Learning path recommendation method fusing hierarchical knowledge graph and deep knowledge tracking

Abstract

The invention discloses a learning path recommendation method fusing a hierarchical knowledge graph and deep knowledge tracking, and belongs to the technical field of data analysis, and the method comprises the following steps: constructing a KGDKT model, and constructing a deep knowledge tracking model capable of considering a logic relationship between knowledge points; searching a learning path, and generating all paths pointing from the starting point knowledge point to the target knowledge point based on the hierarchical course knowledge graph and the KGDKT; and optimal path selection: predicting the mastering degrees of the target knowledge points of the learner on different paths through KGDKT, and selecting one path with the best learning effect to be recommended to the learner. The path generated by the method accords with the logical relationship between the knowledge points, can adapt to the dynamically changing knowledge level of the learner, and improves the learning effect of the learner on the target knowledge point.

Description

Technical Field

[0001] The present invention belongs to the field of data analysis technology, and in particular to a learning path recommendation method that integrates hierarchical knowledge graphs and deep knowledge tracking. Background Art

[0002] Personalized learning path recommendation aims to recommend the most suitable learning path according to learners' interests, knowledge level, learning ability, etc., so as to improve learners' learning effect and satisfaction. There are two main methods for learning path recommendation: global optimal path recommendation and local iterative path recommendation. Among them, global optimal path recommendation can give a complete learning path at one time, but does not consider the changes in learners' knowledge status during the learning process. Local iterative path recommendation is based on the learners' changing knowledge status during the learning process. However, since the learning path is constructed during the learning process, learners cannot have a holistic grasp of the learning content before learning. Global path recommendation is divided into graph theory-based recommendation method, evolutionary algorithm-based recommendation method and clustering-based recommendation method according to the recommendation method. Graph theory-based method represents the relationship between knowledge points through graphs and assists in generating paths; evolutionary algorithm-based method regards the learning process as an evolutionary process and optimizes the learning path through genetic algorithms; clustering-based method clusters learners according to their characteristics and recommends the same learning path for similar learners.

[0003] Defects and shortcomings of existing technologies: First, the overall defect of the global path recommendation method is that the existing method does not take into account the changes in the learner's knowledge level during the learning process, resulting in low accuracy of path recommendation. The knowledge level reflects the learner's mastery of the knowledge point and is an important factor in the recommendation of personalized learning paths. Due to the different knowledge levels of learners, different learners may need different learning paths for the same target knowledge point. Therefore, ignoring the changes in the learner's knowledge level during the learning process, it is impossible to accurately recommend personalized learning paths.

[0004] Secondly, in terms of characterizing the learner's knowledge level, Deep Learning based Knowledge Tracing (DLKT) is the current mainstream method, which is used to obtain the learner's mastery of knowledge points and thus reflect the learner's knowledge level. However, the existing DLKT model rarely considers the relationship with other knowledge points when predicting the mastery of a knowledge point, and ignores the mutual influence between the mastery of different knowledge points, resulting in inaccurate results of the predicted mastery of knowledge points, affecting the measurement of the learner's knowledge level, and thus affecting the recommendation effect of the learning path.

[0005] Based on this, the present invention designs a learning path recommendation method that integrates hierarchical knowledge graph and deep knowledge tracking to solve the above problems. Summary of the invention

[0006] The purpose of the present invention is to solve the problems in the above-mentioned background technology and to propose a learning path recommendation method that integrates hierarchical knowledge graph and deep knowledge tracking.

[0007] In order to achieve the above object, the present invention adopts the following technical solution: a learning path recommendation method integrating hierarchical knowledge graph and deep knowledge tracking, comprising the following steps:

[0008] KGDKT model construction, building a deep knowledge tracking model that can consider the logical relationship between knowledge points;

[0009] Learning path search algorithm, based on the hierarchical course knowledge graph and KGDKT, generates all paths from the starting knowledge point to the target knowledge point;

[0010] The optimal path selection algorithm uses the KGDKT model to predict the improvement effect of each path on the target knowledge point, and selects a path that can maximize the learner's mastery of the target knowledge point.

[0011] Preferably, before constructing the KGDKT model, a hierarchical course knowledge graph needs to be constructed;

[0012] The construction of the KGDKT model specifically includes: using knowledge graph embedding technology to represent the knowledge graph as a vector e q , e q The features of the DTransformer model are integrated to complete the integration of the hierarchical course knowledge graph and DTransformer. A custom AttentionLayer layer is implemented and the hidden state of the last encoder layer of DTransformer is extracted as the feature f. This layer receives two input feature vectors f and e. q , and fuse them into a new vector a through the attention mechanism. Specifically, first define the weight matrix, then calculate the attention score and normalize it, and finally obtain the fused vector a through weighted summation. The process is shown in formula (1), where W a is the weight matrix used to transform f, U a is used to convert e q The weight matrix of is the weight vector used to calculate the attention score, s is f and e q Attention scores between

[0013]

[0014] Preferably, the learning path search algorithm specifically includes determining the starting knowledge point set SKP according to the learner's mastery of the knowledge point, and generating all paths {path a ,path b ,…,path n};

[0015] The algorithm first determines the target knowledge point and the starting knowledge point, where the target knowledge point is specified by the learner, and the starting knowledge point is obtained based on the learner's mastery of the knowledge point. Specifically, the KGDKT model is used to predict the learner's mastery of all knowledge points, and the knowledge points that the learner has mastered are taken as starting knowledge points and added to the knowledge point set SKP. In order to find all possible paths from all starting knowledge points in SKP to the target knowledge point t and ensure that all its prerequisite knowledge points have been learned before learning a certain knowledge point, it is necessary to consider the learning order of knowledge points with multiple prerequisite knowledge points;

[0016] Based on this, a backward breadth-first search algorithm (Backward BFS, referred to as B-BFS) is proposed. Starting from the target knowledge point, the algorithm traces back HKG to the starting knowledge point along the prerequisite relationship. If the starting knowledge point cannot be found or there is no starting knowledge point that meets the conditions, it will continue to trace back to the knowledge point without a prerequisite knowledge point and use it as the starting point of the learning path;

[0017] The learning order of multiple parallel prerequisite knowledge points is determined by the difficulty. They are learned in a learning order from easy to difficult. The difficulty is determined by the learner's mastery of these multiple knowledge points. The higher the difficulty of the knowledge point, the lower the corresponding mastery. The algorithm can ensure that all necessary prerequisite knowledge points have been learned before starting to learn a certain knowledge point by reversely searching for prerequisite knowledge points, and multiple parallel prerequisite knowledge points conform to the learning order from easy to difficult.

[0018] Finally, the path is constructed in the order of queue dequeue, and after reversing the direction of the path, it is added to the candidate path set Paths.

[0019] Preferably, the optimal path selection algorithm specifically includes: target knowledge point learning effect E p As shown in Formula 2, where E s Indicates the learner's initial mastery of the target knowledge point, E e Indicates the learner's final mastery of the target knowledge point;

[0020] The goal of the algorithm is to select a path from the candidate learning path set Paths that can maximize the learning effect of the learner's target knowledge point and recommend the path to the learner;

[0021] First, through the learning path search algorithm based on the hierarchical course knowledge graph, all learning paths from the knowledge points in the starting knowledge point set SKP to the target knowledge point t are searched from HKG to obtain the candidate learning path set Paths;

[0022] Secondly, for each path in Paths, KGDKT is used to predict the learner's learning effect on each path to select the path. Specifically, KGDKT is first used to predict the initial mastery level E of each path before learning t s , Secondly, according to the order of knowledge points in the path, predict the learner's mastery of each knowledge point in the path in turn, and form a new interaction record with the predicted results, which is added to the historical interaction record H to complete the update of H, and then complete the prediction of the mastery of other knowledge points, and finally obtain the mastery of the target knowledge point t after learning each path E e , the E obtained through learning of each path is calculated by formula (2) p Value, select E p The largest path is recommended to learners;

[0023]

[0024] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are:

[0025] 1. In the present invention, in terms of modeling the learner's knowledge level, by integrating the knowledge graph, the KGDKT model can improve the accuracy of predicting the learner's knowledge level, thereby improving the accuracy of path recommendation; in terms of the relationship between knowledge points recommended in the learning path, the path generated by the learning path search algorithm based on the hierarchical course knowledge graph proposed in this patent is more in line with the logical relationship between knowledge points, and adapts to the learner's knowledge level, avoiding the impact of the target knowledge point learning effect due to the failure to learn the prerequisite knowledge points; in terms of the optimal path selection for learning path recommendation, the KGDKT model takes into account the changes in the learner's knowledge level during the learning process along the path and the knowledge level at the target knowledge point, so that the path is more adapted to the learner's knowledge level and can improve the learner's learning effect at the target knowledge point. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A framework diagram of the learning path recommendation method proposed in the present invention that integrates hierarchical knowledge graph and deep knowledge tracking;

[0027] Figure 2 The KGDKT model diagram of the learning path recommendation method integrating hierarchical knowledge graph and deep knowledge tracking proposed in the present invention;

[0028] Figure 3A flow chart of a learning path search algorithm based on a hierarchical course knowledge graph for the learning path recommendation method that integrates hierarchical knowledge graph and deep knowledge tracking proposed in the present invention. DETAILED DESCRIPTION

[0029] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0030] Please refer to the attached Figure 1 -Attached Figure 3 The present invention provides a technical solution: a learning path recommendation method integrating hierarchical knowledge graph and deep knowledge tracking, comprising the following steps:

[0031] KGDKT model construction, building a deep knowledge tracking model that can consider the logical relationship between knowledge points;

[0032] Learning path search algorithm, based on the hierarchical course knowledge graph and KGDKT, generates all paths from the starting knowledge point to the target knowledge point;

[0033] The optimal path selection algorithm uses the KGDKT model to predict the improvement effect of each path on the target knowledge point, and selects a path that can maximize the learner's mastery of the target knowledge point.

[0034] Specifically, before constructing the KGDKT model, it is necessary to construct a hierarchical course knowledge graph;

[0035] The construction of the KGDKT model specifically includes: using knowledge graph embedding technology to represent the knowledge graph as a vector e q , e q The features of the DTransformer model are integrated to complete the integration of the hierarchical course knowledge graph and DTransformer. A custom AttentionLayer layer is implemented and the hidden state of the last encoder layer of DTransformer is extracted as the feature f. This layer receives two input feature vectors f and e. q , and fuse them into a new vector a through the attention mechanism. Specifically, first define the weight matrix, then calculate the attention score and normalize it, and finally obtain the fused vector a through weighted summation. The process is shown in formula (1), where W a is the weight matrix used to transform f, U a is used to convert e q The weight matrix of is the weight vector used to calculate the attention score, s is f and e q Attention scores between

[0036] h1=W a f

[0037] h q =U a e q

[0038]

[0039]

[0040] a=α(W f f+U f e q ).

[0041] Specifically, the learning path search algorithm specifically includes determining the starting knowledge point set SKP according to the learner's mastery of the knowledge point, and generating all paths {path a ,path b ,…,path n};

[0042] The algorithm first determines the target knowledge point and the starting knowledge point, where the target knowledge point is specified by the learner, and the starting knowledge point is obtained based on the user's mastery of the knowledge point. Specifically, the KGDKT model is used to predict the learner's mastery of all knowledge points, and the knowledge points that the learner has mastered are taken as starting knowledge points and added to the knowledge point set SKP. In order to find all possible paths from all starting knowledge points in SKP to the target knowledge point t and ensure that all prerequisite knowledge points have been learned before learning a certain knowledge point, it is necessary to consider the learning order of knowledge points with multiple prerequisite knowledge points;

[0043] Based on this, a backward breadth-first search algorithm (Backward BFS, referred to as B-BFS) is proposed. Starting from the target knowledge point, the algorithm traces back HKG to the starting knowledge point along the prerequisite relationship. If the starting knowledge point cannot be found or there is no starting knowledge point that meets the conditions, it will continue to trace back to the knowledge point without a prerequisite knowledge point and use it as the starting point of the learning path;

[0044] The learning order of multiple parallel prerequisite knowledge points is determined by the difficulty. They are learned in a learning order from easy to difficult. The difficulty is determined by the learner's mastery of these multiple knowledge points. The higher the difficulty of the knowledge point, the lower the corresponding mastery. The algorithm can ensure that all necessary prerequisite knowledge points have been learned before starting to learn a certain knowledge point by reversely searching for prerequisite knowledge points, and multiple parallel prerequisite knowledge points conform to the learning order from easy to difficult.

[0045] Finally, the path is constructed in the order of queue dequeue, and after reversing the direction of the path, it is added to the candidate path set Paths.

[0046] Specifically, the optimal path selection algorithm specifically includes: target knowledge point learning effect E p As shown in Formula 2, where E s Indicates the learner's initial mastery of the target knowledge point, E e Indicates the learner's final mastery of the target knowledge point;

[0047] The goal of the algorithm is to select a path from the candidate learning path set Paths that can maximize the learning effect of the learner's target knowledge point and recommend the path to the learner;

[0048] First, through the learning path search algorithm based on the hierarchical course knowledge graph, all learning paths from the knowledge points in the starting knowledge point set SKP to the target knowledge point t are searched from HKG to obtain the candidate learning path set Paths;

[0049] Secondly, for each path in Paths, KGDKT is used to predict the learner's learning effect on each path to select the path. Specifically, KGDKT is first used to predict the initial mastery level E of each path before learning t s , Secondly, according to the order of knowledge points in the path, predict the learner's mastery of each knowledge point in the path in turn, and form a new interaction record with the predicted results, which is added to the historical interaction record H to complete the update of H, and then complete the prediction of the mastery of other knowledge points, and finally obtain the mastery of the target knowledge point t after learning each path E e , the E obtained through learning of each path is calculated by formula (2) p Value, select E p The largest path is recommended to learners;

[0050]

[0051] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A learning path recommendation method integrating hierarchical knowledge graph and deep knowledge tracking, characterized in that: The following steps are involved: KGDKT model construction, building a deep knowledge tracking model that can consider the logical relationship between knowledge points; Learning path search algorithm, based on the hierarchical course knowledge graph and KGDKT, generates all paths from the starting knowledge point to the target knowledge point; The optimal path selection algorithm uses the KGDKT model to predict the improvement effect of each path on the target knowledge point, and selects a path that can maximize the learner's mastery of the target knowledge point.

2. The learning path recommendation method integrating hierarchical knowledge graph and deep knowledge tracing according to claim 1 is characterized in that: Before constructing the KGDKT model, it is necessary to construct a hierarchical course knowledge graph; The construction of the KGDKT model specifically includes: using knowledge graph embedding technology to represent the knowledge graph as a vector e q , e q The features of the DTransformer model are integrated to complete the integration of the hierarchical course knowledge graph and DTransformer. A custom AttentionLayer layer is implemented and the hidden state of the last encoder layer of DTransformer is extracted as the feature f. This layer receives two input feature vectors f and e. q , and fuse them into a new vector a through the attention mechanism. Specifically, first define the weight matrix, then calculate the attention score and normalize it, and finally obtain the fused vector a through weighted summation. The process is shown in formula (1), where W a is the weight matrix used to transform f, U a is used to convert e q The weight matrix of is the weight vector used to calculate the attention score, s is f and e q Attention scores between h1=W a f h q =U a have been q a=α(W f f+U f e q )。 3. The learning path recommendation method integrating hierarchical knowledge graph and deep knowledge tracking according to claim 1 is characterized in that: The learning path search algorithm specifically includes determining the starting knowledge point set SKP according to the learner's mastery of the knowledge point, and generating all paths {path a ,path b ,…,path n }; The algorithm first determines the target knowledge point and the starting knowledge point, where the target knowledge point is specified by the learner, and the starting knowledge point is obtained based on the learner's mastery of the knowledge point. Specifically, the KGDKT model is used to predict the learner's mastery of all knowledge points, and the knowledge points that the learner has mastered are taken as starting knowledge points and added to the knowledge point set SKP. In order to find all possible paths from all starting knowledge points in SKP to the target knowledge point t and ensure that all its prerequisite knowledge points have been learned before learning a certain knowledge point, it is necessary to consider the learning order of knowledge points with multiple prerequisite knowledge points; Based on this, a backward breadth-first search algorithm (Backward BFS, referred to as B-BFS) is proposed. Starting from the target knowledge point, the algorithm traces back HKG to the starting knowledge point along the prerequisite relationship. If the starting knowledge point cannot be found or there is no starting knowledge point that meets the conditions, it will continue to trace back to the knowledge point without a prerequisite knowledge point and use it as the starting point of the learning path; The learning order of multiple parallel prerequisite knowledge points is determined by the difficulty. They are learned in a learning order from easy to difficult. The difficulty is determined by the learner's mastery of these multiple knowledge points. The higher the difficulty of the knowledge point, the lower the corresponding mastery. The algorithm can ensure that all necessary prerequisite knowledge points have been learned before starting to learn a certain knowledge point by reversely searching for prerequisite knowledge points, and multiple parallel prerequisite knowledge points conform to the learning order from easy to difficult. Finally, the path is constructed in the order of queue dequeue, and after reversing the direction of the path, it is added to the candidate path set Paths.

4. The learning path recommendation method integrating hierarchical knowledge graph and deep knowledge tracking according to claim 1 is characterized in that: The optimal path selection algorithm specifically includes predicting the learner's mastery of target knowledge points on different paths through KGDKT, selecting a path with the best learning effect and recommending it to the learner, and the target knowledge point learning effect E p As shown in Formula 2, where E s Indicates the learner's initial mastery of the target knowledge point, E e Indicates the learner's final mastery of the target knowledge point; The goal of the algorithm is to select a path from the candidate learning path set Paths that can maximize the learning effect of the learner's target knowledge point and recommend the path to the learner; First, through the learning path search algorithm based on the hierarchical course knowledge graph, all learning paths from the knowledge points in the starting knowledge point set SKP to the target knowledge point t are searched from HKG to obtain the candidate learning path set Paths; Secondly, for each path in Paths, KGDKT is used to predict the learner's learning effect on each path to select the path. Specifically, KGDKT is first used to predict the initial mastery level E of each path before learning t s , Secondly, according to the order of knowledge points in the path, predict the learner's mastery of each knowledge point in the path in turn, and form a new interaction record with the predicted results, which is added to the historical interaction record H to complete the update of H, and then complete the prediction of the mastery of other knowledge points, and finally obtain the mastery of the target knowledge point t after learning each path E e , the E obtained through learning of each path is calculated by formula (2) p Value, select E p The largest path is recommended to learners;

Images