knowledge point features.

[0026] Here, by introducing behavioral feedback from similar learners, we can fine-tune individual scores and enhance the generalization ability of the model.

[0027] The third step is to use the cognitive diagnostic model to conduct explicit and implicit assessments of the learner's knowledge mastery status, and dynamically revise the initial learning path to obtain the revised learning path. Specifically: No. The learners The mastery status of knowledge points The expression is: in, Indicates the The first part of the fusion feedback vector of the knowledge point dimensions; Indicates the The learners The mastery threshold of a knowledge point.

[0028] Here, an implicit cognitive difficulty evaluation function is established, and its expression is: in, represents a nonlinear activation function; Represents the global weight coefficient for regulating the difficulty of knowledge points; Indicates the global weight coefficient of control efficiency / time consumption; n represents the number of samples participating in weighted averaging; Indicates the Individual learner's personalized sensitivity coefficient to difficulty signals; Indicates the Individual learner's individual sensitivity coefficient to efficiency / time consumption signals; Indicates the The learners The cognitive difficulty of each knowledge point; Indicates the The learners Learning each knowledge point takes time.

[0029] This embodiment uses the maximum likelihood and maximum a posteriori estimation algorithms to dynamically estimate the learner's ability value and dynamically correct the initial learning path, which includes the following steps: Step (31) uses an S-shaped function to model the mastery probability of the learner's ability and the question difficulty, and its expression is: where, represents the mastery level of the th learner for the th question; represents the ability parameter of the th learner; represents the difficulty parameter of the th question; represents the slope; e represents the base of the natural logarithm.

[0030] Step (32) uses maximum likelihood estimation to reverse-infer the ability parameter , and its expression is: where, represents the weight of the th question; represents the true classification result of the th question; represents the probability of answering correctly calculated based on ability and question difficulty.

[0031] Here, a prior distribution is introduced, and its expression is: , where, represents the prior intensity parameter; C represents the normalization constant; represents the regularization function for estimating the ability parameter of the th learner.

[0032] Step (33) dynamically corrects the preliminary learning path according to the mastery level of the th learner for the th question and the learner's ability parameter .

[0033] Here, based on the relationship between student ability and question difficulty, the maximum likelihood and maximum a posteriori estimation methods are used to dynamically estimate the student ability value and achieve precise adaptation of the learning path.

[0034] Fourthly, the principal component analysis algorithm is used to reduce the dimension of multi-source behavior data to extract core factors and form an evaluation basis for path optimization. Specifically: Step (41) normalizes the multi-source behavior data.

[0035] Step (42) constructs a covariance matrix and performs eigenvalue decomposition, and its expression is: , where, R represents the covariance matrix; represents the number of samples of multi-source behavior data after normalization; Z represents the matrix of multi-source behavior data after normalization; represents the transpose of the matrix of multi-source behavior data after normalization; represents an orthogonal direction of data variation in the covariance matrix; represents an orthogonal direction of data variation in the covariance matrix variance information content of the above data.

[0036] Step (43) extracts the core factors of the multi-source behavior data after normalization, and its expression is: where, represents the th learner's score on the th principal component; represents the th principal component's weight load of the th learner on the th behavior feature; represents the th standardized behavior feature; p represents the total number of normalized behavior features.

[0037] Step (44) selects the core factors of the multi-source behavior data using the variance contribution rate, and its expression is: where, represents the eigenvalue of the th principal component; represents the th principal component's factor's variance contribution rate; represents the total contribution rate of all principal components; represents the number of core factors of all multi-source behavior data.

[0038] In this embodiment, the expression for the evaluation of this path optimization is: where, represents the comprehensive learning effect scoring function; represents the number of core factors of the selected multi-source behavior data.

[0039] Fifth step, construct a multi-objective path function model, and combine the corrected learning path and core factors, and use a multi-objective optimization path algorithm to obtain the optimal dynamic learning path, including the following steps: where, the expression of this multi-objective path function model is: where, represents the objective function of the 1st optimized learning path vector ; represents the objective function of the 2nd optimized learning path vector The objective function; Indicates the th optimized learning path vector The objective function; Indicates the learning path vector; Indicates the - dimensional space sequence composed of real numbers that constitutes the learning path set; X represents the learning path vector The value set of

[0040] For the multi - objective path function model Perform constraints, and its expression is: Among them, Indicates the th evaluation constraint function of the principal component; c represents the constant threshold; Indicates the constraint function of the ; Indicates at the th principal component, the th learner's preference weight for the th knowledge point; Indicates the participation or weight of the th knowledge point in the path; Indicates the th knowledge point's mastery threshold; Indicates the comprehensive learning effect scoring function The natural logarithm function with base e.

[0041] Definition of Pareto - optimal solution. The preset optimal path solution for multi - objective optimization satisfies the following conditions: Among them, Indicates the set except the feasible learning path ; X represents the set of all feasible learning paths; Indicates the evaluation value of the feasible learning path at the th optimization objective; ; Indicates the evaluation value of the current candidate path at the th optimization objective.

[0042] The above - mentioned embodiments are only the preferred embodiments of the present invention, and do not limit the protection scope of the present invention. Any changes made by using the design principle of the present invention and non - creative labor on this basis shall fall within the protection scope of the present invention.

Claims

1. A multi-source heterogeneous learning path planning method based on personalized constraints, characterized in that It includes the following steps: Obtain the multi-source behavior data of the learner, conduct feature modeling, and construct a high-dimensional behavior portrait; According to the high-dimensional behavior portrait, use the neural collaborative filtering algorithm to predict the learner's interest and mastery probability of any knowledge point, and generate a preliminary learning path; Adopt a cognitive diagnosis model to conduct explicit and implicit evaluations on the learner's knowledge mastery status, and dynamically correct the preliminary learning path to obtain a corrected learning path; Use the principal component analysis algorithm to reduce the dimension of the multi-source behavior data to extract core factors, and form an evaluation basis for path optimization; Construct a multi-objective path function model, combine the corrected learning path and core factors, and use the multi-objective optimization path algorithm to obtain the optimal dynamic learning path.

2. The multi-source heterogeneous learning path planning method based on personalized constraints according to claim 1, wherein According to the high-dimensional behavior portrait, use the neural collaborative filtering algorithm to predict the learner's interest and mastery probability of any knowledge point, and generate a preliminary learning path, which includes the following steps: The high-dimensional behavior portrait contains the states of several learners, and the state of the learner state is composed of the splicing of three types of vectors, and its expression is: , where represents the transposed matrix of the static basic information vector of the -th learner; represents the transposed matrix of the behavioral process feature vector of the -th learner; represents the transposed matrix of the cognitive result feature vector of the -th learner. Predict the interest and mastery probability of the th learner in the th knowledge point using the neural collaborative filtering algorithm , and its expression is: , where represents the -th dimension of the behavior embedding vector of the -th learner; represents the -th dimension of the embedding vector of the -th knowledge point; b represents the system bias term; E represents the identity matrix; N represents the number of dimensions of the embedding vector; For the th learner, the interest and mastery probability for the th knowledge point are sorted from largest to smallest, and a preliminary learning path is generated.

3. The multi-source heterogeneous learning path planning method based on personalized constraints according to claim 2, wherein It also includes: Perform fusion embedding representation and score correction on the states of several learners, and its expression is: , ; Among them, represents the fitness measure of the th knowledge point; M represents the total number of other learners considered in the fusion process; represents the total number of knowledge points; represents the th learner's preference weight for the th knowledge point; represents a non-linear activation function; represents the transposed matrix of the th dimension of the fusion feedback vector of the th knowledge point; represents the corrected embedding vector of the th learner; represents the similarity between the th learner and the th learner; represents the weight coefficient for measuring the relevant features of the th learner and the th knowledge point; represents the original feature of the th knowledge point feature.

4. The multi-source heterogeneous learning path planning method based on personalized constraints according to claim 3, characterized in that Adopt a cognitive diagnosis model to conduct explicit and implicit evaluations on the learner's knowledge mastery status, including the following steps: The th learner's mastery status of the th knowledge point is expressed as: ; among them, represents the -th learner's mastery threshold for the -th knowledge point; Establish a recessive cognitive difficulty evaluation function, and its expression is: ; among them, represents the global weight coefficient for regulating the difficulty of knowledge points; represents the global weight coefficient for regulating efficiency / time consumption; n represents the number of samples participating in the weighted average; represents the th learner's personalized sensitivity coefficient to the difficulty signal; represents the th learner's personalized sensitivity coefficient to the efficiency / time consumption signal; represents the th learner's cognitive difficulty of the th knowledge point, ; represents the th learner's learning time for the th knowledge point.

5. The multi-source heterogeneous learning path planning method based on personalized constraints according to claim 4, wherein It also includes: Use the maximum likelihood and maximum a posteriori estimation algorithms to dynamically estimate the learner's ability value, and dynamically correct the preliminary learning path, which includes the following steps: Use the S-shaped function to model the mastery probability of the learner's ability and the question difficulty, and its expression is: ; among them, represents the mastery level of the th learner for the th question; represents the ability parameter of the th learner; represents the difficulty parameter of the th question; represents the slope; e represents the base of the natural logarithm; Use maximum likelihood estimation to back-calculate the ability parameters of the learner , and its expression is: ; among them, represents the weight of the th question; represents the true classification result of the th question; represents the probability of answering correctly calculated based on ability and question difficulty; Introduce a prior distribution, and its expression is: , ; wherein, represents the prior intensity parameter; C represents the normalization constant; represents the regularization function for estimating the ability parameter of the According to the mastery level of the th learner for the th question and the ability parameter of the learner, dynamically correct the initial learning path.

6. The multi-source heterogeneous learning path planning method based on personalized constraints according to claim 5, wherein, Use the principal component analysis algorithm to reduce the dimension of the multi-source behavior data to extract core factors, and form an evaluation basis for path optimization, which includes: Normalize the multi-source behavior data; Construct a covariance matrix and perform eigenvalue decomposition, and its expression is: , ; where R represents the covariance matrix; represents the number of samples of the multi-source behavior data after normalization; Z represents the matrix of multi-source behavior data after normalization; represents the transpose of the matrix of multi-source behavior data after normalization; represents an orthogonal direction of data variation in the covariance matrix; represents an orthogonal direction of data variation in the covariance matrix the variance information content of the data above Extract the core factors of the multi-source behavior data after normalization processing, and its expression is: , where represents the score of the th learner on the th principal component; represents the weight load of the th learner on the th principal component for the th behavioral characteristic; represents the th behavioral characteristic after standardization; Select the core factors of the multi-source behavior data using the variance contribution rate, and its expression is: , ; Among them, represents the eigenvalue of the th principal component; represents the variance contribution rate of the factor of the th principal component; represents the total contribution rate of all principal components; represents the number of core factors of all multi-source behavior data; p represents the total number of normalized behavior characteristics.

7. The multi-source heterogeneous learning path planning method based on personalized constraints according to claim 6, wherein The expression of the evaluation of the path optimization is: ; among them, represents the comprehensive learning effect scoring function; represents the number of core factors of the selected multi-source behavior data.

8. The multi-source heterogeneous learning path planning method based on personalized constraints according to claim 7, characterized in that, Construct a multi-objective path function model, combine the corrected learning path and core factors, and use the multi-objective optimization path algorithm to obtain the optimal dynamic learning path, including the following steps: The expression of the multi-objective path function model is: ; wherein, represents the objective function of the first optimized learning path vector ; represents the objective function of the second optimized learning path vector ; represents the th optimized learning path vector ; represents the learning path vector; represents the learning path set composed of a sequence of -dimensional spaces composed of real numbers; X represents the value set of the learning path vector ; For the multi-objective path function model to be constrained, its expression is: , where represents the evaluation constraint function of the -th principal component; c represents a constant threshold; represents the constraint function of the s-th knowledge point, ; represents the preference weight of the -th learner for the -th knowledge point on the -th principal component; represents the participation or weight of the -th knowledge point in the path; represents the mastery threshold of the -th knowledge point; represents the comprehensive learning effect scoring function the natural logarithm function with base e; Definition of Pareto optimal solution, it is preset that the optimal path solution of multi-objective optimization satisfies the following conditions: , where represents the set excluding the feasible learning paths other than; X represents the set of all feasible learning paths; represents the evaluation value of the feasible learning path on the optimization objective; ; represents the evaluation value of the current candidate path on the optimization objective.

Images