Man-machine collaborative education resource recommendation method and system for large-scale online learning

Abstract

The invention relates to a man-machine collaborative education resource recommendation method and system for large-scale online learning, and belongs to the technical field of intelligent education. The method comprises the following steps: firstly, modeling video watching interaction behavior information of a learner in a learning process, and respectively converting video interaction data of the learner into a graph structure and a sequence structure so as to describe long-term learning preference and short-term learning attention change of the learner; secondly, through a man-machine cooperation mechanism based on multiple rounds of dialogues, embedded representation of the learner and teaching resources is dynamically updated by utilizing real-time interaction feedback of the learner in the learning process, and continuous perception and adjustment of learning requirements are realized; and finally, in combination with professional background information of the learner, a recommendation strategy is adaptively optimized through a synergistic effect among different recommendation agents, so that a teaching video recommendation result which better conforms to the learning progress and personalized requirements of the learner is provided for the learner.

Description

Technical Field

[0001] This invention belongs to the field of intelligent education technology and relates to a human-computer collaborative education resource recommendation method and system for large-scale online learning. Background Technology

[0002] With the rapid development of information technology, online education has gradually become an important form of education. Especially in the post-pandemic era, the blended learning model combining online and offline learning has become a significant trend in the education field. Large-scale online learning platforms provide learners with a low-cost way to access high-quality educational resources. These platforms offer a rich variety of learning resources, including structured curriculum systems, instructional videos, and supplementary learning content related to specific knowledge points, broadly covering various application scenarios such as basic education, higher education, and vocational education. When using these resources for self-study, learners need to rationally select learning content based on their own knowledge base, learning goals, and learning progress to improve learning efficiency and experience. However, the sheer scale and variety of resources on online learning platforms, without effective guidance and intelligent support, can easily lead to information overload and difficulty in making choices. On the one hand, learners struggle to quickly locate instructional videos that highly match their current learning goals from a vast amount of resources; on the other hand, inappropriate resource recommendations may result in fragmented learning paths, increased cognitive load, and consequently, negatively impacted learning outcomes or even reduced learning motivation. Therefore, how to combine the characteristics of educational scenarios to provide learners with personalized teaching video recommendations that can dynamically respond to changes in their learning needs has become a key issue that urgently needs to be addressed in the field of intelligent education technology, and it is also an important technological foundation for promoting the transformation of online education from "resource-oriented" to "service-oriented". Existing learning resource recommendation methods mainly include traditional recommendation techniques such as collaborative filtering, content-based recommendation, and matrix factorization. These methods usually rely on historical interaction data between learners and resources for modeling, which can reflect learners' long-term interests and preferences to a certain extent, but they have obvious limitations in characterizing learners' real-time changes in needs and dealing with complex learning behaviors across courses and disciplines. Especially in online learning scenarios, learners' learning goals and focus often evolve with the learning process, and simply relying on static historical data is difficult to meet the timeliness and relevance requirements of teaching video recommendations.

[0003] With the development of artificial intelligence technology, dialogue recommendation systems are increasingly being applied in online recommendation scenarios. These systems introduce multi-turn dialogue interaction mechanisms, enabling continuous interaction with users during the recommendation process. They proactively acquire user feedback on recommended content or related attributes and dynamically adjust recommendation strategies based on this feedback, thus achieving human-computer collaborative decision-making. In multi-turn dialogue human-computer collaborative recommendation, the system typically performs attribute queries and resource recommendations in different interaction rounds. Users convey their true needs to the system by accepting or rejecting recommended content and confirming or denying attribute preferences, allowing the recommendation results to gradually approach the user's goals through human-computer interaction. Compared to traditional static recommendation methods that rely on historical data, multi-turn dialogue recommendation, through human-computer collaboration, can more promptly capture changes in users' immediate preferences, reduce invalid recommendations, and improve the flexibility and accuracy of the recommendation process. Therefore, it has significant advantages in application scenarios requiring continuous preference confirmation and dynamic decision-making. Introducing this method into the scenario of educational video recommendation helps the recommendation system continuously revise its judgment of learners' learning needs based on their feedback during the learning process, thereby providing learners with educational video recommendations that better match their learning progress and knowledge requirements.

[0004] However, existing dialogue recommendation methods still have shortcomings when applied to recommending instructional videos. On the one hand, most methods focus on modeling immediate interaction signals, lacking a systematic characterization of learners' long-term professional background and historical learning behavior, making it difficult to balance learners' long-term learning direction and short-term learning priorities. On the other hand, some methods have relatively simple models for learners' real-time feedback, making it difficult to efficiently integrate the feedback information accumulated over multiple rounds of dialogue, resulting in untimely adjustments to recommendation strategies. Furthermore, instructional video recommendations have obvious professional attributes and knowledge-dependent characteristics, while existing methods have limited ability to differentiate learning needs across different professional fields, making it difficult to meet the personalized recommendation needs of diverse learners in a large-scale online learning environment. Summary of the Invention

[0005] In view of this, the purpose of this invention is to provide a human-computer collaborative educational resource recommendation method and system for large-scale online learning, so as to solve the problem that the existing dialogue recommendation methods have poor transfer and application effects in educational scenarios, thereby improving learners' learning interest and interactive experience.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A human-computer collaborative educational resource recommendation method for large-scale online learning includes the following steps: S1. Offline learner behavior representation stage: The interaction data between learners and learning resources is synchronously constructed into graph structure and sequence structure, and fine-grained dependencies and temporal evolution rules are extracted respectively to obtain the offline embedded representation of learners; S2, Online Dialogue Interaction Representation Update Stage: Obtain real-time feedback information from learners in multi-turn dialogues, dynamically fuse the real-time feedback with the offline embedded representation based on the preference-aware attention mechanism, and update the representations of learners and learning resources; S3, Hierarchical Multi-Agent Collaborative Reasoning Stage: Construct multiple professional recommendation agents and one adaptive recommendation agent to collaboratively generate and output personalized teaching video recommendation results based on the learner's professional background and dialogue state.

[0007] Furthermore, step S1 specifically includes: S11. Simultaneously organize the learner-learning resource interaction data into both graph and sequence structures; S12. Employ graph-based representation learning methods to extract fine-grained dependencies on graphs to characterize learners' short-term dynamic preferences; S13. A sequence-based temporal pattern modeling method is used to perform temporal learning on learners' historical interaction sequences in order to capture long-term preferences and their evolutionary patterns, providing a basic representation for the online dialogue interaction reasoning stage.

[0008] Furthermore, step S2 specifically includes: S21. Obtain real-time feedback information from learners during each round of dialogue. The real-time feedback is provided by learners during their interaction with the system, including expressions of interest in the queried knowledge points and feedback on acceptance or rejection of recommended videos. S22. The learner feedback is used as a human-machine collaboration signal. Based on the preference-aware attention mechanism, the real-time feedback is dynamically fused with the embedded representation obtained in step S1 to collaboratively update the learner's current intention representation and learning resource representation. S23. By iteratively executing the above collaborative update process during multi-round dialogue interaction, the recommendation system can continuously perceive changes in learners' intentions and continuously improve the matching degree between learning resources and learners' real-time learning needs.

[0009] Furthermore, step S3 specifically includes: S31. Construct multiple recommendation agents for specific majors based on learners' academic background information. Each recommendation agent for a specific major is responsible for mining learners' preferences in that major field and learning corresponding recommendation strategies under the guidance of learner interaction feedback. S32. Construct an adaptive recommendation agent to receive and integrate the dialogue recommendation strategies formed by various professional recommendation agents during the collaborative reasoning process. S33. The adaptive recommendation agent dynamically integrates and learns the strategies of various professional agents based on the learner's cross-disciplinary learning needs during the dialogue process, forming a unified recommendation decision. S34. Output instructional video recommendation results to learners based on the integrated recommendation strategy.

[0010] Furthermore, in step S1, the construction of the offline learner behavior representation includes two parts: graph structure modeling and sequence structure modeling. First, the relationship between learners, video resources, and the knowledge point attributes contained in the video is modeled as a heterogeneous interaction graph, where learners, video resources, and knowledge points are nodes, and other interaction behaviors are edges. Triples are constructed to store the structural dependencies between nodes. Second, the constructed interaction graph is encoded using a TransD-based graph embedding method to obtain initial node representations. Furthermore, a graph convolutional neural network is used to capture the fine-grained topological preference dependencies of learners in the heterogeneous interaction graph, enabling the system to quickly respond to changes in learners' short-term viewing behavior. In addition, the learner's historical interaction sequence is input into the Longformer sequence model, which uses its long-sequence time-dependent attention mechanism to learn the characteristics of learner preferences evolving over time, thereby representing the learner's long-term preference patterns. Through the joint modeling of the above graph representation and sequence representation, a comprehensive offline embedding representation is formed, providing robust basic feature support for the online dialogue stage.

[0011] Furthermore, in step S2, the implementation of the online dialogue interaction representation update includes two steps: real-time feedback modeling and preference fusion. First, through multi-round dialogue interaction, the learner's immediate feedback information generated during the collaboration with the recommendation system is collected, including the expression of preference for the queried knowledge points and the acceptance or rejection of recommended videos. Then, the feedback is used as an explicit constraint on the learning intention in the human-computer collaboration process to update the learner's dialogue state representation in real time. The updated representation is then input into the preference-aware attention mechanism, which dynamically fuses it with the offline preference embedding representation obtained in S1. Attention weights are used to adjust the contribution of short-term interests and long-term preferences in the current collaborative intention representation and their relevance to learning resources. As the dialogue progresses, the above fusion and update process is iteratively executed, so that the learner's preference representation and the learning resource representation gradually align with the learner's evolving needs and intentions in continuous interaction, thereby realizing dynamic teaching video matching driven by human-computer collaboration.

[0012] Furthermore, in step S3, the implementation of hierarchical multi-agent collaborative reasoning includes a human-machine collaborative decision-making process between a specific professional recommendation agent and an adaptive recommendation agent: First, multiple specific professional recommendation agents are constructed based on the learner's academic background information. Each specific professional agent, guided by the learner's interactive feedback and dialogue intent, focuses on analyzing the learner's interests and preferences in the corresponding professional field and learns the corresponding recommendation strategy to explore the potential demand for learning resources in different professional fields. Then, an adaptive recommendation agent is constructed to receive the recommendation strategies generated by each specific professional agent during the collaborative reasoning process, and to integrate and adjust the strategies from different professional fields in a unified manner based on the learner's current dialogue state to adapt to cross-professional learning scenarios. Finally, personalized teaching video recommendation results that match the learner's current learning intent are output to the learner through the integrated collaborative strategy.

[0013] The present invention also provides a human-computer collaborative educational resource recommendation system for large-scale online learning, which adopts the method described above.

[0014] The present invention also provides an electronic device, including a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the method described above.

[0015] The beneficial effects of this invention are as follows: The method provided by this invention first models learners' video viewing interaction behavior information during the learning process, transforming learners' video interaction data into graph structures and sequence structures to characterize learners' long-term learning preferences and short-term changes in learning attention. Second, through a human-computer collaboration mechanism based on multi-turn dialogue, the embedded representations of learners and teaching resources are dynamically updated using real-time interactive feedback during the learning process, enabling continuous perception and adjustment of learning needs. Finally, by further combining learners' professional background information and through the collaborative effect between different recommendation agents, the recommendation strategy is adaptively optimized, thereby providing learners with teaching video recommendation results that better match their learning progress and personalized needs.

[0016] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0017] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1This is a framework diagram of a human-computer collaborative educational resource recommendation method for large-scale online learning as described in this invention. Detailed Implementation

[0018] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0019] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0020] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0021] Figure 1 The figure shows a framework diagram of a human-computer collaborative educational resource recommendation method for large-scale online learning according to the present invention. The specific steps of the human-computer collaborative educational resource recommendation method for large-scale online learning provided by the present invention are as follows: S1: Offline Learner Behavior Representation Stage: First, the interaction data between learners and learning resources is synchronously constructed into a heterogeneous graph and time series structure; second, a graph basis representation learning method is used to extract fine-grained dependencies from the graph structure to characterize the learner's short-term dynamic preferences; at the same time, a time series analysis is performed on the learner's historical interaction sequence through a sequence time modeling method to capture its long-term preferences and evolutionary patterns, ultimately providing basic representation support for the online dialogue interaction reasoning stage, specifically including the following two sub-steps.

[0022] 1) Heterogeneous topology graph representation and modeling stage First, a heterogeneous graph containing three types of entity nodes and four types of edge relationships is constructed based on the original interaction data. The three types of entities are learner nodes, learning video nodes, and knowledge point nodes. The four types of edge relationships are: "friend" relationships between learners, used to characterize potential similarity in learning behaviors; "interaction" relationships between learners and videos, used to represent learners' viewing behavior on videos; "belonging" relationships between videos and knowledge points, used to characterize the knowledge structure of learning resources; and "liking" relationships between learners and knowledge points, used to represent learners' preference tendencies towards certain knowledge points. Next, based on the constructed heterogeneous graph, graph embedding pre-training is performed using the TransD model. By minimizing the boundary loss function, the pre-trained embedding vectors of entities are obtained. Finally, considering that triple-based embedding learning cannot fully utilize the neighborhood structure in the graph, a graph convolutional neural network is further used to enhance the local structure of the pre-trained embeddings. The graph convolutional neural network updates node representations by aggregating information from neighboring nodes, so that each node not only contains its own features but also integrates the local dependencies of surrounding learners, resources, and knowledge points, thereby further enhancing the graph's structural feature representation capabilities. Learner embeddings, video embeddings, and knowledge point embeddings of the heterogeneous graph were obtained respectively. , and .

[0023] 2) Temporal preference evolution modeling stage First, the learner's historical interactions are represented as sequences. Each learning resource is associated with several knowledge points. Multi-modal features of the learning resources and their associated knowledge points are integrated into the sequence to capture fine-grained content associations and temporal relationships. Specifically, five types of embeddings are constructed: knowledge point semantic embedding, using the textual content representation of knowledge points extracted by the BERT model; video semantic embedding, obtained by average pooling multiple semantic embeddings of knowledge points associated with the video; location embedding, representing the different temporal positions of the video in the sequence; professional type embedding, used to distinguish knowledge points from different disciplines; and video-knowledge point relationship embedding, derived from the edge relationship vectors in the heterogeneous topology graph representation modeling stage, used to measure the difference in contribution of different knowledge points to the resources. After layer normalization, the final representation is input into the Longformer model. The sparse attention mechanism of Longformer, combined with local window attention and a small amount of global attention, is used to capture long-distance dependencies while maintaining the computational efficiency of sequence processing. Learner embeddings, video embeddings, and knowledge point embeddings for temporal modeling are obtained respectively. , and .

[0024] S2: Online Dialogue Interaction Representation Update Stage: First, real-time feedback data from learners' multi-turn dialogues is collected, covering knowledge point query preferences and recommended video feedback information; second, relying on the preference-aware attention mechanism, the real-time feedback is dynamically fused with the embedded representation generated in stage S1 to achieve synchronous updates of learners' current preference intentions and learning resource representations; finally, by iteratively executing the above fusion update process in the dialogue process, the matching accuracy between learning resources and learners' immediate needs is continuously optimized.

[0025] 1) Multi-round dialogue information feedback collection stage First, we obtain real-time feedback information generated by learners in each round of dialogue, mainly including knowledge point query preferences and recommended video feedback. Knowledge point query preferences refer to the set of knowledge point attributes that learners explicitly indicate are acceptable and the set of knowledge point attributes they are not interested in, denoted as […]. and Recommended video feedback information refers to the learner's immediate feedback on the videos recommended by the system, including acceptance and rejection, denoted as [insert values ​​here]. and .

[0026] 2) Dynamic Representation Fusion Stage Based on Preference-Aware Attention Mechanism To achieve deep fusion of the offline learner features obtained in S1 and the real-time learner feedback obtained in the previous sub-step, a preference-aware attention mechanism is used to update the representations of three types of entities—learners, resources, and knowledge points—in real time. Specifically, through... Compare the learner’s current set of acceptable attributes with the short-term heterogeneous graph preferences obtained in stage S1. With long-term sequence preferences By fusion, a dynamic representation of the learner's current dialogue turn can be obtained. ;pass For each candidate resource, construct its dynamic representation. This allows candidate video resources to reflect learners' latest learning intentions in real time; similarly, the same mechanism is used to dynamically update the representation of knowledge points. This enhances the distinguishability of knowledge points in the process of dialogue reasoning.

[0027] 3) Feedback-driven candidate resource scoring stage To achieve higher accuracy in resource set selection, a dual-mechanism scoring strategy is introduced based on dynamic representation, specifically including similarity-based resource score calculation and attribute information content evaluation based on weighted entropy. The similarity-based resource score calculation method is as follows ( , , (Adjustable hyperparameters)

[0028] The attribute information content evaluation mechanism based on weighted entropy is as follows:

[0029]

[0030] S3: Hierarchical Multi-Agent Collaborative Reasoning Stage: Building upon S2, a hierarchical multi-agent collaborative reasoning structure is further introduced to adapt to the dialogue preferences and knowledge needs of learners from different disciplinary backgrounds. First, multi-disciplinary dedicated recommendation agents are constructed based on learner interests. Each agent focuses on resources within its corresponding discipline and optimizes its strategy through a deep Q-network to maximize the cumulative value estimation within the domain. Subsequently, an adaptive recommendation agent is introduced as the upper-level agent. Its structure is consistent with the professional agents, but it does not directly rely on learner feedback. Instead, it integrates the experience pools of agents from different domains to update strategies, achieving cross-disciplinary knowledge transfer and experience absorption, forming a more generalized global strategy. To address the imbalance in professional interactions, an adaptive sampling mechanism is used to dynamically adjust the proportion of experience from each domain in joint training, avoiding strategy degradation caused by insufficient experience in less popular disciplines.

[0031] 1) Specific professional recommendations for training within the intelligent agent domain During the system initialization phase, all majors are divided into categories based on the learners' professional backgrounds. There are several categories. The system builds a dedicated, domain-specific recommendation agent for each category. Its training process includes state input construction, action selection decision-making, and optimization of the specialized Q-network loss function. The first step is state input construction, where the specialized agent, in each round of dialogue, constructs the state input based on the learner's dialogue history vector. The attribute information entropy vector in the current candidate video set and the current size of the video candidate set Constructing state vectors The calculation method is as follows Then comes the action selection decision; the agent, based on the output of the deep Q-network, makes decisions on two types of actions (attribute query). With video recommendations Value estimation and adoption The -greedy strategy achieves a balance between action space exploration and utilization. Finally, the loss function of the specialized Q-network is optimized. After each action is performed by the agent and feedback is received from the learner, a four-tuple experience is formed. Stored in the corresponding professional experience pool In this study, small-batch sampling experience is used to optimize the Q-network loss function.

[0032]

[0033]

[0034] 2) Adaptive recommendation agent cross-domain policy generalization learning To enable the system to generalize across professional scenarios, this sub-step constructs an adaptive recommendation agent. Compared to professional agents, adaptive agents maintain a consistent structure with the training objective, but do not directly rely on learners' online feedback as the source of experience for optimizing the policy network. Instead, they draw from all professional experience pools through a cross-experience pool sampling mechanism. Training samples are extracted from the dataset. To mitigate the experience imbalance caused by differences in the number of video interactions across different specialties, the adaptive recommendation agent is based on the number of professional experience samples. Total sample size N and class imbalance Calculate the dynamic sampling probability.

[0035]

[0036] in, The sampling weights are used to automatically balance the amount of data across different disciplines, ensuring that each discipline receives due attention during the training of the recommendation agent, thus enabling cross-disciplinary policy learning and transfer. The sampled quadruple experience... The global Q-network parameters used to update the adaptive agent are optimized by minimizing the loss function during training.

[0037]

[0038]

[0039] Example: This invention primarily addresses the problem of educational resource recommendation in the field of intelligent education technology. The method comprises four functional modules: viewing behavior regularization, dual-structure representation learning, dynamic learning resource matching, and hierarchical collaborative reasoning. First, starting from learners' viewing history, the interaction records collected on the platform are regularized and segmented to construct the heterogeneous information graph and time-series data required for subsequent processing. Then, through the dual-structure representation learning module, which jointly models the heterogeneous graph and sequence structure, a stable initial semantic representation is obtained. Further, during online human-computer collaborative dialogue, the dynamic learning resource matching module continuously receives real-time feedback from learners and constructs representations relevant to the current dialogue context based on a preference-aware attention mechanism. It dynamically adjusts the ranking of candidate resources by combining similarity scores and entropy-based attribute weights to adapt to the evolving learner intent throughout the interaction process. Finally, the hierarchical collaborative reasoning module integrates cross-domain experience through collaborative reasoning between multiple professional domain agents and adaptive agents, forming a unified recommendation decision strategy, thereby outputting more accurate teaching video recommendation results to learners.

[0040] Step 1: The actual viewing behavior data of learners on the platform was preprocessed uniformly. First, anomalies and missing records were filtered out (each learner must have watched at least 100 videos, each video must have been watched by at least 10 different learners, and each knowledge point must belong to at least 6 different video resources). The viewing sequences were then segmented and aligned, resulting in 8051 learners, 17420 learning video resources, and 4977 knowledge point attributes. Subsequently, a heterogeneous infographic and time-series data containing "learner—video resource—knowledge point attribute" were constructed to ensure the subsequent training of representations. A total of 672735 valid sequences were obtained, with each resource associated with an average of 8.04 attributes, and each learner completing an average of 139.63 resource views.

[0041] Step Two: The dual-structure representation learning module employs a dual-structure collaborative learning mechanism combining graph neural networks and sequence models to simultaneously characterize learners' long-term learning preferences and short-term learning needs. First, the graph neural network propagates resource attributes, knowledge point associations, and learner preference information within the graph, obtaining short-term semantic embeddings of learners and resources. Then, a Longformer-based sequence encoder is used to sequentially model learners' long-term viewing behavior segments, obtaining their long-term learning preference vector.

[0042] Step 3: Dynamic learning resource matching. During online interaction, the ranking of candidate resources is continuously adjusted based on learners' real-time feedback. First, a preference-aware attention mechanism is used to calculate interaction attention weights by combining the learner's current contextual intent with the candidate resource representations. This yields resource scores based on similarity and attribute information based on weighted entropy, automatically increasing the contribution of highly discriminative resource attributes in the ranking. Finally, candidate resources are ranked according to their comprehensive similarity scores.

[0043] Step 4: Hierarchical Collaborative Reasoning. After obtaining the dynamic candidate set, the hierarchical collaborative reasoning module automatically invokes the corresponding domain agent based on the learner's professional category, and estimates the action value by combining the current state vector. Simultaneously, the adaptive agent supplements the estimation of these candidate resources based on cross-domain experience to enhance generalization ability. Finally, a comprehensive reasoning score is obtained for each candidate resource, and the top 3 learning resources with the highest scores are selected as the final recommendation results.

[0044] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A human-computer collaborative educational resource recommendation method for large-scale online learning, characterized in that, The method specifically includes the following steps: S1. Offline learner behavior representation stage: The interaction data between learners and learning resources is synchronously constructed into graph structure and sequence structure, and fine-grained dependencies and temporal evolution rules are extracted respectively to obtain the offline embedded representation of learners; S2, Online Dialogue Interaction Representation Update Stage: Obtain real-time feedback information from learners in multi-turn dialogues, dynamically fuse the real-time feedback with the offline embedded representation based on the preference-aware attention mechanism, and update the representations of learners and learning resources; S3, Hierarchical Multi-Agent Collaborative Reasoning Stage: Construct multiple professional recommendation agents and one adaptive recommendation agent to collaboratively generate and output personalized teaching video recommendation results based on the learner's professional background and dialogue state.

2. The human-computer collaborative educational resource recommendation method for large-scale online learning according to claim 1, characterized in that, Step S1 specifically includes: S11. Simultaneously organize the learner-learning resource interaction data into both graph and sequence structures; S12. Employ graph-based representation learning methods to extract fine-grained dependencies on graphs to characterize learners' short-term dynamic preferences; S13. A sequence-based temporal pattern modeling method is used to perform temporal learning on learners' historical interaction sequences in order to capture long-term preferences and their evolutionary patterns, providing a basic representation for the online dialogue interaction reasoning stage.

3. The human-computer collaborative educational resource recommendation method for large-scale online learning according to claim 2, characterized in that, Step S2 specifically includes: S21. Obtain real-time feedback information from learners during each round of dialogue. The real-time feedback is provided by learners during their interaction with the system, including expressions of interest in the queried knowledge points and feedback on acceptance or rejection of recommended videos. S22. The learner feedback is used as a human-machine collaboration signal. Based on the preference-aware attention mechanism, the real-time feedback is dynamically fused with the embedded representation obtained in step S1 to collaboratively update the learner's current intention representation and learning resource representation. S23. By iteratively executing the above collaborative update process during multi-round dialogue interaction, the recommendation system can continuously perceive changes in learners' intentions and continuously improve the matching degree between learning resources and learners' real-time learning needs.

4. The human-computer collaborative educational resource recommendation method for large-scale online learning according to claim 3, characterized in that, Step S3 specifically includes: S31. Construct multiple recommendation agents for specific majors based on learners' academic background information. Each recommendation agent for a specific major is responsible for mining learners' preferences in that major field and learning corresponding recommendation strategies under the guidance of learner interaction feedback. S32. Construct an adaptive recommendation agent to receive and integrate the dialogue recommendation strategies formed by various professional recommendation agents during the collaborative reasoning process. S33. The adaptive recommendation agent dynamically integrates and learns the strategies of various professional agents based on the learner's cross-disciplinary learning needs during the dialogue process, forming a unified recommendation decision. S34. Output instructional video recommendation results to learners based on the integrated recommendation strategy.

5. The human-computer collaborative educational resource recommendation method for large-scale online learning according to claim 4, characterized in that, In step S1, the construction of the offline learner behavior representation includes two parts: graph structure modeling and sequence structure modeling. First, the relationship between learners, video resources, and the knowledge point attributes contained in the video is modeled as a heterogeneous interaction graph, where learners, video resources, and knowledge points are nodes, and other interaction behaviors are edges. Triples are constructed to store the structural dependencies between nodes. Second, the constructed interaction graph is encoded using a TransD-based graph embedding method to obtain initial node representations. Furthermore, a graph convolutional neural network is used to capture the fine-grained topological preference dependencies of learners in the heterogeneous interaction graph, enabling the system to quickly respond to changes in learners' short-term viewing behavior. In addition, the learner's historical interaction sequence is input into the Longformer sequence model, which uses its long-sequence time-dependent attention mechanism to learn the characteristics of learner preferences evolving over time, thereby representing the learner's long-term preference patterns. Through the joint modeling of the above graph representation and sequence representation, a comprehensive offline embedding representation is formed, providing robust basic feature support for the online dialogue stage.

6. The human-computer collaborative educational resource recommendation method for large-scale online learning according to claim 5, characterized in that, In step S2, the implementation of the online dialogue interaction representation update includes two steps: real-time feedback modeling and preference fusion. First, through multi-round dialogue interaction, the learner's real-time feedback information generated during the collaboration with the recommendation system is collected, including the expression of preference for the queried knowledge points and the acceptance or rejection of the recommended videos. Subsequently, the feedback is used as an explicit constraint on the learning intention during the human-computer collaboration process. The learner's dialogue state representation is updated in real time, and the updated representation is input into the preference perception attention mechanism, so that it is dynamically fused with the offline preference embedding representation obtained in S1. The contribution of short-term interest and long-term preference in the current collaborative intention representation and its correlation with learning resources are adjusted by attention weight. As the dialogue progresses, the aforementioned fusion and update process is iteratively executed, enabling learner preference representations and learning resource representations to gradually align with the learner's evolving needs and intentions through continuous interaction, thereby achieving dynamic teaching video matching driven by human-computer collaboration.

7. The human-computer collaborative educational resource recommendation method for large-scale online learning according to claim 5, characterized in that, In step S3, the implementation of hierarchical multi-agent collaborative reasoning includes a human-machine collaborative decision-making process between a specific professional recommendation agent and an adaptive recommendation agent: First, multiple specific professional recommendation agents are constructed based on the learner's academic background information. Each specific professional agent, guided by the learner's interactive feedback and dialogue intent, focuses on analyzing the learner's interests and preferences in the corresponding professional field and learns the corresponding recommendation strategy to explore the potential demand for learning resources in different professional fields. Subsequently, an adaptive recommendation agent is constructed to receive the recommendation strategies generated by each specific professional agent during the collaborative reasoning process, and to integrate and adjust the weights of the strategies from different professional fields in conjunction with the learner's current dialogue state to adapt to cross-professional learning scenarios. Ultimately, the integrated collaborative strategy delivers personalized instructional video recommendations to learners that align with their current learning objectives.

8. A human-computer collaborative educational resource recommendation system for large-scale online learning, characterized in that, The system employs the method described in any one of claims 1 to 7.

9. An electronic device, characterized in that, It includes a processor and a memory, the memory storing a computer program that, when executed by the processor, implements the method as described in any one of claims 1-7.

Images