A knowledge-driven personalized resource recommendation method and system for a collaborative learning space

Abstract

The application is a knowledge-driven personalized resource recommendation method and system for collaborative learning space, belonging to the field of intelligent education technology.The method comprises the following steps: acquiring multi-modal learning resources and preprocessing;extracting knowledge concepts based on a double-agent cooperation mechanism, and constructing a learning resource vector library by using contrastive learning;realizing concept perception knowledge tracking by fusing cognitive rules, combining a time decay mechanism, a behavior perception gating mechanism and a item response theory model, and dynamically modeling and updating the knowledge state of learners;and realizing personalized recommendation based on reinforcement learning and cross-modal retrieval.The system constructs corresponding functional modules to realize the method.The application alleviates the problems in the prior art, such as the existence of field illusion in knowledge extraction, rough cognitive modeling and insufficient interpretability, and the lack of long-term planning and single form in resource recommendation, and can realize knowledge state-driven and support multi-modal resource collaborative personalized recommendation.

Description

Technical Field

[0001] This invention belongs to the field of intelligent education technology, and relates to educational data mining and personalized resource recommendation technology, specifically to a knowledge-driven personalized resource recommendation method and system for collaborative learning spaces. Background Technology

[0002] With the deepening development of educational informatization, collaborative learning spaces that deeply integrate online and offline learning have become an important carrier for daily teaching in universities. By integrating online learning platforms, offline smart classrooms, and various teaching tools, collaborative learning spaces can support the synchronous or asynchronous conduct of various teaching activities, including classroom teaching, self-directed learning, and collaborative learning. During the teaching process, collaborative learning spaces not only continuously collect behavioral log data generated by learners but also accumulate a large amount of multimodal learning resources, including heterogeneous data such as mixed text and image exercises, teaching materials, and classroom audio and video recordings. To provide learners with targeted guidance and support, it is necessary to perform semantic analysis of these heterogeneous resources at the knowledge concept level and, combined with the learners' dynamic knowledge status, conduct adaptive resource recommendations.

[0003] In existing technologies, knowledge concept extraction for multimodal teaching resources mainly relies on two approaches: manual annotation by domain experts and open extraction using general-purpose large language models. The former requires significant manpower and its annotation efficiency struggles to keep up with the rapid updates and continuous iterations of teaching resources; the latter is prone to "domain knowledge illusion," where the model generates conceptual representations that deviate from the actual curriculum system based on general pre-trained corpora, and the granularity of concept extraction is unstable, failing to align with structured standard curriculum outlines. Regarding the semantic representation of learning resources, since teaching resource texts often contain a lot of redundant background information and irrelevant content, directly using general-purpose pre-trained text embedding models to vectorize resource texts leads to feature dilution and semantic defocusing, making it difficult to provide high-quality vector representations for downstream cognitive modeling.

[0004] In the field of knowledge tracing technology, existing mainstream models, such as the Deep Knowledge Tracing (DKT) model based on recurrent neural networks and the Dynamic Key-Value Memory Network (DKVMN) based on external memory enhancement, mostly rely solely on students' historical correct and incorrect answer sequences and discrete interaction steps for modeling. This simplified modeling paradigm ignores two cognitive laws in the real learning process: first, learners' memory of knowledge decays with increasing time intervals, i.e., the Ebbinghaus forgetting curve; second, the same answer often reflects different cognitive states. For example, a correct answer in a very short time may stem from guessing, while an incorrect answer after a long period of thought may be due to carelessness rather than a lack of mastery. Therefore, existing models struggle to accurately recreate the dynamic evolution of knowledge states that conforms to objective human cognitive laws, and they also suffer from cold-start problems when facing new semesters or new exercises.

[0005] In personalized learning resource recommendation, existing technologies mostly employ recommendation methods based on collaborative filtering, content matching, or single-step greedy strategies. These methods typically perform static matching based on learners' interests or historical behavior, and their optimization objectives are mostly focused on short-term interaction effects, lacking direct modeling and explicit optimization of learners' knowledge status improvement and long-term learning benefits. Furthermore, existing knowledge-driven recommendation research is largely limited to single-question formats, and its recommendation decisions heavily rely on structured question-knowledge concept annotation relationships, making it difficult to directly extend to unstructured multimodal resource scenarios such as instructional videos and lecture notes.

[0006] In summary, the existing technologies for knowledge extraction, cognitive state modeling, and personalized recommendation are not yet fully coordinated, failing to form a closed-loop technological chain for collaborative learning spaces. On the one hand, the lack of accurate and reliable multimodal knowledge extraction and vectorized representation methods makes it difficult to transform teaching resources into structured cognitive diagnostic evidence. On the other hand, knowledge tracking models that ignore real cognitive patterns struggle to provide credible and interpretable knowledge state assessments. Furthermore, recommendation strategies based on static rules or single-step optimization fail to achieve collaborative delivery of cross-modal resources and maximize long-term knowledge gain. Due to these shortcomings, existing technologies struggle to provide learners with accurate, interpretable, and sustainably beneficial personalized learning services in collaborative learning scenarios. Summary of the Invention

[0007] This invention provides a knowledge-driven personalized resource recommendation method and system for collaborative learning spaces. It adopts a closed-loop technical solution of "multimodal knowledge extraction and vector library construction - concept perception knowledge tracking based on cognitive laws - personalized recommendation based on reinforcement learning and cross-modal retrieval". This solution alleviates the technical problems of existing technologies, such as domain illusion in knowledge extraction, rough cognitive state modeling with insufficient interpretability, and single recommendation form without long-term planning.

[0008] The knowledge-driven personalized resource recommendation method for collaborative learning spaces of the present invention includes the following steps:

[0009] Step 1: Acquire multimodal learning resources and learner behavior data in the collaborative learning space, and preprocess and standardize the acquired data; perform unified preprocessing and standardization on data from different sources and of different types to form the data foundation required for subsequent knowledge extraction, knowledge state modeling and resource recommendation.

[0010] Step 2: Construct a dual-agent collaborative mechanism for knowledge concept extraction and review experts. Extract knowledge concepts from multimodal learning resources, and review and standardize the extraction results. Represent the standardized knowledge concepts as semantic vectors and establish the association between learning resources and knowledge concepts. Optimize the semantic alignment between learning resources and knowledge concepts using contrastive learning methods. Store the aligned semantic vectors of learning resources and knowledge concepts, as well as the resource-concept association matrix, into the learning resource vector library.

[0011] Step 3: Establish a concept perception knowledge tracking model that integrates cognitive laws to describe the learner's knowledge status.

[0012] The knowledge tracing model consists of a concept-aware key-value memory framework, a time-aware forgetting module, a behavior-aware gating module, and a prediction layer. Learners' answer records and behavioral characteristics at each time step are input into the model. Answer records include question feature vectors, interaction features indicating correctness, and semantic vectors of the knowledge concepts contained in the questions. Behavioral characteristics include answer time, number of attempts, and number of prompts.

[0013] The construction of the concept-aware key-value memory framework includes: each slot of the key matrix corresponds to a semantic vector of a normalized knowledge concept; the value matrix represents the learner's mastery status of each knowledge concept; and based on the learner's mastery of each knowledge concept at a given time step... The answer records, and the feature vectors of the questions. Interactive features of answer correctness Separately related to the semantic vectors of the knowledge concepts contained in the question. The result is a query vector for attention addressing. and the content vector used for knowledge state updates ; Calculate attention weights using the correspondence between questions and concepts. ;Utilizing attention weights The learner obtains the value matrix of the previous time step after decay by performing a weighted summation. The questions encompass a comprehensive range of knowledge and concepts, testing the candidate's overall abilities. ; in acquiring time step After receiving the actual response results, based on content vectors Generate erase vector With the write vector ,right and Learning gain coefficients obtained using the behavior-aware gating module Adaptive amplitude control is performed, and the erase-increment operation is applied to the value matrix using the controlled erase and write vectors to update the time step. The value matrix.

[0014] The construction of the time-aware forgetting module includes: calculating the time interval between the current answer and the last answer. After undergoing a nonlinear smoothing transformation, it is mapped to a high-dimensional temporal feature embedding vector. Then, compared with the learner in the previous time step, the first... The knowledge states of each knowledge concept are concatenated and input into a decaying network constructed using a multilayer perceptron. This network calculates the learner's understanding of the knowledge state of the first knowledge concept. Memory retention coefficient of a knowledge concept , Then, the memory retention coefficient is used to dynamically decay the value matrix of the previous time step element by element.

[0015] The construction of the behavior-aware gating module includes: setting time steps behavioral characteristics The input is a behavior encoder composed of a multilayer perceptron, which is mapped to a latent behavior semantic vector. ,Will With content vector The learned gain coefficients are obtained by concatenating the input gain gate network. , , It represents the level of credibility of the current interaction behavior.

[0016] The construction of the prediction layer includes: integrating the overall capability status With query vector The data is pieced together and mapped to potential capability values ​​via a multilayer perceptron. Then, the two-parameter logic model from item response theory is introduced to predict the learner's mastery probability.

[0017] Step 4: Model the exercise recommendation as a Markov decision process, and use the reinforcement learning method D3QN to construct an exercise recommendation agent. Based on the learner's current knowledge state, generate exercise recommendation results oriented towards knowledge gain. Then, based on the recommended exercises, query other similar learning resources from the learning resource vector library, and together they constitute the recommendation results to be displayed to the learner. This realizes the generation of personalized learning resource recommendations oriented towards knowledge gain based on reinforcement learning and cross-modal retrieval.

[0018] Accordingly, the present invention also provides a knowledge-driven personalized resource recommendation system for collaborative learning spaces, which, based on the above method, includes the following modules:

[0019] The data acquisition and preprocessing module is used to acquire multimodal learning resources and learners' learning behavior data in the collaborative learning space, and to preprocess and standardize the acquired data.

[0020] The dual-agent collaborative knowledge extraction framework utilizes a knowledge concept extraction agent and an expert review agent to extract knowledge concepts from multimodal learning resources, and reviews and standardizes the extraction results, representing the standardized knowledge concepts as semantic vectors, and establishing the association between learning resources and knowledge concepts.

[0021] The learning resource vector library construction module uses a contrastive learning method to optimize the semantic alignment between learning resources and knowledge concepts, and stores the aligned semantic vectors of learning resources and knowledge concepts, as well as the resource-concept association matrix, into the learning resource vector library.

[0022] The concept-aware knowledge tracking model consists of a concept-aware key-value memory framework, a time-aware forgetting module, a behavior-aware gating module, and a prediction layer. It inputs the learner's answer records and behavioral characteristics at each time step into the model to track the learner's knowledge status.

[0023] In the concept-aware key-value memory framework, each slot in the key matrix corresponds to a semantic vector of a standardized knowledge concept, and the value matrix represents the learner's mastery status of each knowledge concept; based on the learner's progress at each time step... The answer records, and the feature vectors of the questions. Interactive features of answer correctness Separately related to the semantic vectors of the knowledge concepts contained in the question. The result is a query vector for attention addressing. and the content vector used for knowledge state updates ; Calculate attention weights using the correspondence between questions and concepts. ;Utilizing attention weights The learner obtains the value matrix of the previous time step after decay by performing a weighted summation. The questions encompass a comprehensive range of knowledge and concepts, testing the candidate's overall abilities. ; in acquiring time step After receiving the actual response results, based on content vectors Generate erase vector With the write vector ,right and Learning gain coefficients obtained using the behavior-aware gating module Adaptive amplitude control is performed, and the erase-increment operation is applied to the value matrix using the controlled erase and write vectors to update the time step. The value matrix;

[0024] In the time-aware forgetting module, the time interval between the current answer and the last answer is recorded. After undergoing a nonlinear smoothing transformation, it is mapped to a high-dimensional temporal feature embedding vector. Then, compared with the learner in the previous time step, the first... The knowledge states of each knowledge concept are concatenated and input into a decaying network constructed using a multilayer perceptron. This network calculates the learner's understanding of the knowledge state of the first knowledge concept. Memory retention coefficient of a knowledge concept , Then, the memory retention coefficient is used to dynamically decay the value matrix of the previous time step element by element;

[0025] In the behavior-aware gating module, the time step behavioral characteristics The input is a behavior encoder composed of a multilayer perceptron, which is mapped to a latent behavior semantic vector. ,Will With content vector The learned gain coefficients are obtained by concatenating the input gain gate network. , , It represents the level of credibility of the current interaction behavior;

[0026] The prediction layer will integrate the capability status. With query vector The data is pieced together and mapped to potential capability values ​​via a multilayer perceptron. Then, the two-parameter logical model from Project Response Theory is introduced to predict the learner's probability of knowledge mastery;

[0027] A D3QN-based question recommendation agent generates question recommendations based on the learner's current knowledge state, with a focus on knowledge gain.

[0028] Multimodal learning resource recommendation involves querying similar learning resources from the learning resource vector library based on the recommended questions output by the question recommendation agent, and then combining these to form the recommendation results to be displayed to the learner.

[0029] Compared with the prior art, the advantages of the present invention are:

[0030] (1) The method and system of this invention realize multimodal knowledge extraction and vector library construction based on dual-agent collaboration and contrastive learning. Through dual-agent collaboration, knowledge concept extraction, verification and standardization are completed, and contrastive learning is used to enhance the semantic alignment between learning resources and knowledge concepts, so that learning resources of different modalities can be mapped to a unified vector space. The method and system of this invention can alleviate the knowledge illusion and granularity divergence problems of general large models in vertical education fields, and construct a learning resource vector library with resource-concept mapping relationship and semantic alignment, providing a data foundation for downstream knowledge tracking and recommendation.

[0031] (2) The method and system of this invention realize a concept-aware knowledge tracking model that integrates cognitive laws. Based on a key-value memory network, it introduces semantic information of knowledge concepts, a time decay mechanism, a behavioral perception gating mechanism, and an item response theory model, enabling the knowledge status update process to simultaneously reflect the relationships between knowledge concepts, learners' forgetting patterns, and differences in learning behavior. Compared to traditional knowledge tracking methods that update knowledge status solely based on response results, the method and system of this invention can more accurately track learners' knowledge mastery, improving the accuracy, reliability, and interpretability of knowledge status assessment.

[0032] (3) The method and system of this invention realizes a knowledge-driven hybrid recommendation based on reinforcement learning and cross-modal retrieval. First, it generates exercise recommendation results oriented towards long-term knowledge gain through reinforcement learning, and then recalls related lecture notes, videos and other multimodal resources based on the recommended exercises, which can realize the collaborative recommendation of different types of learning resources. Compared with traditional recommendation methods based on static rules, collaborative filtering or single-step greedy strategies, the method and system of this invention can take into account both current learning needs and long-term knowledge mastery improvement, and can realize the collaborative push of multimodal resources from exercises to teaching videos, lecture notes and other resources. Attached Figure Description

[0033] Figure 1 This is an overall flowchart of the knowledge-driven personalized resource recommendation method for collaborative learning spaces according to the present invention;

[0034] Figure 2 This is a schematic diagram of the dual-agent collaborative knowledge extraction framework of the present invention;

[0035] Figure 3 This is a schematic diagram illustrating the construction of the resource vector library based on contrastive learning according to the present invention;

[0036] Figure 4 This is a diagram of the concept-aware knowledge tracking model architecture of the present invention;

[0037] Figure 5 This is a diagram of the problem recommendation framework based on D3QN of the present invention. Detailed Implementation

[0038] The technical solution of the present invention will be described below with reference to the accompanying drawings and embodiments.

[0039] This invention uses an online SPOC (Small Private Online Course) platform and offline smart classroom recordings from a university's undergraduate "Computer Organization" course as data sources. The multimodal learning resources specifically include: online platform exercises, offline classroom recordings, and standard course syllabus knowledge concepts pre-compiled by domain experts. Learners' online answering behavior records include behavioral fields such as answer results, answering time, number of attempts, and number of prompt requests.

[0040] Step 1: Acquire and preprocess multimodal learning resources. Acquire multimodal learning resources such as exercises, teaching materials, and classroom videos from the collaborative learning space, as well as learning behavior data. Perform unified preprocessing and standardized representation on data from different sources and of different types to form the data foundation required for subsequent knowledge extraction, knowledge state modeling, and resource recommendation. Learning behavior data mainly includes learners' online answering behavior records.

[0041] For online exercises that combine text and images, regular expressions are used to parse the original text, locate image reference tags, and replace them with structured placeholders to preserve the context of the text and images. Subsequently, the image files are converted into Base64 lossless encoded strings and embedded into standardized character input for subsequent processing by the visual semantic model.

[0042] For offline classroom recording videos, the audio track is separated to extract the pure audio stream. The automatic speech recognition interface of the Tongyi Listening API is called to convert the audio into text records with precise timestamps. At the same time, a chapter overview based on timestamps is generated, and finally a structured text corpus with time segment annotations is obtained.

[0043] Step 2: Extracting Knowledge Concepts and Constructing a Learning Resource Vector Library. A dual-agent collaborative mechanism is used to extract knowledge concepts from multimodal learning resources, and the extraction results are reviewed and standardized. The association between learning resources and knowledge concepts is established. Furthermore, a contrastive learning method is used to optimize the semantic alignment between the representation of learning resources and the representation of knowledge concepts, constructing a unified learning resource vector library to achieve a unified vector representation of learning resources across different modalities.

[0044] This step first constructs a knowledge concept extraction agent and a review expert agent. The knowledge concept extraction agent identifies candidate knowledge concepts from multimodal learning resources such as exercises, handouts, and classroom videos. Then, the review expert agent verifies the candidate knowledge concepts for course relevance, knowledge granularity, and name standardization, and corrects and standardizes the candidate knowledge concepts based on the review results. Subsequently, a mapping relationship is established between learning resources and standardized knowledge concepts, forming resource-knowledge concept association results. Based on this, associated resources and knowledge concepts are used as semantic positive samples, while irrelevant or weakly related knowledge concepts are used as negative samples. Through comparative learning, the semantic distance between resource vectors and knowledge concept vectors is optimized, thereby constructing a semantically aligned learning resource vector library and improving the accuracy and consistency of multimodal resource knowledge representation.

[0045] Step 2.1: Extract knowledge concepts from multimodal learning resources based on dual-agent collaboration.

[0046] The dual-agent collaborative knowledge extraction framework includes a knowledge concept extraction agent and a review expert agent. In this embodiment, both are built based on the GLM-4-plus large language model. The overall workflow of this framework is as follows: Figure 2 As shown.

[0047] The knowledge concept extraction agent is configured with thought chain reasoning prompts and forced to execute according to the following logic: The first stage analyzes the learning resources' intent, inputting course constraints to enable the agent to distinguish between core professional mechanisms and general background knowledge, outputting an intermediate representation with course constraints. Taking the exercise "Implementing Narcissistic Numbers Based on MIPS Assembly Language" as an example, the prompts input to the agent constrain the course background of knowledge extraction to "Computer Organization," minimizing the output of knowledge concepts outside the professional course background. The knowledge concept extraction agent identifies that the knowledge point tested in the "Computer Organization" course is MIPS assembly language, while marking general programming domain concepts such as "recursion" or "loop control" as irrelevant interference and eliminating them, outputting the intermediate representation: MIPS assembly language. This intermediate representation is the set of textual knowledge concepts. The second stage summarizes key mechanisms based on the intermediate representation, mapping them to candidate knowledge concepts. The process of summarizing key mechanisms from the intermediate representation strictly internalizes two constraints: First, candidate knowledge concepts must belong to the course knowledge system, further requiring the filtering of general knowledge concepts. Second, the granularity of candidate knowledge concepts needs to be kept moderate. Concepts that are too broad need to be further refined; content that is biased towards program procedures or implementation details should be eliminated or replaced.

[0048] The review expert agent performs the following multi-dimensional rule verification: (a) course relevance check, eliminating general domain concepts that are detached from the course background; (b) granularity control, refining overly macro concepts and merging or deleting overly detailed implementation process descriptions; (c) name standardization, mapping different expressions to a unified standard form. Based on the above rule verification, the review expert agent judges each candidate knowledge concept, and the review fails if any of the following conditions are met: (1) no other knowledge concepts exist after eliminating general domain concepts; (2) granularity is unbalanced and cannot be automatically corrected, that is, the concept is too macro and cannot be mapped to specific knowledge points in the course after refinement, or it is too trivial and still deviates from the knowledge concept level after merging; (3) name standardization fails, that is, the concept has multiple ambiguous expressions. If the review fails, the review opinion and reflection feedback are sent back to the context of the knowledge concept extraction agent to drive its self-correction; if it fails after more than K consecutive retries, the learning resource is transferred to the manual review queue. K is a positive integer greater than 2.

[0049] Step 2.2: Adaptive Normalization of Knowledge Concepts Based on Prior Awareness. Depending on whether the course has a standard syllabus, adaptive normalization of knowledge concepts is performed.

[0050] For scenarios with a standard curriculum syllabus, such as in this embodiment, the names and descriptions of standard knowledge concepts are input into the ZhipuAI Embedding-3 pre-trained text embedding model to generate 1024-dimensional semantic vectors, thus constructing a standard syllabus vector library. After generating candidate concepts, the knowledge concept extraction agent maps them to query vectors and performs Top-5 nearest neighbor recall in the standard syllabus vector library to obtain the five most similar standard concepts as a candidate subset. The agent, in conjunction with the teaching objectives of the learning resources, performs logical reasoning within this candidate subset to select the standard knowledge concepts with the best semantic match and outputs a structured alignment result.

[0051] For scenarios lacking a standard outline, the Embedding-3 model is used to vectorize the approved knowledge concept descriptions. Cosine similarity is used to measure semantic distance, and the K-means algorithm is employed for clustering. The clustering parameters are analyzed using the elbow method to assess the inertia weight change curve, and the silhouette coefficient is used to evaluate intra-cluster compactness and inter-cluster separation, automatically determining the optimal number of clusters. The centroid vector of each cluster is used as the standardized semantic representation of the knowledge concept, and the most frequent and semantically complete natural language description within the cluster is selected as the representative label, forming a binary mapping of "semantic vector - standardized description".

[0052] Step 2.3: Construction of a learning resource vector library based on contrastive learning. The contrastive learning method is used to optimize the semantic alignment between learning resource representations and knowledge concept representations, constructing a unified learning resource vector library to achieve unified vector representation of learning resources from different modalities.

[0053] By processing each learning resource text through steps 2.1 and 2.2, the association between learning resources and normative knowledge concepts can be established, resulting in resource-knowledge concept mapping pairs. The construction process of the learning resource vector library based on contrastive learning is as follows: Figure 3 As shown. First, positive and negative samples are constructed. The "resource-knowledge concept" mapping pairs produced in steps 2.1 and 2.2 are used as positive samples, and negative samples are constructed according to the following three-level strategy: the preferred negative samples are concepts that have not been selected by the review expert agent; the difficult negative samples are other knowledge concepts in the same module, that is, different knowledge concepts under the same topic. For example, "trigger", "register" and "state machine" all belong to the "sequential logic" module, so these knowledge concepts are difficult negative samples for each other; the remaining negative samples are irrelevant concepts randomly selected from the global knowledge concept space.

[0054] Next, the resources in the samples, i.e., the summaries of learning resource text or video clips, are input into the Embedding-3 model to extract a 1024-dimensional basic resource vector. This embodiment of the invention constructs a projection layer embedding-contrast network composed of a two-layer perceptron. The encoder of the contrast network is a pre-trained Embedding-3 model. Positive and negative sample pairs are input into the contrast network, and the projection layer parameters are optimized using the InfoNCE loss function. After alignment training is completed, the embedding vectors of learning resources and knowledge concepts are persistently stored and imported into the FAISS high-performance vector database, completing the construction process of the learning resource vector library for the collaborative learning space.

[0055] Step 3: Achieve concept-aware knowledge tracking that integrates cognitive patterns. Using a key-value memory network as the foundation of the knowledge tracking model, semantic information of knowledge concepts is introduced to establish a correspondence between knowledge concepts and memory slots. Combining time decay mechanisms, behavioral perception gating mechanisms, and item response theory models, the learner's knowledge state is dynamically modeled and updated to obtain the learner's current knowledge mastery.

[0056] This step uses a key-value memory network as the basic model, introducing semantic information of knowledge concepts into the key matrix construction process to establish an explicit correspondence between memory slots and specific knowledge concepts. During the memory retrieval phase, the association between questions and knowledge concepts guides attention addressing, enabling the model to focus on retrieving the knowledge state corresponding to the knowledge concepts involved in the current question. In the memory update phase, historical knowledge states are decayed based on the time interval between adjacent learning behaviors, and the memory writing amplitude is adjusted according to learning behavior characteristics, allowing the knowledge state update process to simultaneously reflect forgetting patterns and behavioral differences. In the prediction phase, a two-parameter model from item response theory is combined to jointly model learner ability levels, question difficulty, and discrimination, thereby improving the accuracy and interpretability of knowledge state assessment.

[0057] Step 3.1: Construction of knowledge concept perception structure.

[0058] Figure 4 This paper demonstrates the overall structure of a concept-aware knowledge tracking model that integrates cognitive patterns. The model consists of four parts: a concept-aware key-value memory framework, a time-aware forgetting module, a behavior-aware gating module, and a prediction layer. The concept-aware key-value memory framework is the core of the model. The time-aware forgetting module and the behavior-aware gating module optimize the intermediate representation of the model by acting on the key-value memory network. The prediction layer obtains the final prediction result through the output of the key-value memory network.

[0059] In the concept-aware key-value memory framework, the key matrix Built upon standard knowledge concepts The number of slots is equal to the total number of standard knowledge concepts in the canonical dataset. Each slot corresponds to a real knowledge concept. In this embodiment, the dimension of each slot is set to 128. The textual description of the standard knowledge concept is mapped to a 128-dimensional semantic vector through a projection adaptation layer, which consists of a linear projection layer, layer normalization, and a GELU activation function. The key matrix in the memory network can be obtained through the mapping of the projection adaptation layer. In subsequent steps, the key matrix... It will be used to establish a mapping relationship between learners' memory slots and real knowledge concepts.

[0060] Acquire learners at time steps The input is the corresponding time step in the learner's historical answer record sequence. The recorded information includes question information, answer results, and the set of knowledge concepts contained in the question. From this input, the feature vector of the question is obtained through an embedding layer. Interactive feature embeddings that include the correctness of answers and knowledge concept semantic vector The embedding layer is the Embedding-3 model pre-trained in step 2.2. Through... Figure 4 The question-concept feature fusion structure in the text yields a query vector for attention addressing. and the content vector used for knowledge state updates :

[0061]

[0062]

[0063] in, This represents element-wise multiplication. Based on this fusion strategy, the model can retain the zero-shot generalization ability brought by the pre-trained semantic model, and can also learn different question features, thereby achieving complementary fusion of question and concept features.

[0064] Subsequently, attention weights are calculated using the correspondence between questions and concepts. This focuses greater attention on the knowledge concepts actually being tested in the current question. The calculation formula is as follows:

[0065]

[0066] Where β is a learnable gating parameter, For the Sigmoid function, This indicates that the knowledge concept slots involved in the current question are subjected to Softmax normalization. This indicates that the slots not involved are normalized. query vector AND key matrix The dot product of the vectors of each concept. By introducing explicit constraints on the relationship between the question and the concept, noise interference from irrelevant concepts can be suppressed, laying the foundation for subsequent prediction and knowledge state updates.

[0067] The value matrix is ​​a knowledge state matrix, recording the learner's mastery status of each knowledge concept. Let the value matrix at the previous time step be... The value matrix of the first The line represents the learner's response to the first... The mastery status of each knowledge concept. Initial value matrix. It can be randomly initialized.

[0068] Step 3.2: Memory update modeling based on time decay and behavior gating.

[0069] A time-aware forgetting module is introduced before memory retrieval. This module utilizes a multilayer perceptron (MLP) to implement a memory decay network, tracking the time interval between the current response and the previous response. After undergoing a nonlinear smoothing transformation, it is mapped to a high-dimensional temporal feature embedding vector. The embedding is compared with the learner's input at the previous time step. The knowledge state of a knowledge concept The components are spliced ​​together and input into the attenuation network to obtain the attenuation factor, which is the memory retention coefficient. :

[0070]

[0071] in, This represents a vector concatenation operation. and To decay the learnable parameters of the network, . use The value matrix of the previous time step is dynamically decayed element by element, that is, each memory slot is multiplied by its corresponding retention coefficient. The smaller the value, the deeper the learner's forgetting of the knowledge concept.

[0072] A behavior-aware memory update gating module is introduced during memory writes, i.e., a behavior-aware gating module. This involves setting time steps... Multidimensional behavioral characteristics The input is a behavior encoder composed of a multilayer perceptron, which is mapped to a latent behavior semantic vector. Behavioral characteristics This includes the time taken to answer the question, the number of attempts, and the number of hints received. Combine this vector with the content vector. Concatenation, input gain gating network:

[0073]

[0074] in, , , , For learnable parameters of the gain-gated network, For learning gain coefficient, This is a linear rectified unit function. When the model detects low-confidence interactions, such as multiple attempts and short response times, If the value approaches 0, the writing process of this interaction will be suppressed in subsequent knowledge state update processes; conversely, when the interaction behavior exhibits normal cognitive characteristics, When the value approaches 1, the model updates the learner's knowledge state normally. The gain-gated network in this embodiment consists of a multilayer perceptron.

[0075] Step 3.3: Dynamic memory reading and writing process.

[0076] Based on attention weights By performing a weighted summation on the attenuated value matrix, the learner's comprehensive ability status regarding the knowledge concepts involved in the current question can be obtained. ,Will With query vector The data is pieced together and mapped to potential capability values ​​via a multilayer perceptron. This is used to characterize the ability to map the learner's potential knowledge level within the context of the current question. Then, a two-parameter logistic model from item response theory is introduced to calculate the probability of a correct answer. :

[0077]

[0078] in, Indicates time step The question number, The question discrimination parameter, The difficulty parameter of the question. It is the Sigmoid activation function. This represents the corresponding time step in the learner's historical answer sequence. Recorded answer results A value of 0 indicates an incorrect answer, and a value of 1 indicates a correct answer.

[0079] After obtaining the actual response results, based on the content vector Generate erase vector With the write vector Among them, the erasure vector Constrained by the Sigmoid activation function The interval represents the proportion of knowledge that needs to be erased from the current knowledge state in this interaction. Write vector. The calculation method and Similarly, the difference lies in replacing the activation function with the Tanh function, which represents the new knowledge representation acquired in this interaction.

[0080] The erase vector is calculated. With the write vector Then, the learning gain coefficients obtained from the behavior-aware gating module are used. Adaptive amplitude control is performed by multiplying both values ​​by the gain coefficient to obtain the controlled erase vector. and write vector This is done to suppress the update intensity of low-confidence interactions. Finally, an erase-insert operation is performed on the value matrix to complete the knowledge state update process.

[0081]

[0082] in, These represent the learner's understanding of the first... Each knowledge concept is located at time step The state of control.

[0083] The final model training employs a binary cross-entropy loss function, calculating the loss at each effective prediction time step and ignoring padding positions through a masking mechanism. By minimizing this loss function, the learnable parameters in the key-value memory framework, time-aware forgetting module, behavior-aware gating module, and prediction layer network are optimized.

[0084] Step 4: Implement personalized recommendations based on reinforcement learning and cross-modal retrieval. Construct a question recommendation strategy based on the learner's current knowledge state, and use reinforcement learning methods to generate personalized question recommendations oriented towards knowledge gain.

[0085] This step first identifies weak knowledge concepts based on the learner's knowledge state obtained in step 3, and generates a candidate exercise set based on the resource-knowledge concept association. Then, the exercise recommendation process is modeled as a sequential decision-making process, employing reinforcement learning to learn a recommendation strategy. The optimization objective is to improve the learner's knowledge state or gain knowledge, selecting suitable target exercises from the candidate exercises. After obtaining recommended exercises, the exercise is used as a semantic anchor to calculate its vector similarity with other resources such as teaching materials and classroom videos in the learning resource vector library, recalling relevant multimodal resources. Finally, the candidate resources are ranked based on resource similarity and the learner's current knowledge state, generating personalized recommendation results containing exercises, teaching materials, and teaching videos.

[0086] Step 4.1: Model the Markov decision process as recommended in the question.

[0087] In the reinforcement learning framework of the present invention, the learner and their knowledge state are abstracted as the environment, and the exercise recommendation algorithm interacts with it as an agent. The overall recommendation process can be modeled as a Markov decision process.

[0088] state space The knowledge concept mastery vector output by the knowledge tracking model trained in step 3 constitutes the knowledge concept mastery vector, denoted as . ,in For the number of knowledge concepts, Indicates learner's understanding of the first... The probability estimate of mastery of each knowledge concept. In specific implementation, this embodiment uses the knowledge tracking model trained in step 3 to maintain the learner's dynamic memory state, and obtains the mastery estimate of each knowledge concept based on the memory state. This mastery can be interpreted as: the expected accuracy rate of the learner when answering questions containing only a single knowledge concept under the current cognitive state of the learner.

[0089] Action space Defined as a subset of candidate questions covering learners' weak knowledge concepts. Specifically, the 20% of knowledge concepts with the lowest learner mastery are defined as the weak set. Based on the resource-concept association matrix constructed in step 1, unrecommended questions covering at least one knowledge concept in this set are selected to form the effective action space.

[0090] Instant rewards Defined as the difference in average knowledge mastery before and after answering a question. The state transition process is implemented by the write operation of the knowledge tracking model: after the learner receives the recommended questions and generates answers, the model updates the knowledge state vector, completing the transition from the previous state to the current state. arrive The transfer.

[0091] Step 4.2: Construction of the recommendation agent based on D3QN.

[0092] After establishing the Markov decision process model, this embodiment uses the D3QN (Dueling Double DeepQ-Network) algorithm as the core value network of the recommending agent. First, the main value network is constructed. With the target network The network endpoint is decoupled into a state-value flow. With the advantage of action flow ,in These are the parameters of the main value network; This is common knowledge about target networks, and it is regularly updated based on... renew; These are state-value stream parameters. These are the action dominance flow parameters. It is the input state. This is a decision-making action. The formula for calculating the value of the final action is:

[0093]

[0094] in, It is the set of all selectable actions for the agent in the current environment. At time step The agent's current state Decision-making actions The Q value is calculated by adding the state value to the dominance function and subtracting the average dominance value. Let the time step be... State Select Action Learners perform actions Received instant reward ,state Transfer to .

[0095] During the training of the value network, quadruples are randomly sampled and transferred from the experience replay pool D. Utilize the main network to select the most valuable action for the next state. The target value is calculated using the target network:

[0096]

[0097] in, The discount factor is set to 0.90 in this embodiment to calculate the expected long-term learning benefits. The training optimization objective is to minimize the mean square error between the main network's predicted value and the target value.

[0098] After establishing the D3QN value network architecture, it is necessary to build a recommendation agent. The overall construction process is as follows: Figure 5 As shown, a learner simulator is constructed using the concept-aware knowledge tracking model trained in step 3. First, the learner simulator initializes the current knowledge state based on the target learner's historical answer sequence through the knowledge tracking model, obtaining a K-dimensional knowledge concept mastery vector, i.e., the current knowledge state. Subsequently, the system enters a sequence recommendation loop: at time step... The D3QN-based recommendation agent relies on the current knowledge state. For questions not yet recommended in the question bank, the optimal question is selected using the following action scoring formula:

[0099]

[0100] in, Indicates the topic The average degree of weakness of the covered knowledge concepts under the current mastery vector; The knowledge guidance coefficient is set to 0.2 in this embodiment.

[0101] The selected optimal question is then input into the learner simulator. The learner simulator calculates the probability of answering based on the knowledge tracking model in step 3 and samples the results to generate the answer. This result, along with the question information, drives the knowledge tracking model to complete a memory write operation, transferring the knowledge state from... Updated to Repeat this process until N steps of recommendation are completed.

[0102] Step 4.3: Multimodal learning resource recommendation based on vector retrieval.

[0103] Recommended exercises output from step 4.2 As semantic anchors, their representation vectors are extracted from the learning resource vector library constructed in step 2.3. This will be used as a query vector. The query vector will then be compared with the learning resource vectors in the learning resource vector library. The cosine similarity is used as the semantic relevance score between the two resources, and they are sorted in descending order of score. The top-M learning resources are recalled to form a candidate resource set. In this embodiment... .

[0104] In the personalized ranking stage, for each candidate resource Calculation and learner's current knowledge state Match score The calculation formula is as follows:

[0105]

[0106] in, Candidate resources The core set of knowledge concepts covered For learners to understand concepts The degree of mastery. The comprehensive score of each candidate learning resource is the sum of the semantic relevance score and the knowledge state matching score. The final recommendation results are sorted in descending order of comprehensive score, and the Top-K recommended resources are output. In this embodiment... .

[0107] The knowledge-driven personalized resource recommendation system for collaborative learning spaces implemented in this invention includes a data acquisition and preprocessing module, a dual-agent collaborative knowledge extraction framework, a learning resource vector library construction module, a concept-aware knowledge tracking model, a D3QN-based exercise recommendation agent, and a multimodal learning resource recommendation module.

[0108] The data acquisition and preprocessing module is used to acquire multimodal learning resources and learners' learning behavior data in the collaborative learning space, and to preprocess and standardize the acquired data. The data acquisition and preprocessing module is as described in step 1 above, and will not be repeated here.

[0109] The dual-agent collaborative knowledge extraction framework utilizes two expert agents—one for knowledge concept extraction and the other for review—to extract knowledge concepts from multimodal learning resources. The extracted results are then reviewed and standardized, with the standardized knowledge concepts represented as semantic vectors. This establishes a relationship between learning resources and knowledge concepts. The implementation of the dual-agent collaborative knowledge extraction framework is described in steps 2.1-2.2 above.

[0110] The learning resource vector library construction module utilizes a contrastive learning method to optimize the semantic alignment between learning resources and knowledge concepts. It stores the aligned semantic vectors of learning resources and knowledge concepts, as well as the resource-concept association matrix, into the learning resource vector library. The implementation of the learning resource vector library construction module is as described in step 2.3 above.

[0111] The concept-aware knowledge tracing model tracks the learner's knowledge status based on their answer records and behavioral characteristics at each time step. The model consists of a concept-aware key-value memory framework, a time-aware forgetting module, a behavior-aware gating module, and a prediction layer, as described in step 3 above.

[0112] The D3QN-based question recommendation agent generates question recommendations based on the learner's current knowledge state, oriented towards knowledge gain. The multimodal learning resource recommendation module queries other similar learning resources from the learning resource vector library based on the recommended questions output by the question recommendation agent, and together they form the recommendation results displayed to the learner. The implementation of these two functional modules is as described in step 4 above.

[0113] Result verification:

[0114] In an offline simulation environment, the method of this invention was compared and validated with three baseline recommendation methods: random recommendation, popularity recommendation, and greedy mastery recommendation. The following metrics were used to quantitatively evaluate the recommendation performance:

[0115] Mastery Gain (MG): Measures the extent to which students' overall knowledge of all knowledge concepts improves after N recommended interventions.

[0116] Weak Knowledge Concept Mastery Gain (WKC-MG): Measures the improvement in a student's mastery of weak knowledge concepts after recommendations. For each student, the 20% of knowledge concepts with the lowest initial mastery are selected as the weak knowledge set, and the improvement in mastery of this set is calculated.

[0117] The experiment used students' historical answer records from the fourth week prior to the start of the "Computer Organization Course" dataset to initialize their knowledge state and simulate subsequent N-step recommendation interventions. The "Computer Organization Course" dataset is a private dataset built based on a real teaching scenario of a university's computer science major. The learners' online answer behavior records in this dataset come from learning logs on the online SPOC platform. The multimodal teaching resources come from the online SPOC platform and offline smart classroom recordings, specifically including after-class exercises from the online platform, offline classroom recording videos, and standard course syllabus knowledge concepts pre-compiled by domain experts.

[0118] Table 1. Comparative experimental results of the method of this invention and the baseline recommended method ("Computer Organization Course" dataset)

[0119]

[0120] The experimental results are shown in Table 1. and At a given recommendation length, the method of this invention outperforms the baseline recommendation method in both MG and WKC-MG metrics, demonstrating that the reinforcement learning agent can more effectively capture changes in the learner's knowledge state during the short-to-medium-term recommendation process and form a recommendation sequence superior to single-step rule-based strategies. At that time, the method of the present invention still outperforms greedy recommendation in terms of MG index and maintains a comparable effect to greedy recommendation in terms of WKC-MG index, proving that the method of the present invention can maintain a stable knowledge state improvement effect in long-term recommendation.

[0121] Building upon the experiments on the aforementioned private datasets, further experiments were conducted on the publicly available educational dataset ASSIST2009 to verify the applicability of the proposed method to public educational datasets. The ASSIST2009 dataset originates from the US-based ASSISTments online educational tutoring platform and is one of the classic benchmark datasets in the field of educational data mining. The experimental results are shown in Table 2.

[0122] Table 2. Comparison of experimental results between the method of this invention and the baseline recommendation method (ASSIST2009 dataset)

[0123]

[0124] As shown in Table 2, the method of this invention achieves the highest MG and WKC-MG values ​​across all recommended length settings in ASSIST2009, indicating that the method is effective not only on self-built course datasets but also has certain generalization potential on publicly available educational datasets. Especially in... In short sequence recommendation scenarios, the MG of the method of this invention reaches 0.0439, which is significantly higher than the 0.0266 of greedy recommendation, indicating that the method of this invention can generate effective knowledge intervention more quickly.

[0125] The scope of protection of this invention is not limited to the specific embodiments described above. In the above embodiments, the GLM-4-plus large language model on which the dual-agent framework is based can be replaced with GPT-4 or other large language models with tool invocation and thought chain reasoning capabilities; the pre-trained text embedding model can be replaced with other models with semantic encoding capabilities, such as BERT; the vector database can be replaced with Milvus or Pinecone or other approximate nearest neighbor retrieval tools; the D3QN network can be replaced with other deep reinforcement learning architectures; the hyperparameters in each step can be adaptively adjusted according to the data scale and business needs of the actual application scenario; the form of the loss function, the type of activation function, the regularization method, and the transformation function in preprocessing can also be conventionally modified according to the specific task characteristics. All modifications, equivalent substitutions, improvements, etc., within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A knowledge-driven personalized resource recommendation method for collaborative learning spaces, characterized in that, include: Step 1: Acquire multimodal learning resources and learners' learning behavior data in the collaborative learning space, and preprocess and standardize the acquired data; Step 2: Construct a dual-agent collaborative mechanism for knowledge concept extraction and review experts. Extract knowledge concepts from multimodal learning resources, review and standardize the extraction results, represent the standardized knowledge concepts as semantic vectors, and establish the association between learning resources and knowledge concepts. Utilize contrastive learning methods to optimize the semantic alignment between learning resources and knowledge concepts, and store the aligned semantic vectors of learning resources and knowledge concepts, as well as the resource-concept association matrix, into a learning resource vector library. Step 3: Establish a concept perception knowledge tracking model that integrates cognitive patterns to describe the learner's knowledge mastery status; The knowledge tracing model consists of a concept-aware key-value memory framework, a time-aware forgetting module, a behavior-aware gating module, and a prediction layer. The learner's answer records and behavioral features at each time step are input into the model. The answer records include question feature vectors, interactive features of answer correctness, and semantic vectors of knowledge concepts contained in the questions. The behavioral features include answer time, number of attempts, and number of prompts. The construction of the concept-aware key-value memory framework includes: each slot of the key matrix corresponds to a semantic vector of a normalized knowledge concept; the value matrix represents the learner's mastery status of each knowledge concept; and based on the learner's mastery of each knowledge concept at a given time step... The answer records, and the feature vectors of the questions. Interactive features of answer correctness Separately related to the semantic vectors of the knowledge concepts contained in the question. The result is a query vector for attention addressing. and the content vector used for knowledge state updates ; Calculate attention weights using the correspondence between questions and concepts. ;Utilizing attention weights The learner obtains the value matrix of the previous time step after decay by performing a weighted summation. The questions encompass a comprehensive range of knowledge and concepts, testing the candidate's overall abilities. ; in acquiring time step After receiving the actual response results, based on content vectors Generate erase vector With the write vector ,right and Learning gain coefficients obtained using the behavior-aware gating module Adaptive amplitude control is performed, and the erase-increment operation is applied to the value matrix using the controlled erase and write vectors to update the time step. The value matrix; The construction of the time-aware forgetting module includes: calculating the time interval between the current answer and the last answer. After undergoing a nonlinear smoothing transformation, it is mapped to a high-dimensional temporal feature embedding vector. Then, compared with the learner in the previous time step, the first... The knowledge states of each knowledge concept are concatenated and input into a decaying network constructed using a multilayer perceptron. This network calculates the learner's understanding of the knowledge state of the first knowledge concept. Memory retention coefficient of a knowledge concept , Then, the memory retention coefficient is used to dynamically decay the value matrix of the previous time step element by element; The construction of the behavior-aware gating module includes: setting time steps behavioral characteristics The input is a behavior encoder composed of a multilayer perceptron, which is mapped to a latent behavior semantic vector. ,Will With content vector The learned gain coefficients are obtained by concatenating the input gain gate network. , , It represents the level of credibility of the current interaction behavior; The construction of the prediction layer includes: integrating the overall capability status With query vector The data is pieced together and mapped to potential capability values ​​via a multilayer perceptron. Then, the two-parameter logical model from Project Response Theory is introduced to predict the learner's probability of knowledge mastery; Step 4: Model the exercise recommendation as a Markov decision process, and use the reinforcement learning method D3QN to construct an exercise recommendation agent. Based on the learner's current knowledge state, generate question recommendation results oriented towards knowledge gain, and then query similar learning resources from the learning resource vector library based on the recommended questions. Together, they constitute the recommendation results and are displayed to the learner. Here, D3QN represents a competitive dual deep Q network.

2. The method according to claim 1, characterized in that, Step 2 includes: The knowledge concept extraction agent is configured with thought chain reasoning prompts and executes according to the following logic: In the first stage, the agent performs intent analysis on the learning resources and inputs course constraints, enabling the agent to distinguish between core professional mechanisms and general background knowledge, and outputs an intermediate representation with course constraints; In the second stage, the agent summarizes key mechanisms based on the intermediate representation and maps them to candidate knowledge concepts. The review expert agent performs verification on the candidate knowledge concepts extracted by the knowledge concept extraction agent. The verification rules include: (a) course relevance check, eliminating general domain concepts that are detached from the course background; (b) control of the granularity of candidate knowledge concepts. If the granularity is unbalanced and cannot be automatically corrected, the candidate knowledge concept will not pass the review; (c) name standardization, mapping different expressions of the concept to a unified standard form.

3. The method according to claim 1 or 2, characterized in that, In step 2, positive sample pairs of resource-knowledge concept are constructed for the approved candidate knowledge concepts and associated learning resources, where the resources are learning resource text or video clip summaries, and negative sample pairs are constructed. The encoder of the contrast network is a pre-trained text embedding model, and the projection layer is a two-layer perceptron. The positive and negative sample pairs are input into the contrast network, and the projection layer parameters are optimized by combining the InfoNCE loss function.

4. The method according to claim 1, characterized in that, In step 3, the learner takes a time step... Calculate the query vector from the answer records. and content vector as follows: ； ； in, This represents the element-wise multiplication operation; Calculate attention weights as follows: ； Where β is a learnable gating parameter, For the Sigmoid function, This indicates that the knowledge concept slots involved in the current question are subjected to Softmax normalization. This indicates that the slots not involved are normalized. query vector AND key matrix The dot product of the vectors of each concept in the equation.

5. The method according to claim 1, characterized in that, Step 3 includes: updating the key-value memory framework time step. The value matrix is ​​as follows: ； in, These represent the learner's understanding of the first... Each knowledge concept is located at time step Time step The state of mastery; For the first Attention weights for each knowledge concept; and These are the modified erase vector and write vector, respectively; The prediction layer calculates the learner's mastery probability. as follows: ； in, Indicates time step The question number, The question discrimination parameter, The difficulty parameter of the question. It is the Sigmoid activation function. Indicates the learner's time step Recorded answer results A value of 0 indicates an incorrect answer. A value of 1 indicates that the answer is correct.

6. The method according to claim 1, characterized in that, Step 4 involves modeling the exercise recommendation as a Markov decision process, where the state space consists of the learner's correct answers to all knowledge concepts predicted by the knowledge tracking model trained in step 3. Let the time step be... The learner's knowledge status is , For the number of knowledge concepts, Indicates learner's understanding of the first... The probability of mastering a knowledge concept; the action space is a subset of candidate questions covering the learner's weak knowledge concepts; the knowledge concepts with the lowest mastery probability of the learner are grouped into the learner's weak set, and based on the correlation between learning resources and knowledge concepts, unrecommended questions covering at least one knowledge concept in the learner's weak set are selected to form the effective action space; instant reward. Defined as the difference in average knowledge mastery before and after answering; the state transition process is implemented by the writing operation of the knowledge tracking model trained in step 3: after the learner receives the recommended questions and generates the answer results, the knowledge tracking model updates the knowledge state vector to realize the state transition.

7. The method according to claim 1 or 6, characterized in that, Step 4 involves constructing a D3QN-based problem recommendation agent, in time step... Based on the learner's current knowledge status For questions not yet recommended in the question bank, the optimal question is selected using the following action scoring formula: ； in, The main value network of D3QN is based on the current knowledge state. and decision-making actions The calculated Q value, These are the main value network parameters and decision-making actions. Indicates the selected topic; Indicates the topic The average degree of weakness of the covered knowledge concepts under the learner's current knowledge mastery probability vector; This is the knowledge guidance coefficient.

8. The method according to claim 1, characterized in that, In step 4, the recommended questions are output by the question recommendation agent. Extract the corresponding representation vector from the learning resource vector library. ,Will Using the query vector, the top M learning resources with the highest semantic relevance scores are searched from the learning resource vector library to form a candidate resource set. For each candidate resource in the set... Calculation and current knowledge state Match score as follows: ； in, Candidate resources The set of knowledge concepts covered; , For learners to understand knowledge concepts The degree of mastery; Each candidate resource The overall score is the sum of the semantic relevance score and the current knowledge state matching score. Finally, the top K learning resources with the highest scores are recommended to the learner based on the overall score; where M and K are both positive integers.

9. The method according to claim 1, characterized in that, In step 1, the multimodal learning resources include exercises, teaching materials, and classroom videos in the collaborative learning space, and the learner's learning behavior data includes online answering behavior records. For exercises with mixed text and images, the original text is parsed using regular expressions, image reference tags are located, images are replaced with structured placeholders, and image files are converted into Base64 lossless encoded strings and embedded into standardized character input. For videos, audio is extracted and automatic speech recognition is performed, which converts the audio into timestamped text records and generates a timestamp-based chapter overview, forming structured text with time segment annotations.

10. A knowledge-driven personalized resource recommendation system for collaborative learning spaces, based on the knowledge-driven personalized resource recommendation method as described in any one of claims 1-2, 4-6, and 8-9, implements the following functional modules: The data acquisition and preprocessing module is used to acquire multimodal learning resources and learners' learning behavior data in the collaborative learning space, and to preprocess and standardize the acquired data. The dual-agent collaborative knowledge extraction framework utilizes a knowledge concept extraction agent and an expert review agent to extract knowledge concepts from multimodal learning resources, and reviews and standardizes the extraction results, representing the standardized knowledge concepts as semantic vectors, and establishing the association between learning resources and knowledge concepts. The learning resource vector library construction module uses a contrastive learning method to optimize the semantic alignment between learning resources and knowledge concepts, and stores the aligned semantic vectors of learning resources and knowledge concepts, as well as the resource-concept association matrix, into the learning resource vector library. The concept-aware knowledge tracking model consists of a concept-aware key-value memory framework, a time-aware forgetting module, a behavior-aware gating module, and a prediction layer. It inputs the learner's answer records and behavioral characteristics at each time step into the model to track the learner's knowledge status. In the concept-aware key-value memory framework, each slot in the key matrix corresponds to a semantic vector of a standardized knowledge concept, and the value matrix represents the learner's mastery status of each knowledge concept; based on the learner's progress at each time step... The answer records, and the feature vectors of the questions. Interactive features of answer correctness Separately related to the semantic vectors of the knowledge concepts contained in the question. The result is a query vector for attention addressing. and the content vector used for knowledge state updates ; Calculate attention weights using the correspondence between questions and concepts. ;Utilizing attention weights The learner obtains the value matrix of the previous time step after decay by performing a weighted summation. The questions encompass a comprehensive range of knowledge and concepts, testing the candidate's overall abilities. ; In acquiring time steps After receiving the actual response results, based on content vectors Generate erase vector With the write vector ,right and Learning gain coefficients obtained using the behavior-aware gating module Adaptive amplitude control is performed, and the erase-increment operation is applied to the value matrix using the controlled erase and write vectors to update the time step. The value matrix; In the time-aware forgetting module, the time interval between the current answer and the last answer is recorded. After undergoing a nonlinear smoothing transformation, it is mapped to a high-dimensional temporal feature embedding vector. Then, compared with the learner in the previous time step, the first... The knowledge states of each knowledge concept are concatenated and input into a decaying network constructed using a multilayer perceptron. This network calculates the learner's understanding of the knowledge state of the first knowledge concept. Memory retention coefficient of a knowledge concept , Then, the memory retention coefficient is used to dynamically decay the value matrix of the previous time step element by element; In the behavior-aware gating module, the time step behavioral characteristics The input is a behavior encoder composed of a multilayer perceptron, which is mapped to a latent behavior semantic vector. ,Will With content vector The learned gain coefficients are obtained by concatenating the input gain gate network. , , It represents the level of credibility of the current interaction behavior; The prediction layer will integrate the capability status. With query vector The data is pieced together and mapped to potential capability values ​​via a multilayer perceptron. Then, the two-parameter logical model from Project Response Theory is introduced to predict the learner's probability of knowledge mastery; A D3QN-based question recommendation agent generates question recommendations based on the learner's current knowledge state, with a focus on knowledge gain. The multimodal learning resource recommendation module queries similar learning resources from the learning resource vector library based on the recommended questions output by the exercise recommendation agent, and together they form the recommendation results to be displayed to the learner.

Images