A scene self-adaption-based course intelligent recommendation method and device and storage medium

Abstract

The application provides a scene self-adaptive course intelligent recommendation method, which comprises the following steps: constructing a four-dimensional knowledge graph of "course-knowledge point-ability target-application scene", generating a multi-modal scene data matrix based on obtained user physical environment data, obtaining a current application scene of a user, using a Transform encoder to encode the current application scene of the user to obtain a current application scene semantic vector, and using a Transform decoder to determine an application scene intention of the user based on the current application scene semantic vector; using an intelligent self-adaptive recommendation model to recommend a first course set based on the application scene intention; obtaining a second course set based on the current scene semantic vector, fusing the second course set by using an attention fusion mechanism, and obtaining a recommended course. The application improves the speed and accuracy of scene self-adaptive course recommendation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence processing technology, specifically to a method, apparatus, and storage medium for intelligent course recommendation based on scene adaptation. Background Technology

[0002] Driven by both lifelong learning and the digital transformation of education, personalized course recommendations have become a core support for solving educational pain points such as "information overload" and "supply-demand mismatch." Traditional recommendation technologies largely rely on users' historical behavior data, which has limitations such as a lack of context awareness, insufficient dynamic adaptation, and weak knowledge connections, making it difficult to meet learners' personalized needs at different learning stages, task scenarios, and environmental conditions. Traditional course recommendation technologies mainly rely on collaborative filtering algorithms or single content feature matching, and cannot achieve context-adaptive course recommendations.

[0003] Current knowledge graph-based recommendation systems primarily use two-dimensional or three-dimensional knowledge graphs for recommendations, failing to adequately consider dynamic adaptation to application scenarios in course recommendations. Furthermore, recommendations using a single knowledge graph suffer from poor accuracy and slow speed. Summary of the Invention

[0004] In view of one or more technical defects in the prior art, the present invention proposes the following technical solution.

[0005] A scenario-adaptive intelligent course recommendation method, comprising:

[0006] The construction steps involve building a four-dimensional knowledge graph consisting of "courses, knowledge points, ability goals, and application scenarios." The nodes of the four-dimensional knowledge graph are composed of courses, core knowledge points, ability levels, and application scenarios. The edges between the nodes are defined as multiple relationships such as "prerequisite dependencies," "ability correspondences," and "scenario adaptations."

[0007] The generation step involves generating a multimodal scene data matrix based on the acquired user physical environment data, task target data, cognitive state data, and social interaction data. The Transformer encoder, based on the self-attention mechanism and cross-modal fusion mechanism, transforms the discrete scene features in the modal scene data matrix into continuous semantic vectors, and then uses the continuous semantic vectors of all users to form a semantic vector space.

[0008] The steps are as follows: obtain the user's current application scenario, use the Transformer encoder to encode the user's current application scenario to obtain the current application scenario semantic vector, and use the Transformer decoder to determine the user's application scenario intent based on the current application scenario semantic vector and the semantic vector space.

[0009] The recommended steps are as follows: First, a set of courses is recommended using an intelligent adaptive recommendation model based on the application scenario intent; Second, based on the similarity calculation between the semantic vector of the current scenario and the application scenario nodes in the knowledge graph, knowledge points with a fit greater than a first threshold for the current application scenario are obtained; Third, a second set of courses is obtained based on the knowledge points; Fourth, an attention fusion mechanism is used to perform a weighted fusion of the first set of courses and the second set of courses to obtain recommended courses that are fed back to the user.

[0010] Furthermore, the user's physical environment data includes at least the learning location, device type, and network status; the task goal data includes at least career advancement needs, exam preparation goals, and skill improvement directions; the cognitive state data includes at least learning duration, focus level, and knowledge mastery level; and the social interaction data includes at least instant messaging interaction data, online community interaction data, and microblog interaction data.

[0011] Furthermore, scene-aware technology is used to acquire the physical environment data, and the light and noise levels and available time segments of the learning environment can be acquired simultaneously.

[0012] Furthermore, when the user's current application scenario changes, the trained Long Short-Term Memory Neural Network is used to obtain the current application scenario features in real time, and the semantic weights in the Transformer encoder are dynamically adjusted to update the semantic vector of the current application scenario.

[0013] Furthermore, the intelligent adaptive recommendation model is a graph attention neural network built based on a four-dimensional knowledge graph. The graph used in the graph attention neural network is structured as follows: course-knowledge point-ability goal-application scenario; with M courses, N knowledge points, K ability goals, and L application scenarios as nodes. The feature vector of the M courses as nodes is: Clesson(i) = (lessonname_i, lessonfield_i, lessontype_i); the feature vector of the N knowledge points as nodes is: Cknowledgepoint(j) = (knowledgefield_j, knowledgetime_j, knowledgetype_j); the feature vector of the K ability goals as nodes is: Cskillgoal(k) = (career_k, test_k, skillimprove_k); and the feature vector of the L application scenarios as nodes is: Cscene(p) = (cscenetype_p, devicetype_p, timescene_p, taskscene_p). ); where 1≤i≤M, 1≤j≤N, 1≤k≤K, 1≤p≤L, lessonname_i, lessonfield_i, and lessontype_i represent the course name, course field, and course type of the i-th course, respectively; knowledgefield_j, knowledgetime_j, and knowledgetype_j represent the knowledge field, knowledge time, and knowledge field of the j-th knowledge point, respectively; career_k, test_k, and skillimprove_k represent the career advancement goal, exam preparation goal, and skill improvement goal of the k-th ability goal, respectively; and scenetype_p, devicetype_p, timescene_p, and taskscene_p represent the scene type, device type, time scene, and target scene of the p-th application scenario, respectively.

[0014] Furthermore, the way edges are formed between nodes in the graph of the graph attention neural network is as follows: if there is a "prerequisite dependency" relationship between two courses, then there is an edge between the nodes of the two courses; if a course includes knowledge points, then there is an edge between the course node and the knowledge point node; if obtaining a skill goal requires learning the corresponding knowledge points, then there is an edge between the skill goal node and the knowledge point node; if obtaining a matching skill goal in an application scenario, then there is an edge between the skill goal node and the application scenario node; otherwise, there is no edge between any two nodes.

[0015] Furthermore, if there is an edge between two nodes in the graph of the graph attention neural network, the attention coefficient Att_mn between the two nodes is calculated as follows: Att_mn = cosine similarity(p_m, p_n), where cosine similarity(p_m, p_n) represents the cosine similarity between two nodes with an edge in the graph, p_m represents the feature vector of node m, p_n represents the feature vector of node n, 1≤m≤M+N+K+L, 1≤n≤M+N+K+L.

[0016] This invention also proposes a scenario-adaptive intelligent course recommendation device, which includes:

[0017] The construction unit constructs a four-dimensional knowledge graph consisting of "course-knowledge point-ability goal-application scenario". The nodes of the four-dimensional knowledge graph are composed of course, core knowledge point, ability level and application scenario. The edges between the nodes are defined as multiple relationships such as "prerequisite dependency", "ability correspondence" and "scenario adaptation".

[0018] The generation unit generates a multimodal scene data matrix based on the acquired user physical environment data, task target data, cognitive state data, and social interaction data. Using the Transformer encoder, based on the self-attention mechanism and cross-modal fusion mechanism, it transforms the discrete scene features in the modal scene data matrix into continuous semantic vectors, and combines the continuous semantic vectors of all users into a semantic vector space.

[0019] The unit determines the user's current application scenario, encodes the user's current application scenario using the Transformer encoder to obtain the current application scenario semantic vector, and uses the Transformer decoder to determine the user's application scenario intent based on the current application scenario semantic vector and the semantic vector space.

[0020] The recommendation unit recommends a first set of courses based on the application scenario intent using an intelligent adaptive recommendation model; it calculates the similarity between the semantic vector of the current scenario and the application scenario nodes in the knowledge graph, obtains knowledge points whose adaptability to the current application scenario is greater than a first threshold, obtains a second set of courses based on the knowledge points, and uses an attention fusion mechanism to perform weighted fusion of the first set of courses and the second set of courses to obtain recommended courses and feeds them back to the user.

[0021] Furthermore, the user's physical environment data includes at least the learning location, device type, and network status; the task goal data includes at least career advancement needs, exam preparation goals, and skill improvement directions; the cognitive state data includes at least learning duration, focus level, and knowledge mastery level; and the social interaction data includes at least instant messaging interaction data, online community interaction data, and microblog interaction data.

[0022] Furthermore, scene-aware technology is used to acquire the physical environment data, and the light and noise levels and available time segments of the learning environment can be acquired simultaneously.

[0023] Furthermore, when the user's current application scenario changes, the trained Long Short-Term Memory Neural Network is used to obtain the current application scenario features in real time, and the semantic weights in the Transformer encoder are dynamically adjusted to update the semantic vector of the current application scenario.

[0024] Furthermore, the intelligent adaptive recommendation model is a graph attention neural network built based on a four-dimensional knowledge graph. The graph used in the graph attention neural network is structured as follows: course-knowledge point-ability goal-application scenario; with M courses, N knowledge points, K ability goals, and L application scenarios as nodes. The feature vector of the M courses as nodes is: Clesson(i) = (lessonname_i, lessonfield_i, lessontype_i); the feature vector of the N knowledge points as nodes is: Cknowledgepoint(j) = (knowledgefield_j, knowledgetime_j, knowledgetype_j); the feature vector of the K ability goals as nodes is: Cskillgoal(k) = (career_k, test_k, skillimprove_k); and the feature vector of the L application scenarios as nodes is: Cscene(p) = (cscenetype_p, devicetype_p, timescene_p, taskscene_p). ); where 1≤i≤M, 1≤j≤N, 1≤k≤K, 1≤p≤L, lessonname_i, lessonfield_i, and lessontype_i represent the course name, course field, and course type of the i-th course, respectively; knowledgefield_j, knowledgetime_j, and knowledgetype_j represent the knowledge field, knowledge time, and knowledge field of the j-th knowledge point, respectively; career_k, test_k, and skillimprove_k represent the career advancement goal, exam preparation goal, and skill improvement goal of the k-th ability goal, respectively; and scenetype_p, devicetype_p, timescene_p, and taskscene_p represent the scene type, device type, time scene, and target scene of the p-th application scenario, respectively.

[0025] Furthermore, the way edges are formed between nodes in the graph of the graph attention neural network is as follows: if there is a "prerequisite dependency" relationship between two courses, then there is an edge between the nodes of the two courses; if a course includes knowledge points, then there is an edge between the course node and the knowledge point node; if obtaining a skill goal requires learning the corresponding knowledge points, then there is an edge between the skill goal node and the knowledge point node; if obtaining a matching skill goal in an application scenario, then there is an edge between the skill goal node and the application scenario node; otherwise, there is no edge between any two nodes.

[0026] Furthermore, if there is an edge between two nodes in the graph of the graph attention neural network, the attention coefficient Att_mn between the two nodes is calculated as follows: Att_mn = cosine similarity(p_m, p_n), where cosine similarity(p_m, p_n) represents the cosine similarity between two nodes with an edge in the graph, p_m represents the feature vector of node m, p_n represents the feature vector of node n, 1≤m≤M+N+K+L, 1≤n≤M+N+K+L.

[0027] The present invention also proposes a computer-readable storage medium storing computer program code, which, when executed by a computer, performs any of the methods described above.

[0028] The technical effect of this invention is as follows: This invention provides a scenario-adaptive intelligent course recommendation method, device, and storage medium. The method includes: a construction step S101, constructing a four-dimensional knowledge graph of "course-knowledge point-ability goal-application scenario," wherein the nodes of the four-dimensional knowledge graph consist of courses, core knowledge points, ability levels, and application scenarios, and the edges between the nodes are defined as multi-dimensional relationships of "prerequisite dependency," "ability correspondence," and "scenario adaptation"; and a generation step S102, generating a multimodal scenario data matrix based on acquired user physical environment data, task goal data, cognitive state data, and social interaction data, and using a Transformer encoder based on a self-attention mechanism and a cross-modal fusion mechanism to transform the discrete scenario features in the modal scenario data matrix into continuous semantic vectors, and combining the continuous semantic features of all users... The vectors form a semantic vector space; in step S103, the user's current application scenario is obtained, and the current application scenario is encoded using the Transformer encoder to obtain a semantic vector of the current application scenario. The user's application scenario intent is determined using the Transformer decoder based on the semantic vector of the current application scenario and the semantic vector space; in step S104, a first set of courses is recommended using an intelligent adaptive recommendation model based on the application scenario intent; based on the similarity calculation between the current scenario semantic vector and the application scenario nodes in the knowledge graph, knowledge points with a fit greater than a first threshold to the current application scenario are obtained, a second set of courses is obtained based on the knowledge points, and the first set of courses and the second set of courses are weighted and fused using an attention fusion mechanism to obtain recommended courses and feed them back to the user. The key inventive concept of this invention is the proposal to construct a four-dimensional knowledge graph consisting of "courses, knowledge points, ability objectives, and application scenarios." This four-dimensional knowledge graph can achieve an organic unity between "knowledge logical rationality" and "scenario demand adaptability," avoiding the recommendation lag problem caused by traditional static modeling. It enables real-time response to dynamic scenarios and further uses a Transformer decoder to determine the user's application scenario intent based on the semantic vector of the current application scenario and the semantic vector space, identifying complex scenarios and accurately capturing the user's core needs. This ensures that corresponding courses can be recommended adaptively according to changes in the scenario, improving the speed and accuracy of scenario-adaptive course recommendation.

[0029] Furthermore, in the specific recommendation stage, based on the application scenario intent, an intelligent adaptive recommendation model is used to recommend a first set of courses. That is, a graph attention neural network built on a four-dimensional knowledge graph is used for structural reasoning to mine the pre- and post-relationships between knowledge points, which conforms to the cognitive pattern of "from shallow to deep". Based on the similarity calculation between the semantic vector of the current scenario and the application scenario nodes in the knowledge graph, knowledge points with a fit greater than a first threshold with the current application scenario are obtained. Based on the knowledge points, a second set of courses is obtained. Then, an attention fusion mechanism is used to perform a weighted fusion of the first and second set of courses to obtain recommended courses that are fed back to the user. That is, two recommendation methods are used in the recommendation stage. One is an artificial intelligence-based recommendation method, which has high accuracy but the results may be relatively concentrated. The second recommendation method locates the set of knowledge points in the corresponding scenario, which has a large and comprehensive recommendation result. The weighted fusion of the two results yields a more objective and accurate recommendation result that meets the user's needs. This is another important inventive point of this invention.

[0030] It should be understood that the above general description and the following detailed description are merely exemplary and do not limit the invention. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below. Obviously, the drawings described below only relate to some embodiments of the present invention, and are not intended to limit the present invention.

[0032] Figure 1 This is a flowchart of a scenario-adaptive intelligent course recommendation method according to an embodiment of the present invention.

[0033] Figure 2 This is a structural diagram of a scenario-adaptive intelligent course recommendation device according to an embodiment of the present invention. Detailed Implementation

[0034] The present application will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0035] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0036] Figure 1 This invention illustrates a scenario-adaptive intelligent course recommendation method, which includes:

[0037] In step S101, a four-dimensional knowledge graph of “course-knowledge point-ability goal-application scenario” is constructed. The nodes of the four-dimensional knowledge graph consist of course, core knowledge point, ability level and application scenario. The edges between the nodes are defined as multiple relationships of “prerequisite dependency”, “ability correspondence” and “scenario adaptation”.

[0038] In step S102, a multimodal scene data matrix is ​​generated based on the acquired user physical environment data, task target data, cognitive state data, and social interaction data. The Transformer encoder is used to transform the discrete scene features in the modal scene data matrix into continuous semantic vectors based on the self-attention mechanism and cross-modal fusion mechanism, and the continuous semantic vectors of all users are combined into a semantic vector space.

[0039] In step S103, the user's current application scenario is obtained, the Transformer encoder is used to encode the user's current application scenario to obtain the current application scenario semantic vector, and the Transformer decoder is used to determine the user's application scenario intent based on the current application scenario semantic vector and the semantic vector space.

[0040] In step S104, a first set of courses is recommended using an intelligent adaptive recommendation model based on the application scenario intent; based on the similarity calculation between the semantic vector of the current scenario and the application scenario nodes in the knowledge graph, knowledge points with a fit greater than a first threshold for the current application scenario are obtained; a second set of courses is obtained based on the knowledge points; and an attention fusion mechanism is used to perform weighted fusion of the first set of courses and the second set of courses to obtain recommended courses that are fed back to the user.

[0041] The key inventive concept of this invention is the proposal to construct a four-dimensional knowledge graph encompassing "courses, knowledge points, competency objectives, and application scenarios." This four-dimensional knowledge graph achieves an organic unity between "logical rationality of knowledge" and "adaptability to scenario requirements." Based on acquired user physical environment data, task objective data, cognitive state data, and social interaction data, a multimodal scenario data matrix is ​​generated. Using a Transformer encoder based on self-attention and cross-modal fusion mechanisms, discrete scenario features in the modal scenario data matrix are transformed into continuous semantic vectors. These continuous semantic vectors from all users are combined to form a semantic vector space. Therefore, when the learner's (i.e., the user's) scenario elements change (e.g., switching from "in-office self-study" to "offline intensive training camp group learning"), the Transformer encoder can respond within 50ms. By completing scene semantic recoding, the recommendation lag problem caused by traditional static modeling is avoided, enabling real-time response to dynamic scenes. Furthermore, the Transformer decoder is used to determine the user's application scene intent based on the current application scene semantic vector and the semantic vector space. For example, Transformer uses a self-attention mechanism and a cross-modal fusion module to transform discrete scene features into continuous semantic vectors, achieving a deep understanding of "scene intent"—for example, identifying composite scenes such as "commuting (physical scene) + fragmented time (time scene) + basic English listening and speaking (task scene)" and accurately capturing their core needs of "short duration, high portability, and strong practicality," thereby ensuring that corresponding courses can be recommended adaptively according to changes in the scene. This is one of the important inventive concepts of this invention.

[0042] Furthermore, in the specific recommendation stage, based on the application scenario intent, an intelligent adaptive recommendation model is used to recommend a first set of courses. That is, a graph attention neural network built on a four-dimensional knowledge graph is used for structural reasoning to mine the pre- and post-relationships between knowledge points, which conforms to the cognitive pattern of "from shallow to deep". Based on the similarity calculation between the semantic vector of the current scenario and the application scenario nodes in the knowledge graph, knowledge points with a fit greater than a first threshold with the current application scenario are obtained. Based on the knowledge points, a second set of courses is obtained. Then, an attention fusion mechanism is used to perform a weighted fusion of the first and second set of courses to obtain recommended courses that are fed back to the user. That is, two recommendation methods are used in the recommendation stage. One is an artificial intelligence-based recommendation method, which has high accuracy but the results may be relatively concentrated. The second recommendation method locates the set of knowledge points in the corresponding scenario, which has a large and comprehensive recommendation result. The weighted fusion of the two results yields a more objective and accurate recommendation result that meets the user's needs. This is another important inventive point of this invention.

[0043] In one embodiment, the user's physical environment data includes at least the learning location, device type, and network status; the task objective data includes at least career advancement needs, exam preparation goals, and skill improvement directions; the cognitive state data includes at least learning duration, focus level, and knowledge mastery; and the social interaction data includes at least instant messaging interaction data, online community interaction data, and microblog interaction data. In one embodiment, scene-aware technology is used to acquire the physical environment data, and the light and noise levels of the learning environment, as well as available time segments, can be acquired simultaneously.

[0044] Based on ubiquitous computing technology, the system collects environmental data through the built-in sensors of smart devices and combines computer vision and speech recognition technologies to automatically identify the physical environment characteristics of learners. For example, by analyzing ambient light and noise levels, the system can determine whether a learner is in a public place and thus recommend a suitable course format for that environment—recommending text-based courses rather than video courses in noisy environments is another important aspect of this invention.

[0045] With the rapid development of artificial intelligence technology, especially large-scale model technology, the text and video courses in this invention can be generated using large-scale model technology. Text courses are generated using large language models, such as GPT-4 (which has strong semantic understanding and generation capabilities, accurately captures user needs, and generates logically coherent text) and LLaMA series models (which are lightweight, easily deployable open-source models with a rich community ecosystem. These models have flexible parameter scales, support multilingual processing, and have extensive training data coverage, making them suitable for secondary development and customization, and applicable to text generation, automated content creation, and other scenarios). Video courses can be generated using text-to-video models such as SORA, Tencent Hunyuan (which leads in text-to-video consistency, motion quality, and image quality, achieving ultra-realistic image quality and generating video footage that highly matches prompts), and Wan 2.1 (which supports tasks such as text-to-video generation, image-to-video generation, and video editing). Utilizing large-scale model technology can significantly improve the development speed of courses.

[0046] In one embodiment, when the user's current application scenario changes, the trained Long Short-Term Memory Neural Network is used to obtain the current application scenario features in real time, and the semantic weights in the Transformer encoder are dynamically adjusted to update the current application scenario semantic vector.

[0047] This invention uses a Transformer encoder based on a self-attention mechanism and a cross-modal fusion mechanism to transform discrete scene features in the modal scene data matrix into continuous semantic vectors. The continuous semantic vectors of all users are combined into a semantic vector space. Thus, when the scene elements of the learner (i.e., the user) change (such as switching from "in-office self-study" to "offline training camp group learning"), the Transformer encoder can complete the scene semantic recoding within 50ms. This avoids the recommendation lag problem caused by traditional static modeling and achieves an instant response to dynamic scenes. This enables scene-adaptive course recommendations, which is another important inventive concept of this invention.

[0048] In one embodiment, the intelligent adaptive recommendation model is a graph attention neural network built on a four-dimensional knowledge graph. This network is used to mine pre- and post-knowledge connections between knowledge points, ensuring that the recommended first set of courses conforms to the cognitive pattern of "from shallow to deep." The graph used in the graph attention neural network is constructed as follows: M courses, N knowledge points, K skill objectives, and L application scenarios are used as nodes. The feature vectors for the M courses as nodes are: Clesson(i) = (lessonname_i, lessonfield_i, lessontype_i); the feature vectors for the N knowledge points as nodes are: Cknowledgepoint(j) = (knowledgefield_j, knowledgetime_j, knowledgetype_j); the feature vectors for the K skill objectives as nodes are: Cskillgoal(k) = (career_k, test_k, skillimprove_k); and the feature vectors for the L application scenarios as nodes are: Cscene(p) = (cscenetype_p, devicetype_p, timescene_p, taskscene_p). ); where 1≤i≤M, 1≤j≤N, 1≤k≤K, 1≤p≤L, lessonname_i, lessonfield_i, and lessontype_i represent the course name, course field, and course type of the i-th course, respectively; knowledgefield_j, knowledgetime_j, and knowledgetype_j represent the knowledge field, knowledge time, and knowledge field of the j-th knowledge point, respectively; career_k, test_k, and skillimprove_k represent the career advancement goal, exam preparation goal, and skill improvement goal of the k-th ability goal, respectively; and scenetype_p, devicetype_p, timescene_p, and taskscene_p represent the scene type, device type, time scene, and target scene of the p-th application scenario, respectively.

[0049] This invention combines a four-dimensional knowledge graph with a graph attention neural network to construct an intelligent adaptive recommendation model. This model can automatically sense changes in the user's scenario and recommend corresponding courses based on the user's learning goals, already learned courses, mastered knowledge points, and learning goals to be achieved. This is another important inventive concept of this invention when the learner's (i.e., the user's) scenario elements change (such as switching from "in-office self-study" to "offline training camp group learning").

[0050] Furthermore, the way edges are formed between nodes in the graph of the graph attention neural network is based on the definition of the relationship between nodes in a four-dimensional knowledge graph of "course-knowledge point-ability goal-application scenario". The core idea is that the edges between nodes are defined as multiple relationships such as "prerequisite dependency", "ability correspondence" and "scenario adaptation". Specifically: if there is a "prerequisite dependency" relationship between two courses, then there is an edge between the nodes of the two courses; if a course includes knowledge points, then there is an edge between the course node and the knowledge point node; if obtaining an ability goal requires learning the corresponding knowledge points, then there is an edge between the ability goal node and the knowledge point node; if obtaining a matching ability goal in an application scenario, then there is an edge between the ability goal node and the application scenario node. Apart from these, there is no edge between any two nodes.

[0051] Furthermore, if there is an edge between two nodes in the graph of the graph attention neural network, the attention coefficient Att_mn between the two nodes is calculated as follows: Att_mn = cosine similarity(p_m, p_n), where cosine similarity(p_m, p_n) represents the cosine similarity between two nodes with an edge in the graph, p_m represents the feature vector of node m, p_n represents the feature vector of node n, 1≤m≤M+N+K+L, 1≤n≤M+N+K+L.

[0052] This invention combines a four-dimensional knowledge graph with a graph attention neural network. To ensure the accurate and reliable operation of the graph attention neural network, this invention proposes a specific method for calculating the relationship between nodes and edges in the graph and the attention coefficient. This is another important inventive concept of this invention.

[0053] Of course, graph attention neural networks need to be trained after they are generated. The specific training methods and parameter settings of neural networks are fundamental technologies in this field.

[0054] Figure 2 This invention illustrates a scenario-adaptive intelligent course recommendation device, which includes:

[0055] Construction unit 201 constructs a four-dimensional knowledge graph of "course-knowledge point-ability goal-application scenario". The nodes of the four-dimensional knowledge graph consist of course, core knowledge point, ability level and application scenario. The edges between the nodes are defined as multiple relationships of "prerequisite dependency", "ability correspondence" and "scenario adaptation".

[0056] The generation unit 202 generates a multimodal scene data matrix based on the acquired user physical environment data, task target data, cognitive state data and social interaction data. It uses a Transformer encoder based on self-attention mechanism and cross-modal fusion mechanism to transform the discrete scene features in the modal scene data matrix into continuous semantic vectors, and combines the continuous semantic vectors of all users into a semantic vector space.

[0057] The determination unit 203 obtains the user's current application scenario, encodes the user's current application scenario using the Transformer encoder to obtain the current application scenario semantic vector, and uses the Transformer decoder to determine the user's application scenario intent based on the current application scenario semantic vector and the semantic vector space.

[0058] Recommendation unit 204 recommends a first set of courses based on the application scenario intent using an intelligent adaptive recommendation model; it calculates the similarity between the semantic vector of the current scenario and the application scenario nodes in the knowledge graph, obtains knowledge points with a fit greater than a first threshold to the current application scenario, obtains a second set of courses based on the knowledge points, and uses an attention fusion mechanism to perform weighted fusion of the first set of courses and the second set of courses to obtain recommended courses and feeds them back to the user.

[0059] The key inventive concept of this invention is the proposal to construct a four-dimensional knowledge graph encompassing "courses, knowledge points, competency objectives, and application scenarios." This four-dimensional knowledge graph achieves an organic unity between "logical rationality of knowledge" and "adaptability to scenario requirements." Based on acquired user physical environment data, task objective data, cognitive state data, and social interaction data, a multimodal scenario data matrix is ​​generated. Using a Transformer encoder based on self-attention and cross-modal fusion mechanisms, discrete scenario features in the modal scenario data matrix are transformed into continuous semantic vectors. These continuous semantic vectors from all users are combined to form a semantic vector space. Therefore, when the learner's (i.e., the user's) scenario elements change (e.g., switching from "in-office self-study" to "offline intensive training camp group learning"), the Transformer encoder can respond within 50ms. By completing scene semantic recoding, the recommendation lag problem caused by traditional static modeling is avoided, enabling real-time response to dynamic scenes. Furthermore, the Transformer decoder is used to determine the user's application scene intent based on the current application scene semantic vector and the semantic vector space. For example, Transformer uses a self-attention mechanism and a cross-modal fusion module to transform discrete scene features into continuous semantic vectors, achieving a deep understanding of "scene intent"—for example, identifying composite scenes such as "commuting (physical scene) + fragmented time (time scene) + basic English listening and speaking (task scene)" and accurately capturing their core needs of "short duration, high portability, and strong practicality," thereby ensuring that corresponding courses can be recommended adaptively according to changes in the scene. This is one of the important inventive concepts of this invention.

[0060] Furthermore, in the specific recommendation stage, based on the application scenario intent, an intelligent adaptive recommendation model is used to recommend a first set of courses. That is, a graph attention neural network built on a four-dimensional knowledge graph is used for structural reasoning to mine the pre- and post-relationships between knowledge points, which conforms to the cognitive pattern of "from shallow to deep". Based on the similarity calculation between the semantic vector of the current scenario and the application scenario nodes in the knowledge graph, knowledge points with a fit greater than a first threshold with the current application scenario are obtained. Based on the knowledge points, a second set of courses is obtained. Then, an attention fusion mechanism is used to perform a weighted fusion of the first and second set of courses to obtain recommended courses that are fed back to the user. That is, two recommendation methods are used in the recommendation stage. One is an artificial intelligence-based recommendation method, which has high accuracy but the results may be relatively concentrated. The second recommendation method locates the set of knowledge points in the corresponding scenario, which has a large and comprehensive recommendation result. The weighted fusion of the two results yields a more objective and accurate recommendation result that meets the user's needs. This is another important inventive point of this invention.

[0061] In one embodiment, the user's physical environment data includes at least the learning location, device type, and network status; the task objective data includes at least career advancement needs, exam preparation goals, and skill improvement directions; the cognitive state data includes at least learning duration, focus level, and knowledge mastery; and the social interaction data includes at least instant messaging interaction data, online community interaction data, and microblog interaction data. In one embodiment, scene-aware technology is used to acquire the physical environment data, and the light and noise levels of the learning environment, as well as available time segments, can be acquired simultaneously.

[0062] Based on ubiquitous computing technology, the system collects environmental data through the built-in sensors of smart devices and combines computer vision and speech recognition technologies to automatically identify the physical environment characteristics of learners. For example, by analyzing ambient light and noise levels, the system can determine whether a learner is in a public place and thus recommend a suitable course format for that environment—recommending text-based courses rather than video courses in noisy environments is another important aspect of this invention.

[0063] With the rapid development of artificial intelligence technology, especially large-scale model technology, the text and video courses in this invention can be generated using large-scale model technology. Text courses are generated using large language models, such as GPT-4 (which has strong semantic understanding and generation capabilities, accurately captures user needs, and generates logically coherent text) and LLaMA series models (which are lightweight, easily deployable open-source models with a rich community ecosystem. These models have flexible parameter scales, support multilingual processing, and have extensive training data coverage, making them suitable for secondary development and customization, and applicable to text generation, automated content creation, and other scenarios). Video courses can be generated using text-to-video models such as SORA, Tencent Hunyuan (which leads in text-to-video consistency, motion quality, and image quality, achieving ultra-realistic image quality and generating video footage that highly matches prompts), and Wan 2.1 (which supports tasks such as text-to-video generation, image-to-video generation, and video editing). Utilizing large-scale model technology can significantly improve the development speed of courses.

[0064] In one embodiment, when the user's current application scenario changes, the trained Long Short-Term Memory Neural Network is used to obtain the current application scenario features in real time, and the semantic weights in the Transformer encoder are dynamically adjusted to update the current application scenario semantic vector.

[0065] This invention uses a Transformer encoder based on a self-attention mechanism and a cross-modal fusion mechanism to transform discrete scene features in the modal scene data matrix into continuous semantic vectors. The continuous semantic vectors of all users are combined into a semantic vector space. Thus, when the scene elements of the learner (i.e., the user) change (such as switching from "in-office self-study" to "offline training camp group learning"), the Transformer encoder can complete the scene semantic recoding within 50ms. This avoids the recommendation lag problem caused by traditional static modeling and achieves an instant response to dynamic scenes. This enables scene-adaptive course recommendations, which is another important inventive concept of this invention.

[0066] In one embodiment, the intelligent adaptive recommendation model is a graph attention neural network built on a four-dimensional knowledge graph. This network is used to mine pre- and post-knowledge connections between knowledge points, ensuring that the recommended first set of courses conforms to the cognitive pattern of "from shallow to deep." The graph used in the graph attention neural network is constructed as follows: M courses, N knowledge points, K skill objectives, and L application scenarios are used as nodes. The feature vectors for the M courses as nodes are: Clesson(i) = (lessonname_i, lessonfield_i, lessontype_i); the feature vectors for the N knowledge points as nodes are: Cknowledgepoint(j) = (knowledgefield_j, knowledgetime_j, knowledgetype_j); the feature vectors for the K skill objectives as nodes are: Cskillgoal(k) = (career_k, test_k, skillimprove_k); and the feature vectors for the L application scenarios as nodes are: Cscene(p) = (cscenetype_p, devicetype_p, timescene_p, taskscene_p). ); where 1≤i≤M, 1≤j≤N, 1≤k≤K, 1≤p≤L, lessonname_i, lessonfield_i, and lessontype_i represent the course name, course field, and course type of the i-th course, respectively; knowledgefield_j, knowledgetime_j, and knowledgetype_j represent the knowledge field, knowledge time, and knowledge field of the j-th knowledge point, respectively; career_k, test_k, and skillimprove_k represent the career advancement goal, exam preparation goal, and skill improvement goal of the k-th ability goal, respectively; and scenetype_p, devicetype_p, timescene_p, and taskscene_p represent the scene type, device type, time scene, and target scene of the p-th application scenario, respectively.

[0067] This invention combines a four-dimensional knowledge graph with a graph attention neural network to construct an intelligent adaptive recommendation model. This model can automatically sense changes in the user's scenario and recommend corresponding courses based on the user's learning goals, already learned courses, mastered knowledge points, and learning goals to be achieved. This is another important inventive concept of this invention when the learner's (i.e., the user's) scenario elements change (such as switching from "in-office self-study" to "offline training camp group learning").

[0068] Furthermore, the way edges are formed between nodes in the graph of the graph attention neural network is based on the definition of the relationship between nodes in a four-dimensional knowledge graph of "course-knowledge point-ability goal-application scenario". The core idea is that the edges between nodes are defined as multiple relationships such as "prerequisite dependency", "ability correspondence" and "scenario adaptation". Specifically: if there is a "prerequisite dependency" relationship between two courses, then there is an edge between the nodes of the two courses; if a course includes knowledge points, then there is an edge between the course node and the knowledge point node; if obtaining an ability goal requires learning the corresponding knowledge points, then there is an edge between the ability goal node and the knowledge point node; if obtaining a matching ability goal in an application scenario, then there is an edge between the ability goal node and the application scenario node. Apart from these, there is no edge between any two nodes.

[0069] Furthermore, if there is an edge between two nodes in the graph of the graph attention neural network, the attention coefficient Att_mn between the two nodes is calculated as follows: Att_mn = cosine similarity(p_m, p_n), where cosine similarity(p_m, p_n) represents the cosine similarity between two nodes with an edge in the graph, p_m represents the feature vector of node m, p_n represents the feature vector of node n, 1≤m≤M+N+K+L, 1≤n≤M+N+K+L.

[0070] This invention combines a four-dimensional knowledge graph with a graph attention neural network. To ensure the accurate and reliable operation of the graph attention neural network, this invention proposes a specific method for calculating the relationship between nodes and edges in the graph and the attention coefficient. This is another important inventive concept of this invention.

[0071] Of course, graph attention neural networks need to be trained after they are generated. The specific training methods and parameter settings of neural networks are fundamental technologies in this field.

[0072] Simulation tests were conducted using data from a large online education platform, containing over 3,000 course resources, 20,000 users, and over 50,000 user learning behavior records. A dataset of course knowledge dependencies was constructed through manual annotation and used as the experimental baseline.

[0073] Evaluation metrics: Accuracy and recommended response time when the scenario changes.

[0074] Experimental Results: The experiment compares the proposed technique with traditional collaborative filtering algorithms and single knowledge graph recommendation algorithms. The results are as follows:

[0075] Accuracy: The accuracy of this invention reaches 79.3%, which is significantly higher than traditional collaborative filtering and single knowledge graph recommendation methods.

[0076] Recommendation response time when the scene changes: The technology of this invention achieves 3 seconds, while traditional collaborative filtering takes 10 seconds and single knowledge graph recommendation takes 13.7 seconds, which greatly improves the recommendation speed.

[0077] One embodiment of the present invention provides a computer storage medium storing a computer program. When the computer program on the computer storage medium is executed by a processor, the above-described method is implemented. The computer storage medium may be a hard disk, DVD, CD, flash memory, or other storage device.

[0078] For ease of description, the above apparatus is described by dividing it into various functional units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0079] As can be seen from the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the apparatus described in various embodiments or some parts of the embodiments of this application.

[0080] Finally, it should be noted that the above embodiments are for illustration only and not for limiting the technical solutions of the present invention. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention without departing from the spirit and scope of the present invention. Any modifications or partial substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for scene-adaptive course intelligent recommendation, characterized in that, The method includes: The construction steps involve building a four-dimensional knowledge graph consisting of "courses, knowledge points, ability goals, and application scenarios". The nodes of the four-dimensional knowledge graph are composed of courses, core knowledge points, ability levels, and application scenarios. The edges between the nodes are defined as multiple relationships such as "prerequisite dependencies", "ability correspondence", and "scenario adaptation". The generation step involves generating a multimodal scene data matrix based on the acquired user physical environment data, task target data, cognitive state data, and social interaction data. The Transformer encoder, based on the self-attention mechanism and cross-modal fusion mechanism, transforms the discrete scene features in the modal scene data matrix into continuous semantic vectors, and then uses the continuous semantic vectors of all users to form a semantic vector space. The steps are as follows: obtain the user's current application scenario, use the Transformer encoder to encode the user's current application scenario to obtain the current application scenario semantic vector, and use the Transformer decoder to determine the user's application scenario intent based on the current application scenario semantic vector and the semantic vector space. The recommended steps are as follows: First, a set of courses is recommended using an intelligent adaptive recommendation model based on the application scenario intent; Second, based on the similarity calculation between the semantic vector of the current application scenario and the application scenario nodes in the knowledge graph, knowledge points with a fit greater than a first threshold for the current application scenario are obtained; Third, a second set of courses is obtained based on the knowledge points; Fourth, an attention fusion mechanism is used to perform a weighted fusion of the first set of courses and the second set of courses to obtain recommended courses and feed them back to the user. In this study, a four-dimensional knowledge graph was combined with a graph attention neural network to construct an intelligent adaptive recommendation model. The way edges are formed between nodes in the graph of the graph attention neural network is based on the relationship definition between nodes in the four-dimensional knowledge graph of "course-knowledge point-ability goal-application scenario". The edges between nodes are defined as multiple relationships of "prerequisite dependency", "ability correspondence" and "scenario adaptation". Specifically, if there is a "prerequisite dependency" relationship between two courses, then there is an edge between the nodes of the two courses; if a course includes knowledge points, then there is an edge between the course node and the knowledge point node; if obtaining an ability goal requires learning the corresponding knowledge points, then there is an edge between the ability goal node and the knowledge point node; if obtaining a matching ability goal in an application scenario, then there is an edge between the ability goal node and the application scenario node. Specifically, based on the application scenario intent, an intelligent adaptive recommendation model is used to recommend a first set of courses. This involves using a graph attention neural network built on a four-dimensional knowledge graph for structural reasoning, mining the pre- and post-relationships between knowledge points, and calculating the similarity between the semantic vector of the current application scenario and the application scenario nodes in the knowledge graph to obtain knowledge points with a fit greater than a first threshold. Based on these knowledge points, a second set of courses is obtained. Then, an attention fusion mechanism is used to weightedly fuse the first and second set of courses to obtain recommended courses that are fed back to the user. In other words, two recommendation methods are used: one is an artificial intelligence-based recommendation method, and the second recommendation method locates the knowledge point set in the corresponding scenario. The weighted fusion of the two results yields a more objective and accurate recommendation result that meets the user's needs.

2. The method according to claim 1, characterized in that, The user's physical environment data includes at least the learning location, device type, and network status; the task goal data includes at least career advancement needs, exam preparation goals, and skill improvement directions; the cognitive status data includes at least learning duration, focus level, and knowledge mastery level; and the social interaction data includes at least instant messaging interaction data, online community interaction data, and microblog interaction data.

3. The method according to claim 2, characterized in that, The physical environment data is obtained using scene-aware technology, and the light and noise levels and available time segments of the learning environment can be obtained simultaneously.

4. The method according to claim 3, characterized in that, When the user's current application scenario changes, the trained Long Short-Term Memory Neural Network is used to obtain the current application scenario features in real time, and the semantic weights in the Transformer encoder are dynamically adjusted to update the semantic vector of the current application scenario.

5. The method according to claim 4, characterized in that, The intelligent adaptive recommendation model is a graph attention neural network built on a four-dimensional knowledge graph. The graph used in the graph attention neural network is structured as follows: course-knowledge point-skill goal-application scenario; with M courses, N knowledge points, K skill goals, and L application scenarios as nodes. The feature vector of the M courses as nodes is: Clesson(i) = (lessonname_i, lessonfield_i, lessontype_i); the feature vector of the N knowledge points as nodes is: Cknowledgepoint(j) = (knowledgefield_j, knowledgetime_j, knowledgetype_j); the feature vector of the K skill goals as nodes is: Cskillgoal(k) = (career_k, test_k, skillimprove_k); and the feature vector of the L application scenarios as nodes is: Cscene(p) = (cscenetype_p, devicetype_p, timescene_p, taskscene_p). ); where 1≤i≤M, 1≤j≤N, 1≤k≤K, 1≤p≤L, lessonname_i, lessonfield_i, and lessontype_i represent the course name, course field, and course type of the i-th course, respectively; knowledgefield_j, knowledgetime_j, and knowledgetype_j represent the knowledge field, knowledge time, and knowledge field of the j-th knowledge point, respectively; career_k, test_k, and skillimprove_k represent the career advancement goal, exam preparation goal, and skill improvement goal of the k-th ability goal, respectively; and scenetype_p, devicetype_p, timescene_p, and taskscene_p represent the scene type, device type, time scene, and target scene of the p-th application scenario, respectively.

6. A scenario-adaptive intelligent course recommendation device, characterized in that, The device includes: The construction unit constructs a four-dimensional knowledge graph consisting of "course-knowledge point-ability goal-application scenario". The nodes of the four-dimensional knowledge graph are composed of courses, core knowledge points, ability levels and application scenarios. The edges between the nodes are defined as multiple relationships such as "prerequisite dependency", "ability correspondence" and "scenario adaptation". The generation unit generates a multimodal scene data matrix based on the acquired user physical environment data, task target data, cognitive state data, and social interaction data. Using the Transformer encoder, based on the self-attention mechanism and cross-modal fusion mechanism, it transforms the discrete scene features in the modal scene data matrix into continuous semantic vectors, and combines the continuous semantic vectors of all users into a semantic vector space. The unit determines the user's current application scenario, encodes the user's current application scenario using the Transformer encoder to obtain the current application scenario semantic vector, and uses the Transformer decoder to determine the user's application scenario intent based on the current application scenario semantic vector and the semantic vector space. The recommendation unit recommends a first set of courses based on the application scenario intent using an intelligent adaptive recommendation model; it calculates the similarity between the semantic vector of the current application scenario and the application scenario nodes in the knowledge graph, obtains knowledge points with a fit greater than a first threshold for the current application scenario, obtains a second set of courses based on the knowledge points, and uses an attention fusion mechanism to perform a weighted fusion of the first set of courses and the second set of courses to obtain recommended courses and feeds them back to the user. In this study, a four-dimensional knowledge graph was combined with a graph attention neural network to construct an intelligent adaptive recommendation model. The way edges are formed between nodes in the graph of the graph attention neural network is based on the relationship definition between nodes in the four-dimensional knowledge graph of "course-knowledge point-ability goal-application scenario". The edges between nodes are defined as multiple relationships of "prerequisite dependency", "ability correspondence" and "scenario adaptation". Specifically, if there is a "prerequisite dependency" relationship between two courses, then there is an edge between the nodes of the two courses; if a course includes knowledge points, then there is an edge between the course node and the knowledge point node; if obtaining an ability goal requires learning the corresponding knowledge points, then there is an edge between the ability goal node and the knowledge point node; if obtaining a matching ability goal in an application scenario, then there is an edge between the ability goal node and the application scenario node. Specifically, based on the application scenario intent, an intelligent adaptive recommendation model is used to recommend a first set of courses. This involves using a graph attention neural network built on a four-dimensional knowledge graph for structural reasoning, mining the pre- and post-relationships between knowledge points, and calculating the similarity between the semantic vector of the current application scenario and the application scenario nodes in the knowledge graph to obtain knowledge points with a fit greater than a first threshold. Based on these knowledge points, a second set of courses is obtained. Then, an attention fusion mechanism is used to weightedly fuse the first and second set of courses to obtain recommended courses that are fed back to the user. In other words, two recommendation methods are used: one is an artificial intelligence-based recommendation method, and the second recommendation method locates the knowledge point set in the corresponding scenario. The weighted fusion of the two results yields a more objective and accurate recommendation result that meets the user's needs.

7. The apparatus according to claim 6, characterized in that, The user's physical environment data includes at least the learning location, device type, and network status; the task goal data includes at least career advancement needs, exam preparation goals, and skill improvement directions; the cognitive status data includes at least learning duration, focus level, and knowledge mastery level; and the social interaction data includes at least instant messaging interaction data, online community interaction data, and microblog interaction data.

8. The apparatus according to claim 7, characterized in that, The physical environment data is obtained using scene-aware technology, and the light and noise levels and available time segments of the learning environment can be obtained simultaneously.

9. The apparatus according to claim 8, characterized in that, When the user's current application scenario changes, the trained Long Short-Term Memory Neural Network is used to obtain the current application scenario features in real time, and the semantic weights in the Transformer encoder are dynamically adjusted to update the semantic vector of the current application scenario.

10. A computer storage medium, characterized in that, The computer storage medium stores a computer program, which, when executed by a processor, implements the method described in any one of claims 1-5.

Images