Learning path recommendation method and device, computer device and storage medium

Abstract

The application provides a learning path recommendation method and device, computer equipment and a storage medium, applied to the technical field of computers, and the method comprises the following steps: acquiring a constructed global knowledge graph, wherein the global knowledge graph comprises a plurality of nodes, connection edges between different nodes, and weights of each connection edge, and the plurality of nodes comprise ability nodes and knowledge nodes; determining a local learning graph matched with a target user from the global knowledge graph according to a current learning task determined by the target user; generating current multi-modal portrait information matched with the local learning graph according to historical learning behavior data of the target user, wherein the current multi-modal portrait information is used to reflect the learning state of the target user for the local learning graph; and generating a current learning path of the target user for the current learning task by using a determined path planning algorithm based on the current multi-modal portrait information and the local learning graph, and sending the current learning path to the target user.

Description

Technical Field

[0001] This application relates to the field of computer technology, and more specifically, to a learning path recommendation method, apparatus, computer device, and storage medium. Background Technology

[0002] With the development of educational informatization and intelligentization, personalized teaching has gradually become an important direction for improving learning outcomes and educational equity. Among related technologies, personalized teaching systems mainly rely on static knowledge point recommendations to achieve path delivery. Common approaches include using collaborative filtering, rule matching, or shallow semantic models to rank knowledge points. However, the knowledge graphs used in these methods only focus on the logical dependencies between knowledge points, neglecting the formation and evolution of ability elements (such as understanding, reasoning, and creativity) during the learning process, resulting in low accuracy in path delivery. Summary of the Invention

[0003] Based on this, this application provides a learning path recommendation method, apparatus, computer device, and storage medium.

[0004] Specifically, this application is implemented through the following technical solution: Firstly, embodiments of this application provide a learning path recommendation method, including: Obtain the constructed global knowledge graph, wherein the global knowledge graph includes multiple nodes, connection edges between different nodes, and weight of each connection edge, and the multiple nodes include capability nodes and knowledge nodes; Based on the current learning task determined by the target user, determine the local learning graph that matches the target user from the global knowledge graph; Based on the target user's historical learning behavior data, a current multimodal profile information matching the local learning graph is generated, wherein the current multimodal profile information is used to reflect the target user's current learning status with respect to the local learning graph; Based on the current multimodal profile information and the local learning graph, a determined path planning algorithm is used to generate the current learning path for the target user for the current learning task, and the current learning path is sent to the target user.

[0005] Optionally, the step of generating the current learning path for the target user for the current learning task based on the current multimodal profile information and the local learning graph using a determined path planning algorithm includes: Based on the current multimodal profile information and the structured attribute information of each node in the local learning graph, the correlation degree between the current multimodal profile information and the node is determined; When the relevance is greater than the set relevance threshold, the node is added to the path candidate set; Using the path planning algorithm and the multiple nodes included in the path candidate set, a path cost function is used to generate the current learning path for the target user for the current learning task. The path cost function includes a cumulative path cost and a learning state cost. The cumulative path cost is determined based on the weights of the connecting edges, and the learning state cost is determined based on the target user's phased learning state data. The learning state data includes one or more of the following state parameters: completion rate, error rate, learning duration, and interest change rate.

[0006] Optionally, the method further includes: During the process of the target user performing the current learning task according to the current learning path, the rate of change of the state parameters in the learning state data corresponding to the target user is determined; When the rate of change of the state parameters in the learning state data is detected to be greater than the set change threshold, a target knowledge node is determined to compensate the learning ability of the target user. Identify the remaining unlearned learning paths in the current learning path, insert the target knowledge node into the remaining learning paths, generate updated unlearned paths, and send the updated unlearned paths to the target user.

[0007] Optionally, the method further includes: When the target user completes the current learning task according to the current learning path, the target user's stage learning data is periodically acquired, and a capability performance vector reflecting the target user's execution ability is generated based on the target user's stage learning data. Based on the capability performance vector, the current multimodal profile information and the local learning graph of the target user are updated to generate updated multimodal profile information and updated learning graph; Based on the updated multimodal profile information and the updated learning graph, a new learning path is generated using the path planning algorithm, and the new learning path is sent to the target user.

[0008] Optionally, the capability performance vector includes the capability score of each capability node in the local learning graph. The step of updating the current multimodal profile information and the local learning graph of the target user based on the capability performance vector to generate updated multimodal profile information and an updated learning graph includes: The ability score of each ability node in the ability performance vector is determined as the score of the ability node in the local learning graph; For the target connection edge between the capability node and the knowledge node, the weight of the target connection edge is updated using the score of the capability node, the updated weight of the target connection edge is determined, and the weight of the target connection edge in the local learning graph is adjusted to the updated weight to generate an updated learning graph. Furthermore, the capability score vector in the current multimodal profile information of the target user is updated using the capability score corresponding to each capability node in the capability performance vector, thereby generating updated multimodal profile information.

[0009] Optionally, the weights of the connecting edges in the global knowledge graph are determined according to the following steps: For the first connection edge between two knowledge nodes, the weight information of the first connection edge is determined based on the knowledge level distance between the course structures corresponding to the two knowledge nodes and the error association rate between the two knowledge nodes, wherein the knowledge level distance is inversely proportional to the dependency relationship between the two knowledge nodes. For the second connection edge between two capability nodes, feature extraction is performed on the text information describing the capability elements of the capability node to generate feature data corresponding to the capability node; based on the feature data corresponding to the two capability nodes respectively, the semantic similarity between the two capability nodes is determined; based on the semantic similarity between the two capability nodes and the capability level distance, the weight information of the second connection edge is determined. For the third connection edge between the knowledge node and the capability node, semantic vector processing is performed on the structured attribute information of the capability node to generate a first semantic feature vector of the capability node; semantic vector processing is also performed on the structured attribute information of the knowledge node to generate a second semantic feature vector of the knowledge node; based on the first semantic feature vector of the capability node and the second semantic feature vector of the knowledge node, the semantic similarity between the capability node and the knowledge node is determined; and the weight information of the third connection edge is determined according to the semantic similarity and the co-occurrence frequency between the capability node and the knowledge node.

[0010] Optionally, the current multimodal profile information includes: knowledge mastery vector, ability score vector, learning preference set, and interest distribution vector; The knowledge mastery vector includes the mastery score corresponding to each knowledge node in the local learning graph; The capability score vector includes the capability score corresponding to each capability node in the local learning graph.

[0011] Optionally, after generating the current learning path, the method further includes: When it is determined, based on the current multimodal profile information, that the target user has a capability node that needs to be improved with a capability score lower than the threshold, the compensation knowledge node with the highest correlation to the capability node to be improved is determined from the local learning graph. When the current learning path does not include the compensating knowledge node, the compensating knowledge node is inserted into the current learning path to generate an updated current learning path; When the current learning path includes the compensating knowledge node, other knowledge nodes with the same learning theme as the compensating knowledge node are determined from the local learning graph, and the other knowledge nodes are inserted into the current learning path to generate an updated current learning path; The updated current learning path is sent to the target user.

[0012] Secondly, embodiments of this application also provide a learning path recommendation device, comprising: The acquisition module is used to acquire the constructed global knowledge graph, wherein the global knowledge graph includes multiple nodes, connection edges between different nodes, and weights of each connection edge, and the multiple nodes include capability nodes and knowledge nodes; The first determining module is used to determine the local learning graph that matches the target user from the global knowledge graph based on the current learning task determined by the target user. The generation module is used to generate current multimodal profile information that matches the local learning graph based on the target user's historical learning behavior data, wherein the current multimodal profile information is used to reflect the target user's current learning status in relation to the local learning graph; The second determining module is used to generate a current learning path for the target user for the current learning task based on the current multimodal profile information and the local learning graph using a determined path planning algorithm, and send the current learning path to the target user.

[0013] This application provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described learning path recommendation method.

[0014] This application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the above-described learning path recommendation method.

[0015] The method provided in this application constructs a global knowledge graph, which includes multiple nodes, connecting edges between different nodes, and the weight of each connecting edge. The multiple nodes include capability nodes and knowledge nodes, thus constructing a structured graph that reflects the knowledge logic and capability formation path, providing strong data support for subsequent path planning. After determining the local learning graph matching the target user from the global knowledge graph based on the target user's current learning task, the system uses the target user's historical learning behavior data to generate current multimodal profile information that matches the local learning graph. This current multimodal profile information reflects the target user's current learning status with respect to the local learning graph. Then, based on the current multimodal profile information and the local learning graph, a defined path planning algorithm is used to generate the target user's current learning path for the current learning task. This ensures that the current learning path not only matches the current learning task but also accurately recommends a learning path based on the target user's current learning status with respect to the local learning graph. This allows for adapting different learning paths for users with different abilities to the same learning task. After sending the current learning path to the target user, the target user can better complete the current learning task based on the current learning path, improving learning efficiency and effectiveness. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this specification and form part of this specification, illustrate exemplary embodiments and are used to explain this specification, but do not constitute an undue limitation thereof. In the drawings: Figure 1 This is a flowchart illustrating a learning path recommendation method in an exemplary embodiment of this application; Figure 2 This is a flowchart illustrating another learning path recommendation method in an exemplary embodiment of this application; Figure 3 This is a schematic diagram of a learning path recommendation device illustrated in an exemplary embodiment of this application; Figure 4 This is a schematic diagram of the structure of a computer device provided in this application. Detailed Implementation

[0017] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0018] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0019] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0020] Research has found that personalized teaching systems in related technologies mainly rely on static knowledge point recommendations to achieve path delivery. Common practices include using collaborative filtering, rule matching, or shallow semantic models to rank knowledge points. However, the knowledge graphs used in these methods only focus on the logical dependencies between knowledge points, neglecting the formation and evolution of ability elements (such as understanding, reasoning, and creativity) during the learning process, resulting in low accuracy in path delivery.

[0021] To alleviate the above problems, this application proposes a learning path recommendation method. By constructing a global knowledge graph, which includes multiple nodes, connecting edges between different nodes, and the weight of each connecting edge, the multiple nodes include capability nodes and knowledge nodes. This constructs a structured graph that reflects the knowledge logic and capability formation path, providing strong data support for subsequent path planning. After determining the local learning graph matching the target user from the global knowledge graph based on the target user's current learning task, the system uses the target user's historical learning behavior data to generate current multimodal profile information that matches the local learning graph. This current multimodal profile information reflects the target user's current learning status with respect to the local learning graph. Then, based on the current multimodal profile information and the local learning graph, a defined path planning algorithm is used to generate the target user's current learning path for the current learning task. This ensures that the current learning path not only matches the current learning task but also accurately recommends a learning path based on the target user's current learning status with respect to the local learning graph. This allows for adapting different learning paths for users with different abilities to the same learning task. After sending the current learning path to the target user, the target user can better complete the current learning task based on the current learning path, improving learning efficiency and effectiveness.

[0022] To facilitate understanding of this embodiment, a detailed description of the learning path recommendation method disclosed in this disclosure is provided first. The execution entity of the learning path recommendation method provided in this disclosure is generally a computer device with a certain computing capability. This computer device may include, for example, a terminal device, a server, or other processing devices. The terminal device may be a user equipment (UE), a mobile device, a user terminal, a personal digital assistant (PDA), a handheld device, a computing device, an embedded device, etc. In some possible implementations, this learning path recommendation method can be implemented by the processor calling computer-readable instructions stored in memory.

[0023] See Figure 1 The diagram shows a flowchart of a learning path recommendation method provided in this embodiment of the present disclosure. The method includes steps S101 to S104, wherein: S101. Obtain the constructed global knowledge graph, wherein the global knowledge graph includes multiple nodes, connection edges between different nodes, and weight of each connection edge, and the multiple nodes include capability nodes and knowledge nodes. S102. Based on the current learning task determined by the target user, determine the local learning graph matching the target user from the global knowledge graph; S103. Based on the target user's historical learning behavior data, generate current multimodal profile information that matches the local learning graph, wherein the current multimodal profile information is used to reflect the target user's current learning status in relation to the local learning graph; S104. Based on the current multimodal profile information and the local learning graph, a determined path planning algorithm is used to generate the current learning path for the target user for the current learning task, and the current learning path is sent to the target user.

[0024] Regarding S101: The global knowledge graph is a directed weighted graph G=(V,E,W) constructed based on knowledge points and required learning ability information for various learning tasks. The global knowledge graph includes multiple nodes V, with connecting edges E between different nodes, and each connecting edge has a corresponding weight W. The multiple nodes include ability nodes and knowledge nodes, i.e. A collection of knowledge nodes Let m and n be the set of capability nodes. Knowledge nodes are determined based on the knowledge points corresponding to the learning task, while capability nodes are capability-related nodes set according to actual needs.

[0025] In the global knowledge graph of this application, each node corresponds to a structured attribute information. The structured attribute information corresponding to a knowledge node includes, but is not limited to: name, definition text, knowledge module to which it belongs, difficulty level, list of prerequisite knowledge, and typical learning task scenarios. The definition text can be a text description added to the knowledge node. The knowledge module to which it belongs can be, for example, the curriculum system to which the knowledge node belongs. The difficulty level can be assessed based on the curriculum system or Bloom cognitive level. The list of prerequisite knowledge is determined according to the relationship between knowledge points. For example, before learning an economics case, one needs to learn the basics of economics. That is, learning the knowledge points of the economics case depends on learning the prerequisites of the basics of economics. Typical learning task scenarios can be, for example, the corresponding learning tasks.

[0026] The structured attribute information of a capability node includes, but is not limited to: capability type, capability level, capability performance indicators and observable behavior descriptions (i.e., capability element description text), typical learning task scenarios (such as corresponding learning tasks), and evaluation methods. The capability type can be set according to actual needs; for example, capability types include, but are not limited to, understanding, application, analysis, evaluation, and creation. The capability level can be between 1 and 5 based on a Bloom hierarchy mapping.

[0027] Performance metrics are used to quantify learners' achievement of abilities at different levels and are key parameters for system evaluation of ability status. For example, performance metrics for the comprehension level may include: concept recognition accuracy; for the application level, they may include: task completion rate, operational accuracy, etc.; for the analysis level, they may include: logical reasoning structure in reports, variable recognition rate, etc.; for the evaluation level, they may include: model evaluation consistency, self-reflection score, etc.; and for the creation level, they may include: original solutions, frequency of new model generation, etc.

[0028] Observable behavior descriptions refer to behavioral characteristics that the system can automatically identify through objective data such as learning logs, assignment texts, and test results, used to support the quantitative calculation of the aforementioned performance indicators. For example, observable behavior descriptions at the understanding level may include: correctly selecting definitions in quizzes and restating core concepts in summary texts; observable behavior descriptions at the application level may include: completing all steps in simulation experiments and submitting validated code or result files; observable behavior descriptions at the analysis level may include: comparing two schemes in a report and explaining the reasons for differences; observable behavior descriptions at the evaluation level may include: pointing out experimental limitations and providing a reasonable judgment in assignments; and observable behavior descriptions at the creation level may include: proposing improvement schemes in texts, generating new models or research hypotheses, etc.

[0029] This application extracts semantic features, performs key indicator statistics and weight calculations from the aforementioned observable behavioral data, and can subsequently update the score vector of capability nodes in real time, thereby achieving dynamic modeling of the capability formation and evolution process.

[0030] The global knowledge graph includes connection edges between different nodes, such as connection edges between knowledge nodes, connection edges between capability nodes, and connection edges between knowledge nodes and capability nodes. Each type of connection edge has a different weight calculation method so that dynamic updates and interpretable modeling can be achieved in the future.

[0031] The first connection edge between knowledge nodes can be determined based on the logical dependencies and prerequisite order between knowledge nodes. For example, if knowledge nodes... For knowledge points Given the prior knowledge, a directed edge e is established. , ).

[0032] In implementation, for the first connection edge between two knowledge nodes, the weight information of the first connection edge is determined based on the knowledge level distance between the course structures corresponding to the two knowledge nodes and the error association rate between the two knowledge nodes, wherein the knowledge level distance is inversely proportional to the dependency relationship between the two knowledge points.

[0033] The weight of the first connecting edge represents the dependency strength between different knowledge nodes, which can be determined by combining the hierarchical distance between course structures and the error association rate. For example, knowledge nodes can be determined according to the following formula (1). and knowledge nodes Weight information of the first connecting edge between :

[0034] in, For knowledge nodes With knowledge nodes The knowledge hierarchy distance in the course structure can range from [0,1]. It is inversely proportional to the dependency between two knowledge nodes; a smaller distance indicates a stronger dependency. The knowledge hierarchy distance between two knowledge nodes can be determined based on the course syllabus, knowledge structure tree, or the hierarchy and prerequisite relationships of knowledge points in the teaching knowledge graph. For example, it can be calculated using structured rules based on the above. The knowledge hierarchy distance between two knowledge nodes can be flexibly configured according to different teaching systems.

[0035] For knowledge nodes With knowledge nodes The error correlation rate between them indicates the learner's mastery of the knowledge points. Next, learn the knowledge points The error rate occurring at any given time ranges from [0,1]. The error correlation rate between two knowledge nodes can be obtained statistically from historical learning behavior data; for example, the error rate when a learner masters knowledge nodes... After the corresponding knowledge points, learn the knowledge nodes. The percentage of errors occurring when applying to the corresponding knowledge points is used to reflect the degree of dependence in the actual learning process.

[0036] , The weighting coefficient can be set based on prior experience; it is used to balance the contribution of knowledge hierarchy distance and error correlation rate, for example... It can be 0.6. It can be 0.4. For example, when the hierarchical distance between two knowledge nodes is close and the error association rate is low, A larger value indicates a closer relationship to prerequisites and a higher importance to the learning path.

[0037] For the second connecting edge between capability nodes in the global knowledge graph, the second connecting edge can represent the hierarchical progression or interdependence between capability nodes. Directed edges e can be established from low to high according to Bloom's cognitive hierarchy (e.g., understanding → application → analysis → evaluation → creation). , ).

[0038] Weight information of the second connecting edge The strength of capability progression is determined by a combination of semantic similarity and capability hierarchy distance.

[0039] In implementation, for the second connection edge between two capability nodes, feature extraction is performed on the text information describing the capability elements of the capability node to generate feature data corresponding to the capability node; based on the feature data corresponding to the two capability nodes respectively, the semantic similarity between the two capability nodes is determined; based on the semantic similarity between the two capability nodes and the capability level distance, the weight information of the second connection edge is determined.

[0040] Each capability node corresponds to a capability element description text (i.e., text information used to describe the capability elements of the capability node). Feature extraction is performed on the capability element description texts of the two capability nodes to obtain the feature data corresponding to each capability node. Based on the feature data corresponding to the two capability nodes, the semantic similarity between the two capability nodes is determined. For example, this semantic similarity can be calculated using a semantic embedding model. The weight information of the second connecting edge can then be calculated using the following formula (2). :

[0041] in, The semantic similarity of the text describing the capability element, with values ​​ranging from [0,1]; This represents the capability hierarchy distance between two capability nodes, with a value between 1 and 4. For example, it can be the Bloom hierarchy difference. This represents the maximum difference between the levels, for example, it can be 4, which is used for normalization. , These are weighting coefficients used to control the relative importance between semantic similarity and capability hierarchy distance. They can be set based on prior experience, for example... It can be 0.7. It can be 0.3.

[0042] When two capability nodes have similar semantics and are adjacent in level, weight information A higher value indicates a stronger correlation with the capability development path.

[0043] The third connection edge between knowledge nodes and capability nodes represents the supporting role of the knowledge node in capability formation and is a core structure of the global knowledge graph. In practice, the semantic similarity between knowledge nodes and capability nodes can be used to determine whether a third connection edge exists. If the semantic similarity between a knowledge node and a capability node is greater than a set similarity threshold, a third connection edge is established; otherwise, it is not established.

[0044] In practice, the first semantic feature vector of the capability node and the second semantic feature vector of the knowledge node can be determined separately, and the semantic similarity between the capability node and the knowledge node can be determined based on the first semantic feature vector and the second semantic feature vector.

[0045] For example, semantic vectorization is performed on the structured attribute information of capability nodes to generate the first semantic feature vector of capability nodes; and semantic vectorization is performed on the structured attribute information of knowledge nodes to generate the second semantic feature vector of knowledge nodes. For example, capability nodes are calculated according to the following formula (3). First semantic feature vector And calculate knowledge nodes according to the following formula (4). The second semantic feature vector : (3) (4) in This represents the semantic embedding layer.

[0046] In this application, to achieve a computable unified representation, semantic vectorization processing is performed on the textual information of the structured attribute information of each node. After the vector space is unified, knowledge nodes and capability nodes can perform similarity calculations in the same semantic space to support semantic computation and graph retrieval. In other words, a node is both a graph structural element and a carrier of semantic vectors, enabling knowledge and capability information to be computed in the same space.

[0047] The capability nodes are determined according to the following formula (5). and knowledge nodes Semantic similarity between them: (5) In implementation, for the third connection edge between the knowledge node and the capability node, semantic vector processing is performed on the structured attribute information of the capability node to generate a first semantic feature vector of the capability node; semantic vector processing is also performed on the structured attribute information of the knowledge node to generate a second semantic feature vector of the knowledge node; based on the first semantic feature vector of the capability node and the second semantic feature vector of the knowledge node, the semantic similarity between the capability node and the knowledge node is determined; and the weight information of the third connection edge is determined according to the semantic similarity and the co-occurrence frequency between the capability node and the knowledge node.

[0048] The co-occurrence frequency between capability nodes and knowledge nodes refers to the frequency at which capability nodes and knowledge nodes appear together in historical learning materials. It is obtained by statistical analysis of historical learning materials and takes a value between [0,1].

[0049] For example, the weight information of the third connecting edge can be determined according to the following formula (6). : (6) in, For semantic similarity, Represents capability nodes and knowledge nodes Co-occurrence frequency between them , These are weighting coefficients used to control the contribution ratio of semantic and empirical co-occurrence; they can be set as needed, for example... It can be 0.6. It can be 0.4.

[0050] When a knowledge node semantically strongly supports a certain ability (i.e., an ability node), and that ability is frequently triggered in actual teaching, This will increase significantly, reflecting its crucial role in the capacity-building process.

[0051] This application, by determining the weight information of the three types of connection edges mentioned above, can construct a multi-layered directed weighted graph structure, i.e., a global knowledge graph. This graph structure can express both the logical relationships of knowledge and the hierarchical relationships of capabilities within the same semantic space, and can also quantify the support strength between knowledge and capabilities, providing a computational foundation for subsequent graph evolution, adaptive learning path planning, and feedback optimization.

[0052] For S102 and S103: During implementation, the target user can determine the current learning task. Based on the current learning task, at least a portion of the knowledge graph is selected from the global knowledge graph as the local learning graph that matches the target user's current learning task.

[0053] After determining the local learning graph, based on the target user's historical learning behavior data, current multimodal profile information matching the local learning graph is generated. The current multimodal profile information is used to reflect the target user's current learning status with respect to the local learning graph. For example, the multimodal profile information may include, but is not limited to, the knowledge mastery vector corresponding to the knowledge node in the local learning graph, the ability score corresponding to the ability node, interest distribution, learning preferences, etc.

[0054] The learner profile of a target user, also known as a multimodal profile, can be generated through the fusion of multimodal data. This means that historical learning behavior data constitutes multimodal data, which can be any learning-related data of the target user. For example, historical learning behavior data may include, but is not limited to: user personal information (such as major, school, and grades), user learning logs, user behavior records, user assignment texts, and user search records. User learning logs may include, but are not limited to, learning duration (e.g., learning time per unit of time), access frequency (e.g., frequency of learning activities), and task error rate (e.g., the error rate of tasks corresponding to knowledge points). For instance, features can be extracted from historical learning behavior data, normalized, semantically vectorized, and then combined with time series modeling to form a unified representation, thus obtaining the current multimodal profile information of the target user.

[0055] Optionally, the current multimodal profile information includes: a knowledge mastery vector, an ability score vector, a learning preference set, and an interest distribution vector. The knowledge mastery vector includes the mastery score corresponding to each knowledge node in the local learning graph; the ability score vector includes the ability score corresponding to each ability node in the local learning graph.

[0056] For example, the current multimodal profile information can be U=(S, A, P, I), where S is the knowledge mastery vector, S=(s1, s2, ..., s...). mEach of the m knowledge nodes in the local learning graph contains a mastery score corresponding to each node; A is the ability score vector, A = (a1, a2, ..., a...). n Each of the n capability nodes in the local learning graph includes the capability score corresponding to each node. P represents the learning preference set, which can be determined by statistically analyzing the learning content of the target user; I represents the interest distribution vector, which can be determined by analyzing the target user's historical search records. The structure of the multimodal profile information in this application realizes a one-to-one mapping between the profile and the attributes of the learning graph nodes, enabling changes in the profile to directly drive the evolution of the graph. In other words, U is both a representation model of the learner's state and a driving signal for adaptive adjustment of the learning path.

[0057] Knowledge mastery vector, ability score vector, learning preference set, and interest distribution vector together constitute the multimodal profile information of the target user, enabling subsequent dynamic path planning and ability optimization. For example, for each knowledge node, data such as the target user's course grades, test results, or self-assessment questionnaire scores for that knowledge node can be extracted to determine the initial mastery score for that knowledge node. Using the set maximum and minimum mastery scores, this initial mastery score is normalized to obtain the corresponding mastery score for that knowledge node. The mastery scores corresponding to multiple knowledge nodes constitute the knowledge mastery vector, which quantifies the mastery level of each knowledge node.

[0058] For example, the original ability scores for each ability node of the target user can be obtained based on the target user's historical learning behavior data (such as learning duration and frequency) or historical task evaluation results (such as evaluations and scores of historical learning tasks). Based on these original ability scores, an ability score vector can be determined. Alternatively, semantic parsing can be performed on the text content of the target user's learning process, such as assignments, reports, or code text, and then a semantic vector v_d can be generated using language model embedding. The similarity between the semantic vector v_d and the feature data corresponding to the ability element description text of the ability node is calculated to determine the matching degree between the text content and each ability node. When the matching degree is greater than a preset threshold, it is determined that the text content reflects the ability performance of the corresponding ability node, and an ability increment value Δa is generated based on the matching degree. Then, the ability increment value Δa is fused and updated with the original ability scores obtained based on the target user's learning behavior data or task evaluation results to obtain the ability score of the ability node. The fusion method can be weighted average or exponential smoothing update, etc.

[0059] For example, a learning preference set P can be generated by analyzing data such as the target user's learning path selection, task type, operating habits, and learning time of the learning content. The preference intensity of tasks in the learning preference set can be determined by the task selection frequency. For example, the selection frequency and value of multiple types of tasks can be determined, and the preference intensity of each type of task can be obtained by dividing the selection frequency and value of each type of task by the selection frequency and value.

[0060] For example, a latent Dirichlet distribution can be applied to the target user's learning logs, assignment texts, and interaction data. Topic modeling to obtain the topic distribution for each document: ,in Indicates the first The probability distribution of documents across various topics, where K represents the number of specific topics.

[0061] In this application, multiple topics obtained through topic modeling correspond to different dimensions of user learning interest content. It can be seen that the topic distribution vector can be used to characterize the degree of attention of the target user on each interest topic. Therefore, by determining the topic modeling results, the interest distribution vector of the target user can be obtained.

[0062] According to the following formula (7), the document set D of the target user before time t is... t Calculate the time-weighted average of multiple documents within the current topic to obtain the current topic modeling result (i.e., the current interest distribution vector): (7) in For document timestamps, This is the time decay factor, used to control the document's impact; for example, its value can be between 0.05 and 0.2. This represents the time decay weight.

[0063] Considering that interests evolve and update over time, the current topic modeling results can be updated exponentially to obtain the target user's interest distribution vector. Specifically, the smoothing update is performed using the following formula (8): (8) in, Let be the interest distribution vector from the previous time step. This is a smoothing coefficient used to control the proportion of historical interests retained; for example, its value can be between 0.6 and 0.8.

[0064] For example, the rate of change of interest can be calculated based on the interest distribution vector at the current time and the interest distribution vector at the previous time. : .

[0065] If the rate of change of interest is determined in the subsequent process If the change exceeds the set threshold, it is determined that the target user's interest topic has shifted, and path reconstruction can be automatically triggered to update the learning path; alternatively, relevant knowledge nodes can be added to the learning path based on the shift status of the interest topic to achieve dynamic updates of the learning path.

[0066] In practical implementation, the multimodal profile information of the target user can be stored in a JSON structure to facilitate real-time system calls and updates. An example of its format is as follows: { "Knowledge Mastery": {"Atmospheric Chemistry": 0.85, "Numerical Simulation": 0.60}, "Ability": {"Experimental Operation": 0.40, "Modeling and Reasoning": 0.80}, "Preference": ["Experimental"], "Interest": {"Modeling": 0.70, "Climate": 0.55} }

[0067] Regarding S104: After determining the local learning graph and the current multimodal profile information, a defined path planning algorithm can be used to select nodes from the local learning graph, generate a current learning path corresponding to the target user and matching the current learning task, and send the current learning path to the target user for learning. For example, the path planning algorithm can select... algorithm.

[0068] In this application, to enable path planning to reflect the learner's real-time state, traditional methods can be used... The algorithm is improved by adding a dynamic heuristic driven by a user profile. Algorithms to leverage improved The algorithm generates the current learning path for the current learning task.

[0069] In one optional implementation, the step of generating the current learning path for the target user regarding the current learning task based on the current multimodal profile information and the local learning graph using a determined path planning algorithm includes: Step a1: Based on the current multimodal profile information and the structured attribute information of each node in the local learning graph, determine the correlation between the current multimodal profile information and the nodes; Step a2: When the relevance is greater than the set relevance threshold, add the node to the path candidate set; Step a3: Using the path planning algorithm and the multiple nodes included in the path candidate set, a path cost function is used to generate the current learning path for the target user for the current learning task. The path cost function includes a cumulative path cost and a learning state cost. The cumulative path cost is determined based on the weights of the connecting edges, and the learning state cost is determined based on the target user's phased learning state data. The learning state data includes one or more of the following state parameters: completion rate, error rate, learning duration, and interest change rate.

[0070] In practice, we can first extract the structured attribute information of each node in the local learning graph to determine the feature vector of the node; then, using the current multimodal profile information and the feature vector of the node, we can calculate the correlation between the current multimodal profile information of the target user and the node.

[0071] If the correlation degree is greater than the set correlation degree threshold, it means that the node matches the current multimodal profile, and the node is added to the path candidate set; otherwise, if the correlation degree is less than or equal to the set correlation degree threshold, it means that the node does not match the current multimodal profile, and no further processing is performed.

[0072] The path cost function of this application includes a cumulative path cost and a learning state cost. The cumulative path cost is determined based on the weights of the connecting edges, while the learning state cost is a dynamic heuristic term set in this application. It is determined based on the target user's phased learning state data and can be dynamically adjusted as the target user profile changes in real time, allowing path planning to converge with the learning state and thus achieving continuous adaptive optimization. The learning state data includes one or more of the following: completion rate, error rate, learning duration, and interest change rate.

[0073] Specifically, the completion rate (cr) represents the degree of completion of learning tasks within the current learning stage, and it is defined as follows: , This represents a change in knowledge mastery over a period of time. For time intervals.

[0074] Error rate er represents the percentage of errors in the current learning phase, defined as follows: , The score represents the actual ability to complete the current learning task. This represents the historical average ability score.

[0075] For example, the actual ability score of the current learning task The historical average ability score can be calculated based on data such as the target user's test scores, assignment evaluation results, experiment operation logs, or text / code semantic analysis results during the current learning stage, using preset ability evaluation mapping rules or models. Statistics can be obtained by taking the average or weighted average of the ability evaluation results corresponding to multiple historical learning stages.

[0076] The learning duration t is the cumulative learning time within the current learning phase, used to reflect the learning engagement of the target user.

[0077] Rate of change of interest Used to measure how a target user's interests change over time. For example, for an interest distribution vector... The overall rate of change is calculated using the following formula (9): (9) By calculating the average absolute rate of change of interest scores across multiple topic dimensions in the interest distribution vector, the trend of interest migration can be quantified.

[0078] For example, when the learning state data includes completion rate, error rate, and interest change rate, the learning state cost is as shown in the following formula (10): (10) in, , , These are weighting coefficients used to control the relative influence of each indicator in the path cost. Their values ​​can be set according to actual needs, for example... It can be 0.4. It can be 0.4. It can be 0.2.

[0079] By setting a heuristic function, i.e., the cost of the learning state, which is updated in real time with the learning state, the search process can prioritize optimizing nodes with "low completion rate, high error rate, and declining interest," thereby achieving dynamic correction of the learning path.

[0080] When determining the current learning path for the first time, if the target user does not have no node-specific learning state data, the learning state cost can be initialized based on preset values ​​or the target user's user information.

[0081] After determining the path cost function, a path planning algorithm is used to select the target node from multiple nodes included in the path candidate set, and generate the current learning path for the target user for the current learning task.

[0082] Through the above process, the current learning path can be correlated with the target user's stage of learning, thereby improving the accuracy of the current learning path and ensuring user experience and learning effectiveness.

[0083] In one optional implementation, the method further includes: Step b1: During the process of the target user performing the current learning task according to the current learning path, determine the rate of change of the state parameters in the learning state data corresponding to the target user; Step b2: When the rate of change of the state parameters in the learning state data is detected to be greater than the set change threshold, the target knowledge node used to compensate the target user's learning ability is determined. Step b3: Determine the remaining unlearned learning paths in the current learning path, insert the target knowledge node into the remaining learning paths, generate updated unlearned paths, and send the updated unlearned paths to the target user.

[0084] As the target user completes the current learning task according to the current learning path, the rate of change of status parameters in the target user's learning status data can be determined in real time. Status parameters include completion rate, error rate, learning time, and rate of change of interest.

[0085] When there are multiple state parameters, if the rate of change of any state parameter is detected to be greater than the change threshold corresponding to that state parameter, a target node is determined to compensate the learning ability of the target user. The change thresholds are different for different state parameters.

[0086] For example, when the learning status meets any of the following conditions: completion rate is less than 0.7, error rate is greater than 0.3, and interest change rate is less than -0.2, the system can automatically trigger a path reconstruction strategy, such as inserting a supplementary node between the current node and the next node in the current learning path to achieve path optimization.

[0087] For example, capability nodes with capability scores below a threshold can be selected from multimodal profile information as compensation targets. Then, knowledge nodes with a connectivity degree greater than a preset value (e.g., connectivity degree determined by the weight of the connecting edges) to the capability node can be retrieved from the local learning graph to form a knowledge node set. Target knowledge nodes can be selected from this knowledge node set, such as the knowledge node with the highest connectivity degree. Alternatively, when a high error rate is detected, knowledge nodes that reduce task difficulty can be used as target knowledge nodes, such as the low-level knowledge nodes corresponding to the currently learned knowledge nodes.

[0088] After identifying the target knowledge node, it can be inserted into the remaining unlearned learning paths. If the generated current learning path is... After determining the current learning node When that happens, the remaining learning path is Then, the target knowledge node can be determined. Next, the target knowledge node is inserted after the current learning node, generating an updated unlearned path, i.e. And send the updated, unlearned paths to the target users.

[0089] In the above process, by monitoring the changes in the target user's learning status data in real time, the remaining learning path of the target user is adjusted in real time according to the changes, rather than global rollback or random replanning, which improves the accuracy of the learning path recommendation and ensures the learning effect of the target user.

[0090] In one optional implementation, after generating the current learning path, the method further includes: when, based on the current multimodal profile information, it is determined that the target user has a capability node with a capability score lower than a threshold that needs improvement, determining from the local learning graph the compensation knowledge node with the highest correlation to the capability node to be improved. If the current learning path does not include the compensation knowledge node, inserting the compensation knowledge node into the current learning path to generate an updated current learning path; if the current learning path includes the compensation knowledge node, determining from the local learning graph other knowledge nodes with the same learning theme as the compensation knowledge node, and inserting the other knowledge nodes into the current learning path to generate an updated current learning path. The updated current learning path is then sent to the target user.

[0091] After generating the current learning path, it is determined whether there are any capability nodes in the current multimodal profile information whose capability scores are below a threshold and need to be improved. For example, if any capability score is less than 0.5, then the capability node corresponding to that score is a capability node that needs to be improved, and a supplementary task to compensate for that capability can be inserted into the current learning path.

[0092] During implementation, the compensating knowledge node with the highest correlation to the node whose capability needs to be improved can be identified from the local learning graph. If the current learning path does not include the compensating knowledge node, it is inserted into the current learning path to generate an updated current learning path. If the current learning path already includes the compensating knowledge node, other knowledge nodes with the same learning theme as the compensating knowledge node are identified from the local learning graph, and these other knowledge nodes are inserted into the current learning path to generate an updated current learning path. If no other knowledge nodes exist, a substitute node under the shortest dependency link can be inserted.

[0093] The updated learning path is then sent to the target user so that the target user can complete the current learning task according to the updated learning path, achieve the synergistic development of knowledge and ability, and thus realize the adaptive adjustment of the learning path and maximize learning efficiency.

[0094] In one optional implementation, the method further includes: Step c1: When the target user completes the current learning task according to the current learning path, periodically acquire the target user's stage learning data, and generate a capability performance vector that reflects the target user's execution ability based on the target user's stage learning data; Step c2: Based on the capability performance vector, update the current multimodal profile information and the local learning graph of the target user to generate updated multimodal profile information and updated learning graph; Step c3: Based on the updated multimodal profile information and the updated learning graph, generate a new learning path using the path planning algorithm, and send the new learning path to the target user.

[0095] During implementation, as the target user completes the current learning task according to the current learning path (or the updated current learning path), ability evaluation can be continuously performed throughout the learning process, generating an ability performance vector, such as E = (e1, e2, ..., e...). n ), where e j Represents the capability node c j The ability scores for each stage are calculated, and the ability nodes in the ability performance vector can be relevant ability nodes of the current learning path. For example, the ability nodes in the ability performance vector can be determined based on the knowledge nodes corresponding to the current learning task, through the relationship between knowledge nodes and ability nodes in the learning graph. The ability performance reflected by quantitative or unstructured data is used to update the energy scores of the aforementioned relevant ability nodes.

[0096] It can periodically acquire stage learning data from target users. This stage learning data includes, but is not limited to, quantitative data such as test scores and task completion rates, and unstructured data such as lab reports, text assignments, and operation logs. During implementation, the quantitative data can be normalized to generate capability scores for relevant capability nodes. Alternatively, a language model can be used to perform semantic parsing on the unstructured data, extract capability-related features, and then quantify the extracted features into scoring items using a mapping function to obtain capability scores for capability nodes.

[0097] By determining the capability performance vector, we can not only correct the capability scores of capability nodes in the target user profile, but also use the capability performance vector as a driving signal to drive learning graph updates and path optimization.

[0098] After determining the capability performance vector, the current multimodal profile information and local learning graph of the target user are updated based on the capability performance vector, generating updated multimodal profile information and updated learning graph. For example, the capability score vector in the current multimodal profile information can be updated, and the weight information of the connecting edges related to capability nodes in the local learning graph can be updated.

[0099] The capability performance vector includes the capability score of each capability node in the local learning graph.

[0100] Optionally, updating the current multimodal profile information and the local learning graph of the target user based on the capability performance vector to generate updated multimodal profile information and updated learning graph includes: determining the capability score of each capability node in the capability performance vector as the score of the capability node in the local learning graph; updating the weight of the target connection edge between the capability node and the knowledge node using the score of the capability node to determine the updated weight of the target connection edge, and adjusting the weight of the target connection edge in the local learning graph to the updated weight to generate the updated learning graph; and updating the capability score vector in the current multimodal profile information of the target user using the capability score corresponding to each capability node in the capability performance vector to generate updated multimodal profile information.

[0101] In implementation, the capability score of each capability node in the capability performance vector is determined as the score of that capability node in the local learning graph. For the target connection edge between the capability node and the knowledge node, the weight of the target connection edge is updated using the capability node's score according to the following formula (11), thus determining the updated weight of the target connection edge: (11) in, For knowledge node k i With capability node c j The updated weights of the target connection edges between them. e represents the weights before the update. j For capability node c j The score, This is the fusion coefficient, used to control the weighting of historical structure and latest performance; for example, its value can be between 0.6 and 0.8.

[0102] The aforementioned update mechanism achieves positive feedback for capability improvement. That is, when capability performance improves, the strength of the connection edge between related knowledge and capability is automatically strengthened; when capability declines, the weight of the related connection edge is weakened.

[0103] After obtaining the updated weights of the target connection edges, the weights of the target connection edges in the local learning graph can be adjusted to the updated weights to generate an updated learning graph. Furthermore, the capability score vector in the current multimodal profile information of the target user can be updated using the capability score corresponding to each capability node in the capability performance vector, generating updated multimodal profile information.

[0104] After obtaining the updated multimodal profile information and the updated learning graph, a path planning algorithm is used to re-execute path search and sorting to generate new learning paths, which are then sent to the target user. The path planning process of the algorithm can be referred to the previous explanation and will not be repeated here.

[0105] The above process enables phased global calibration of the learning path. After the learning cycle ends or a phased task is completed, the execution system can modify the graph structure and user profile based on the ability evaluation results so as to generate new learning paths in the future.

[0106] In practical implementation, during the learning execution phase, the rate of change of state parameters in the learning state data corresponding to the target user can be monitored in real time to fine-tune the current learning path. After the accumulated change in learning exceeds a threshold or a phase task is completed, the target user's ability performance vector can be periodically determined to trigger closed-loop feedback and generate a new learning path, such as... This forms an adaptive loop. A specific periodic capability representation vector E is determined, and the learning graph G is updated using E to obtain the updated learning graph G′. The user profile U is also updated to obtain the updated user profile U′ (i.e., the updated multimodal profile information). Path planning is then performed using the updated G′ and U′ to obtain a new learning path P′, and learning continues.

[0107] The closed-loop mechanism of "evaluation-evolution-planning-re-evaluation" described in this application enables the execution system to continuously optimize the knowledge structure and teaching path, achieving dynamic and self-learning personalized learning path recommendations. Therefore, the closed-loop feedback mechanism of this application possesses the following technical advantages: It achieves dynamic self-correction, meaning that changes in ability are immediately reflected in the graph structure, enhancing adaptability; it achieves local-global collaboration, meaning that real-time evolution and phased calibration work together to maintain system stability and sensitivity; it has strong interpretability, meaning that changes in edge weights and node updates can be traced back to specific ability evaluation results, facilitating teaching intervention; and it achieves closed-loop learning optimization, meaning that a complete feedback chain is formed from learning performance to graph updates and then to path adjustments, enabling continuously optimized learning decisions.

[0108] See Figure 2 The method of this application is described in detail below. It includes the following steps: S201. Construct a knowledge-capability coupling graph G(V,E,W).

[0109] S202, Graph Evolution and Weight Update.

[0110] This yields a local learning map.

[0111] S203. Generate a learner multimodal profile U=[S,A,P,I].

[0112] S204, Based on improvements The algorithm generates the learning path P.

[0113] S205. Learning status monitoring.

[0114] S206. Should path reconstruction be triggered? If yes, then execute S207; otherwise, execute S209.

[0115] S207, Local backtrack and insert transition node.

[0116] This refers to the process of updating the remaining learning path and inserting compensating knowledge nodes.

[0117] S208, Perform supplementary tasks and enhance capabilities.

[0118] S209, Competency assessment and closed-loop feedback, resulting in updated G', U', and P'.

[0119] The above process achieves a closed-loop mechanism for adaptive adjustment and path reconstruction of the learning path.

[0120] This application's method centers on knowledge points, ability elements, and learner profiles. By constructing a bidirectional coupled graph (G) between knowledge points and ability elements and a multimodal learner profile (U), it forms a closed-loop mechanism of "evaluation—graph evolution—path reconstruction," enabling dynamic planning and real-time intervention of learning paths. Unlike related technologies based on static profiles or knowledge point recommendations, this application introduces ability dimension modeling and weight evolution feedback mechanisms at the graph layer and establishes a time-series profile update model at the learner layer, allowing knowledge learning and ability development to evolve co-evolve. This method can be widely applied in online education, vocational training, intelligent course recommendation, experimental teaching, and other scenarios, providing data-driven intelligent support for personalized teaching.

[0121] In other words, the core objective of this application is to enable the teaching system to have self-learning and self-optimization capabilities, so that the system can automatically adjust the knowledge structure and task sequence according to the learner's real-time performance, so that the learning path always matches the learner's ability development, thereby realizing a dynamic teaching closed loop of "learning and adjusting at the same time".

[0122] The above-described method of this application will be illustrated by specific embodiments below.

[0123] Example 1: Personalized Path Reconstruction and Competency Feedback Optimization for Graduate Students. Assume that a graduate student in environmental science aims to master interdisciplinary knowledge of "air pollution control" and improve their experimental operation and modeling skills.

[0124] (1) Initial state. The learner profile is initialized as follows: ; The initial learning path is defined as follows: .

[0125] (2) Path reconstruction trigger.

[0126] During the learning phase, if the error rate er = 0.32 > 0.3, and the capability node corresponding to the experimental operation is... Score a j If 0.40 < 0.5, then path reconstruction is triggered to adjust the learning paths that have not yet been learned.

[0127] For example, it can be based on the correlation weights w between knowledge and ability in the learning graph. ij Automatically identify capability nodes corresponding to experimental operations The knowledge node with the highest relevance, such as "Fundamentals of Numerical Analysis," is used as a capability compensation node. This capability compensation node is then inserted between the current node and the target node to generate an updated learning path. .

[0128] (3) Capability evaluation and edge weight update.

[0129] Generate phased competency assessment results, such as e j =0.55. The system updates the edge weights between knowledge nodes and ability nodes according to the edge weight update formula, and adds the experiment report writing node as an ability enhancement task. The edge weight update formula is as follows:

[0130] (4) Closed-loop feedback execution.

[0131] Evaluation results updated to capability nodes: .

[0132] Redefining the learner profile and perform path replanning. This generates new learning paths.

[0133] (5) Results analysis.

[0134] This process embodies a complete closed-loop mechanism from error detection to path reconstruction, capability evaluation, and graph update. The system achieves dynamic path optimization and adaptive improvement of capability nodes. Testing showed that the operational capability improved from 0.40 to 0.55, with a learning efficiency increase of approximately 20%.

[0135] Example 2: Cross-domain migration and interest-driven optimization.

[0136] Suppose a computer science undergraduate student plans to enter the field of environmental economics; their initial profile would be as follows: .

[0137] Initial learning path: .

[0138] (1) Monitoring changes in interest.

[0139] The system detects a 25% increase in the weight of the interest topic "Data Modeling and Prediction" through topic modeling. It calculates the interest change rate Δi = 0.25 > 0.20 (the set threshold), which triggers interest-driven path reconstruction.

[0140] (2) Interest-driven path update.

[0141] The system retrieves the knowledge nodes with the highest relevance to the topic, such as "carbon emission factor database modeling," from the learning graph and inserts paths based on interest-driven changes to generate an updated learning path. .

[0142] (3) Capability enhancement and edge weight adjustment.

[0143] Post-learning transferability nodes The score was raised to a j =0.72, the system automatically updates the edge weights between relevant knowledge nodes and ability nodes to achieve interest-driven dynamic adaptive learning.

[0144] (4) System closed loop manifestation.

[0145] The learning graph is updated as abilities change, and the system completes this process sequentially. This closed-loop process ensures that knowledge acquisition and interest transfer are aligned.

[0146] (5) Results analysis.

[0147] This embodiment demonstrates the system's interest perception and path self-adjustment capabilities. When learning interests change significantly, the system can automatically adjust the learning content and sequence to ensure that the learning path always matches the learner's interest development, thereby achieving dynamic optimization of highly transferable learning goals.

[0148] Therefore, the method of this application can perceive changes in learning status in real time and automatically optimize the path during the learning process, realizing dynamic path planning based on profiles, and generating personalized paths for different learner states; through multi-source data fusion and interest modeling, interest-driven path expansion and task reconstruction are realized; and a self-learning optimization loop is formed through a closed-loop feedback mechanism.

[0149] Corresponding to the embodiments of the aforementioned learning path recommendation method, this application also provides embodiments of a learning path recommendation device. Figure 3 A schematic diagram of the learning path recommendation device provided in this application specifically includes: The acquisition module 301 is used to acquire the constructed global knowledge graph, wherein the global knowledge graph includes multiple nodes, connection edges between different nodes, and weights of each connection edge, and the multiple nodes include capability nodes and knowledge nodes. The first determining module 302 is used to determine the local learning graph matching the target user from the global knowledge graph based on the current learning task determined by the target user. The generation module 303 is used to generate current multimodal profile information that matches the local learning graph based on the target user's historical learning behavior data, wherein the current multimodal profile information is used to reflect the target user's current learning status in relation to the local learning graph; The second determining module 304 is used to generate a current learning path for the target user for the current learning task based on the current multimodal profile information and the local learning map using a determined path planning algorithm, and send the current learning path to the target user.

[0150] Optionally, the second determining module 304, when generating the current learning path for the target user regarding the current learning task based on the current multimodal profile information and the local learning graph using a determined path planning algorithm, is used to: Based on the current multimodal profile information and the structured attribute information of each node in the local learning graph, the correlation degree between the current multimodal profile information and the node is determined; When the relevance is greater than the set relevance threshold, the node is added to the path candidate set; Using the path planning algorithm and the multiple nodes included in the path candidate set, a path cost function is used to generate the current learning path for the target user for the current learning task. The path cost function includes a cumulative path cost and a learning state cost. The cumulative path cost is determined based on the weights of the connecting edges, and the learning state cost is determined based on the target user's phased learning state data. The learning state data includes one or more of the following state parameters: completion rate, error rate, learning duration, and interest change rate.

[0151] Optionally, the device further includes: a reconfiguration module 305, used for: During the process of the target user performing the current learning task according to the current learning path, the rate of change of the state parameters in the learning state data corresponding to the target user is determined; When the rate of change of the state parameters in the learning state data is detected to be greater than the set change threshold, a target knowledge node is determined to compensate the learning ability of the target user. Identify the remaining unlearned learning paths in the current learning path, insert the target knowledge node into the remaining learning paths, generate updated unlearned paths, and send the updated unlearned paths to the target user.

[0152] Optionally, the device further includes: a feedback update module 306, used for: When the target user completes the current learning task according to the current learning path, the target user's stage learning data is periodically acquired, and a capability performance vector reflecting the target user's execution ability is generated based on the target user's stage learning data. Based on the capability performance vector, the current multimodal profile information and the local learning graph of the target user are updated to generate updated multimodal profile information and updated learning graph; Based on the updated multimodal profile information and the updated learning graph, a new learning path is generated using the path planning algorithm, and the new learning path is sent to the target user.

[0153] Optionally, the capability performance vector includes the capability score of each capability node in the local learning graph. The feedback update module 306, when updating the current multimodal profile information and the local learning graph of the target user based on the capability performance vector to generate updated multimodal profile information and an updated learning graph, is used for: The ability score of each ability node in the ability performance vector is determined as the score of the ability node in the local learning graph; For the target connection edge between the capability node and the knowledge node, the weight of the target connection edge is updated using the score of the capability node, the updated weight of the target connection edge is determined, and the weight of the target connection edge in the local learning graph is adjusted to the updated weight to generate an updated learning graph. Furthermore, the capability score vector in the current multimodal profile information of the target user is updated using the capability score corresponding to each capability node in the capability performance vector, thereby generating updated multimodal profile information.

[0154] Optionally, the acquisition module 301 is used to determine the weights of the connecting edges in the global knowledge graph according to the following steps: For the first connection edge between two knowledge nodes, the weight information of the first connection edge is determined based on the knowledge level distance between the course structures corresponding to the two knowledge nodes and the error association rate between the two knowledge nodes, wherein the knowledge level distance is inversely proportional to the dependency relationship between the two knowledge nodes. For the second connection edge between two capability nodes, feature extraction is performed on the text information describing the capability elements of the capability node to generate feature data corresponding to the capability node; based on the feature data corresponding to the two capability nodes respectively, the semantic similarity between the two capability nodes is determined; based on the semantic similarity between the two capability nodes and the capability level distance, the weight information of the second connection edge is determined. For the third connection edge between the knowledge node and the capability node, semantic vector processing is performed on the structured attribute information of the capability node to generate a first semantic feature vector of the capability node; semantic vector processing is also performed on the structured attribute information of the knowledge node to generate a second semantic feature vector of the knowledge node; based on the first semantic feature vector of the capability node and the second semantic feature vector of the knowledge node, the semantic similarity between the capability node and the knowledge node is determined; and the weight information of the third connection edge is determined according to the semantic similarity and the co-occurrence frequency between the capability node and the knowledge node.

[0155] Optionally, the current multimodal profile information includes: knowledge mastery vector, ability score vector, learning preference set, and interest distribution vector; The knowledge mastery vector includes the mastery score corresponding to each knowledge node in the local learning graph; The capability score vector includes the capability score corresponding to each capability node in the local learning graph.

[0156] Optionally, after generating the current learning path, the reconstruction module 305 is further configured to: When it is determined, based on the current multimodal profile information, that the target user has a capability node that needs to be improved with a capability score lower than the threshold, the compensation knowledge node with the highest correlation to the capability node to be improved is determined from the local learning graph. When the current learning path does not include the compensating knowledge node, the compensating knowledge node is inserted into the current learning path to generate an updated current learning path; When the current learning path includes the compensating knowledge node, other knowledge nodes with the same learning theme as the compensating knowledge node are determined from the local learning graph, and the other knowledge nodes are inserted into the current learning path to generate an updated current learning path; The updated current learning path is sent to the target user.

[0157] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0158] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this application according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0159] This application also provides a computer-readable storage medium storing a computer program that can be used to execute the learning path recommendation method described in the above embodiments.

[0160] This application also provides a computer device, see [link to relevant documentation] Figure 4 The diagram shown illustrates the structure of the computer device provided in this application. At the hardware level, the computer device includes a processor, an internal bus, a network interface, memory, and non-volatile memory, and may also include other hardware required for various business operations. The processor reads the corresponding computer program from the non-volatile memory into memory and then runs it to implement the learning path recommendation method described in the above embodiments. Of course, besides software implementation, this specification does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. That is to say, the execution entity of the following processing flow is not limited to individual logic units, but can also be hardware or logic devices.

[0161] All user-related data involved in the technical solution of this application are obtained with the user's authorization, such as through pop-up windows or by inviting individuals to upload their own data.

[0162] Those skilled in the art will understand that embodiments of this application can be provided as methods, apparatus, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0163] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0164] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0165] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the functions specified in one or more boxes. In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0166] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0167] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0168] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0169] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0170] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0171] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A learning path recommendation method, characterized in that, The method includes: Obtain the constructed global knowledge graph, wherein the global knowledge graph includes multiple nodes, connecting edges between different nodes, and weights of each connecting edge. The multiple nodes include capability nodes and knowledge nodes. The capability nodes are capability-related nodes set according to actual needs, and the knowledge nodes are nodes determined according to the knowledge points corresponding to the learning task. Based on the current learning task determined by the target user, determine the local learning graph that matches the target user from the global knowledge graph; Based on the target user's historical learning behavior data, a current multimodal profile information matching the local learning graph is generated, wherein the current multimodal profile information is used to reflect the target user's current learning status with respect to the local learning graph; Based on the current multimodal profile information and the local learning graph, a determined path planning algorithm is used to generate the current learning path for the target user regarding the current learning task, and the current learning path is sent to the target user; wherein, generating the current learning path for the target user regarding the current learning task based on the current multimodal profile information and the local learning graph using a determined path planning algorithm includes: Based on the current multimodal profile information and the structured attribute information of each node in the local learning graph, the correlation degree between the current multimodal profile information and the node is determined; When the relevance is greater than the set relevance threshold, the node is added to the path candidate set; Using the path planning algorithm and the multiple nodes included in the path candidate set, a path cost function is used to generate the current learning path for the target user for the current learning task. The path cost function includes a cumulative path cost and a learning state cost. The cumulative path cost is determined based on the weights of the connecting edges, and the learning state cost is determined based on the target user's phased learning state data. The learning state data includes one or more of the following state parameters: completion rate, error rate, learning duration, and interest change rate.

2. The method according to claim 1, characterized in that, The method further includes: During the process of the target user performing the current learning task according to the current learning path, the rate of change of the state parameters in the learning state data corresponding to the target user is determined; When the rate of change of the state parameters in the learning state data is detected to be greater than the set change threshold, a target knowledge node is determined to compensate the learning ability of the target user. Identify the remaining unlearned learning paths in the current learning path, insert the target knowledge node into the remaining learning paths, generate updated unlearned paths, and send the updated unlearned paths to the target user.

3. The method according to claim 1, characterized in that, The method further includes: When the target user completes the current learning task according to the current learning path, the target user's stage learning data is periodically acquired, and a capability performance vector reflecting the target user's execution ability is generated based on the target user's stage learning data. Based on the capability performance vector, the current multimodal profile information and the local learning graph of the target user are updated to generate updated multimodal profile information and updated learning graph; Based on the updated multimodal profile information and the updated learning graph, a new learning path is generated using the path planning algorithm, and the new learning path is sent to the target user.

4. The method according to claim 3, characterized in that, The capability performance vector includes the capability score of each capability node in the local learning graph. The step of updating the current multimodal profile information and the local learning graph of the target user based on the capability performance vector to generate updated multimodal profile information and an updated learning graph includes: The ability score of each ability node in the ability performance vector is determined as the score of the ability node in the local learning graph; For the target connection edge between the capability node and the knowledge node, the weight of the target connection edge is updated using the score of the capability node, the updated weight of the target connection edge is determined, and the weight of the target connection edge in the local learning graph is adjusted to the updated weight to generate an updated learning graph. Furthermore, the capability score vector in the current multimodal profile information of the target user is updated using the capability score corresponding to each capability node in the capability performance vector, thereby generating updated multimodal profile information.

5. The method according to claim 1, characterized in that, The weights of the connecting edges in the global knowledge graph are determined according to the following steps: For the first connection edge between two knowledge nodes, the weight information of the first connection edge is determined based on the knowledge level distance between the course structures corresponding to the two knowledge nodes and the error association rate between the two knowledge nodes, wherein the knowledge level distance is inversely proportional to the dependency relationship between the two knowledge nodes. For the second connection edge between two capability nodes, feature extraction is performed on the text information used to describe the capability elements of the capability node to generate feature data corresponding to the capability node; The semantic similarity between the two capability nodes is determined based on the feature data corresponding to the two capability nodes respectively. The weight information of the second connection edge is determined based on the semantic similarity and capability hierarchy distance between the two capability nodes. For the third connection edge between the knowledge node and the capability node, semantic vector processing is performed on the structured attribute information of the capability node to generate a first semantic feature vector of the capability node; semantic vector processing is also performed on the structured attribute information of the knowledge node to generate a second semantic feature vector of the knowledge node; based on the first semantic feature vector of the capability node and the second semantic feature vector of the knowledge node, the semantic similarity between the capability node and the knowledge node is determined; and the weight information of the third connection edge is determined according to the semantic similarity and the co-occurrence frequency between the capability node and the knowledge node.

6. The method according to claim 1, characterized in that, The current multimodal profile information includes: knowledge mastery vector, ability score vector, learning preference set, and interest distribution vector; The knowledge mastery vector includes the mastery score corresponding to each knowledge node in the local learning graph; The capability score vector includes the capability score corresponding to each capability node in the local learning graph.

7. The method according to claim 6, characterized in that, After generating the current learning path, the method further includes: When it is determined, based on the current multimodal profile information, that the target user has a capability node that needs to be improved with a capability score lower than the threshold, the compensation knowledge node with the highest correlation to the capability node to be improved is determined from the local learning graph. When the current learning path does not include the compensating knowledge node, the compensating knowledge node is inserted into the current learning path to generate an updated current learning path; When the current learning path includes the compensating knowledge node, other knowledge nodes with the same learning theme as the compensating knowledge node are determined from the local learning graph, and the other knowledge nodes are inserted into the current learning path to generate an updated current learning path; The updated current learning path is sent to the target user.

8. A learning path recommendation device, characterized in that, The device includes: The acquisition module is used to acquire the constructed global knowledge graph, wherein the global knowledge graph includes multiple nodes, connecting edges between different nodes, and the weight of each connecting edge. The multiple nodes include capability nodes and knowledge nodes. The capability nodes are capability-related nodes set according to actual needs, and the knowledge nodes are nodes determined according to the knowledge points corresponding to the learning task. The first determining module is used to determine the local learning graph that matches the target user from the global knowledge graph based on the current learning task determined by the target user. The generation module is used to generate current multimodal profile information that matches the local learning graph based on the target user's historical learning behavior data, wherein the current multimodal profile information is used to reflect the target user's current learning status in relation to the local learning graph; The second determining module is configured to generate a current learning path for the target user on the current learning task based on the current multimodal profile information and the local learning graph, using a determined path planning algorithm, and send the current learning path to the target user; the second determining module, when generating the current learning path for the target user on the current learning task based on the current multimodal profile information and the local learning graph, is configured to: Based on the current multimodal profile information and the structured attribute information of each node in the local learning graph, the correlation degree between the current multimodal profile information and the node is determined; When the relevance is greater than the set relevance threshold, the node is added to the path candidate set; Using the path planning algorithm and the multiple nodes included in the path candidate set, a path cost function is used to generate the current learning path for the target user for the current learning task. The path cost function includes a cumulative path cost and a learning state cost. The cumulative path cost is determined based on the weights of the connecting edges, and the learning state cost is determined based on the target user's phased learning state data. The learning state data includes one or more of the following state parameters: completion rate, error rate, learning duration, and interest change rate.

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

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, The processor performs the steps of the method according to any one of claims 1-7.

Images