Student adaptive learning recommendation system based on deep learning

Abstract

The invention belongs to the technical field of learning recommendation, and particularly relates to a student adaptive learning recommendation system based on deep learning, which comprises a data processing module, a learning integration module, a cognition judgment module and an adaptive recommendation module, according to the system, by designing an L2 normalized linear attention mechanism, learning resource information is effectively integrated, and a learning resource theme is clarified; a context perception collaboration graph is constructed, multi-aspect cognitive states of the learner are judged, learning data are deeply analyzed, and learning recommendation accuracy is improved; the method comprises the following steps: establishing an SASRec model as a learning model, introducing an expert mixed self-attention layer, decomposing learning preferences of learners, optimizing and training the learning model by using a global cross entropy loss function, introducing a gradient scaling algorithm to update a parameter set, providing personalized recommendation according to related features of learning resources, continuously optimizing a learning resource recommendation list, and finally, carrying out learning resource recommendation. Therefore, the learning resource recommendation list better fits the actual learning condition, and the learning efficiency of the learner is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of learning recommendation, and specifically refers to a student adaptive learning recommendation system based on deep learning. Background Art

[0002] Adaptive learning recommendation refers to providing personalized learning resource recommendations for students based on their learning information. Existing learning recommendation technologies use relatively simple model architectures. Content-based learning recommendations are only based on the similarity between learning resources and content that students have already been exposed to. Collaborative filtering algorithms mainly rely on the behavioral similarity of student groups and lack comprehensive consideration of students' cognitive states and multi-dimensional characteristics of learning resources. Simple model architectures make it difficult to capture the deep relationship between students' complex learning needs and resources, and cannot provide personalized recommendations that truly meet students' comprehensive needs. Existing learning recommendation technologies show poor dynamic adaptability when facing students' ever-changing learning situations and needs. During the learning process, students' knowledge level, interest preferences and learning abilities change as they learn more, and their subject requirements for learning resources also change. Existing technologies usually cannot capture these changes in time and adjust recommendation strategies accordingly, resulting in the recommended resources gradually deviating from students' actual needs, and failing to continuously provide students with effective learning support, affecting learning efficiency and experience; Existing learning recommendation technologies lack effective means to explore the intrinsic characteristics and themes of learning resources when processing learning resources, and are unable to accurately judge the difficulty, importance, and degree of relevance of resources to different knowledge points. For comprehensive learning resources that contain multiple knowledge points, existing technologies find it difficult to accurately parse their core content and applicable scenarios, and cannot accurately match students' specific needs when making recommendations. In terms of organizing learning resources, relevant resources cannot be effectively integrated and sorted according to students' learning progress and knowledge structure, causing students to face problems of information overload or resource fragmentation when acquiring resources, making it difficult to build a systematic knowledge system. Summary of the invention

[0003] In view of the above problems, the present invention provides a student adaptive learning recommendation system based on deep learning. The system designs an L2 normalized linear attention mechanism, calculates the PMI matrix and vector centrality to determine the resource theme and importance, effectively integrates learning resource information, and clarifies the learning resource theme; constructs a context-aware collaboration graph to judge the learner's various cognitive states, deeply analyzes the learning data, and improves the accuracy of learning recommendations; establishes a SASRec model as a learning model, introduces an expert mixed self-attention layer, decomposes the learner's learning preferences, optimizes the training learning model with a global cross entropy loss function, introduces a gradient scaling algorithm to update the parameter set, provides personalized recommendations based on the relevant features of the learning resources, and continuously optimizes the learning resource recommendation list to make the learning resource recommendation list more in line with the actual learning situation and improve the learner's learning efficiency.

[0004] The student adaptive learning recommendation system based on deep learning includes: data processing module, learning integration module, cognitive judgment module and adaptive recommendation module; The data processing module is connected to the learning integration module, the learning integration module is connected to the cognitive judgment module, and the cognitive judgment module is connected to the adaptive recommendation module.

[0005] The data processing module collects the learner's learning data and the phrase information of the learning resources, and performs preprocessing operations on the learning data and the phrase information; The learning integration module designs an attention mechanism, obtains learning interaction sequences, and establishes connections between learners and learners and learning resources; The cognitive judgment module factors the learner's cognitive state, constructs a context-aware collaboration graph, and judges the learner's cognitive state by deeply analyzing the learning data; The adaptive recommendation module establishes a learning model, introduces an expert mixed self-attention layer, initializes the learning model and optimizes the training, generates a learning resource recommendation list, and predicts the learning resources that the learner wants to learn.

[0006] The student adaptive learning recommendation system based on deep learning adopts a student cognitive state perception method driven by deep learning to perform adaptive learning recommendation. The method includes the following steps: Step S1: Data processing: collect the learner's learning data, as well as the phrase information and word frequency information of the learning resources, and perform preprocessing operations on the learning data, phrase information and word frequency information; Step S2: Interaction sequence construction: construct the learning interaction sequence and embedding lookup table, introduce the normalized linear attention mechanism, and assign importance to the phrase information; Step S3: Cognitive state judgment: introduce a multi-head attention mechanism to calculate the attention score of the learning resource in the learning interaction sequence, reflect the similarity between the current learning resource and other learning resources, and obtain the probability distribution of the relative importance of each learning resource to the learner; Step S4: Adaptive learning recommendation: Establish the SASRec model as the learning model, output the learning resource recommendation list and preference score, optimize the learning model and optimize the learning resource recommendation list.

[0007] Furthermore, step S1 includes the following steps: Step S11: Collecting data: Collecting learners' learning data and phrase information of learning resources from the online learning platform. The learning data includes learning records, learning behaviors, learning results, and learners' personal information; Step S12: Data preprocessing: Apply the text preprocessing pipeline method to perform word segmentation, split the text into word segments, extract relevant features, encode using the label encoding method, and convert into numerical form.

[0008] Further, step S2 includes the following steps: Step S21: Define a node set: define learners with similar cognitive states as a graph structure, and define the learner node set as , where V is the learner node set, the learner node set contains learner nodes, E is the edge set, and the learning resource node set is defined as ,in, is a learning resource node set, which contains learning resource nodes. It is an edge set. When traversing the nodes, the nodes connected to the nodes by edges are called neighbor nodes, and the nodes not connected to the nodes by edges are called non-neighbor nodes. Step S22: Obtaining a learning interaction sequence: setting a time step, obtaining a list of learning resources for each learner arranged in chronological order as a learning interaction sequence; Step S23: Construct an embedding lookup table: Construct two trainable parameter matrices as the embedding lookup table, initialize the embedding vector of the learning resource node, and the trainable parameter matrices are defined as and , and Mapping learning resources and their positions in the learning interaction sequence into a dimensional dense vector; Step S24: Design an attention mechanism: Design an L2 normalized linear attention mechanism, combine it with the LinRec mechanism formula, use the L2 normalization method in the row and column directions to perform position mapping, multiply the dimensional dense vector with the corresponding weight matrix to obtain the query vector, key vector and value vector, and introduce the ELU activation function to control the negative value influence; The formula used by the L2 normalization method is as follows: ; ; The formula used by the ELU activation function is as follows: ; The formulas used to obtain the query vector, key vector, and value vector are as follows: ; ; ; The formula used by the LinRec mechanism is as follows: ; in, represents the i-th row vector of the query matrix, represents the j-th column vector of the key matrix, Indicates the way to normalize the query matrix. represents the normalization method of the key matrix, Q, K, V represent the query matrix, key matrix and value matrix respectively, d represents the dimension of the vector, and denote the L2 norm of the query vector and key vector respectively, x denotes the embedding vector, represents a nonlinear mapping of the embedding vector, , , denote the weight matrices that project the input matrix into query matrix, key matrix, and value matrix, respectively. represents the output matrix of the attention mechanism, Represents a nonlinear mapping of the query matrix, represents a nonlinear mapping of the key matrix, Indicates that the query matrix is ​​normalized after nonlinear mapping. The key matrix is ​​nonlinearly mapped and transposed before normalization; Step S25: assigning importance to phrases: organizing phrase information into a co-occurrence matrix, calculating a PMI matrix based on the co-occurrence matrix, and assigning importance to phrases based on the PMI matrix.

[0009] Further, step S25 includes the following steps: Step S251: Calculate the PMI matrix: organize the phrase information into a co-occurrence matrix. The elements contained in the co-occurrence matrix represent the co-occurrence frequencies of phrases and phrases. The PMI matrix and the PMI values ​​of the phrases are calculated based on the co-occurrence matrix. The formula used is as follows: ; in, Represents a pair of words, Indicates a calculation phrase The PMI matrix, represents the probability of phrases a and b appearing at the same time, and Respectively represent the probability of phrases a and b appearing alone; Step S252: Calculate vector centrality: Use the eigenvector centrality method to calculate vector centrality based on the PMI matrix. For each phrase appearing in the learning resource, perform a weighted summation of the centralities of neighboring nodes. Perform iterative calculation until convergence to obtain the vector centrality value of the phrase. The formula used for weighted summation of the centrality of neighboring nodes is as follows: ; in, Represents the vector centrality value of the current phrase a, represents a constant used to adjust the proportional relationship of the weighted sum of centrality, n represents the total number of phrases, represents the elements contained in the PMI matrix, indicating the strength of the relationship between phrase a and phrase b. The vector centrality value of phrase b representing the neighbors; Step S253: phrase importance distribution: using the vector centrality value as an indicator to measure the importance of the phrase, the phrases in the learning resource are arranged in descending order according to the vector centrality value.

[0010] Further, step S3 includes the following steps: Step S31: Obtaining embedded representation: Introducing a multi-head attention mechanism to predict the embedded representation of each learner node for each learning resource node; Step S32: Calculate the attention score: Calculate the attention score of each pair of learning resources in the learning interaction sequence, reflecting the similarity between the query vector of the current learning resource and the key vectors of other learning resources, and obtain the probability distribution of the relative importance of each learning resource. The formula used is as follows: ; ; in, Represents the attention score, reflecting the current The query vector of the first learning resource and the query vector of the second learning resource are The similarity between the key vectors of learning resources, Indicates the current The query vector of learning resources, Indicates The key vector of learning resources, represents the probability distribution, t represents the length of the learning interaction sequence; Step S33: Introduce expert mixed self-attention layer: Introduce an expert mixed self-attention layer in each attention head, perform weighted sum calculation on the value vectors in the learning interaction sequence, set the weights as attention scores, and obtain the embedding representation of the final learning resource. The formula used is as follows: ; in, Indicates that in the attention head h, from the perspective of expert n The embedding representation of learning resources, Indicates that in the attention head h, the expert n has learning resources relative to The attention score of each learning resource, Indicates A value vector of learning resources.

[0011] Further, step S31 includes the following steps: Step S311: Factorized representation: Initially assign factorized representation to each learner node, use Gaussian mixture initialization to capture the cognitive state of the learner node on the knowledge point, and obtain the initial disentangled cognitive state, initial encoding and disentangled components; Step S312: constructing a collaboration graph: constructing a context-aware collaboration graph, searching for similar neighbor nodes on each learner node through a disentanglement component, and generating a context-aware collaboration graph; Step S313: Update the collaboration graph: find K optimal neighbor nodes for each learner node through the initial disentangled cognitive state, measure the cognitive similarity between non-neighbor nodes and the current knowledge point, and update the context-aware collaboration graph; Step S314: Decoding and reconstruction: align the initial encoding with the collaborative perspective between learners to form a co-disentangled representation, predict the learning performance of each learner node on the learning resource based on the co-disentangled representation, and decode it to obtain the embedded representation of each learning resource node.

[0012] Further, step S4 includes the following steps: Step S41: Establish a SASRec model as a learning model, combine the multi-head perception prediction mechanism and the multi-head attention mechanism, and generate an embedding vector of the final learning resource according to the embedding representation of the final learning resource; Step S42: Decompose the learner's learning preference, use the MoE network in each attention head, the MoE network consists of N experts, each expert is represented by a trainable matrix, and the embedding vector of the learning resource is converted into an expert query vector, an expert key vector, and an expert value vector through the trainable matrix. The attention score of the learning resource in each attention head is calculated again to obtain the preference score of each learning resource, and generate a learning resource recommendation list; Step S43: randomly initialize the learning model, define the training objectives and parameter sets, use the global cross entropy loss function for optimization training, and optimize the learning model by minimizing the global cross entropy loss function; Step S44: Introduce a gradient scaling algorithm to adaptively update the parameter set, output the confidence of each learning resource preference score, update the preference score of each learning resource based on the confidence weighting, and select the top five preference scores as the final predicted learning resource recommendation list.

[0013] Further, step S42 includes the following steps: Step S421: Use a feedforward network with a ReLU activation function for nonlinear processing, use a softmax function for normalization, and use dropout and residual connection methods to reduce overfitting; Step S422: Each attention head independently calculates the dot product of the embedding vector of the learning resource and the learning interaction sequence to obtain the importance weight of each attention head, calculates the preference score of each learning resource by weighted sum, and generates a learning resource recommendation list.

[0014] The beneficial effects achieved by the present invention are as follows: (1) The system establishes the SASRec model, combines the multi-head perception prediction mechanism and the multi-head attention mechanism, and can learn the complex relationship patterns between learners and learning resources from multiple perspectives; through factorized representation and context-aware collaborative graph technology, it deeply analyzes the learner's cognitive state in terms of knowledge point understanding, mastery, learning style and ability; introduces expert mixed self-attention layer to decompose learning preferences, and through the collaboration of multiple experts, more carefully evaluates the matching degree between learning resources and learner preferences; it can achieve more accurate personalized recommendations and provide highly personalized learning resource recommendations based on the learner's specific situation; (2) The learning resource integration module of the system maps learning resources and their positions in the interaction sequence into dimensional dense vectors by constructing an embedding lookup table, designs an L2 normalized linear attention mechanism to capture the relationship between resources, calculates the PMI matrix and vector centrality to determine the resource theme and importance, and can deeply understand the characteristics and value of learning resources. When making recommendations, it can accurately match learners with the most relevant and valuable learning resources, avoiding learners wasting time screening through a large number of irrelevant resources; (3) The system uses a global cross-entropy loss function for optimization training and introduces a gradient scaling algorithm to adaptively update the parameter set. It can continuously adjust and optimize the parameters of the learning model and more accurately predict the learners' preference for different learning resources. In practical applications, as the learners' learning process progresses and their learning behaviors change, the model can update the recommendation list in a timely manner, always providing learners with the learning resources that best meet their current needs, and providing better quality and more accurate learning recommendation services. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 This is a module diagram of the student adaptive learning recommendation system based on deep learning proposed by the present invention; Figure 2 This is a flow chart of the student cognitive state perception method based on deep learning driven proposed by the present invention. DETAILED DESCRIPTION

[0016] The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0017] Example 1: Reference Figure 1 ,This embodiment provides a student adaptive learning recommendation system based on deep learning, the system includes: a data processing module, a learning integration module, a cognitive judgment module and an adaptive recommendation module; The data processing module is connected to the learning integration module, the learning integration module is connected to the cognitive judgment module, and the cognitive judgment module is connected to the adaptive recommendation module.

[0018] Example 2: Reference Figure 2 This embodiment is based on the above embodiment, and is a student adaptive learning recommendation system based on deep learning. It adopts a student cognitive state perception method driven by deep learning to perform adaptive learning recommendation. The method includes the following steps: Step S1: Data processing: collect the learner's learning data, as well as the phrase information and word frequency information of the learning resources, and perform preprocessing operations on the learning data, phrase information and word frequency information; Step S2: Interaction sequence construction: construct the learning interaction sequence and embedding lookup table, introduce the normalized linear attention mechanism, and assign importance to the phrase information; Step S3: Cognitive state judgment: introduce a multi-head attention mechanism to calculate the attention score of the learning resource in the learning interaction sequence, reflect the similarity between the current learning resource and other learning resources, and obtain the probability distribution of the relative importance of each learning resource to the learner; By analyzing learners' learning performance on different knowledge points, the degree of understanding and mastery can be evaluated; By analyzing the learners' learning behavior patterns at different knowledge points, the learning style and learning ability are evaluated. The learning behavior patterns include the distribution of learning time and the number of repeated learning. The learning styles include visual, auditory and kinesthetic. The learning ability includes memory and reasoning ability. Step S4: Adaptive learning recommendation: Establish the SASRec model as the learning model, output the learning resource recommendation list and preference score, optimize the learning model and optimize the learning resource recommendation list.

[0019] Embodiment 3: This embodiment is based on the above embodiment, and step S1 includes the following steps: Step S11: Collecting data: Collecting learners' learning data and phrase information of learning resources from the online learning platform. The learning data includes learning records, learning behaviors, learning results, and learners' personal information; Learning records include: learning time, learning courses, learning progress and practice records; Learning behaviors include: clicking, browsing, and pausing in learning courses; Academic performance includes: test scores, homework scores, and test scores; Personal information includes: grade and major; Step S12: Data preprocessing: Apply the text preprocessing pipeline method to perform word segmentation processing, split the text into word segments, extract relevant features, encode using the label encoding method, and convert into numerical form; Split the text into words, including lowercase conversion, HTML tag removal, separator formatting and punctuation removal; extract relevant features using the LancasterStemmer method for stem extraction, the WordNet-based NLTK lemmatizer for lemmatization, and the part-of-speech tagging method for word classification.

[0020] Embodiment 4: This embodiment is based on the above embodiment, and step S2 includes the following steps: Step S21: Define a node set: define learners with similar cognitive states as a graph structure, and define the learner node set as , where V is the learner node set, the learner node set contains learner nodes, E is the edge set, and the learning resource node set is defined as ,in, is a learning resource node set, which contains learning resource nodes. It is an edge set. When traversing the nodes, the nodes connected to the nodes by edges are called neighbor nodes, and the nodes not connected to the nodes by edges are called non-neighbor nodes. Step S22: Obtain learning interaction sequence: Set the time step and obtain the learning resource list of each learner in chronological order as the learning interaction sequence. The learning interaction sequence is expressed as , s represents the learner, represents the learning resource learning interaction sequence of the learner at the first time step, represents the learning resource learning interaction sequence of the learner at the i-th time step, represents the learning resource learning interaction sequence of the learner at the tth time step, and t represents the length of the learning interaction sequence; Step S23: Construct an embedding lookup table: Construct two trainable parameter matrices as the embedding lookup table, initialize the embedding vector of the learning resource node, and the trainable parameter matrices are defined as and , and Mapping learning resources and their positions in the learning interaction sequence into a dimensional dense vector; Step S24: Design an attention mechanism: Design an L2 normalized linear attention mechanism, combine it with the LinRec mechanism formula, use the L2 normalization method in the row and column directions to perform position mapping, multiply the dimensional dense vector with the corresponding weight matrix to obtain the query vector, key vector and value vector, and introduce the ELU activation function to control the negative value influence; The formula used by the L2 normalization method is as follows: ; ; The formula used by the ELU activation function is as follows: ; The formulas used to obtain the query vector, key vector, and value vector are as follows: ; ; ; The formula used by the LinRec mechanism is as follows: ; in, represents the i-th row vector of the query matrix, represents the j-th column vector of the key matrix, Indicates the way to normalize the query matrix. represents the normalization method of the key matrix, Q, K, V represent the query matrix, key matrix and value matrix respectively, d represents the dimension of the vector, and denote the L2 norm of the query vector and key vector respectively, x denotes the embedding vector, represents a nonlinear mapping of the embedding vector, , , denote the weight matrices that project the input matrix into query matrix, key matrix, and value matrix, respectively. represents the output matrix of the attention mechanism, Represents a nonlinear mapping of the query matrix, represents a nonlinear mapping of the key matrix, Indicates that the query matrix is ​​normalized after nonlinear mapping. The key matrix is ​​nonlinearly mapped and transposed before normalization; Step S25: assigning importance to phrases: organizing phrase information into a co-occurrence matrix, calculating a PMI matrix based on the co-occurrence matrix, and assigning importance to phrases based on the PMI matrix.

[0021] Embodiment 5: This embodiment is based on the above embodiment, and step S25 includes the following steps: Step S251: Calculate the PMI matrix: organize the phrase information into a co-occurrence matrix. The elements contained in the co-occurrence matrix represent the co-occurrence frequencies of phrases and phrases. The PMI matrix and the PMI values ​​of the phrases are calculated based on the co-occurrence matrix. The formula used is as follows: ; in, Represents a pair of words, Indicates a calculation phrase The PMI matrix, represents the probability of phrases a and b appearing at the same time, and Respectively represent the probability of phrases a and b appearing alone; Step S252: Calculate vector centrality: Use the eigenvector centrality method to calculate vector centrality based on the PMI matrix. For each phrase appearing in the learning resource, perform a weighted summation of the centralities of neighboring nodes. Perform iterative calculation until convergence to obtain the vector centrality value of the phrase. The formula used for weighted summation of the centrality of neighboring nodes is as follows: ; in, Represents the vector centrality value of the current phrase a, represents a constant used to adjust the proportional relationship of the weighted sum of centrality, n represents the total number of phrases, represents the elements contained in the PMI matrix, indicating the strength of the relationship between phrase a and phrase b. The vector centrality value of phrase b representing the neighbors; Step S253: phrase importance distribution: using the vector centrality value as an indicator to measure the importance of the phrase, the phrases in the learning resource are arranged in descending order according to the vector centrality value.

[0022] Embodiment 6: This embodiment is based on the above embodiment, and step S3 includes the following steps: Step S31: Obtaining embedded representation: Introducing a multi-head attention mechanism to predict the embedded representation of each learner node for each learning resource node; Step S32: Calculate the attention score: Calculate the attention score of each pair of learning resources in the learning interaction sequence, reflecting the similarity between the query vector of the current learning resource and the key vectors of other learning resources, and obtain the probability distribution of the relative importance of each learning resource. The formula used is as follows: ; ; in, Represents the attention score, reflecting the current The query vector of the first learning resource and the query vector of the second learning resource are The similarity between the key vectors of learning resources, Indicates the current The query vector of learning resources, Indicates The key vector of learning resources, represents the probability distribution, t represents the length of the learning interaction sequence; Step S33: Introduce expert mixed self-attention layer: Introduce an expert mixed self-attention layer in each attention head, perform weighted sum calculation on the value vectors in the learning interaction sequence, set the weights as attention scores, and obtain the embedding representation of the final learning resource. The formula used is as follows: ; in, Indicates that in the attention head h, from the perspective of expert n The embedding representation of learning resources, Indicates that in the attention head h, the expert n has learning resources relative to The attention score of each learning resource, Indicates A value vector of learning resources.

[0023] Embodiment 7: This embodiment is based on the above embodiment, and step S31 includes the following steps: Step S311: Factorized representation: Initially assign factorized representation to each learner node, use Gaussian mixture initialization to capture the cognitive state of the learner node on the knowledge point, and obtain the initial disentangled cognitive state, initial encoding and disentangled components; Step S312: constructing a collaboration graph: constructing a context-aware collaboration graph, searching for similar neighbor nodes on each learner node through a disentanglement component, and generating a context-aware collaboration graph; Step S313: Update the collaboration graph: find K optimal neighbor nodes for each learner node through the initial disentangled cognitive state, measure the cognitive similarity between non-neighbor nodes and the current knowledge point, and update the context-aware collaboration graph; Step S314: Decoding and reconstruction: align the initial encoding with the collaborative perspective between learners to form a co-disentangled representation, predict the learning performance of each learner node on the learning resource based on the co-disentangled representation, and decode it to obtain the embedded representation of each learning resource node.

[0024] Embodiment 8: This embodiment is based on the above embodiment, and step S4 includes the following steps: Step S41: Establish a SASRec model as a learning model, combine the multi-head perception prediction mechanism and the multi-head attention mechanism, and generate an embedding vector of the final learning resource according to the embedding representation of the final learning resource; Step S42: Decompose the learner's learning preference, use the MoE network in each attention head, the MoE network consists of N experts, each expert is represented by a trainable matrix, and the embedding vector of the learning resource is converted into an expert query vector, an expert key vector, and an expert value vector through the trainable matrix. The attention score of the learning resource in each attention head is calculated again to obtain the preference score of each learning resource, and generate a learning resource recommendation list; Step S43: randomly initialize the learning model, define the training objectives and parameter sets, use the global cross entropy loss function for optimization training, and optimize the learning model by minimizing the global cross entropy loss function; Step S44: Introduce a gradient scaling algorithm to adaptively update the parameter set, output the confidence of each learning resource preference score, update the preference score of each learning resource based on the confidence weighting, and select the top five preference scores as the final predicted learning resource recommendation list.

[0025] Embodiment 9: This embodiment is based on the above embodiment, and step S42 includes the following steps: Step S421: Use a feedforward network with a ReLU activation function for nonlinear processing, use a softmax function for normalization, and use dropout and residual connection methods to reduce overfitting; Step S422: Each attention head independently calculates the dot product of the embedding vector of the learning resource and the learning interaction sequence to obtain the importance weight of each attention head, calculates the preference score of each learning resource by weighted sum, and generates a learning resource recommendation list; In this embodiment, the generated learning resource recommendation list is: "Video course on core knowledge points of advanced mathematics", "Mathematical thinking expansion training exercise collection", "Mathematical modeling case analysis and practical guidance", "Python programming from entry to mastery tutorial", "Algorithm and data structure practice project collection", "Java Web development practical guide".

[0026] The present invention and its implementation methods are described above. Such description is not restrictive. If a person skilled in the art is inspired by it and creatively designs structures and implementation methods similar to the technical solution without departing from the purpose of the present invention, they should all fall within the protection scope of the present invention.

Claims

1. A student adaptive learning recommendation system based on deep learning, characterized by: The system includes a data processing module, a learning integration module, a cognitive judgment module and an adaptive recommendation module; The student adaptive learning recommendation system based on deep learning adopts a student cognitive state perception method driven by deep learning to perform adaptive learning recommendation. The method includes the following steps: Step S1: Data processing: collect the learner's learning data, as well as the phrase information and word frequency information of the learning resources, and perform preprocessing operations on the learning data, phrase information and word frequency information; Step S2: Interaction sequence construction: construct the learning interaction sequence and embedding lookup table, introduce the normalized linear attention mechanism, and assign importance to the phrase information; Step S3: Cognitive state judgment: introduce a multi-head attention mechanism to calculate the attention score of the learning resource in the learning interaction sequence, reflect the similarity between the current learning resource and other learning resources, and obtain the probability distribution of the relative importance of each learning resource to the learner; Step S4: Adaptive learning recommendation: Establish the SASRec model as the learning model, output the learning resource recommendation list and preference score, optimize the learning model and optimize the learning resource recommendation list.

2. The student adaptive learning recommendation system based on deep learning according to claim 1, characterized in that: Step S2 includes the following steps: Step S21: define a node set: define learners with similar cognitive states as a graph structure, define a learner node set and a learning resource node set, the learner node set includes learner nodes, the learning resource node set includes learning resource nodes, traverse the nodes, and define nodes that are connected to the nodes by edges as neighbor nodes, and nodes that are not connected to the nodes by edges as non-neighbor nodes; Step S22: Obtaining a learning interaction sequence: setting a time step, obtaining a list of learning resources for each learner arranged in chronological order as a learning interaction sequence; Step S23: construct an embedding lookup table: construct two trainable parameter matrices as an embedding lookup table, initialize the embedding vector of the learning resource node, and map the learning resource and the position of the learning resource in the learning interaction sequence into a dimensional dense vector; Step S24: Design an attention mechanism: Design an L2 normalized linear attention mechanism, use the L2 normalization method in the row and column directions to perform position mapping, and multiply the dimension dense vector with the corresponding weight matrix to obtain the query vector, key vector and value vector; Step S25: assigning importance to phrases: sorting the word frequency information into a co-occurrence matrix, calculating a PMI matrix based on the co-occurrence matrix, and assigning importance to phrases based on the PMI matrix.

3. The student adaptive learning recommendation system based on deep learning according to claim 2, characterized in that: Step S25 includes the following steps: Step S251: Calculate the PMI matrix: organize the word frequency information into a co-occurrence matrix, the elements contained in the co-occurrence matrix represent the co-occurrence frequencies of phrases and phrases, and calculate the PMI matrix and the PMI values ​​of the phrases based on the co-occurrence matrix; Step S252: Calculate vector centrality: Use the eigenvector centrality method to calculate vector centrality based on the PMI matrix. For each phrase appearing in the learning resource, perform a weighted summation of the centralities of neighboring nodes. Perform iterative calculation until convergence to obtain the vector centrality value of the phrase. Step S253: phrase importance distribution: using the vector centrality value as an indicator to measure the importance of the phrase, the phrases in the learning resource are arranged in descending order according to the vector centrality value.

4. The student adaptive learning recommendation system based on deep learning according to claim 3, characterized in that: Step S3 includes the following steps: Step S31: Obtaining embedded representation: Introducing a multi-head attention mechanism to predict the embedded representation of each learner node for each learning resource node; Step S32: Calculate the attention score: Calculate the attention score of each pair of learning resources in the learning interaction sequence, reflecting the similarity between the query vector of the current learning resource and the key vectors of other learning resources, and obtain the probability distribution of the relative importance of each learning resource; Step S33: Introducing an expert mixed self-attention layer: Introduce an expert mixed self-attention layer in each attention head, perform weighted sum calculation on the value vectors in the learning interaction sequence, set the weights as attention scores, and obtain the embedding representation of the final learning resource.

5. The student adaptive learning recommendation system based on deep learning according to claim 4, characterized in that: Step S31 includes the following steps: Step S311: Factorized representation: Initially assign factorized representation to each learner node, use Gaussian mixture initialization to capture the cognitive state of the learner node on the knowledge point, and obtain the initial disentangled cognitive state, initial encoding and disentangled components; Step S312: constructing a collaboration graph: constructing a context-aware collaboration graph, searching for similar neighbor nodes on each learner node through a disentanglement component, and generating a context-aware collaboration graph; Step S313: Update the collaboration graph: find the optimal neighbor node for each learner node through the initial disentangled cognitive state, measure the cognitive similarity between the non-neighbor nodes and the current knowledge point, and update the context-aware collaboration graph; Step S314: Decoding and reconstruction: align the initial encoding with the collaborative perspective between learners to form a co-disentangled representation, predict the learning performance of each learner node on the learning resource based on the co-disentangled representation, and decode it to obtain the embedded representation of each learning resource node.

6. The student adaptive learning recommendation system based on deep learning according to claim 5, characterized in that: Step S4 includes the following steps: Step S41: Establishing a SASRec model as a learning model, combining a multi-head perception prediction mechanism and a multi-head attention mechanism, and generating an embedding vector of the final learning resource according to the embedding representation of the final learning resource; Step S42: Decompose the learner's learning preference, use the MoE network in each attention head, the MoE network consists of N experts, each expert is represented by a trainable matrix, and the embedding vector of the learning resource is converted into an expert query vector, an expert key vector, and an expert value vector through the trainable matrix. The attention score of the learning resource in each attention head is calculated again to obtain the preference score of each learning resource, and generate a learning resource recommendation list; Step S43: randomly initialize the learning model, define the training objectives and parameter sets, use the global cross entropy loss function for optimization training, and optimize the learning model by minimizing the global cross entropy loss function; Step S44: Introduce a gradient scaling algorithm to adaptively update the parameter set, output the confidence level of each learning resource preference score, update the preference score of each learning resource based on the confidence level weight, and select the top five preference scores as the final predicted learning resource recommendation list.

7. The student adaptive learning recommendation system based on deep learning according to claim 6, characterized in that: Step S42 includes the following steps: Step S421: Use a feedforward network with a ReLU activation function for nonlinear processing, use a softmax function for normalization, and finally use dropout and residual connection methods to reduce overfitting; Step S422: Each attention head independently calculates the dot product of the embedding vector of the learning resource and the learning interaction sequence to obtain the importance weight of each attention head, calculates the preference score of each learning resource by weighted sum, and generates a learning resource recommendation list.

Images