Adaptive tutorial recommendation method based on large model intention reasoning and diffusion generation

Abstract

The present application relates to a kind of self-adapting learning recommendation method based on large model intention reasoning and diffusion generation, comprising: S1, multi-modal multi-large language model driven student learning intention and cognitive state reasoning;S2, adaptive knowledge perception's learning map structure topology reconstruction;S3, semantic priori guided conditional diffusion ID embedding generation;S4, heat perception fusion and multi-task adaptive recommendation prediction.The present application is aimed at the existing adaptive learning system in the face of student behavior extremely sparse, new warehouse education resource zero interaction, the cognitive diagnosis caused by inaccuracy, teaching topology is not strict, and new and old resource representation space mismatch etc.

Description

Technical Field

[0001] This application belongs to the field of data mining and intelligent information processing technology, specifically involving an adaptive guided recommendation method based on large model intent reasoning and diffusion generation. Background Technology

[0002] With the popularization of digital education and personalized learning, adaptive learning recommendation systems have become a core technology in the field of intelligent education (EdTech). The core objective of this system is to dynamically and accurately recommend the most suitable teaching resources (such as exercises, micro-lessons, concept explanation videos, etc.) for students' cognitive level and knowledge gaps based on their historical learning behavior and answer records, thereby achieving "teaching according to aptitude".

[0003] Currently, the development of this technology field has mainly gone through the following stages and technical paths: Traditional collaborative filtering and matrix factorization techniques: By mining the historical interaction matrix between students and educational resources (such as correct or incorrect answers, browsing time, etc.), similar students or similar questions are found, and recommendations are made accordingly.

[0004] Graph recommendation technology based on graph neural networks: Given the complex interactions and relationships between educational resources and students, researchers have widely introduced graph neural networks to abstract students and resources into bipartite or heterogeneous graphs. By using the message passing mechanism of graph convolution, high-order collaborative filtering signals can be explicitly captured, thereby improving the accuracy of recommendations.

[0005] Combining cognitive diagnostic methods: Analyze students' answers using cognitive diagnostic models (such as DINA, IRT, etc.) to infer the probability of students' mastery of each sub-knowledge point (i.e., cognitive state vector), and use this vector as the pre-input or regularization constraint of the recommendation system.

[0006] Exploration of the integration of multimodal and large language models (LLM): With the development of multimodal large models, some studies have begun to try to extract multimodal features from exercises (such as illustrations in mathematical geometry problems, text of physics formulas, audio and video of micro-lessons, etc.), or to use the powerful semantic understanding and logical reasoning capabilities of large language models (LLM) to help extract semantic representations of resources and provide semantic support for precise learning guidance.

[0007] While the aforementioned technical approaches have achieved some success in conventional recommendation scenarios, existing technologies still suffer from the following serious shortcomings and bottlenecks that urgently need to be addressed when dealing with "cold-start in adaptive learning" and "dynamic adaptive path evolution": 1. The persistent problem of "extreme sparseness and ambiguity" in the diagnosis of new students' cognitive intentions In actual teaching scenarios, newly registered students or students who have just transferred to a completely new subject (such as a new first-year high school student transferring from junior high school mathematics to high school mathematics) have extremely sparse historical interaction behavior in the system (often only 1-2 wrong questions or a few seconds of click records).

[0008] Failure of traditional collaborative filtering and graph network methods: Due to the severe lack of collaborative signals, graph neural networks cannot establish effective topological relationships through message passing mechanisms, resulting in the degradation or collapse of the calculated student representations (embeddings) and the inability to capture their true preferences.

[0009] Limitations of cognitive diagnostic models: Traditional cognitive diagnostic models (such as DINA and IRT) heavily rely on a large number of answer sequences to approximate students' true cognitive level. In a cold start state, sparse 1-2 records are simply insufficient for the diagnostic model to converge, causing the system to be unable to identify the student's true "deep knowledge blind spots" and "underlying learning motivations" (for example, if a student makes a mistake on a composite function problem, is it because they don't understand the basic concepts or because they lack prior quadratic function solving skills? The system cannot perform deep causal reasoning based on shallow interactions).

[0010] 2. The problems of "semantic drift" and "logical discontinuity" in educational resources To mitigate the cold start effect, existing technologies typically construct resource-to-resource similarity graphs based on the content features of resources (such as word vector similarity of exercise texts and visual features of micro-lecture videos). However, educational resources have a strong scientific rigor and strict pedagogical logic, and this graph construction method leads to serious drawbacks: Semantic drift phenomenon: Traditional contrastive learning or feature embedding methods often align based on surface word frequency or general multimodal representations (e.g., if two questions both mention "apple," the system considers them similar). This message passing based on surface semantic similarity introduces a large amount of noise unrelated to the teaching function, resulting in "semantic drift" after propagation in the graph convolutional network, making it impossible to accurately express the teaching logic.

[0011] A disconnect exists in the teaching logic: Educational resources exhibit strict "pre-requisite relations" and "complementary teaching functions" (e.g., learning conic sections requires prior knowledge of coordinate systems; micro-lesson A introduces the concept, while exercise B reinforces the formula). Existing technologies often lack the ability to reconstruct this rigorous teaching topology, failing to generate a noise-free topological structure that conforms to the principles of scientifically guided learning.

[0012] 3. Barriers of "Inconsistency and Misalignment" in the Characterization Space of Cold and Hot Resources The adaptive system must continuously inject newly developed exercises and newly recorded micro-lessons. This results in the system simultaneously containing a large number of "high-popularity old resources (with tens of thousands of student answer records)" and "new resources with zero interaction (only initial multimodal content features)".

[0013] Spatial heterogeneity: The representation of highly popular resources is dominated by a strong collaborative signal from student groups; while the representation of cold-start new resources is entirely determined by the initial content features. These two belong to completely different modal spaces.

[0014] Recommendation bias imbalance: Existing methods, when integrating these two types of resources, often simply perform linear concatenation or mapping, failing to explicitly bridge the significant gap between the "collaboration space" and the "content space." This leads to a severe underestimation of the score for new resources when calculating similarity predictions. The system tends to continuously recommend popular, repetitive questions, rather than adaptively pushing high-quality, highly relevant new questions or micro-lessons to the students who need them most.

[0015] In summary, the key technical bottlenecks that urgently need to be overcome in the field of adaptive learning are: how to achieve chain-like logical reasoning of deep cognitive intentions and knowledge blind spots under extremely sparse student interaction behaviors; how to eliminate semantic drift caused by content mapping of educational resources and reconstruct scientific learning topology; and how to deeply align cold and hot resources across space and generate representations that conform to collaborative response distribution. Summary of the Invention

[0016] The purpose of this application is to overcome the shortcomings of the existing technology and provide an adaptive learning recommendation method based on large model intention reasoning and diffusion generation. This method addresses the persistent problems of existing adaptive learning systems, such as inaccurate cognitive diagnosis, imprecise teaching topology, and mismatch between the representation space of new and old resources, when faced with extremely sparse student behavior and zero interaction with newly added educational resources.

[0017] The technical problem solved by this application is achieved through the following technical solution: An adaptive guided recommendation method based on large model intent reasoning and diffusion generation is proposed. The architecture of the method is divided into a data preparation and preprocessing layer, a multimodal multi-language model semantic logic reasoning layer, a knowledge topology graph structure reconstruction layer, a conditional diffusion ID embedding generation layer, and a popularity-aware multi-task recommendation layer. The steps of the method are as follows: S1, a multimodal, multi-language model-driven reasoning of student learning intentions and cognitive states, solves the problem of causal cognitive diagnosis when new students' interactive behaviors are extremely sparse; 1) Transformation of multimodal educational resources into teaching functions, for newly added exercises or micro-lessons. It includes illustrations for multimodal mathematical geometry problems. Text formula description First, it is input into a pre-trained multimodal large language model (MLLM). Through fixed functional transformation prompts (Prompt), the multimodal features are transformed into unified functional teaching text descriptions focusing on teaching attributes. : ; The description explicitly includes the core knowledge points to be tested, the required abstract thinking ability, the prerequisite concepts, and common pitfalls. 2) Based on the reasoning of students' learning intentions through the thought chain, students currently in a cold start state are defined as... Its historical sparse interaction sequence is It also includes error markers; extracts resources from the sequence. Constructing thought chain prompts for cognitive diagnosis The prompt word is concatenated with the sequence and input into the large language model LLM, which then guides the model to perform the following chain-like causal inference: Observation phase: Analyze which multimodal resources students interacted with negatively; Diagnostic phase: Analyze the teaching function descriptions corresponding to these wrong questions, and deduce the underlying unmastered prerequisite concepts they all point to; Deduction phase: Predict the student's current virtual cognitive profile and deep-seated learning motivation; LLM inference outputs a semantic description of the student's learning needs. Subsequently, a pre-trained, efficient text encoder was used to map it into a low-dimensional dense vector, which served as the semantic vector representing the student's learning needs. : ; S2. Adaptive knowledge-aware learning graph structure topology reconstruction, eliminating noise introduced by general semantic graph construction, and constructing a science teaching topology graph without semantic drift; 1) Logical discrimination aided by external knowledge of the large model: To determine the strict educational logic between new resources and existing high-popularity anchor resources within the system, a large language model (LLM) is used to perform bidirectional logical evaluation on resource pairs (i, j), with input prompt words. Does the question ask whether there is a "prerequisite knowledge dependency learning" between the two questions? You must first master The relationship between "or 'strong complementary teaching functions'" is used to obtain the explicit logical association score from the output of the large model. ; 2) Graph structure topology denoising and reconstruction: Combining cosine similarity with multimodal features of new resources Logical score of large model Calculate the final integrated logical edge weights. To completely remove semantic drift noise in the graph, an adaptive threshold gating function is designed. Reconstructing an adaptive learning topology for knowledge perception : ; ; in: For resource-based The noise threshold is adaptively calculated based on the connection weight distribution. S3, semantic prior-guided conditional diffusion ID embedding generation, converts semantic signals in the content space into ID representations in the collaborative response space of the recommendation system, enabling the "breakthrough" of cold-start new resources; 1) Extraction of collaborative graph structure representation to enhance the reconstructed learning knowledge graph. Heterogeneous fusion with a bipartite graph composed of real student-resource interactions is performed, and the lightweight graph convolutional network LightGCN is used to perform high-order message passing of adaptive interest and cognitive signals on the graph structure: ; go through After layer aggregation, old resource collaborative ID embeddings containing local topology enhancement features are extracted. And the prior representation of the topology of cold start resources. ; 2) Conditional diffusion model reverse denoising generation, embedding the real collaborative ID of high-popularity resources. As the endpoint distribution of the forward diffusion process, the student guidance demand semantic vector output as S1 in the reverse denoising generation stage. and the a priori representation of the topology derived from S2 Assemble into a joint strong guidance condition Inverse denoising network Receive arbitrary Gaussian noise Under the conditions Under explicit constraints and physical guidance, T-step iterative denoising is performed. The denoising network minimizes the noise prediction loss function. Perform parameter optimization: When iterating to At that time, the x_0 output by the diffusion model is a high-fidelity virtual collaborative ID embedded that is explicitly generated specifically for cold-starting new resources and conforms to the current student group's real collaborative cognitive distribution. ; S4, Heat-Aware Fusion and Multi-Task Adaptive Recommendation Prediction 1) Adaptive heat-gating fusion mechanism: To enable the system to simultaneously handle massive amounts of hot resources and new resources with zero interaction, a mechanism based on interaction frequency is designed. Adaptive soft transition gating: ; ; in: The Sigmoid activation function enables complete trust in the graph-learned collaborative representation of high-popularity old resources, and complete trust in the virtual collaborative representation generated by diffusion for zero-interaction new resources. 2) Joint multi-task loss for end-to-end training: To force a closer relationship between the deep teaching semantic space of large model inference and the collaborative structural space of graph networks, an information bottleneck semantic-structural contrast alignment loss is introduced. : ; Finally, the Bayesian Personalized Ranking Main Loss (BPR) and the Diffuse Noise Prediction Loss are combined. and comparative loss Perform joint multi-task, multi-objective end-to-end optimization: ; Prediction phase calculates target student representation Final representation of all candidate educational resources The inner product of the two is used to generate an adaptive tutoring recommendation list and distribute it to students.

[0018] Compared with the prior art, the present invention has the following beneficial effects: 1. Completely overcome the blind spot of cognitive diagnosis in the cold start state of students, realize precise adaptive guidance of "deepening with less", and overcome the limitations of traditional methods: traditional collaborative filtering or cognitive diagnosis models rely heavily on a large number of long-term question-answering interaction sequences to approximate the students' knowledge mastery state. When facing new students (cold start scenario), the model usually cannot converge due to the extreme sparsity of data.

[0019] This invention, by introducing a multimodal large language model-based thought chain reasoning technique, can deeply analyze and uncover the underlying causal logic of a new student's behavior from only a very small number of 1-2 historical incorrect answers or browsing clues, going beyond superficial word frequency or question categories. The system can accurately infer the student's current "virtual cognitive profile" and "underlying knowledge blind spots," and accurately generate personalized semantic vectors of the student's learning needs under shallow interaction. This results in a significant improvement in the system's adaptive recommendation accuracy (indices such as Recall@K and NDCG@K) under cold-start conditions compared to existing best-in-class baseline models.

[0020] 2. It fundamentally eliminates the "semantic drift" contamination caused by general semantic graph construction, ensuring the scientific rigor of the teaching and learning path; it overcomes the limitations of traditional methods: when faced with new educational resources (new questions, new micro-lessons) without interaction, existing technologies usually construct "resource-resource" graphs based on text keywords or general multimodal feature similarity. This approach is highly susceptible to introducing low-noise noise, such as connecting "red apple" and "red billiard ball" due to surface visual similarity, causing severe "semantic drift" in the graph convolutional neural network during message transmission.

[0021] This invention designs a "knowledge-aware adaptive learning topology reconstruction mechanism." By combining a large language model to explicitly discriminate educational logics such as strict "prerequisite knowledge dependency (e.g., learning B requires learning A first)" and "strong complementarity of teaching functions" between questions and micro-lessons, and with adaptive threshold denoising gating, a logically rigorous and semantically drift-free clean teaching topology is reconstructed. This provides high-value path constraints for message passing in LightGCN, avoiding the blind cross-disciplinary and cross-difficulty cross-contamination of teaching resources, and ensuring the scientific rigor and seriousness of the adaptive learning path.

[0022] 3. It perfectly breaks the heterogeneity barrier of the representation space of hot and cold resources, and realizes the rapid adaptive distribution of newly added educational resources; it overcomes the limitations of traditional methods: in the adaptive system, popular old questions have a strong ID space representation dominated by student collaborative answering signals, while new questions with zero interaction only have a content space representation. The heterogeneity between the two leads to the severe underestimation of the scores of newly added high-quality questions, and the system will fall into a vicious cycle of continuously recommending popular old questions and repeated questions.

[0023] This invention departs from the traditional simple feature splicing and fusion approach. Instead, it innovatively injects the deep intent and knowledge topology inferred from a large model as strong guiding conditions into the conditional diffusion model. Through a reverse denoising process, it explicitly "generates" high-quality virtual ID embeddings for cold-start new resources that conform to the collaborative cognitive distribution of the real student population. Combined with an adaptive heat-gated fusion mechanism, the new resources are seamlessly and fairly integrated into the existing recommendation vector space and accurately aligned with the target students across spatial depths.

[0024] 4. The invention integrates multi-task joint optimization design, balancing system engineering convergence with the efficiency of online inference. It introduces an information bottleneck (semantic-structural contrast alignment loss), maximizing the mutual information between the collaborative structural representation learned by the graph network and the deep learning requirement semantic representation generated by large-scale model inference. This explicit alignment objective, combined with the implicit distributed learning of the diffusion model, significantly accelerates the end-to-end training convergence speed. Simultaneously, high-energy-consuming and high-latency operations such as large-scale model inference and logical semantic graph construction are completed in the offline preprocessing stage or asynchronous cache queue. Online recommendation computation only requires low-latency, low-dimensional dense vector dot product operations, greatly improving cold-start recommendation performance while perfectly meeting the low-latency and high-concurrency requirements of industrial-grade large-scale online adaptive learning systems. Attached Figure Description

[0025] Figure 1 This is a flowchart of the process architecture of the present invention. Detailed Implementation

[0026] The present application will be further described in detail below through specific embodiments. The following embodiments are merely descriptive and not limiting, and should not be used to limit the scope of protection of the present application.

[0027] like Figure 1 The meanings of the numbers shown are as follows: 1: Input terminal for student history error questions and history learning behavior sequence (corresponding to extremely sparse initial answer data during cold start) ); 2: A cognitive diagnosis and personalized learning intention reasoning module based on the Large Language Model (LLM) driven thinking chain; 3: Knowledge-aware adaptive topology reconstruction module (outputs a denoised heterogeneous graph of knowledge dependency relationships in the learning process) ); 4: Semantic vectors of personalized learning needs generated by large-scale model inference Storage unit; 5: A high-order collaborative interest and cognitive signal propagation feature extraction layer based on a lightweight graph convolutional network; 6: Conditional diffusion model inverse denoising explicit generator; 7: Dynamically adaptive heat-gated fusion layer (used to eliminate representation bias between cold-start resources and high-heat resources); 8: Embedding of virtual collaborative IDs of new cold-start resources explicitly output by the diffusion generator under joint conditional constraints. ; 9: Output of a personalized list of adaptive learning exercises and micro-lesson resources recommended for the target students; An adaptive guided learning recommendation method based on large model intent reasoning and diffusion generation is innovative in that the architecture of the method is divided into a data preparation and preprocessing layer, a multimodal multi-language model semantic logic reasoning layer, a knowledge topology graph structure reconstruction layer, a conditional diffusion ID embedding generation layer, and a popularity-aware multi-task recommendation layer. The steps of the method are as follows: S1, a multimodal, multi-language model-driven reasoning of student learning intentions and cognitive states, solves the problem of causal cognitive diagnosis when new students' interactive behaviors are extremely sparse; 1) Transformation of multimodal educational resources into teaching functions, for newly added exercises or micro-lessons. It includes illustrations for multimodal mathematical geometry problems. Text formula description First, it is input into a pre-trained multimodal large language model (MLLM). Through fixed functional transformation prompts (Prompt), the multimodal features are transformed into unified functional teaching text descriptions focusing on teaching attributes. : ; The description explicitly includes the core knowledge points to be tested, the required abstract thinking ability, the prerequisite concepts, and common pitfalls. 2) Based on the reasoning of students' learning intentions through the thought chain, students currently in a cold start state are defined as... Its historical sparse interaction sequence is It also includes error markers; extracts resources from the sequence. Constructing thought chain prompts for cognitive diagnosis The prompt word is concatenated with the sequence and input into the large language model LLM, which then guides the model to perform the following chain-like causal inference: Observation phase: Analyze which multimodal resources students interacted with negatively; Diagnostic phase: Analyze the teaching function descriptions corresponding to these wrong questions, and deduce the underlying unmastered prerequisite concepts they all point to; Deduction phase: Predict the student's current virtual cognitive profile and deep-seated learning motivation; LLM inference outputs a semantic description of the student's learning needs. Subsequently, a pre-trained, efficient text encoder was used to map it into a low-dimensional dense vector, which served as the semantic vector representing the student's learning needs. : ; S2. Adaptive knowledge-aware learning graph structure topology reconstruction, eliminating noise introduced by general semantic graph construction, and constructing a science teaching topology graph without semantic drift; 1) Logical discrimination aided by external knowledge of the large model: To determine the strict educational logic between new resources and existing high-popularity anchor resources within the system, a large language model (LLM) is used to perform bidirectional logical evaluation on resource pairs (i, j), with input prompt words. Does the question ask whether there is a "prerequisite knowledge dependency learning" between the two questions? You must first master The relationship between "or 'strong complementary teaching functions'" is used to obtain the explicit logical association score from the output of the large model. ; 2) Graph structure topology denoising and reconstruction: Combining cosine similarity with multimodal features of new resources Logical score of large model Calculate the final integrated logical edge weights. To completely remove semantic drift noise in the graph, an adaptive threshold gating function is designed. Reconstructing an adaptive learning topology for knowledge perception : ; ; in: For resource-based The noise threshold is adaptively calculated based on the connection weight distribution. S3, semantic prior-guided conditional diffusion ID embedding generation, converts semantic signals in the content space into ID representations in the collaborative response space of the recommendation system, enabling the "breakthrough" of cold-start new resources; 1) Extraction of collaborative graph structure representation to enhance the reconstructed learning knowledge graph. Heterogeneous fusion with a bipartite graph composed of real student-resource interactions is performed, and the lightweight graph convolutional network LightGCN is used to perform high-order message passing of adaptive interest and cognitive signals on the graph structure: ; go through After layer aggregation, old resource collaborative ID embeddings containing local topology enhancement features are extracted. And the prior representation of the topology of cold start resources. ; 2) Conditional diffusion model reverse denoising generation, embedding the real collaborative ID of high-popularity resources. As the endpoint distribution of the forward diffusion process, the student guidance demand semantic vector output as S1 in the reverse denoising generation stage. and the a priori representation of the topology derived from S2 Assemble into a joint strong guidance condition Inverse denoising network Receive arbitrary Gaussian noise Under the conditions Under explicit constraints and physical guidance, T-step iterative denoising is performed. The denoising network minimizes the noise prediction loss function. Perform parameter optimization: When iterating to At that time, the x_0 output by the diffusion model is a high-fidelity virtual collaborative ID embedded that is explicitly generated specifically for cold-starting new resources and conforms to the current student group's real collaborative cognitive distribution. ; S4, Heat-Aware Fusion and Multi-Task Adaptive Recommendation Prediction 1) Adaptive heat-gating fusion mechanism: To enable the system to simultaneously handle massive amounts of hot resources and new resources with zero interaction, a mechanism based on interaction frequency is designed. Adaptive soft transition gating: ; ; in: The Sigmoid activation function enables complete trust in the graph-learned collaborative representation of high-popularity old resources, and complete trust in the virtual collaborative representation generated by diffusion for zero-interaction new resources. 2) Joint multi-task loss for end-to-end training: To force a closer relationship between the deep teaching semantic space of large model inference and the collaborative structural space of graph networks, an information bottleneck semantic-structural contrast alignment loss is introduced. : ; Finally, the Bayesian Personalized Ranking Main Loss (BPR) and the Diffuse Noise Prediction Loss are combined. and comparative loss Perform joint multi-task, multi-objective end-to-end optimization: ; Prediction phase calculates target student representation Final representation of all candidate educational resources The inner product of the two is used to generate an adaptive tutoring recommendation list and distribute it to students.

[0028] Although embodiments and drawings of the present invention have been disclosed for illustrative purposes, those skilled in the art will understand that various substitutions, variations and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the scope of the present invention is not limited to the contents disclosed in the embodiments and drawings.

Claims

1. An adaptive guided recommendation method based on large-model intent reasoning and diffusion generation, characterized in that: The architecture of the method is divided into a data preparation and preprocessing layer, a multimodal multi-language model semantic logic reasoning layer, a knowledge topology graph structure reconstruction layer, a conditional diffusion ID embedding generation layer, and a popularity-aware multi-task recommendation layer. The steps of the method are as follows: S1, a multimodal, multi-language model-driven reasoning of student learning intentions and cognitive states, solves the problem of causal cognitive diagnosis when new students' interactive behaviors are extremely sparse; 1) Transformation of multimodal educational resources into teaching functions, for newly added exercises or micro-lessons. It includes illustrations for multimodal mathematical geometry problems. Text formula description First, it is input into a pre-trained multimodal large language model (MLLM). Through fixed functional transformation prompts (Prompt), the multimodal features are transformed into unified functional teaching text descriptions focusing on teaching attributes. : ; The description explicitly includes the core knowledge points to be tested, the required abstract thinking ability, the prerequisite concepts, and common pitfalls. 2) Based on the reasoning of students' learning intentions through the thought chain, students currently in a cold start state are defined as... Its historical sparse interaction sequence is It also includes error markers; Extracting each resource from the sequence Constructing thought chain prompts for cognitive diagnosis The prompt word is concatenated with the sequence and input into the large language model LLM, which then guides the model to perform the following chain-like causal inference: Observation phase: Analyze which multimodal resources students interacted with negatively; Diagnostic phase: Analyze the teaching function descriptions corresponding to these wrong questions, and deduce the underlying unmastered prerequisite concepts they all point to; Deduction phase: Predict the student's current virtual cognitive profile and deep-seated learning motivation; LLM inference outputs a semantic description of the student's learning needs. Subsequently, a pre-trained, efficient text encoder was used to map it into a low-dimensional dense vector, which served as the semantic vector representing the student's learning needs. : ; S2. Adaptive knowledge-aware learning graph structure topology reconstruction, eliminating noise introduced by general semantic graph construction, and constructing a science teaching topology graph without semantic drift; 1) Logical discrimination aided by external knowledge of the large model: To determine the strict educational logic between new resources and existing high-popularity anchor resources within the system, a large language model (LLM) is used to perform bidirectional logical evaluation on resource pairs (i, j), with input prompt words. Does the question ask whether there is a "prerequisite knowledge dependency" between the two questions? You must first master The relationship of "or strong complementarity in teaching functions" is used to obtain the explicit logical association score of the large model output. ; 2) Graph structure topology denoising and reconstruction: Combining cosine similarity with multimodal features of new resources Logical score of large model Calculate the final integrated logical edge weights. To completely remove semantic drift noise in the graph, an adaptive threshold gating function is designed. Reconstructing an adaptive learning topology for knowledge perception : ; ； in: For resource-based The noise threshold is adaptively calculated based on the connection weight distribution. S3, semantic prior-guided conditional diffusion ID embedding generation, converts semantic signals in the content space into ID representations in the collaborative response space of the recommendation system, enabling "breaking the cold" for cold-starting new resources; 1) Extraction of collaborative graph structure representation to enhance the reconstructed learning knowledge graph. Heterogeneous fusion with a bipartite graph composed of real student-resource interactions is performed, and the lightweight graph convolutional network LightGCN is used to perform high-order message passing of adaptive interest and cognitive signals on the graph structure: ; go through After layer aggregation, old resource collaborative ID embeddings containing local topology enhancement features are extracted. And the prior representation of the topology of cold start resources. ; 2) Conditional diffusion model reverse denoising generation, embedding the real collaborative ID of high-popularity resources. As the endpoint distribution of the forward diffusion process, the student guidance demand semantic vector output as S1 in the reverse denoising generation stage. and the a priori representation of the topology derived from S2 Assemble into a joint strong guidance condition Inverse denoising network Receive arbitrary Gaussian noise Under the conditions Under explicit constraints and physical guidance, T-step iterative denoising is performed. The denoising network minimizes the noise prediction loss function. Perform parameter optimization: When iterating to At that time, the x_0 output by the diffusion model is a high-fidelity virtual collaborative ID embedded that is explicitly generated specifically for cold-starting new resources and conforms to the current student group's real collaborative cognitive distribution. ; S4, Heat-Aware Fusion and Multi-Task Adaptive Recommendation Prediction 1) Adaptive heat-gating fusion mechanism: To enable the system to simultaneously handle massive amounts of hot resources and new resources with zero interaction, a mechanism based on interaction frequency is designed. Adaptive soft transition gating: ; ; in: The Sigmoid activation function enables complete trust in the graph-learned collaborative representation of high-popularity old resources, and complete trust in the virtual collaborative representation generated by diffusion for zero-interaction new resources. 2) Joint multi-task loss for end-to-end training: To force a closer relationship between the deep teaching semantic space of large model inference and the collaborative structural space of graph networks, an information bottleneck semantic-structural contrast alignment loss is introduced. : ; Finally, the Bayesian Personalized Ranking Main Loss (BPR) and the Diffuse Noise Prediction Loss are combined. and comparative loss Perform joint multi-task, multi-objective end-to-end optimization: ; Prediction phase calculates target student representation Final representation of all candidate educational resources The inner product of the two is used to generate an adaptive tutoring recommendation list and distribute it to students.

Images