An Automatic Recommendation Method for Teaching Resources Based on Deep Learning

Abstract

This invention discloses an automatic recommendation method for teaching resources based on deep learning, comprising: acquiring student learning data and preprocessing the data; extracting high-level features from multimodal data using a deep neural network to generate high-dimensional feature vectors; obtaining a unified resource feature representation through cross-modal fusion; generating personalized student demand vectors based on students' historical interaction data and neural attention mechanisms; generating student knowledge point mastery state vectors using a time-decay-based knowledge tracking algorithm; inputting the student's state vectors into an improved neural collaborative filtering network to generate an initial recommendation list and preference scores; constructing a resource semantic association graph and updating resource recommendation scores using a diffusion recommendation model. This invention achieves accurate matching and dynamic recommendation optimization between teaching resources and students' personalized learning needs through multimodal deep learning, knowledge tracking algorithms, improved neural collaborative filtering, and diffusion recommendation techniques.

Description

Technical Field

[0001] This invention relates to the fields of intelligent education and deep learning technology, and in particular to an automatic recommendation method for teaching resources based on deep learning. Background Technology

[0002] With the rapid development of online education, teaching resource recommendation systems have become an important tool for improving learning efficiency and achieving personalized education. Traditional recommendation methods mainly rely on collaborative filtering and content recommendation techniques. Collaborative filtering algorithms make recommendations by analyzing historical interaction data between students and teaching resources, but they face the problem of data sparsity, especially for new users or new resources, where the recommendation effect drops significantly. Content recommendation methods rely on the text descriptions, tags, and other attributes of teaching resources. While they can provide recommendations to a certain extent, they ignore changes in students' cognitive states and learning needs, leading to recommendations that may not match students' actual abilities and interests.

[0003] In recent years, deep learning technology has been widely applied to recommender systems, especially neural collaborative filtering models. These models can automatically extract features from complex student behavior data and optimize recommendation results through deep neural networks. However, existing deep learning recommender systems still have significant shortcomings. First, most methods only process single data modalities, such as text or behavioral data, neglecting the rich semantic information contained in multimodal educational resources such as videos and images. Second, knowledge tracking algorithms often operate independently of the recommender system, failing to effectively combine students' dynamic cognitive states with resource features. Third, traditional neural collaborative filtering models lack modeling of the dynamic matching relationship between student ability and resource difficulty, and cannot adjust recommendation strategies in real time according to students' current abilities to ensure that recommended content matches students' current level while promoting in-depth knowledge development.

[0004] Therefore, how to provide an automatic recommendation method for teaching resources based on deep learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an automatic recommendation method for teaching resources based on deep learning. By combining multimodal data with neural collaborative filtering, knowledge tracking algorithms, and diffusion recommendation models, this invention details how to extract and fuse features through deep neural networks to achieve personalized teaching resource recommendations. The method of this invention has the advantages of accurately matching students' learning needs, dynamically adjusting recommendation strategies, effectively fusing multimodal data, and optimizing recommendation results in real time.

[0006] An automatic recommendation method for teaching resources based on deep learning according to an embodiment of the present invention includes:

[0007] Acquire multimodal educational data, preprocess the multimodal educational data, and extract the low-level features of each modality;

[0008] A pre-defined deep neural network is used to extract high-level features from low-level features, generating high-dimensional feature vectors for each modality.

[0009] Cross-modal fusion is performed on the high-dimensional feature vectors of each modality to generate a unified multimodal resource feature vector;

[0010] Based on students' historical interaction data in multimodal education data, a student-resource interaction matrix is ​​constructed. The feature vectors of multimodal resources that students have interacted with in the past are weighted and aggregated using a neural attention mechanism to generate personalized demand vectors for students.

[0011] Input students’ historical interaction data into a knowledge tracking algorithm based on time decay to generate a vector of students’ knowledge point mastery status.

[0012] Based on the student's personalized demand vector, multimodal resource feature vector, and student's knowledge mastery state vector, an improved neural collaborative filtering network is used to predict the interaction probability between the student and each candidate resource, and generate an initial recommendation list and corresponding preference scores.

[0013] A semantic association graph between resources is constructed. The resources in the initial recommendation list are used as seed nodes. The preference scores corresponding to each resource in the initial recommendation list and the multimodal resource feature vectors are used to form a multidimensional preference representation. Based on the multidimensional preference representation, a diffusion recommendation model is used to propagate preferences on the association graph. During the propagation process, the multidimensional preference representation is updated by combining the multimodal resource feature vectors corresponding to the nodes to generate the recommendation scores corresponding to each resource.

[0014] The resources are reordered based on the updated recommendation scores, and the final recommendation list is output.

[0015] Optionally, acquiring multimodal educational data includes:

[0016] Student historical interaction data, which includes the number of times students clicked on each teaching resource, the viewing time, the accuracy rate of their test answers, the homework completion score, and the type of interactive feedback behavior;

[0017] Text data of teaching resources, including detailed descriptions of course content, learning objectives, syllabus, and course tags;

[0018] The video data of the teaching resources includes the total duration of the teaching video, video frame data, and video subtitle text.

[0019] Acquire image data of teaching resources, including course PPT pages, diagrams, and screenshots of whiteboard writing.

[0020] Optionally, the extraction of low-level features for each modality includes:

[0021] The multimodal educational data is preprocessed, including: cleaning, removing outliers, and standardizing student historical interaction data; segmenting text data, removing stop words, and tagging parts of speech; extracting frames from video data and extracting the timestamp of each video frame; and normalizing image data by size and converting color space.

[0022] The preprocessed data is subjected to low-level feature extraction, including: converting text units into corresponding word vectors using natural language processing techniques; extracting temporal features from video frame sequences and converting them into time series features; and using convolutional neural networks to extract image features and generate image feature vectors.

[0023] The processed student historical interaction data is timestamped and integrated with the underlying feature vectors of each modality to form a multimodal feature representation, which serves as the input data for high-level feature extraction.

[0024] Optionally, generating high-dimensional feature vectors for each modality includes:

[0025] For text modalities, a pre-trained language model is used to extract high-level semantic features;

[0026] A 3D convolutional network is used to extract spatiotemporal features for video modalities;

[0027] Visual semantic features are extracted from image modalities using a visual transformer.

[0028] The high-dimensional feature vectors of each modality are uniformly mapped to the same dimensional space.

[0029] Optionally, constructing the student-resource interaction matrix includes:

[0030] The student historical interaction data is aggregated according to student identifiers and resource identifiers to form a two-dimensional interaction index structure with students as rows and teaching resources as columns;

[0031] Based on the interaction records between each student and each teaching resource, we extract interaction features such as the number of clicks, viewing time, test answer accuracy, homework completion score, and interaction feedback behavior type. We then normalize these interaction features and map them to a unified numerical space.

[0032] The various interaction features are weighted and fused according to preset weights to generate the corresponding interaction intensity value between the student and the teaching resources, and the interaction intensity value is filled into the corresponding position in the two-dimensional interaction index structure;

[0033] For student and teaching resource combinations with no interaction records, the interaction intensity value at the corresponding position is initialized to zero or set as a missing marker, and implicit completion is performed through model training;

[0034] The student-resource interaction matrix is ​​subjected to sparse storage and index optimization processing to generate a standardized student-resource interaction matrix.

[0035] Optionally, generating a student's personalized needs vector includes:

[0036] The multimodal resource features of the current resource to be predicted are represented as a query vector. The multimodal features of each resource that the student has interacted with in the past are represented as keys and values ​​using a neural attention mechanism. The relevance weights between the historically interacted resources and the current resource are calculated by scaling dot product attention. The features of the historically interacted resources are weighted and summed to generate a personalized demand vector for the student.

[0037] Optionally, generating the student's knowledge point mastery state vector includes:

[0038] Obtain students' historical interaction data, including video viewing time, homework completion rate, test accuracy, and interaction feedback;

[0039] The mastery status of each knowledge point is calculated based on students' historical interaction data, and the mastery status is adjusted using the time decay factor and the Ebbinghaus forgetting curve decay formula.

[0040] The knowledge tracking algorithm combines students' multimodal learning data and the time interval of historical interactions to update the mastery status of each knowledge point. The interaction data of video, image and text modalities jointly affect the knowledge point update according to feature weights.

[0041] The decayed knowledge point mastery status is integrated into a student knowledge point mastery status vector.

[0042] Optionally, generating the initial recommendation list and corresponding preference scores includes:

[0043] The vectors of students' personalized needs, multimodal resource features, and students' knowledge mastery status are aligned and standardized in terms of dimensions.

[0044] An improved neural collaborative filtering network was constructed, including a basic neural collaborative filtering layer, a cognitive dynamic bias layer, a difficulty adaptive calibration layer, and a fusion output layer;

[0045] The student's personalized needs vector and multimodal resource feature vector are input into the basic neural collaborative filtering layer. The basic interaction score is obtained through vector concatenation and multilayer perceptron mapping. The student's knowledge mastery state vector and the multimodal resource feature vector of the candidate resources are input into the cognitive dynamic bias layer to generate dynamic bias terms. The dynamic bias terms are then superimposed on the basic interaction score to obtain the cognitively enhanced interaction score.

[0046] The cognitively enhanced interaction score is input into the difficulty adaptive calibration layer. The resource difficulty extraction unit extracts the difficulty coefficient of the candidate resource, and the student ability extraction unit extracts the student ability vector. The difficulty matching score is generated based on the matching relationship between the student ability vector and the resource difficulty coefficient. The cognitively enhanced interaction score is then calibrated based on the difficulty matching score to obtain the final interaction score of the candidate resource.

[0047] The final interaction score is input into the fusion output layer, and the interaction probability between the student and each candidate resource is generated through normalization mapping. The interaction probability is then used as the preference score corresponding to the candidate resource.

[0048] All candidate resources are sorted in descending order according to preference scores, and a preset number of candidate resources at the top of the sort are selected to form an initial recommendation list.

[0049] Output the initial recommendation list and the preference score corresponding to each candidate resource.

[0050] Optionally, the semantic relationship graph between the constructed resources includes:

[0051] Obtain the multimodal resource feature vectors corresponding to each teaching resource, and perform unified normalization processing on the multimodal resource feature vectors as the semantic representation of the resources;

[0052] A graph structure is constructed with each teaching resource as a node, and each teaching resource is associated and mapped with the corresponding multimodal resource feature vector to form a node feature set;

[0053] The semantic similarity between any two teaching resources is calculated based on their multimodal resource feature vectors, and the semantic similarity is obtained by cosine similarity or Euclidean distance transformation.

[0054] The calculated semantic similarity is compared with a preset similarity threshold. When the semantic similarity is greater than the preset threshold, an edge connection is established between the two corresponding teaching resource nodes.

[0055] The semantic similarity is used as the weight value of the established edge, and each edge in the graph is weighted according to the weight value to form a weighted resource semantic association graph;

[0056] The resource semantic association graph is subjected to sparsification and structural optimization, and edge connections with weight values ​​higher than a preset threshold are retained to generate a resource semantic association graph.

[0057] Optionally, the step of updating the preference scores of each resource to generate recommendation scores includes:

[0058] A semantic association graph of resources is constructed with each teaching resource as a node. The semantic similarity between the multimodal resource feature vectors of resources is used as the edge. When the similarity is higher than a preset threshold, an edge connection is established between the corresponding nodes, and the semantic similarity is used as the edge weight.

[0059] A diffusion recommendation model is constructed, with resources in the initial recommendation list as seed nodes. Preference propagation is performed on the resource semantic association graph, and the initial preference scores corresponding to the seed nodes and the corresponding resource feature information are used to form a multi-dimensional preference representation, which is then propagated to adjacent nodes along the edges.

[0060] During the propagation process, the multidimensional preference representations received by each node are associated and mapped with the multimodal resource feature vectors corresponding to the current node, and the multidimensional preference representations are updated to generate the propagation preference representations corresponding to the nodes.

[0061] The propagation preference representations received by each node from different propagation paths are accumulated, and the contributions of different propagation paths are weighted and fused according to the correlation strength of the propagation paths to generate a comprehensive preference representation for each resource.

[0062] The comprehensive preference representation is mapped to a scalar score and fused with the initial preference scores of each resource to obtain the final recommendation score for each resource.

[0063] All candidate resources are sorted according to the final recommendation score to obtain the recommendation results for output.

[0064] The beneficial effects of this invention are:

[0065] By combining multimodal feature fusion with a time-decay-based knowledge tracking algorithm, this invention achieves collaborative modeling of student learning status and semantic information of teaching resources, improving the accuracy and personalization of recommendation results. Specifically, this invention utilizes a pre-defined deep neural network to extract high-level features from multimodal data such as text, video, and images, and generates a unified multimodal resource feature vector through cross-modal fusion. This results in a better comprehensive representation of teaching resource content, overcoming the problem of insufficient utilization of single-modal information in traditional recommendation methods. Simultaneously, by constructing a student personalized needs vector and a student knowledge point mastery state vector, and introducing these vectors into an improved neural collaborative filtering network, this invention achieves joint modeling of student interests and cognitive levels, making the recommendation results more closely match students' actual learning needs.

[0066] After generating the initial recommendation list, a resource semantic association graph is constructed, and a diffusion recommendation model based on multidimensional preference representation is introduced. The preference scores of each resource in the initial recommendation list are combined with multimodal resource feature vectors to form a multidimensional preference representation, which is then propagated and updated within the graph structure. This allows preference information to diffuse among resources according to semantic similarity. Teaching resources that are semantically related to high-preference resources but lack historical interaction can gain propagation gain, improving their recommendation scores, alleviating the cold-start problem, and enhancing the diversity of recommendation results.

[0067] By synergistically combining multimodal feature modeling, learning state modeling, and a graph diffusion mechanism based on multidimensional preference representation, this invention achieves significant improvements in recommendation accuracy, personalization, knowledge adaptability, and the coverage and exploratory nature of the recommendation results. It provides learners with more reasonable, continuous, and targeted personalized learning resource recommendations. Overall, this invention realizes a complete recommendation loop from multimodal feature modeling and learning state perception to multidimensional preference propagation optimization, enabling teaching resource recommendations to achieve high-precision personalized matching, dynamic cognitive adaptation, and enhanced result diversity. Attached Figure Description

[0068] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They explain the invention together with the embodiments of the invention and do not constitute a limitation thereof. In the drawings:

[0069] Figure 1 This is an overall flowchart of a deep learning-based automatic recommendation method for teaching resources proposed in this invention.

[0070] Figure 2 This is a schematic diagram of the neural collaborative filtering network structure of the automatic recommendation method for teaching resources based on deep learning proposed in this invention.

[0071] Figure 3This is a schematic diagram of the diffusion model-based reordering process in the automatic recommendation method for teaching resources based on deep learning proposed in this invention. Detailed Implementation

[0072] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0073] refer to Figure 1 , Figure 2 and Figure 3 Methods for automatically recommending teaching resources based on deep learning include:

[0074] Acquire multimodal educational data, preprocess the multimodal educational data, and extract the low-level features of each modality;

[0075] A pre-defined deep neural network is used to extract high-level features from low-level features, generating high-dimensional feature vectors for each modality.

[0076] Cross-modal fusion is performed on the high-dimensional feature vectors of each modality to generate a unified multimodal resource feature vector;

[0077] Based on students' historical interaction data in multimodal education data, a student-resource interaction matrix is ​​constructed. The feature vectors of multimodal resources that students have interacted with in the past are weighted and aggregated using a neural attention mechanism to generate personalized demand vectors for students.

[0078] Input students’ historical interaction data into a knowledge tracking algorithm based on time decay to generate a vector of students’ knowledge point mastery status.

[0079] Based on the student's personalized demand vector, multimodal resource feature vector, and student's knowledge mastery state vector, an improved neural collaborative filtering network is constructed and input to predict the interaction probability between the student and each candidate resource, and generate an initial recommendation list and corresponding preference scores.

[0080] Construct a semantic relationship graph between resources;

[0081] Using the resources in the initial recommendation list as seed nodes, the preference scores corresponding to each resource in the initial recommendation list and the multimodal resource feature vectors are used to construct a multidimensional preference representation. Based on the multidimensional preference representation, a diffusion recommendation model is used to propagate preferences on the association graph. During the propagation process, the multidimensional preference representation is updated in combination with the multimodal resource feature vectors corresponding to the nodes to generate recommendation scores for each resource.

[0082] The resources are reordered based on the updated recommendation scores, and the final recommendation list is output.

[0083] In this embodiment, acquiring multimodal educational data includes:

[0084] Student historical interaction data, which includes the number of times students clicked on each teaching resource, the viewing time, the accuracy rate of their test answers, the homework completion score, and the type of interactive feedback behavior;

[0085] Text data of teaching resources, including detailed descriptions of course content, learning objectives, syllabus, and course tags;

[0086] The video data of the teaching resources includes the total duration of the teaching video, video frame data, and video subtitle text.

[0087] Acquire image data of teaching resources, including course PPT pages, diagrams, and screenshots of whiteboard writing.

[0088] In this embodiment, the extraction of low-level features for each modality includes:

[0089] The multimodal educational data is preprocessed, including: cleaning, removing outliers, and standardizing student historical interaction data; segmenting text data, removing stop words, and tagging parts of speech; extracting frames from video data and extracting the timestamp of each video frame; and normalizing image data by size and converting color space.

[0090] The preprocessed data is subjected to low-level feature extraction, including: converting text units into corresponding word vectors using natural language processing techniques; extracting temporal features from video frame sequences and converting them into time series features; and using convolutional neural networks to extract image features and generate image feature vectors.

[0091] The processed student historical interaction data is timestamped and integrated with the underlying feature vectors of each modality to form a multimodal feature representation, which serves as the input data for high-level feature extraction.

[0092] The underlying feature extraction specifically includes:

[0093] Low-level physical or statistical features directly extracted from raw data without high-level semantic abstraction; the specific form of text low-level features is: after word segmentation, stop word removal, and part-of-speech tagging, each text unit is converted into a bag-of-words vector or TF-IDF vector, with the vector dimension equal to the number of unique words in the corpus, typically 10,000 to 50,000 dimensions. The specific form of video low-level features is: after extracting frames from the video, each frame is scaled to 224 x 224 pixels, and features of each frame are extracted using a pre-trained convolutional neural network. The output of the last pooling layer is taken to obtain a 1024-dimensional feature vector for each frame. For a video containing N frames, the system obtains an N x 1024 feature matrix. The system calculates the optical flow information between adjacent frames to generate motion feature vectors with a dimension of 10. Finally, the video low-level features are concatenated from the time-averaged pooling result of the frame feature matrix and the motion features, with a total dimension of 1034. The specific form of image low-level features is: for each The system scales each image to 224 x 224 pixels and performs RGB three-channel normalization. It then uses the same ResNet-18 convolutional neural network as the one used for video frame extraction to extract features. The system takes the global average pooling output of the last convolutional layer to obtain a 1024-dimensional feature vector. This vector represents the low-level features of the image, with each element ranging from zero to one. The output of these low-level features is uniformly stored as a JSON-formatted vector file. The filename consists of "Resource ID_Modal Type_Low-Level Feature", for example, "abc123_text_base.json", for easy reading and processing.

[0094] In this embodiment, generating the high-dimensional feature vector for each modality includes:

[0095] For text modalities, a pre-trained language model is used to extract high-level semantic features;

[0096] A 3D convolutional network is used to extract spatiotemporal features for video modalities;

[0097] Visual semantic features are extracted from image modalities using a visual transformer.

[0098] The high-dimensional feature vectors of each modality are uniformly mapped to the same dimensional space through a linear transformation projection method and cross-modal fusion is performed.

[0099] The pre-trained language model extracts high-level semantic features, specifically:

[0100] For the text modality, the pre-trained language model BERT is used to extract high-level semantic features. The pre-trained language model BERT is a deep learning model pre-trained on large-scale text data, which can learn rich language expressions and contextual information. In this step, through the pre-trained language model, the text data is transformed into high-dimensional vectors with deep semantic information. These vectors not only represent the basic information of words, but also capture the grammatical and semantic relationships in the context, generating high-level semantic features for each text.

[0101] In this embodiment, constructing the student-resource interaction matrix includes:

[0102] The student historical interaction data is aggregated according to student identifiers and resource identifiers to form a two-dimensional interaction index structure with students as rows and teaching resources as columns;

[0103] Based on the interaction records between each student and each teaching resource, the number of clicks, viewing time, test answer accuracy, homework completion score and interaction feedback behavior type are extracted, and the various interaction features are normalized and mapped to a unified numerical space.

[0104] Based on preset weights, the various interaction features are weighted and fused to generate the corresponding interaction intensity value between students and teaching resources. The interaction intensity value is then filled into the corresponding position in the two-dimensional interaction index structure. The student-resource interaction matrix is ​​then subjected to sparse storage and index optimization processing to generate a standardized student-resource interaction matrix.

[0105] The preset weights are specifically as follows:

[0106] Fixed weight parameters are set for click count, viewing time, test accuracy, homework completion score, and interactive feedback behavior type to reflect the impact of different interaction features on students' actual learning engagement. The interaction features are normalized: click count is mapped to a uniform numerical range, viewing time is converted into an effective viewing ratio, test accuracy and homework completion score are directly mapped to standardized values, and interactive feedback behaviors such as likes, favorites, rewinds, pauses, and comments are encoded as corresponding behavioral feature values. Preset weights are assigned to each feature, with higher weights for test accuracy, viewing time, and homework completion score, which directly reflect learning effectiveness and engagement. Click count and interactive feedback behavior type have secondary weights. The preset weights for click count are 0.10, viewing time is 0.30, test accuracy is 0.25, homework completion score is 0.25, and interactive feedback behavior type is 0.10.

[0107] The generation of the interaction strength value is specifically as follows:

[0108] The number of clicks, viewing time, test answer accuracy, homework completion score, and interaction feedback behavior type between the same student and the same teaching resource are normalized to obtain corresponding standardized feature values. Each standardized feature value is multiplied by its corresponding preset weight, and the weighted results are summed to obtain the interaction strength value between the student and the teaching resource. If an interaction feature is missing, it is either set to zero or filled in according to preset rules before participating in the weighted calculation. Finally, the obtained interaction strength values ​​are filled into the corresponding positions in the student-resource interaction matrix as input data for subsequent recommendation calculations. For student-teaching resource combinations with no interaction records, the interaction strength value at the corresponding position is initialized to zero or set as a missing marker, and implicit completion is performed through model training.

[0109] In this embodiment, generating a student's personalized needs vector includes:

[0110] The multimodal resource features of the current resource to be predicted are represented as a query vector. The multimodal features of each resource that the student has interacted with in the past are represented as keys and values ​​using a neural attention mechanism. The relevance weights between the historical interaction resources and the current resource are calculated by scaling dot product attention. The features of the historical interaction resources are weighted and summed to generate a personalized demand vector for the student.

[0111] The weighted summation of the features of historical interaction resources is specifically as follows:

[0112] In the process, the multimodal resource feature representation of the current resource to be predicted is input into a linear mapping layer to obtain a query vector. The multimodal resource feature representation of each resource that the student has interacted with in the past is input into two other linear mapping layers to obtain corresponding key vectors and value vectors. The current resource corresponds to a query vector, and each resource in the historical resource set corresponds to a key vector and a value vector. The query vector is then multiplied by the key vector of each historical resource to obtain the relevance score between the current resource and each historical resource. To prevent the dot product result from being too high due to the large vector dimension, the score is divided by the square root of the key vector dimension to perform scaling. All scaled scores are then Softmax normalized and converted into weight coefficients. The sum of all weights is 1. The larger the value, the stronger the relevance between the corresponding historical resource and the current resource to be predicted. The value vectors corresponding to each historical resource are then weighted and summed using the above weight coefficients. The value vector of each historical resource is multiplied by the corresponding weight and then summed to obtain a comprehensive representation vector, which is the student's personalized demand vector. This vector represents the student's interests, preferences, and demand features that are most relevant to the current resource to be predicted, formed during the historical learning process.

[0113] In this embodiment, generating the student's knowledge point mastery state vector includes:

[0114] Obtain students' historical interaction data, including video viewing time, homework completion rate, test accuracy, and interaction feedback;

[0115] The mastery status of each knowledge point is calculated based on students' historical interaction data, and the mastery status is adjusted using the time decay factor and the Ebbinghaus forgetting curve decay formula.

[0116] The knowledge tracking algorithm combines students' multimodal learning data and the time interval of historical interactions to update the mastery status of each knowledge point. The interaction data of video, image and text modalities jointly affect the knowledge point update according to the feature weights, and integrate the decayed knowledge point mastery status into a student knowledge point mastery status vector.

[0117] The update of the mastery status of each knowledge point is specifically as follows:

[0118] Acquire students' historical interaction data, including video viewing time, homework completion rate, test answer accuracy, and interactive feedback. Calculate the mastery status of each knowledge point based on the historical interaction data, and adjust the mastery status of each knowledge point using a time decay factor. Specifically, the student's mastery status of the knowledge point is updated according to the time difference and the Ebbinghaus forgetting curve decay formula. Integrate the time-decayed updated student knowledge point mastery status data into a student knowledge point mastery status vector, and pass it to the generation and recommendation calculation of personalized demand vectors. Calculate the difference between the student's last interaction time with the knowledge point and the current time using a simplified time decay algorithm based on the Ebbinghaus forgetting curve, and adjust the weights using the decay factor.

[0119] The method of adjusting the mastery status of each knowledge point using a time decay factor is as follows:

[0120] The time decay factor controls the decay rate of knowledge point memory. The algorithm assumes that students' mastery of knowledge points will gradually decline over time. The system updates students' mastery status of knowledge points according to the time decay factor and generates students' knowledge tracking vectors. For each knowledge point, if students have new interactions, the mastery status of knowledge points is adjusted according to the results of the student's interactions.

[0121] In this embodiment, generating the initial recommendation list and corresponding preference scores includes:

[0122] The vectors of students' personalized needs, multimodal resource features, and students' knowledge mastery status are aligned and standardized in terms of dimensions.

[0123] An improved neural collaborative filtering network was constructed, including a basic neural collaborative filtering layer, a cognitive dynamic bias layer, a difficulty adaptive calibration layer, and a fusion output layer;

[0124] The student's personalized needs vector and multimodal resource feature vector are input into the basic neural collaborative filtering layer. The basic interaction score is obtained through vector concatenation and multilayer perceptron mapping. The student's knowledge mastery state vector and the multimodal resource feature vector of the candidate resources are input into the cognitive dynamic bias layer to generate dynamic bias terms. The dynamic bias terms are then superimposed on the basic interaction score to obtain the cognitively enhanced interaction score.

[0125] The cognitively enhanced interaction score is input into the difficulty adaptive calibration layer. The resource difficulty extraction unit extracts the difficulty coefficient of the candidate resource, and the student ability extraction unit extracts the student ability vector. The difficulty matching score is generated based on the matching relationship between the student ability vector and the resource difficulty coefficient. The cognitively enhanced interaction score is then calibrated based on the difficulty matching score to obtain the final interaction score of the candidate resource.

[0126] The final interaction score is input into the fusion output layer, and the interaction probability between the student and each candidate resource is generated through normalization mapping. The interaction probability is then used as the preference score corresponding to the candidate resource.

[0127] All candidate resources are sorted in descending order according to preference scores, and a preset number of candidate resources at the top of the sort are selected to form an initial recommendation list.

[0128] Output the initial recommendation list and the preference score for each candidate resource.

[0129] The construction of the basic neural collaborative filtering layer specifically involves:

[0130] The basic neural collaborative filtering layer adopts a standard multilayer perceptron structure. The student's personalized demand vector and the resource feature vector are concatenated and input into the multilayer perceptron, outputting a basic interaction score. Specifically, the student's personalized demand vector and the resource feature vector are both 256-dimensional, resulting in a 512-dimensional joint vector. This joint vector passes through three fully connected layers with 256, 128, and 64 neurons respectively. Each fully connected layer is followed by a batch normalization layer and a ReLU activation function. During training, some neurons are randomly dropped at a dropout rate of 0.2 to prevent overfitting. The final fully connected layer outputs a scalar value as the basic interaction score, representing the student's initial preference for the resource.

[0131] The construction of the cognitive dynamic bias layer specifically involves:

[0132] In traditional neural collaborative filtering networks, recommendation systems typically score and recommend based on the student's interaction history with resources. This invention introduces a cognitive dynamic bias layer, which dynamically adjusts the recommendation score based on the student's mastery of each knowledge point. Through gating mechanisms and time decay modulation, the recommendation score is adjusted to ensure students receive more recommendations for their weaker areas, resulting in higher personalization and more accurate recommendations considering the student's dynamic mastery. The cognitive dynamic bias layer is responsible for dynamically adjusting the basic interaction score based on the student's current knowledge point mastery. The cognitive dynamic bias layer receives the output student knowledge point mastery vector, whose dimensions correspond to the dimensions in the knowledge graph. The total number of knowledge points is the same; the core of the cognitive dynamic bias layer is a gating mechanism. The specific calculation process is as follows: the student's knowledge point mastery state vector and the resource feature vector are concatenated to obtain a joint vector of 1200 + 256 = 1456 dimensions; the joint vector is input into a fully connected layer with a Sigmoid activation function. The output dimension of the fully connected layer is the same as the output dimension of the basic interaction score, which is 64 dimensions, and the output is a gating vector; at the same time, the system maintains a learnable bias basis vector, which is also 64 dimensions; the gating vector and the bias basis vector are multiplied element-wise to obtain the dynamic bias term, and then added to the basic interaction score to obtain the cognitively enhanced interaction score;

[0133] The construction of the adaptive calibration layer for difficulty is specifically as follows:

[0134] Traditional neural collaborative filtering methods mostly do not consider the matching between resource difficulty and student ability, and make recommendations based on users' historical interaction data. This invention adds a difficulty adaptive calibration layer, which dynamically adjusts the recommendation score by calculating the matching degree between the resource difficulty coefficient and the student's ability vector. Using a bilinear interaction function and a sigmoid function, the difficulty matching score is used as a calibration coefficient to adjust the basic interaction score, matching the student's learning level with the recommended content. This ensures that the recommended resources are both suitable for the student's current ability and adaptable to their learning progress. The layer is responsible for assessing the matching degree between the student's current ability level and the resource difficulty, and calibrating the interaction score after cognitive enhancement accordingly. The difficulty adaptive calibration layer contains three sub-units: a resource difficulty extraction unit that inputs a multimodal resource feature representation into an independent difficulty level. The regression network consists of two stacked fully connected layers, outputting a scalar value between 0 and 1 as the resource difficulty coefficient. The student ability extraction unit inputs the student's personalized need vector into a fully connected mapping layer and outputs the student ability vector, with the dimension being the same as the total number of knowledge points. The difficulty matching calculation unit uses a bilinear interaction function to calculate the matching score. The specific formula is that the matching score equals the transpose of the student ability vector multiplied by the transformation matrix and then multiplied by the resource difficulty coefficient. The transformation matrix is ​​a learnable parameter matrix with the dimension being the number of knowledge points multiplied by one. The matching score is mapped by the Sigmoid function to obtain the calibration coefficient, which means that when the student's ability and resource difficulty are highly matched, the calibration coefficient is close to 1, and when the ability and difficulty are mismatched, the calibration coefficient approaches 0. Finally, the interaction score after cognitive enhancement is multiplied by the calibration coefficient to obtain the final interaction score after difficulty calibration.

[0135] The construction of the fusion output layer specifically involves:

[0136] The system weights and fuses the basic interaction score, dynamic bias term, and difficulty matching score to output the final predicted interaction probability. The synergistic effect of the aforementioned cognitive dynamic bias layer and difficulty adaptive calibration layer enables the improved neural collaborative filtering network to simultaneously achieve three goals: capturing students' stable interest preferences through the basic layer; adjusting the recommendation direction based on students' real-time cognitive state through the dynamic bias layer, focusing on knowledge gaps; and ensuring that the recommended content is within the students' capabilities through the difficulty calibration layer. The organic integration of these three aspects makes the recommendation results superior to traditional methods in terms of personalization, adaptability, and pedagogical rationality. The system sorts all candidate resources in descending order according to the final interaction score after difficulty calibration, selects the top K resources (K is set to 20 in this embodiment) to form the initial recommendation list, and outputs it along with the corresponding preference scores.

[0137] In this embodiment, constructing the semantic relationship graph between resources includes:

[0138] Obtain the multimodal resource feature vectors corresponding to each teaching resource, and perform unified normalization processing on the multimodal resource feature vectors as the semantic representation of the resources;

[0139] A graph structure is constructed with each teaching resource as a node, and each teaching resource is associated with and mapped to the corresponding multimodal resource feature vector to form a node feature set;

[0140] The semantic similarity between any two teaching resources is calculated based on their multimodal resource feature vectors, and the semantic similarity is obtained by cosine similarity or Euclidean distance transformation.

[0141] The calculated semantic similarity is compared with a preset similarity threshold. When the semantic similarity is greater than the preset threshold, an edge connection is established between the two corresponding teaching resource nodes.

[0142] The semantic similarity is used as the weight value of the established edge, and each edge in the graph is weighted according to the weight value to form a weighted resource semantic association graph;

[0143] The resource semantic association graph is subjected to sparsification and structural optimization, and edge connections with weight values ​​higher than a preset threshold are retained to generate a resource semantic association graph.

[0144] In this embodiment, the step of updating the preference scores of each resource to generate recommendation scores includes:

[0145] A semantic association graph of resources is constructed with each teaching resource as a node. The semantic similarity between the multimodal resource feature vectors of resources is used as the edge. When the similarity is higher than a preset threshold, an edge connection is established between the corresponding nodes, and the semantic similarity is used as the edge weight.

[0146] A diffusion recommendation model is constructed, with resources in the initial recommendation list as seed nodes. Preference propagation is performed on the resource semantic association graph, and the initial preference scores corresponding to the seed nodes and the corresponding resource feature information are used to form a multi-dimensional preference representation, which is then propagated to adjacent nodes along the edges.

[0147] During the propagation process, the multidimensional preference representations received by each node are associated and mapped with the multimodal resource feature vectors corresponding to the current node, and the multidimensional preference representations are updated to generate the propagation preference representations corresponding to the nodes.

[0148] The propagation preference representations received by each node from different propagation paths are accumulated, and the contributions of different propagation paths are weighted and fused according to the correlation strength of the propagation paths to generate a comprehensive preference representation for each resource.

[0149] The comprehensive preference representation is mapped to a scalar score and fused with the initial preference scores of each resource to obtain the final recommendation score for each resource.

[0150] All candidate resources are sorted according to the final recommendation score to obtain the recommendation results for output.

[0151] Example 1: In a continuously running online course platform, the system receives a batch of teaching resource data and student learning behavior data for the same knowledge unit. The total number of resources is 1860: 620 video resources, 430 text handouts, 310 image resources, and 500 exercise and diagram resources. Correspondingly, there are 4280 student account records with a cumulative historical interaction record of 312,640. The platform's original recommendation method primarily relies on collaborative filtering based on click and viewing history, exhibiting significant data sparsity and cold start problems. For example, some newly uploaded resources only average 7 exposures in the first 3 days after going live, with a click-through rate of only 4.8%. Although some students have high learning activity, the recommended content remains focused on basic resources they have already learned, failing to reflect students' knowledge gaps and skill enhancement needs.

[0152] In this embodiment, the system first preprocesses the raw data. Student interaction data includes click counts, viewing duration, test accuracy, homework completion scores, and interaction feedback behavior types. After cleaning, 2891 abnormal interaction records are removed, 1734 records with viewing durations exceeding 1.2 times the total resource duration are removed, and 521 abnormal records with 10 consecutive test results of 0 or 100% are removed. 309749 valid interaction records are retained after cleaning. Text resources, after word segmentation, stop word removal, and part-of-speech tagging, yield an average of 1420 valid terms per lecture note, which are then mapped to 300-dimensional word vectors. For video resources, after keyframe extraction, an average of 19 frames are extracted from each video, with 1024-dimensional frame features and 10-dimensional motion features, which are then concatenated to form 1034-dimensional video low-level features. Image resources are normalized and then have 1024-dimensional image low-level features extracted. All low-level features are timestamped according to resource identifiers to form a unified multimodal input sample.

[0153] To verify the feasibility of this invention, the platform extracted a complete training sample from the aforementioned data to simulate the entire process. The student account studied 12 resources within a knowledge unit, including 5 video resources, 3 lecture notes, 2 diagrams, and 2 exercises. In the early stages, the student's average completion rate for basic lecture notes was 0.91, with a test accuracy rate of 0.88. For medium-difficulty video resources, the average completion rate was 0.74, with a test accuracy rate of 0.63. Repeated regressions and concentrated incorrect answers were observed in the exercises related to knowledge points K17, K23, and K31. The system formed a row vector in the student-resource interaction matrix, connecting the student and the interacted resources. Click counts, viewing duration, accuracy rate, completion rate, and behavioral feedback features were extracted from the 12 interacted resource records. After normalization and weighted fusion, the interaction intensity value was obtained. Non-interactive resource positions were initialized to 0. Using the multimodal resource feature vector of the current resource to be predicted R256 as the query vector, the scaling dot product attention is calculated on the multimodal resource feature vectors of resources that students have interacted with in the past. The attention weights for resources R103, R118, R141, and R188 are 0.21, 0.18, 0.24, and 0.16, respectively, and the sum of the weights for the remaining resources is 0.21. After weighted summation, a 256-dimensional personalized demand vector for students is formed.

[0154] The system then initiates a knowledge tracking algorithm based on time decay. The student's knowledge graph contains 1200 knowledge points, with the current knowledge unit associated with 48 knowledge points. The system updates the student's mastery status based on recent viewing time, homework completion rate, and test accuracy. For example, regarding knowledge point K23, the student's most recent effective learning occurred a considerable time after the current recommendation, resulting in an initial mastery of 0.71, which decreased to 0.58 after time decay adjustment. For knowledge point K31, due to two recent incorrect answers accompanied by backtracking, the mastery was adjusted from 0.49 to 0.42. For knowledge point K12, due to repeated practice and stable accuracy, the mastery increased from 0.83 to 0.87. This ultimately forms a 1200-dimensional vector of student knowledge point mastery status, with 17 weak knowledge points below 0.6 and 21 stable knowledge points above 0.8.

[0155] When generating the initial recommendation list, the system inputs the student's personalized demand vector, multimodal resource feature vector, and student's knowledge point mastery state vector into an improved neural collaborative filtering network. For candidate resource R256, the student's personalized demand vector and resource feature vector are concatenated to form a 512-dimensional joint vector, which, after three fully connected layers, yields a basic interaction score of 0.67. The cognitive dynamic bias layer receives the student's 1200-dimensional knowledge point mastery state vector and the 256-dimensional resource feature vector of resource R256, forming a 1456-dimensional joint vector. After gating, a 64-dimensional gating vector with a mean of 0.61 is obtained, corresponding to a dynamic bias term with a mean of 0.09. After superposition, the cognitive enhancement interaction score is increased to 0.76. The difficulty adaptive calibration layer further calculates the difficulty coefficient of resource R256 to be 0.64, and the matching score between the student's ability vector and the difficulty is 0.73. After mapping, a calibration coefficient of 0.69 is obtained, resulting in a final interaction score of 0.524. After calculating all candidate resources, the system selects the top 20 resources to form an initial recommendation list. The preference scores of the top 5 resources are 0.811, 0.784, 0.752, 0.731 and 0.524, respectively.

[0156] After the initial recommendation list is formed, the system further constructs a resource semantic association graph. All 1860 resources are used as nodes, and semantic similarity is calculated based on multimodal resource feature vectors. Edges with similarity higher than 0.72 are retained, ultimately generating a weighted graph with 21874 edges. Using 20 resources from the initial recommendation list as seed nodes, their preference scores are combined with corresponding resource feature information to form a multidimensional preference representation, which is then propagated within the graph. Taking resource R256 as an example, its semantic similarities with resources R271, R309, and R412 are 0.83, 0.79, and 0.76, respectively. After one round of propagation, R271 receives the highest weighted propagated preference representation. After further multi-path accumulation, resource R309, which was not initially included in the initial recommendation list but is highly relevant to weak knowledge points, receives a propagation gain. Its comprehensive preference representation-mapped recommendation score increases from 0.418 to 0.663, and its ranking rises from 38th to 11th. Another newly uploaded illustrated resource, R412, had only received 2 clicks previously and would have been unable to enter the top 50 using traditional methods. However, in the method of this invention, due to its strong semantic association with multiple highly preferred resources, its recommendation score increased from 0.267 to 0.614, placing it 16th in the final recommendation list.

[0157] On the same batch of data, the method of this invention was compared with traditional collaborative filtering methods and traditional neural collaborative filtering methods. The training samples used interaction data from 3400 student accounts, the validation samples used 420 student accounts, and the test samples used 460 student accounts, maintaining a consistent candidate resource size. Test results show that the traditional collaborative filtering method has a HitRate@20 of 0.612, an NDCG@20 of 0.487, a weak interaction resource recall rate of 0.214, and a cold-start resource entry rate of 6.3% in the top 20. The traditional neural collaborative filtering method has a HitRate@20 of 0.684, an NDCG@20 of 0.553, a weak interaction resource recall rate of 0.301, and a cold-start resource entry rate of 9.8% in the top 20. The method of this invention achieves a HitRate@20 of 0.791, an NDCG@20 of 0.671, a weak interaction resource recall rate of 0.462, and a cold-start resource entry rate of 18.7% in the top 20. Regarding personalized adaptation, based on the hit rate of weak knowledge points among students in the test set, the traditional collaborative filtering method achieved 41.5%, the traditional neural collaborative filtering method achieved 53.8%, and the method of this invention achieved 71.2%. In terms of the diversity of recommendation results, based on the average semantic coverage of the first 20 recommended resources, the traditional collaborative filtering method achieved 0.46, the traditional neural collaborative filtering method achieved 0.52, and the method of this invention achieved 0.68. Simulated tracking results of improved learning outcomes showed that the average accuracy of students in the test group who received the recommendations of this invention increased from 0.64 to 0.76 in the subsequent round of practice, while the traditional neural collaborative filtering recommendation group only increased from 0.65 to 0.70.

[0158] As can be seen from this embodiment, in a realistic online teaching recommendation scenario, the present invention can complete the entire recommendation process from multimodal feature modeling, student learning state perception, initial preference scoring to multidimensional preference propagation optimization. It not only effectively solves the problems of insufficient utilization of single modality, lack of inclusion of cognitive state in recommendation, and difficulty in exposing cold-start resources in traditional methods, but also has good effects in terms of recommendation accuracy, personalization, knowledge adaptability, and result diversity, verifying the engineering feasibility and practical effectiveness of the method of the present invention in automatic recommendation of teaching resources.

[0159] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for automatically recommending teaching resources based on deep learning, characterized in that, include: Acquire multimodal educational data, preprocess the multimodal educational data, and extract the low-level features of each modality; A pre-defined deep neural network is used to extract high-level features from low-level features, generating high-dimensional feature vectors for each modality. Cross-modal fusion is performed on the high-dimensional feature vectors of each modality to generate a unified multimodal resource feature vector; Based on students' historical interaction data in multimodal education data, a student-resource interaction matrix is ​​constructed. The feature vectors of multimodal resources that students have interacted with in the past are weighted and aggregated using a neural attention mechanism to generate personalized demand vectors for students. Input students’ historical interaction data into a knowledge tracking algorithm based on time decay to generate a vector of students’ knowledge point mastery status. Based on the student's personalized demand vector, multimodal resource feature vector, and student's knowledge mastery state vector, an improved neural collaborative filtering network is used to predict the interaction probability between the student and each candidate resource, and generate an initial recommendation list and corresponding preference scores. A semantic association graph between resources is constructed. The resources in the initial recommendation list are used as seed nodes. The preference scores corresponding to each resource in the initial recommendation list and the multimodal resource feature vectors are used to form a multidimensional preference representation. Based on the multidimensional preference representation, a diffusion recommendation model is used to propagate preferences on the association graph. During the propagation process, the multidimensional preference representation is updated by combining the multimodal resource feature vectors corresponding to the nodes to generate the recommendation scores corresponding to each resource. The resources are reordered based on the updated recommendation scores, and the final recommendation list is output.

2. The method for automatically recommending teaching resources based on deep learning according to claim 1, characterized in that, The acquisition of multimodal education data includes: Student historical interaction data, which includes the number of times students clicked on each teaching resource, the viewing time, the accuracy rate of their test answers, the homework completion score, and the type of interactive feedback behavior; Text data of teaching resources, including detailed descriptions of course content, learning objectives, syllabus, and course tags; The video data of the teaching resources includes the total duration of the teaching video, video frame data, and video subtitle text. Acquire image data of teaching resources, including course PPT pages, diagrams, and screenshots of whiteboard writing.

3. The method for automatically recommending teaching resources based on deep learning according to claim 1, characterized in that, The extraction of low-level features for each modality includes: The multimodal educational data is preprocessed, including: cleaning, removing outliers, and standardizing student historical interaction data; segmenting text data, removing stop words, and tagging parts of speech; extracting frames from video data and extracting the timestamp of each video frame; and normalizing image data by size and converting color space. The preprocessed data is subjected to low-level feature extraction, including: converting text units into corresponding word vectors using natural language processing techniques; extracting temporal features from video frame sequences and converting them into time series features; and using convolutional neural networks to extract image features and generate image feature vectors. The processed student historical interaction data is timestamped and integrated with the underlying feature vectors of each modality to form a multimodal feature representation, which serves as the input data for high-level feature extraction.

4. The method for automatically recommending teaching resources based on deep learning according to claim 1, characterized in that, The generation of high-dimensional feature vectors for each modality includes: For text modalities, a pre-trained language model is used to extract high-level semantic features; A 3D convolutional network is used to extract spatiotemporal features for video modalities; Visual semantic features are extracted from image modalities using a visual transformer. The high-dimensional feature vectors of each modality are uniformly mapped to the same dimensional space.

5. The method for automatically recommending teaching resources based on deep learning according to claim 1, characterized in that, The construction of the student-resource interaction matrix includes: The student historical interaction data is aggregated according to student identifiers and resource identifiers to form a two-dimensional interaction index structure with students as rows and teaching resources as columns; Based on the interaction records between each student and each teaching resource, we extract interaction features such as the number of clicks, viewing time, test answer accuracy, homework completion score, and interaction feedback behavior type. We then normalize these interaction features and map them to a unified numerical space. The various interaction features are weighted and fused according to preset weights to generate the corresponding interaction intensity value between the student and the teaching resources, and the interaction intensity value is filled into the corresponding position in the two-dimensional interaction index structure; For student and teaching resource combinations with no interaction records, the interaction intensity value at the corresponding position is initialized to zero or set as a missing marker, and implicit completion is performed through model training; The student-resource interaction matrix is ​​subjected to sparse storage and index optimization processing to generate a standardized student-resource interaction matrix.

6. The method for automatically recommending teaching resources based on deep learning according to claim 1, characterized in that, The generated student personalized needs vector includes: The multimodal resource features of the current resource to be predicted are represented as a query vector. The multimodal features of each resource that the student has interacted with in the past are represented as keys and values ​​using a neural attention mechanism. The relevance weights between the historically interacted resources and the current resource are calculated by scaling dot product attention. The features of the historically interacted resources are weighted and summed to generate a personalized demand vector for the student.

7. The method for automatically recommending teaching resources based on deep learning according to claim 1, characterized in that, The generated student knowledge point mastery state vector includes: Obtain students' historical interaction data, including video viewing time, homework completion rate, test accuracy, and interaction feedback; The mastery status of each knowledge point is calculated based on students' historical interaction data, and the mastery status is adjusted using the time decay factor and the Ebbinghaus forgetting curve decay formula. The knowledge tracking algorithm combines students' multimodal learning data and the time interval of historical interactions to update the mastery status of each knowledge point. The interaction data of video, image and text modalities jointly affect the knowledge point update according to the feature weights. The decayed knowledge point mastery status is integrated into a student knowledge point mastery status vector.

8. The method for automatically recommending teaching resources based on deep learning according to claim 1, characterized in that, The generation of the initial recommendation list and corresponding preference scores includes: The vectors of students' personalized needs, multimodal resource features, and students' knowledge mastery status are aligned and standardized in terms of dimensions. An improved neural collaborative filtering network was constructed, including a basic neural collaborative filtering layer, a cognitive dynamic bias layer, a difficulty adaptive calibration layer, and a fusion output layer; The student's personalized needs vector and multimodal resource feature vector are input into the basic neural collaborative filtering layer. The basic interaction score is obtained through vector concatenation and multilayer perceptron mapping. The student's knowledge mastery state vector and the multimodal resource feature vector of the candidate resources are input into the cognitive dynamic bias layer to generate dynamic bias terms. The dynamic bias terms are then superimposed on the basic interaction score to obtain the cognitively enhanced interaction score. The cognitively enhanced interaction score is input into the difficulty adaptive calibration layer. The resource difficulty extraction unit extracts the difficulty coefficient of the candidate resource, and the student ability extraction unit extracts the student ability vector. The difficulty matching score is generated based on the matching relationship between the student ability vector and the resource difficulty coefficient. The cognitively enhanced interaction score is then calibrated based on the difficulty matching score to obtain the final interaction score of the candidate resource. The final interaction score is input into the fusion output layer, and the interaction probability between the student and each candidate resource is generated through normalization mapping. The interaction probability is then used as the preference score corresponding to the candidate resource. All candidate resources are sorted in descending order according to preference scores, and a preset number of candidate resources at the top of the sort are selected to form an initial recommendation list. Output the initial recommendation list and the preference score corresponding to each candidate resource.

9. The method for automatically recommending teaching resources based on deep learning according to claim 1, characterized in that, The semantic relationship graph between the constructed resources includes: Obtain the multimodal resource feature vectors corresponding to each teaching resource, and perform unified normalization processing on the multimodal resource feature vectors as the semantic representation of the resources; A graph structure is constructed with each teaching resource as a node, and each teaching resource is associated and mapped with its corresponding multimodal resource feature vector to form a node feature set; The semantic similarity between teaching resources is calculated based on the multimodal resource feature vectors, and the semantic similarity is obtained by cosine similarity or Euclidean distance transformation. The calculated semantic similarity is compared with a preset similarity threshold. When the semantic similarity is greater than the preset threshold, an edge connection is established between the two corresponding teaching resource nodes. The semantic similarity is used as the weight value of the established edge, and each edge in the graph is weighted according to the weight value to form a weighted resource semantic association graph; The resource semantic association graph is subjected to sparsification and structural optimization, and edge connections with weight values ​​higher than a preset threshold are retained to generate a resource semantic association graph.

10. The method for automatically recommending teaching resources based on deep learning according to claim 1, characterized in that, The process of updating the preference scores for each resource to generate recommendation scores includes: A semantic association graph of resources is constructed with each teaching resource as a node. The semantic similarity between the multimodal resource feature vectors of resources is used as the edge. When the similarity is higher than a preset threshold, an edge connection is established between the corresponding nodes, and the semantic similarity is used as the edge weight. A diffusion recommendation model is constructed, with resources in the initial recommendation list as seed nodes. Preference propagation is performed on the resource semantic association graph, and the initial preference scores corresponding to the seed nodes and the corresponding resource feature information are used to form a multi-dimensional preference representation, which is then propagated to adjacent nodes along the edges. During the propagation process, the multidimensional preference representations received by each node are associated and mapped with the multimodal resource feature vectors corresponding to the current node, and the multidimensional preference representations are updated to generate the propagation preference representations corresponding to the nodes. The propagation preference representations received by each node from different propagation paths are accumulated, and the contributions of different propagation paths are weighted and fused according to the correlation strength of the propagation paths to generate a comprehensive preference representation for each resource. The comprehensive preference representation is mapped to a scalar score and fused with the initial preference scores of each resource to obtain the final recommendation score for each resource. All candidate resources are sorted according to the final recommendation score to obtain the recommendation results for output.

Images