Video content on-demand method and system applied to school education

Abstract

The invention provides a video content on-demand method and system applied to school education, and relates to the technical field of school education streaming media. Fusing student personalized attributes and behavior features to generate student feature representations, and predicting applicability scores in combination with video tags; grouping the student feature data and calculating a group feature vector; fusing the personalized prediction score and the group feature vector to generate a recommendation list; and time sequence preference representation is constructed based on student watching history, and a recommendation strategy is dynamically optimized through an adaptive scoring and calibration mechanism. According to the invention, accurate personalized recommendation of the education video is realized, and the learning effect and the resource utilization rate are improved.

Description

Technical Field

[0001] The present invention relates to school education streaming media technology, and particularly to a method and system for video content on-demand in school education. Background Art

[0002] Traditional educational video on-demand systems usually adopt content-based recommendation methods or collaborative filtering techniques to recommend resources based on users' historical behavior data and content similarity. Some systems have also started to introduce machine learning algorithms to try to model students' learning preferences to improve the accuracy of recommendations. Teachers can specify learning resources for students through these systems, and students can also select appropriate video content according to their interests and needs. However, the existing educational video on-demand technologies have the following defects and deficiencies: The existing technologies have relatively simple understanding and label generation of video content, mostly relying on manual annotation or basic text analysis, lacking in-depth understanding and accurate description of the educational value of video content, resulting in a low matching degree between content and educational goals.

[0003] The existing personalized recommendation algorithms fail to fully consider the cognitive characteristics of students and the complexity of learning behaviors. They often adopt a single feature representation method, unable to effectively capture the competitive relationships and dynamic changes among students' features, and it is difficult to provide video content truly suitable for the individual needs of students. Summary of the Invention

[0004] Embodiments of the present invention provide a method and system for video content on-demand in school education, which can solve the problems in the existing technologies.

[0005] In the first aspect of the embodiments of the present invention, A method for video content on-demand in school education is provided, including: Establishing a video source library, performing content analysis using natural language processing technology, and generating video tags in combination with educational value; Processing students' personalized attributes and learning behavior characteristics through a main path and a competitive path network, calculating competitive signals, then performing dynamic feature suppression and fusion to generate student feature representations, and generating video applicability prediction scores for personalized recommendation in combination with the video tags; clustering the student feature data by decision trees and adaptively adjusting the splitting threshold, calculating group features in combination with decision path weights, introducing a time series attenuation mechanism for dynamic update and optimization, and generating group feature vectors for group-level recommendation; using feature fusion and multi-task learning methods to fuse the video applicability prediction scores and the group feature vectors to obtain a final video recommendation list; Dynamically optimize the generated video recommendation list, including: constructing a temporal preference representation based on the viewing history data of students, calculating an initial score based on an adaptive scoring network, generating a final score through a bias compensation and double-branch calibration mechanism, and dynamically adjusting the video recommendation strategy.

[0006] In an alternative embodiment, Use natural language processing technology for content analysis, and generate video tags in combination with educational value, including: Convert the audio stream information in the video into text information, and fuse it with the subtitle text information to obtain the text to be processed; Perform semantic analysis on the text to be processed, introduce a context attention mechanism based on knowledge progression to enhance semantic representation. The context attention mechanism is enhanced by calculating the weighted combination of the context vector of knowledge concepts and the degree of knowledge progression, where the degree of knowledge progression is quantified based on the prerequisite relationship and depth relationship between concepts, and the dynamic weight coefficient is adaptively adjusted according to the context relevance to obtain the enhanced semantic representation information; Construct an educational knowledge graph based on the enhanced semantic representation information, identify concepts in the education field through named entity recognition technology and semantic dependency analysis, and use relationship extraction technology to establish hierarchical and dependency relationships between knowledge points. Combine educational theory to label the difficulty level and applicable age range for knowledge points, and use an iterative optimization algorithm to dynamically update the concept system and relationship network to form a structured knowledge system; generate video tags according to the structured knowledge system, and the video tags include applicable age, teaching theme, and cultivation ability.

[0007] In an alternative embodiment, Process the student's personalized attributes and learning behavior characteristics through the main path and competitive path networks, calculate the competition signal, and then perform dynamic feature suppression and fusion to generate a student feature representation. Combine the video tags to generate a video applicability prediction score for personalized recommendation, including: Decompose the student's basic information into personalized attribute feature vectors and learning behavior feature vectors; Input the personalized attribute feature vector into the main path network to obtain a first feature representation, and input the learning behavior feature vector into the competitive path network to obtain a second feature representation; Calculate the competition signal based on the ratio of the difference norm to the sum norm of the first feature representation and the second feature representation; perform dynamic feature suppression on the first feature representation and the second feature representation according to the competition signal to obtain a dominant feature representation and a suppressed feature representation; Calculate the feature fusion weight based on the dominant feature representation and the suppressed feature representation, and adaptively fuse the dominant feature representation and the suppressed feature representation according to the feature fusion weight to obtain a student feature representation; The student feature representation and the video tag are respectively converted into feature vectors of the same dimension through a mapping layer, and the similarity between the two feature vectors is calculated to obtain a matching score; the matching score is input into the output layer of the neural network, and the video applicability prediction score is generated through a fully connected layer and an activation function.

[0008] In an alternative implementation, The student feature data is clustered by a decision tree and the splitting threshold is adaptively adjusted. The group features are calculated by combining the decision path weights, and a time series decay mechanism is introduced for dynamic update and optimization, generating a group feature vector for group-level recommendation, including: The student feature data is normalized to obtain normalized data, and the information gain of each feature node is calculated; a density clustering method is used to search for the optimal splitting point of the normalized data, the splitting threshold is determined based on the local density estimation, and the splitting threshold is adaptively adjusted according to the sample distribution skewness to obtain an optimized feature splitting rule; The decision path features from the root node to the leaf node are extracted, and the weights of the decision path features are calculated by the weighted sum of the node importance and the node information gain, and the group center value and the degree of dispersion are calculated based on the decision path features; the group center value and the degree of dispersion are combined to generate a group representation vector, and the group representation vector includes a feature center value component and a feature dispersion degree component; A time series decay factor is introduced into the group representation vector, and the time series decay factor is calculated through an exponential decay function. Based on the optimized feature splitting rule and the group representation vector, the time series weighted distance between the sample and the group center is calculated. When the time series weighted distance is greater than the first preset threshold, the weights of the decision path features are incrementally updated within a sliding time window; based on the updated group representation vector, a group feature distribution statistic is constructed, and when the time series change rate of the statistic exceeds the drift determination threshold, group feature reconstruction is triggered to obtain an optimized group feature vector.

[0009] In an alternative implementation, Feature fusion and multi-task learning methods are used to fuse the video applicability prediction score and the group feature vector to obtain a final video recommendation list, including: Map the video applicability score and the group feature vector into matching feature vectors of the same dimension respectively; use the multi-head self-attention mechanism to calculate the interaction weight matrix of the matching feature vectors, generate a fusion coefficient based on the interaction weight matrix, and perform weighted fusion on the matching feature vectors according to the fusion coefficient to obtain a fusion feature; input the fusion feature into a momentum encoder to generate a query feature, calculate a comprehensive score based on the path similarity and difficulty gradient of the curriculum knowledge graph to select positive samples, select the sample with the highest similarity to the query feature from the unviewed videos as negative samples, and perform contrastive learning through the InfoNCE loss function; Construct a Gaussian mixture model based on the feature representation obtained by contrastive learning to characterize the student interest distribution, calculate the upper confidence bound by combining Thompson sampling and the upper confidence bound algorithm, and adaptively adjust the exploration rate through an annealing strategy; input the student interest distribution and the upper confidence bound into a shared Transformer encoder containing task-specific attention heads, construct a task graph network based on the task gradient similarity, and use an uncertainty weighting mechanism to adaptively allocate task weights for multi-task optimization to generate a video recommendation list.

[0010] In an optional implementation manner, Inputting the fusion feature into a momentum encoder to generate a query feature, calculating a comprehensive score based on the path similarity and difficulty gradient of the curriculum knowledge graph to select positive samples, and selecting the sample with the highest similarity to the query feature from the unviewed videos as negative samples, and performing contrastive learning through the InfoNCE loss function includes: Input the fusion feature into a momentum encoder. The momentum encoder includes a query encoder and a key encoder. Maintain the parameters of the key encoder through a momentum update mechanism. The momentum update mechanism includes: performing an exponential moving average on the parameters of the query encoder with a preset momentum coefficient to obtain the parameters of the key encoder; generating a query feature based on the query encoder, and generating a key feature based on the key encoder; constructing a feature representation dictionary based on the key feature, and calculating the similarity between the query feature and the samples in the feature representation dictionary to obtain a similarity distribution; Calculate the path similarity between video nodes in the curriculum knowledge graph. The path similarity is obtained by calculating the cosine similarity of the node representations after message passing of the nodes by a graph neural network; extract the difficulty level information of the video, and calculate the difficulty gradient based on the difference in difficulty levels; input the path similarity, the difficulty gradient, and the similarity distribution into a multi-layer perceptron to obtain the comprehensive score of the sample; Select the sample with the highest score as the positive sample based on the comprehensive score, and select the negative sample from the samples with the lowest score based on the hard negative sample mining strategy; assign corresponding weights to the positive and negative samples according to the comprehensive score; Input the query feature, positive sample feature, negative sample feature and their corresponding weights into the InfoNCE loss function for contrastive learning. The InfoNCE loss function adjusts the feature distribution through a temperature parameter and calculates the joint optimization objective of maximizing the mutual information of positive sample pairs and minimizing the mutual information of negative sample pairs.

[0011] In an alternative embodiment, Construct a temporal preference representation based on the student's viewing history data, calculate an initial score based on an adaptive scoring network, and generate a final score through a bias compensation and double-branch calibration mechanism. Dynamically adjusting the video recommendation strategy includes: Extract temporal features from the student's viewing history data, construct a temporal encoder based on the attention mechanism. The temporal encoder performs long-short-term memory modeling on the video sequence in the viewing history to obtain a temporal context representation; perform cross-modal interaction on the temporal context representation and the video content feature, and control the information flow through an adaptive gating unit to generate a video preference representation; Calculate the historical score preference distribution based on the video preference representation, and construct an adaptive scoring network. The adaptive scoring network includes a feature interaction layer and a score prediction layer. The feature interaction layer uses the attention mechanism to calculate the interaction intensity between the student feature and the video feature, and the score prediction layer generates an initial score based on the interaction feature; Perform learnable bias compensation on the initial score, including: calculating the mean bias and variance bias of the historical score, constructing a compensation vector based on the mean bias and variance bias, and adaptively fusing the compensation vector with the initial score to obtain the compensated score; Construct a double-branch calibration network. The first branch calculates the score bias based on the historical score distribution, and the second branch calculates the viewing satisfaction based on the viewing completion degree. Input the score bias and the viewing satisfaction into the self-attention layer to obtain the calibration weight, and adaptively adjust the compensated score according to the calibration weight to generate the final video score; Dynamically adjust the recommendation strategy based on the final video score, including updating the sorting weight of the recommendation sequence and adjusting the display probability of the recommended video.

[0012] In the second aspect of the embodiments of the present invention, Provide a video content on-demand system applied to school education, including: The first unit is used to establish a video source library, perform content analysis using natural language processing technology, and generate video tags in combination with educational value; A second unit, configured to process the personalized attributes and learning behavior characteristics of students through the main path and the competitive path network, calculate competitive signals, perform dynamic feature suppression and fusion, generate student feature representations, and generate video applicability prediction scores for personalized recommendation in combination with the video tags; cluster student feature data through a decision tree and adaptively adjust the splitting threshold, calculate group features in combination with decision path weights, introduce a time series decay mechanism for dynamic update and optimization, and generate group feature vectors for group-level recommendation; use feature fusion and multi-task learning methods to fuse video applicability prediction scores and group feature vectors to obtain a final video recommendation list; A third unit, configured to dynamically optimize the generated video recommendation list, including: constructing a time series preference representation based on the viewing history data of students, calculating an initial score based on an adaptive scoring network, generating a final score through a bias compensation and dual-branch calibration mechanism, and dynamically adjusting the video recommendation strategy.

[0013] In a third aspect of the embodiments of the present invention, there is provided an electronic device, including: a processor; a memory for storing instructions executable by the processor; wherein, the processor is configured to call the instructions stored in the memory to execute the foregoing method.

[0014] In a fourth aspect of the embodiments of the present invention, there is provided a computer-readable storage medium, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the foregoing method is implemented.

[0015] By establishing a video source library and using natural language processing technology to perform content analysis, and generating video tags in combination with educational values, the present invention can accurately grasp the educational characteristics of video content, laying a foundation for subsequent personalized recommendation. At the same time, by deeply analyzing the personalized attributes and learning behavior characteristics of students, accurate student feature representations are generated, which helps to improve the pertinence and applicability of video recommendations.

[0016] The present invention adopts the method of clustering by a decision tree and adaptively adjusting the splitting threshold, calculates group features in combination with decision path weights, and introduces a time series decay mechanism for dynamic update and optimization, which can effectively capture the common features of student groups and improve the accuracy of group-level recommendation. Through feature fusion and multi-task learning methods, personalized recommendation and group-level recommendation are organically combined, further improving the overall performance of the recommendation system.

[0017] The present invention also introduces a dynamic optimization mechanism, constructs a temporal preference representation based on the viewing history data of students, and generates a final score through an adaptive scoring network, deviation compensation, and a dual-branch calibration mechanism, realizing the dynamic adjustment of the recommendation strategy. This method can respond in a timely manner to changes in students' interests and needs, continuously optimize the recommendation results, and improve students' learning experience and effects. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a schematic flowchart of a video content on-demand method according to an embodiment of the present invention applied to school education; Figure 2 It is a schematic diagram of learning duration distribution and splitting threshold; Figure 3 It is a performance comparison diagram between the present invention and the benchmark decision tree method; Figure 4 It is a cosine similarity distribution diagram of positive and negative samples of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only some of the embodiments of the present invention, rather than all of them. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0020] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0021] Figure 1 It is a schematic flowchart of a video content on-demand method according to an embodiment of the present invention applied to school education, as Figure 1 shown, the method includes: Establish a video source library, perform content analysis using natural language processing technology, and generate video tags in combination with educational value; Process the student's personalized attributes and learning behavior characteristics through the main path and competitive path networks, calculate the competitive signals, perform dynamic feature suppression and fusion, generate a student feature representation, and generate a video applicability prediction score for personalized recommendation in combination with the video tags; cluster the student feature data through a decision tree and adaptively adjust the splitting threshold, calculate the group features in combination with the decision path weights, introduce a temporal decay mechanism for dynamic update and optimization, and generate a group feature vector for group-level recommendation; use feature fusion and multi-task learning methods to fuse the video applicability prediction score and the group feature vector to obtain a final video recommendation list; Dynamically optimize the generated video recommendation list, including: constructing a temporal preference representation based on the viewing history data of students, calculating an initial score based on an adaptive scoring network, generating a final score through a deviation compensation and double-branch calibration mechanism, and dynamically adjusting the video recommendation strategy.

[0022] In an optional implementation manner, use natural language processing technology for content analysis, and generate video tags in combination with educational value, including: Convert the audio stream information in the video into text information, and fuse it with the subtitle text information to obtain the text to be processed; Perform semantic analysis on the text to be processed, introduce a context attention mechanism based on knowledge progression to enhance the semantic representation. The context attention mechanism is enhanced by calculating the weighted combination of the context vector of the knowledge concept and the degree of knowledge progression, where the degree of knowledge progression is quantified based on the prerequisite relationship and depth relationship between concepts, and the dynamic weight coefficient is adaptively adjusted according to the context relevance to obtain the enhanced semantic representation information; Construct an educational knowledge graph based on the enhanced semantic representation information, identify concepts in the education field through named entity recognition technology and semantic dependency analysis, and use relation extraction technology to establish hierarchical relationships and dependency relationships between knowledge points. Combine educational theory to label the difficulty level and applicable age range for knowledge points, and use an iterative optimization algorithm to dynamically update the concept system and relationship network to form a structured knowledge system; generate video tags according to the structured knowledge system, and the video tags include applicable age, teaching theme, and cultivation ability.

[0023] Exemplarily, the step of converting the audio stream information in the video into text information can be achieved through speech recognition technology, such as using an end-to-end speech recognition model based on deep learning. Specifically, a speech recognition model based on the Transformer architecture can be adopted. This model consists of two main parts: an encoder and a decoder. The encoder is responsible for mapping the audio feature sequence into a hidden representation, and the decoder generates the corresponding text sequence based on the hidden representation. During the training phase, a large-scale speech-text pair dataset can be used for model training, such as the LibriSpeech dataset. During the inference phase, the audio stream of the video is segmented into several audio clips, which are input into the trained model one by one for recognition, and finally the complete audio-to-text result is obtained. The audio-to-text result obtained in the previous step is fused with the original subtitle text information of the video to obtain the text to be processed. The possible time alignment problem between the two text sources needs to be considered during the fusion process. The dynamic time warping algorithm can be used to achieve the best alignment of the two text sequences. Specifically, first, the two text sequences are respectively segmented into sentence-level segments, then the similarity score between each pair of sentence segments is calculated (such as using cosine similarity), and finally, the dynamic programming algorithm is used to find the alignment scheme with the highest overall similarity. Based on the alignment, the text in the overlapping part can be fused in a weighted average manner, and the weights can be determined according to their respective confidence levels.

[0024] Semantic analysis is performed on the fused text to be processed. This step introduces a context attention mechanism based on knowledge progression to enhance semantic representation. Specifically, when implementing, a domain knowledge base needs to be constructed first, which contains the core concepts in the education field and the relationships between them. The initial knowledge base can be constructed manually based on existing educational resources (such as textbooks, curriculum outlines, etc.), and then continuously expanded and optimized through machine learning methods.

[0025] When performing semantic analysis, first use a pre-trained language model (such as BERT) to encode the input text to obtain an initial semantic representation. Then, for each educational concept that appears in the text, retrieve the relevant context concepts from the knowledge base. Next, calculate the degree of knowledge progression between the current concept and these context concepts. The calculation of the degree of knowledge progression takes into account the prerequisite relationship and depth relationship between concepts. The prerequisite relationship can be determined through the predefined concept dependency graph in the knowledge base, and the depth relationship can be quantified by the hierarchical difference of concepts in the knowledge system. For example, assume the current concept is "quadratic function", and its relevant context concepts include "linear function", "function graph", and "derivation". According to the information in the knowledge base, "linear function" is a prerequisite concept for "quadratic function", and the depth relationship between them is relatively close; "function graph" and "quadratic function" are in a parallel relationship with equivalent depth; while "derivation" is a more advanced concept with a relatively distant depth relationship from "quadratic function". Based on this, different progression degree weights can be assigned to these concepts, such as 0.8 for "linear function", 0.5 for "function graph", and 0.2 for "derivation".

[0026] Calculate the semantic relevance between the current concept and each context concept. This can be achieved by comparing their vector representations in the language model, such as using cosine similarity. Then, multiply the semantic relevance by the progression degree weight calculated earlier to obtain the final dynamic weight coefficient. Finally, use these weight coefficients to perform a weighted sum of the semantic representations of the context concepts and fuse them with the original representation of the current concept to obtain an enhanced semantic representation.

[0027] After obtaining the enhanced semantic representation, the next step is to construct an educational knowledge graph. First, use named entity recognition technology to identify educational domain concepts in the text. A sequence labeling model based on BiLSTM-CRF can be adopted and trained using a large-scale labeled dataset. Then, identify the relationships between concepts through semantic dependency analysis technology. Here, a semantic dependency analysis model based on graph neural networks can be used, which can effectively capture long-distance dependencies. Adopt relation extraction technology to establish hierarchical and dependency relationships between knowledge points. The method of distant supervision can be used, and the existing knowledge base can be used as weak supervision signals to train the relation extraction model. Specifically, a relation classification model based on BERT can be adopted, taking each pair of concepts and their contexts as inputs to predict the type of relationship between them. To further improve the knowledge graph, it is necessary to annotate the difficulty level and applicable age range of knowledge points in combination with educational theory. This can be achieved by designing a multi-task learning model that simultaneously predicts the difficulty level and applicable age range of knowledge points. The inputs of the model include the semantic representation of the knowledge point itself, its structural features in the knowledge graph (such as in-degree, out-degree, etc.), and statistical features extracted from educational resources (such as frequency of occurrence, usage in exams, etc.). An iterative optimization algorithm is used to dynamically update the concept system and relationship network. Specifically, a model based on graph attention network can be designed, which can continuously adjust the concept relationships in the knowledge graph according to new text data and user feedback. In each iteration, the model calculates the importance scores of each node and edge in the graph and fine-tunes the graph structure accordingly.

[0028] Based on the structured knowledge system constructed in the above steps, video tags including applicable age, teaching theme, and cultivation ability can be generated. The specific generation process is as follows: Match relevant knowledge points in the knowledge graph according to the video content. A text-graph matching model based on the attention mechanism can be used, which can find the corresponding relationships between video text and concept nodes in the knowledge graph. Based on the matched knowledge points, use a graph traversal algorithm to find the most representative upper-level concept in the knowledge graph as the teaching theme. At the same time, according to the positions and connection relationships of these knowledge points in the graph, infer the abilities that the video content may cultivate. Considering the difficulty level and applicable age range information of the matched knowledge points, and combining the age stage division criteria in educational theory, determine the applicable age range of the video.

[0029] The multi-dimensional concept association of the present invention is carried out through a text-graph matching model with an attention mechanism, ensuring the semantic accuracy and knowledge coverage of label generation; based on the graph traversal algorithm and the hierarchical organization of knowledge points, representative teaching topics can be automatically extracted, avoiding the problems of overly scattered or repetitive labels; by combining the difficulty level of knowledge points and the applicable age information, and analyzing the connection relationships in the graph structure, the core capabilities cultivated by the video content can be accurately identified; adopting a dynamic update mechanism enables the label system to be continuously optimized with the addition of new resources and user feedback, improving the timeliness and adaptability of the labels; through a structured label generation scheme, richer feature representations are provided for subsequent personalized recommendations.

[0030] In an alternative embodiment, the student's personalized attributes and learning behavior characteristics are processed through the main path and competitive path networks, and after calculating the competitive signal, dynamic feature suppression and fusion are performed to generate a student feature representation. Combining the video labels to generate a video applicability prediction score for personalized recommendation includes: Decompose the student's basic information into a personalized attribute feature vector and a learning behavior feature vector; Input the personalized attribute feature vector into the main path network to obtain a first feature representation, and input the learning behavior feature vector into the competitive path network to obtain a second feature representation; Calculate the competitive signal based on the ratio of the difference norm to the sum norm of the first feature representation and the second feature representation; perform dynamic feature suppression on the first feature representation and the second feature representation according to the competitive signal to obtain a dominant feature representation and a suppressed feature representation; Calculate the feature fusion weight based on the dominant feature representation and the suppressed feature representation, and adaptively fuse the dominant feature representation and the suppressed feature representation according to the feature fusion weight to obtain a student feature representation; Convert the student feature representation and the video label into feature vectors of the same dimension through a mapping layer respectively, and calculate the similarity of the two feature vectors to obtain a matching score; input the matching score into the output layer of the neural network, and generate a video applicability prediction score through a fully connected layer and an activation function.

[0031] Exemplarily, the student's basic information is decomposed into a personalized attribute feature vector and a learning behavior feature vector. The personalized attribute feature vector includes static information such as the student's age, gender, and educational background, with a dimension of 50. The learning behavior feature vector includes dynamic information such as the student's learning duration, homework completion situation, and test scores, with a dimension of 100. The personalized attribute feature vector is input into the main path network, and the learning behavior feature vector is input into the competitive path network. The main path network consists of 3 fully connected layers, and the number of neurons in the hidden layers is 256, 128, and 64 respectively. The competitive path network consists of 2 convolutional layers and 1 fully connected layer. The convolutional kernel size is 3×3, the stride is 1, the padding is same, and the number of output channels is 32 and 64 respectively. The output dimension of the fully connected layer is 64. The outputs of the two networks respectively obtain a first feature representation and a second feature representation, both with a dimension of 64.

[0032] Calculate the competition signal: Calculate the difference vector between the first feature representation and the second feature representation, and calculate the L2 norm of the difference vector to obtain the difference norm. Then calculate the sum vector of the first feature representation and the second feature representation, and calculate the L2 norm of the sum vector to obtain the sum norm. The competition signal is equal to the difference norm divided by the sum norm. For example, if the first feature representation is [0.1, 0.2, 0.3,...] and the second feature representation is [0.2, 0.3, 0.4,...], the calculated competition signal is 0.15.

[0033] Perform dynamic feature suppression on the first feature representation and the second feature representation according to the competition signal. The dynamic feature suppression includes: setting the suppression threshold to 0.2. When the competition signal is less than the threshold, the first feature representation is kept unchanged as the dominant feature representation, and the second feature representation is multiplied by the attenuation coefficient (1 - competition signal) to obtain the suppressed feature representation; when the competition signal is greater than or equal to the threshold, the second feature representation is kept unchanged as the dominant feature representation, and the first feature representation is multiplied by the attenuation coefficient (1 - competition signal) to obtain the suppressed feature representation. For example, when the competition signal is 0.15, the first feature representation is kept unchanged as the dominant feature representation, and each element of the second feature representation is multiplied by 0.85 to obtain the suppressed feature representation.

[0034] Use the softmax function to calculate the fusion weights, and the input is the L2 norm of the dominant feature representation and the suppressed feature representation. For example, if the L2 norm of the dominant feature representation is 1.2 and the L2 norm of the suppressed feature representation is 0.8, the fusion weights are 0.62 and 0.38 respectively. Weighted sum the dominant feature representation and the suppressed feature representation according to the fusion weights to obtain a 64-dimensional student feature representation.

[0035] The student feature representation and video tags are respectively converted into feature vectors of the same dimension through a mapping layer. The mapping layer is a single-layer fully connected network with an output dimension of 128. The video tags are represented using multi-hot encoding with a dimension of 1000. For example, if the tags of a certain video are "mathematics, high school, function", then the corresponding positions in the multi-hot encoding vector are 1, and the rest are 0. Calculate the cosine similarity between the student feature vector and the video feature vector to obtain a matching score. The value range of the matching score is [-1, 1], and the larger the value, the more matching. Input the matching score into the output layer of the neural network, pass through a fully connected layer with 64 hidden units and a ReLU activation function, and finally obtain a video applicability prediction score between 0 and 1 through a Sigmoid activation function. The closer the prediction score is to 1, the more suitable the video is for recommending to the student.

[0036] By introducing a competition mechanism, this method can adaptively adjust the importance of personalized attributes and learning behavior characteristics, thereby generating a more accurate student feature representation. At the same time, by calculating the similarity between the student features and the video features, the applicability of the video to the student can be effectively predicted. This method can be applied to the personalized video recommendation system of an online education platform to improve the accuracy of recommendations and learning effects.

[0037] In specific implementation, the model can be built using the Python language and the TensorFlow framework. The main path network and the competition path network can be implemented using the Dense layer and the Conv1D layer. Feature suppression and fusion can be achieved through custom layers. The final matching degree calculation and prediction score generation can be implemented using the Dense layer and activation functions. When training the model, binary cross-entropy can be used as the loss function, and the Adam optimizer can be used to update the parameters. The training data includes student information, video information, and the actual viewing situation of students for videos. The leave-one-out method can be used to divide the training set and the test set, and the performance of the model can be evaluated on the test set. This method can be further optimized. For example, an attention mechanism can be introduced to assign different importance to different dimensions of the student features. The method of transfer learning can be adopted to use a pre-trained language model to extract the semantic features of video text descriptions. Traditional recommendation algorithms such as collaborative filtering can also be combined to integrate multiple recommendation results to further improve the recommendation effect.

[0038] In an alternative implementation, the student feature data is grouped by a decision tree and the splitting threshold is adaptively adjusted. The group features are calculated by combining the decision path weights, and a time-series decay mechanism is introduced for dynamic update and optimization. The group feature vector generated for group-level recommendation includes: Standardize the student feature data to obtain standardized data, and calculate the information gain of each feature node; use the density clustering method to search for the optimal splitting point of the standardized data, determine the splitting threshold based on local density estimation, and adaptively adjust the splitting threshold according to the sample distribution skewness to obtain an optimized feature splitting rule; Extract the decision path features from the root node to the leaf node. The weight of the decision path features is calculated by the weighted sum of the node importance and the node information gain, and the group center value and the degree of dispersion are calculated based on the decision path features; combine the group center value and the degree of dispersion to generate a group representation vector, and the group representation vector includes a feature center value component and a feature dispersion degree component; Introduce a time series decay factor to the group representation vector. The time series decay factor is calculated by an exponential decay function. Calculate the time series weighted distance between the sample and the group center based on the optimized feature splitting rule and the group representation vector. When the time series weighted distance is greater than the first preset threshold, incrementally update the weight of the decision path features within the sliding time window; construct a group feature distribution statistic based on the updated group representation vector, and trigger group feature reconstruction when the time series change rate of the statistic exceeds the drift determination threshold to obtain an optimized group feature vector.

[0039] Exemplarily, the student feature data is standardized. Taking the data of 10,000 students on an online education platform as an example, the features include learning duration, number of exercises completed, interaction frequency, knowledge point mastery, etc. For each feature, the min-max normalization method is adopted to map the values to the interval [0, 1]. For example, the original value range of the learning duration is 0 - 300 minutes, and after normalization, it is mapped to between 0 and 1. The normalized value of a student with a learning duration of 150 minutes is 0.5. Taking the knowledge point mastery as an example, by calculating the discrimination ability of this feature for the students' learning effects, the information gain value is 0.42, indicating that this feature has a high discrimination value. The density clustering method is used to search for the optimal split points for the standardized data. Taking the learning duration feature as an example, the density curve is generated through the kernel density estimation method, and it is found that there are density valley points at 0.3 and 0.7. These two points can be used as the initial split thresholds to divide the students into three categories: "low learning duration" (0 - 0.3), "medium learning duration" (0.3 - 0.7), and "high learning duration" (0.7 - 1). When determining the split threshold based on the local density estimation, an interval with a window width of 0.05 is taken near each candidate split point, and the sample density within the interval is calculated. Taking the learning duration feature as an example, the density value near 0.3 is 0.15, and the density value near 0.7 is 0.12, both of which are lower than the average density of 0.2 and are confirmed as valid split points. The split threshold is adaptively adjusted according to the sample distribution skewness. When it is found that the learning duration feature distribution is right-skewed (skewness coefficient is 1.2), the original split points 0.3 are adjusted to 0.25, and 0.7 is adjusted to 0.65 to balance the number of samples in each group. After adjustment, the sample ratio of the low, medium, and high learning duration groups is adjusted from the original 20%:50%:30% to approximately 33%:34%:33%, achieving a more balanced distribution.

[0040] Extract the decision path features from the root node to the leaf node. Suppose a student is classified into the "high learning effect" leaf node by a decision tree, and its decision path is: learning duration (>0.65) → interaction frequency (>0.5) → knowledge point mastery (>0.8). This path contains three judgment conditions, which constitute the decision path features of this student. The weights of the decision path features are calculated by the weighted sum of the node importance and the node information gain. Taking the above path as an example, the information gains of the three nodes are 0.38, 0.25, and 0.42 respectively, the node depths are 1, 2, and 3 respectively, and the node importance decreases with the depth, which are 1.0, 0.8, and 0.6 respectively. Through comprehensive calculation, the weights of the three features are 0.38, 0.2, and 0.252 respectively. Calculate the group center value and the degree of dispersion based on the decision path features. For the "high learning effect" group, which contains 2,000 students, the center value of the learning duration feature is 0.82 (indicating that most students have a long learning duration), and the degree of dispersion is 0.12 (indicating that the group is relatively concentrated on this feature); the center value of the interaction frequency feature is 0.75, and the degree of dispersion is 0.18; the center value of the knowledge point mastery feature is 0.88, and the degree of dispersion is 0.09. Combine the group center value and the degree of dispersion to generate a group representation vector. The representation vector of the "high learning effect" group is: [0.82, 0.12, 0.75, 0.18, 0.88, 0.09]. The first two components represent the center value and the degree of dispersion of the learning duration, the middle two components represent the center value and the degree of dispersion of the interaction frequency, and the last two components represent the center value and the degree of dispersion of the knowledge point mastery.

[0041] Introduce a time-series decay factor to the group representation vector. Suppose the current time is t, and a certain historical data was generated 10 days ago at t - 10, and the decay half-life is set to 7 days, then the time-series decay factor of this data is 0.5 (10 / 7) ≈0.39, indicating that the weight of this historical data in the current group feature calculation is reduced to 39% of the original.

[0042] Calculate the time-series weighted distance between the sample and the group center based on the optimized feature splitting rule and the group representation vector. For example, the feature vector of the new student A is [0.79, 0.68, 0.85], and the weighted distance from the "high learning effect" group center is 0.15, which is less than the preset threshold of 0.25, so it is determined to be a member of this group; while the feature vector of student B is [0.45, 0.72, 0.62], and the weighted distance from the group center is 0.38, which is greater than the preset threshold and does not belong to this group. When the time-series weighted distance is greater than the first preset threshold (such as 0.25), incrementally update the weights of the decision path features within the sliding time window. For example, within the recent 7 days, it is found that the discrimination ability of the interaction frequency feature has increased, and its information gain has increased from 0.25 to 0.32. Correspondingly, increase its weight in the decision path from 0.2 to 0.256.

[0043] Construct population feature distribution statistics based on the updated population representation vector. For the "high learning effect" population, calculate the mean, variance, and their change trends of each feature in the past 30 days. For example, the 30-day mean of the learning duration feature is 0.82, the variance is 0.014, the 7-day mean is 0.79, and the variance is 0.016. The calculated mean change rate is -3.7%. When the time series change rate of the statistic exceeds the drift determination threshold (such as 5%), population feature reconstruction is triggered. For example, it is found that the mean change rate of the interaction frequency feature is -6.8%, exceeding the threshold, triggering the reconstruction process. The feature splitting and population representation vector generation are re-executed to obtain the updated population feature vector [0.80, 0.13, 0.70, 0.20, 0.87, 0.10], which reflects the latest change trend of the population features.

[0044] Figure 2 It is a schematic diagram of the learning duration distribution and splitting threshold. This figure shows the data distribution of student features (taking learning duration as an example) in the [0, 1] interval after min-max normalization. Through the simulated bimodal distribution, it can be seen that there are two valleys with lower densities in the normalized data, corresponding to the optimal splitting thresholds (0.25 and 0.65). The determination of these splitting thresholds helps to achieve more balanced data grouping, providing accurate data support for decision tree clustering and subsequent population feature construction. This solution can effectively balance the number of samples in each group, avoid classification errors caused by data skew, and thus improve the overall stability and prediction accuracy of the model.

[0045] Figure 3 It is a performance comparison chart between the present invention and the benchmark decision tree method. This radar chart compares the performance of the improved method proposed in this paper and the benchmark decision tree method in four key indicators: population division balance, decision path weight optimization, dynamic update response of time series data, and final classification accuracy. The figure shows that the technology in this paper is superior to the benchmark decision tree method in each evaluation index, especially in the time series update response and decision path weight calculation. This solution can achieve more accurate and dynamic population feature extraction and update, thus significantly improving the accuracy and response speed of the recommendation system, and further optimizing the user experience and system efficiency in application scenarios such as online education platforms.

[0046] Through the above implementation manners, this method can effectively capture the feature distribution and its dynamic changes of the student population, providing accurate population-level feature representations for the educational recommendation system.

[0047] In an alternative implementation manner, use feature fusion and multi-task learning methods to fuse the video applicability prediction score and the population feature vector, and the resulting final video recommendation list includes: Map the video applicability score and the group feature vector into matching feature vectors of the same dimension respectively; use the multi-head self-attention mechanism to calculate the interaction weight matrix of the matching feature vectors, generate the fusion coefficient based on the interaction weight matrix, and perform weighted fusion on the matching feature vectors according to the fusion coefficient to obtain the fused feature; input the fused feature into the momentum encoder to generate the query feature, calculate the comprehensive score based on the path similarity and difficulty gradient of the curriculum knowledge graph to select the positive samples, select the sample with the highest similarity to the query feature from the unviewed videos as the negative samples, and perform contrastive learning through the InfoNCE loss function; Construct a Gaussian mixture model based on the feature representation obtained by contrastive learning to characterize the student interest distribution, calculate the upper confidence bound by combining Thompson sampling and the upper confidence bound algorithm, and adaptively adjust the exploration rate through the annealing strategy; input the student interest distribution and the upper confidence bound into the shared Transformer encoder containing task-specific attention heads, construct a task graph network based on the task gradient similarity, and use the uncertainty weighting mechanism to adaptively allocate task weights for multi-task optimization to generate a video recommendation list.

[0048] Exemplarily, the specific implementation of mapping the video applicability score and the group feature vector into matching feature vectors of the same dimension is to expand the video applicability score (scalar) into a 64-dimensional vector through two independent fully-connected neural network layers, and at the same time convert the group feature vector into a 64-dimensional vector. For example, the original applicability score of 0.85 is mapped into a 64-dimensional vector F1, and the original group feature vector [0.8, 0.6, 0.4, 0.7] is mapped into a 64-dimensional vector F2.

[0049] The system uses the multi-head self-attention mechanism to calculate the interaction weight matrix of the matching feature vectors. Specifically, 8 attention heads are used, and the dimension of each head is 8. The attention weights are generated through three linear transformations of query, key, and value. For example, the first attention head may focus on the matching degree between the video content and the user's knowledge level, and the second head may focus on the matching degree between the video difficulty and the user's learning progress. The dimension of the finally generated interaction weight matrix W is 64×64.

[0050] Based on the interaction weight matrix, the system generates the fusion coefficient α, and the value range is from 0 to 1. For example, the calculated fusion coefficient α = 0.7 indicates that in the final fused feature, the video applicability feature accounts for 70% of the weight, and the group feature accounts for 30% of the weight. The system performs weighted fusion on the matching feature vectors according to the fusion coefficient: F_fusion = α×F1 + (1-α)×F2, to obtain a 64-dimensional fused feature vector.

[0051] Generate the query feature based on the fused feature vector, select the positive and negative samples and perform contrastive learning.

[0052] Based on the feature representation obtained from contrastive learning, the system constructs a Gaussian mixture model to characterize the student interest distribution. Specifically, 3 Gaussian components are used, and the model parameters are estimated by the expectation maximization algorithm. For example, for a certain student, their interest distribution may consist of three Gaussian components, corresponding to science experiment (weight 0.5), history and culture (weight 0.3), and art creation (weight 0.2) type videos respectively.

[0053] The system combines Thompson sampling and the upper confidence bound algorithm to calculate the upper confidence bound. For each candidate video, calculate the upper bound of its expected reward: UCB = expected reward + c × uncertainty, where c is the exploration parameter, and the initial value is set to 1.0. The system adaptively adjusts the exploration rate through an annealing strategy. As the number of user interactions increases, the value of c linearly decreases from 1.0 to 0.1, achieving a smooth transition from exploration to exploitation.

[0054] The student interest distribution and the upper confidence bound are input into a shared Transformer encoder that includes task-specific attention heads. This encoder has 3 layers, with 4 attention heads in each layer, and the hidden layer dimension is 256. Different tasks (such as click prediction, completion rate prediction, satisfaction prediction) use independent attention heads, but share the same feed-forward network parameters. Based on the task gradient similarity, the system constructs a task graph network, calculates the cosine similarity of different task gradient vectors, and tasks with a similarity greater than 0.6 are connected. For example, the gradient similarity between click prediction and completion rate prediction is 0.75, so these two tasks are connected in the graph. The system adopts an uncertainty-weighted mechanism to adaptively allocate task weights for multi-task optimization. The weight of each task is inversely proportional to its uncertainty, and the uncertainty is estimated by the variance of the task loss. For example, the weights of click prediction, completion rate prediction, and satisfaction prediction may be 0.4, 0.35, and 0.25 respectively. Finally, the system generates a personalized video recommendation list according to the comprehensive score of multi-task learning. The recommendation list is sorted in descending order of the comprehensive score and contains information such as video ID, title, preview image, and estimated suitability. For example, the top three videos recommended for a certain student may be V005 (score 0.92), V008 (score 0.85).

[0055] The present invention realizes feature fusion through the multi-head self-attention mechanism, improves the quality of feature representation by combining the momentum encoder and contrastive learning, adopts the Gaussian mixture model and Thompson sampling to balance exploration and exploitation, and through a multi-task learning framework with task-specific attention heads and uncertainty weighting, realizes the unity of accuracy, diversity, and interpretability of personalized recommendation, effectively improving the accuracy and learning effect of educational video recommendation.

[0056] In an alternative embodiment, the fused feature is input into a momentum encoder to generate a query feature. A comprehensive score is calculated based on the path similarity and difficulty gradient of the curriculum knowledge graph to select positive samples, and the sample with the highest similarity to the query feature is selected as the negative sample from the unviewed videos. The contrastive learning through the InfoNCE loss function includes: The fused feature is input into a momentum encoder, which includes a query encoder and a key encoder. The parameters of the key encoder are maintained through a momentum update mechanism, which includes: performing an exponential moving average on the parameters of the query encoder with a preset momentum coefficient to obtain the parameters of the key encoder; generating a query feature based on the query encoder and a key feature based on the key encoder; constructing a feature representation dictionary based on the key feature, and calculating the similarity between the query feature and the samples in the feature representation dictionary to obtain a similarity distribution; Calculate the path similarity between video nodes in the curriculum knowledge graph, which is obtained by calculating the cosine similarity of node representations after message passing of nodes by a graph neural network; extract the difficulty level information of the video, and calculate the difficulty gradient based on the difference in difficulty levels; input the path similarity, the difficulty gradient, and the similarity distribution into a multi-layer perceptron to obtain the comprehensive score of the sample; Select the sample with the highest score as the positive sample based on the comprehensive score, and select the negative sample from the samples with the lowest score based on the hard negative sample mining strategy; assign corresponding weights to the positive and negative samples according to the comprehensive score; Input the query feature, the positive sample feature, the negative sample feature, and their corresponding weights into the InfoNCE loss function for contrastive learning. The InfoNCE loss function adjusts the feature distribution through a temperature parameter, and calculates the joint optimization objective of maximizing the mutual information of positive sample pairs and minimizing the mutual information of negative sample pairs.

[0057] Exemplarily, the fused feature is input into a momentum encoder to generate a query feature. The momentum encoder consists of two parts: a query encoder and a key encoder. The query encoder is composed of three fully connected layers, with the number of neurons in each layer being 512, 256, and 128 respectively. The activation function uses ReLU and is used to map the fused feature to the query feature space. The structure of the key encoder is the same as that of the query encoder, but the parameters are maintained through a momentum update mechanism.

[0058] The momentum update mechanism updates the key encoder parameters using exponential moving average. The specific implementation is as follows: key encoder parameters = momentum coefficient × old key encoder parameters + (1 - momentum coefficient) × current query encoder parameters. In practical applications, the momentum coefficient is set to 0.999 and updated once after each batch of training. For example, when an element in the weight matrix of the first layer of the query encoder is updated from 0.5 to 0.52, the corresponding element of the key encoder will be updated from 0.5 to 0.5 × 0.999 + 0.52 × 0.001 = 0.50002, maintaining the smoothness of parameter changes.

[0059] The query features are obtained by processing the fused features through the query encoder, with a dimension of 128; the key features are obtained by processing historical samples through the key encoder, also with a dimension of 128. The feature representation dictionary constructed by the key features contains the feature representations of the last 10,000 samples. Calculate the cosine similarity between the query features and all samples in the dictionary to obtain the similarity distribution. For example, the cosine similarity calculation result between the query feature [0.1, 0.2,..., 0.3] and the feature of a certain sample [0.2, 0.1,..., 0.25] in the dictionary is 0.85.

[0060] Calculate the path similarity between video nodes in the curriculum knowledge graph. The curriculum knowledge graph consists of video nodes and knowledge point nodes, and the edges represent the association relationships between them. The graph attention network (GAT) is used for node message passing, and each video node is initially represented as a 64-dimensional vector. The message passing process contains two layers of GAT, with 8 attention heads in each layer and an output dimension of 64. After two rounds of message passing, calculate the cosine similarity between node representations to obtain the path similarity. For example, the node representations of video A and video B are [0.1, 0.3,..., 0.2] and [0.15, 0.28,..., 0.18] respectively, and the calculated path similarity is 0.92.

[0061] The video difficulty level is divided into five levels from 1 to 5, representing from beginner to expert level. The difficulty gradient is calculated as the normalized value of the difference between the current video difficulty and the target video difficulty. For example, if the current video difficulty is 2 and the target video difficulty is 3, the difficulty gradient is (3 - 2) / 5 = 0.2, indicating a moderate difficulty increase.

[0062] Input the path similarity, difficulty gradient, and similarity distribution into a multi-layer perceptron (MLP) to calculate the comprehensive score. This MLP contains two layers, with 64 neurons in the hidden layer and 1 neuron in the output layer, and the activation function is Sigmoid. The input features are [path similarity, difficulty gradient, similarity], such as [0.92, 0.2, 0.85], and the output comprehensive score after being processed by the MLP is 0.88.

[0063] Samples are selected based on the comprehensive score. Specifically, one positive sample is randomly selected from the top 5% of samples with the highest score. A hard negative mining strategy is used to select the sample with the highest cosine similarity to the query feature from the bottom 10% of samples as the negative sample. This strategy ensures that negative samples are both semantically irrelevant to the query and relatively close in feature space, increasing the model's discrimination difficulty and improving learning effectiveness.

[0064] Weights are assigned to positive and negative samples based on their comprehensive scores. The weight of a positive sample is set to its comprehensive score, and the weight of a negative sample is set to 1 minus its comprehensive score. For example, if the comprehensive score of a positive sample is 0.88, its weight is 0.88; if the comprehensive score of a negative sample is 0.15, its weight is 0.85.

[0065] Finally, the query features, positive sample features, negative sample features, and their corresponding weights are input into the InfoNCE loss function for comparative learning. The temperature parameter of the InfoNCE loss function is set to 0.07 to adjust the smoothness of the feature distribution. This loss function calculates the negative logarithm of the ratio of the similarity score between the query feature and the positive sample features to the sum of the similarity scores between the query feature and all negative sample features, multiplied by the corresponding sample weight. By minimizing this loss function, the model learns to bring semantically similar videos closer together in feature space and push dissimilar videos further apart.

[0066] Existing contrastive learning methods mainly use random data augmentation or simple heuristic rules to construct positive and negative sample pairs, which fails to fully utilize domain knowledge and learning rules in educational scenarios. Figure 4 This is a distribution diagram of the cosine similarity of positive and negative samples from the present invention. As shown, the cosine similarity of positive samples (represented by black solid circles) is generally concentrated between 0.8 and 1.0, with an average cosine similarity of approximately 0.90 and a maximum value of approximately 0.98. This indicates a high degree of feature match between the positive samples and the query sample. In contrast, the cosine similarity of negative samples (represented by hollow circles) ranges primarily between 0.1 and 0.6, with an average of approximately 0.38. This data fluctuates significantly and is significantly lower than that of positive samples, reflecting a significant difference between the query and negative samples. These data effectively demonstrate the significant difference in feature matching between positive and negative samples. This application proposes a comparative learning method based on a course knowledge graph and difficulty perception. This method incorporates the course knowledge graph to calculate path similarity between video nodes and captures the associations between knowledge points through a graph neural network. It also considers a learning difficulty gradient to ensure a reasonable progression in the difficulty of recommended content. Finally, a comprehensive scoring mechanism is designed to guide the selection and weighting of positive and negative samples, incorporating the experience of educational experts into the model training process. The present invention shows obvious advantages in processing video recommendations across chapters and difficulty levels.

[0067] In an alternative implementation, a temporal preference representation is constructed based on the viewing history data of students. The initial score is calculated based on an adaptive scoring network, and the final score is generated through a bias compensation and a dual-branch calibration mechanism. The dynamic adjustment of the video recommendation strategy includes: Extract temporal features from the viewing history data of students, and construct a temporal encoder based on the attention mechanism. The temporal encoder performs long short-term memory modeling on the video sequence in the viewing history to obtain a temporal context representation; perform cross-modal interaction on the temporal context representation and the video content features, and control the information flow through an adaptive gating unit to generate a video preference representation; Calculate the historical score preference distribution based on the video preference representation, and construct an adaptive scoring network. The adaptive scoring network includes a feature interaction layer and a score prediction layer. The feature interaction layer calculates the interaction intensity between the student features and the video features using the attention mechanism, and the score prediction layer generates the initial score based on the interaction features; Perform learnable bias compensation on the initial score, including: calculating the mean bias and variance bias of the historical scores, constructing a compensation vector based on the mean bias and variance bias, and adaptively fusing the compensation vector with the initial score to obtain the compensated score; Construct a dual-branch calibration network. The first branch calculates the score bias based on the historical score distribution, and the second branch calculates the viewing satisfaction based on the viewing completion degree. Input the score bias and the viewing satisfaction into the self-attention layer to obtain the calibration weight, and adaptively adjust the compensated score according to the calibration weight to generate the final video score; Dynamically adjust the recommendation strategy based on the final video score, including updating the sorting weight of the recommendation sequence and adjusting the display probability of the recommended videos.

[0068] Exemplarily, the video sequence viewed by the student in the past 30 days is used as the input, and each video is represented by a 128-dimensional feature vector, which includes information such as video category, duration, and difficulty. The temporal encoder adopts a long short-term memory network structure, sets the number of hidden layer units to 256, and captures the temporal dependence relationship in the video sequence through a three-layer stacked structure. During the encoding process, the attention mechanism is introduced to assign weights to the videos at different time points, and the attention weights are normalized by the softmax function, and the weight values range from 0 to 1. For example, for a sequence of 10 historical viewing videos, the recently viewed "Advanced Mathematics" video may obtain a high attention weight of 0.25, while the earlier viewed "English Listening" video may obtain a low weight of 0.05. Through weighted aggregation, a 512-dimensional temporal context representation vector is generated.

[0069] Perform cross-modal interaction between the temporal context representation and video content features. The video content features include 256-dimensional visual features (extracted from video key frames) and 128-dimensional text features (extracted from video titles and descriptions). Design an adaptive gating unit to control the information flow. The gating unit is implemented by a two-layer neural network with 384 nodes in the input layer, 192 nodes in the hidden layer, and 128 nodes in the output layer. The gating value ranges from 0 to 1, and the larger the value, the higher the importance of the corresponding feature. For example, for videos of the "physical experiment" category, the gating value of the visual feature may be 0.8, while the gating value of the text feature is 0.6. Through the gating mechanism, fuse different modal information to finally generate a 384-dimensional video preference representation vector.

[0070] Calculate the historical rating preference distribution based on the video preference representation. First, count the student's historical rating data and convert the ratings from 1 to 5 into a probability distribution. For example, the rating distribution of a student for science videos may be: 1 point (0.05), 2 points (0.1), 3 points (0.2), 4 points (0.4), 5 points (0.25). Then construct an adaptive scoring network, including a feature interaction layer and a rating prediction layer. The feature interaction layer uses the multi-head attention mechanism with the number of heads set to 8 and the dimension of each attention head being 64 to calculate the interaction intensity between the student features and the video features. The dimension of the interaction intensity matrix is the student feature dimension (384) × the video feature dimension (384), and each element value in the matrix represents the correlation between the corresponding feature dimensions, with the value range from -1 to 1. The rating prediction layer consists of three fully connected networks with the number of nodes in each layer being 512, 256, and 128 respectively. Finally, map it to the range of 1 to 5 points through the sigmoid activation function to generate the initial rating.

[0071] Calculate the mean deviation and variance deviation of the historical ratings. For example, the average rating of a student for all videos is 3.8 points, and the standard deviation is 0.7 points. Based on these statistics, construct a 64-dimensional compensation vector, where the first 32 dimensions represent the mean deviation information and the last 32 dimensions represent the variance deviation information. The compensation vector is processed by a LeakyReLU activation function with a parameter of 0.3 to ensure that the compensation value is within a reasonable range. Adaptively fuse the compensation vector and the initial rating through a residual connection method, and the fusion weight is dynamically adjusted by a gating mechanism with a parameter of 0.2 to obtain the compensated rating.

[0072] Construct a dual-branch calibration network to further optimize the scoring. The first branch calculates the scoring deviation based on the historical scoring distribution. By comparing the difference between the current score and the historical average score, a 32-dimensional scoring deviation vector is generated. For example, if the average score of a student for "programming tutorial" videos in the past was 4.2 points, and the current predicted score is 3.5 points, a negative deviation is generated. The second branch calculates the viewing satisfaction based on the viewing completion rate. The average viewing completion rate of the student for this type of video (such as 85%) is converted into a 32-dimensional satisfaction vector. The output vectors of the two branches are input into a self-attention layer. The self-attention layer contains 4 attention heads, each with a dimension of 16, to generate a 64-dimensional calibration weight vector. The compensated score is adaptively adjusted according to the calibration weight, and the adjustment range is limited within ±15% of the original score to generate the final video score.

[0073] Dynamically adjust the recommendation strategy based on the final video score. First, update the sorting weights of the recommendation sequence. For every 0.5-point increase in the score, the sorting weight increases by 0.1. For example, the sorting weight of a video with a score of 4.5 points is 0.9, and the sorting weight of a video with a score of 3.5 points is 0.7. Secondly, adjust the display probability of the recommended videos. The softmax function is used to convert the score into a display probability, and high-score videos have a higher chance of being displayed. For example, among 5 candidate videos with scores [3.2, 3.8, 4.5, 3.5, 4.0], the converted display probabilities are approximately [0.12, 0.18, 0.36, 0.15, 0.19]. Finally, the system updates the student's preference model every 24 hours to ensure that the recommendation results can timely reflect the changes in the student's interests.

[0074] The present invention can accurately grasp the changes in the students' learning interests and knowledge mastery levels, provide video content that better suits the individual learning rhythms and ability levels of the students. At the same time, through the intelligent scoring calibration and dynamic adjustment mechanism, it effectively avoids the monotony and repetition of the recommended content, significantly improves the learning efficiency and learning experience, and realizes the precision, personalization, and high efficiency of educational resource recommendation.

[0075] The present invention also provides: Build a video source library, and add at least one video applicability label to each video in the video source library; obtain multiple matching videos from the video source library according to one or more of the video applicability labels matched with the basic information of the current student, and obtain a video recommendation list provided for the student to watch; adjust the video recommendation list according to the viewing history data and feedback data of the current student. The basic information may include any one or more of physical development, eyesight, academic performance, labor performance, physical exercise time, moral discipline performance, and extracurricular interest preferences; the video applicability labels include any one or more of age, theme, and ability. Calculate the comprehensive evaluation scores of multiple characteristics of each student based on all the characteristics of each student and the weights of each characteristic, and obtain the videos in the video source library whose applicable labels match the comprehensive evaluation scores of multiple characteristics of each student; predict all the basic information of each student group through a neural network to obtain one or more of the video applicability labels of each student group. Use the basic information as the input and the video applicability labels as the output, and train a video recommendation model through a neural network; add multiple hidden layers to the video recommendation model; use the cross-validation method to optimize the parameters of the video recommendation model; input the comprehensive evaluation scores of all the characteristics of the student into the hidden layer of the video recommendation model for training; use the collaborative filtering algorithm to refine the video applicability labels output by the video recommendation model; display the corresponding multiple videos to the student according to the output video applicability labels. Analyze the preference degree of the student for each video according to the viewing duration, viewing frequency, and viewing completion degree of the student.

[0076] In the second aspect of the embodiments of the present invention, Provide a video content on-demand system applied to school education, including: The first unit is used to build a video source library, perform content analysis by using natural language processing technology, and generate video labels in combination with educational values; The second unit is used to process the personalized attributes and learning behavior characteristics of the student through the main path and competitive path networks, calculate the competitive signal and then perform dynamic feature suppression and fusion to generate a student feature representation, and generate a video applicability prediction score for personalized recommendation in combination with the video label; cluster the student feature data through a decision tree and adaptively adjust the splitting threshold, calculate the group feature in combination with the decision path weight, introduce a time series decay mechanism for dynamic update and optimization, and generate a group feature vector for group-level recommendation; use the feature fusion and multi-task learning methods to fuse the video applicability prediction score and the group feature vector to obtain a final video recommendation list; The third unit is used for dynamically optimizing the generated video recommendation list, including: constructing a temporal preference representation based on the viewing history data of students, calculating an initial score based on an adaptive scoring network, generating a final score through a deviation compensation and double-branch calibration mechanism, and dynamically adjusting the video recommendation strategy.

[0077] In the third aspect of the embodiments of the present invention, A kind of electronic device is provided, including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0078] In the fourth aspect of the embodiments of the present invention, A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0079] The present invention can be a method, a device, a system and / or a computer program product. The computer program product can include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0080] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it is still possible to modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A video content on-demand method applied to school education, characterized in that, Including: Establish a video source library, perform content analysis using natural language processing technology, and generate video tags in combination with educational value; Process the personalized attributes and learning behavior characteristics of students through the main path and competitive path networks, calculate the competitive signal, perform dynamic feature suppression and fusion, generate student feature representations, and generate video applicability prediction scores for personalized recommendation in combination with the video tags; Cluster the student feature data through decision trees and adaptively adjust the splitting threshold, calculate the group features in combination with the decision path weights, introduce a time series decay mechanism for dynamic update and optimization, and generate group feature vectors for group-level recommendation; Use feature fusion and multi-task learning methods to fuse the video applicability prediction scores and group feature vectors to obtain the final video recommendation list; Dynamically optimize the generated video recommendation list, including: constructing a time series preference representation based on the viewing history data of students, calculating the initial score based on the adaptive scoring network, generating the final score through the bias compensation and double-branch calibration mechanism, and dynamically adjusting the video recommendation strategy.

2. The method according to claim 1, characterized in that Performing content analysis using natural language processing technology and generating video tags in combination with educational value includes: Converting the audio stream information in the video into text information and fusing it with the subtitle text information to obtain the text to be processed; Perform semantic analysis on the text to be processed, introduce a context attention mechanism based on knowledge progression to enhance semantic representation, and the context attention mechanism is enhanced by calculating the weighted combination of the context vector of the knowledge concept and the degree of knowledge progression, where the degree of knowledge progression is quantified based on the prerequisite relationship and depth relationship between concepts, and the dynamic weight coefficient is adaptively adjusted according to the context relevance to obtain the enhanced semantic representation information; Construct an educational knowledge graph based on the enhanced semantic representation information, identify concepts in the education field through named entity recognition technology and semantic dependency analysis, establish hierarchical relationships and dependency relationships between knowledge points using relation extraction technology, annotate the difficulty level and applicable age range for knowledge points in combination with educational theory, and use an iterative optimization algorithm to dynamically update the concept system and relationship network to form a structured knowledge system; Generate video tags according to the structured knowledge system, and the video tags include applicable age, teaching theme, and cultivation ability.

3. The method according to claim 1, characterized in that Processing the personalized attributes and learning behavior characteristics of students through the main path and competitive path networks, calculating the competitive signal, performing dynamic feature suppression and fusion, generating student feature representations, and generating video applicability prediction scores for personalized recommendation in combination with the video tags includes: Decompose the student basic information into personalized attribute feature vectors and learning behavior feature vectors; Input the personalized attribute feature vector into the main path network to obtain the first feature representation, and input the learning behavior feature vector into the competitive path network to obtain the second feature representation; Calculate the competitive signal based on the ratio of the difference norm to the sum norm of the first feature representation and the second feature representation; perform dynamic feature suppression on the first feature representation and the second feature representation according to the competitive signal to obtain the dominant feature representation and the suppressed feature representation; Calculate the feature fusion weights based on the dominant feature representation and the suppressed feature representation, and adaptively fuse the dominant feature representation and the suppressed feature representation according to the feature fusion weights to obtain the student feature representation; Convert the student feature representation and the video label into feature vectors of the same dimension through the mapping layer respectively, calculate the similarity of the two feature vectors to obtain the matching score; input the matching score into the output layer of the neural network, and generate the video applicability prediction score through the fully connected layer and the activation function.

4. The method according to claim 1, wherein Cluster the student feature data by the decision tree and adaptively adjust the splitting threshold, calculate the group features in combination with the decision path weights, introduce the time series decay mechanism for dynamic update and optimization, and generate the group feature vector for group-level recommendation, including: Perform normalization processing on the student feature data to obtain the normalized data, and calculate the information gain of each feature node; use the density clustering method to search for the optimal splitting point of the normalized data, determine the splitting threshold based on the local density estimation, and adaptively adjust the splitting threshold according to the sample distribution skewness to obtain the optimized feature splitting rule; Extract the decision path features from the root node to the leaf node, the weights of the decision path features are calculated by the weighted sum of the node importance and the node information gain, and calculate the group center value and the degree of dispersion based on the decision path features; combine the group center value and the degree of dispersion to generate the group representation vector, and the group representation vector includes the feature center value component and the feature dispersion component; Introduce the time series decay factor into the group representation vector, the time series decay factor is calculated by the exponential decay function, calculate the time series weighted distance between the sample and the group center based on the optimized feature splitting rule and the group representation vector, when the time series weighted distance is greater than the first preset threshold, incrementally update the weights of the decision path features within the sliding time window; construct the group feature distribution statistic based on the updated group representation vector, and trigger the group feature reconstruction when the time series change rate of the statistic exceeds the drift determination threshold to obtain the optimized group feature vector.

5. The method according to claim 1, wherein Use the feature fusion and multi-task learning methods to fuse the video applicability prediction score and the group feature vector to obtain the final video recommendation list, including: Map the video applicability score and the group feature vector into matching feature vectors of the same dimension respectively; use the multi-head self-attention mechanism to calculate the interaction weight matrix of the matching feature vectors, generate the fusion coefficient based on the interaction weight matrix, and perform weighted fusion on the matching feature vectors according to the fusion coefficient to obtain the fusion feature; input the fusion feature into the momentum encoder to generate the query feature, calculate the comprehensive score based on the path similarity and difficulty gradient of the curriculum knowledge graph to select the positive sample, select the sample with the highest similarity to the query feature from the unviewed videos as the negative sample, and perform contrastive learning through the InfoNCE loss function; Construct a Gaussian mixture model based on the feature representation obtained by contrastive learning to characterize the student interest distribution, combine Thompson sampling and upper confidence bound algorithm to calculate the upper confidence bound, and adaptively adjust the exploration rate through an annealing strategy; input the student interest distribution and the upper confidence bound into a shared Transformer encoder containing task-specific attention heads, construct a task graph network based on task gradient similarity, and adopt an uncertainty weighting mechanism to adaptively allocate task weights for multi-task optimization to generate a video recommendation list.

6. The method according to claim 5, characterized in that, Input the fused features into a momentum encoder to generate query features, calculate a comprehensive score based on the path similarity and difficulty gradient of the course knowledge graph to select positive samples, and select the sample with the highest similarity to the query features from the unwatched videos as negative samples. The contrastive learning through the InfoNCE loss function includes: Input the fused features into a momentum encoder, where the momentum encoder includes a query encoder and a key encoder. The key encoder parameters are maintained through a momentum update mechanism, and the momentum update mechanism includes: performing an exponential moving average on the query encoder parameters with a preset momentum coefficient to obtain the key encoder parameters; generating query features based on the query encoder and key features based on the key encoder; constructing a feature representation dictionary based on the key features, and calculating the similarity between the query features and the samples in the feature representation dictionary to obtain a similarity distribution; Calculate the path similarity between video nodes in the course knowledge graph, where the path similarity is obtained by calculating the cosine similarity of node representations after message passing of nodes through a graph neural network; extract the difficulty level information of the video, and calculate the difficulty gradient based on the difference in difficulty levels; input the path similarity, the difficulty gradient, and the similarity distribution into a multi-layer perceptron to obtain the comprehensive score of the sample; Select the sample with the highest score as the positive sample based on the comprehensive score, and select negative samples from the samples with the lowest score based on the hard negative sample mining strategy; assign corresponding weights to the positive and negative samples according to the comprehensive score; Input the query features, positive sample features, negative sample features, and their corresponding weights into the InfoNCE loss function for contrastive learning. The InfoNCE loss function adjusts the feature distribution through a temperature parameter and calculates the joint optimization objective of maximizing the mutual information of positive sample pairs and minimizing the mutual information of negative sample pairs.

7. The method according to claim 1, characterized in that Construct a temporal preference representation based on the student's viewing history data, calculate an initial score based on an adaptive scoring network, and generate a final score through a bias compensation and double-branch calibration mechanism. The dynamic adjustment of the video recommendation strategy includes: Extract temporal features from the student's viewing history data, construct a temporal encoder based on the attention mechanism, where the temporal encoder performs long-short-term memory modeling on the video sequence in the viewing history to obtain a temporal context representation; perform cross-modal interaction between the temporal context representation and the video content features, and control the information flow through an adaptive gating unit to generate a video preference representation; Calculate the historical rating preference distribution based on the video preference representation, and construct an adaptive scoring network. The adaptive scoring network includes a feature interaction layer and a rating prediction layer. The feature interaction layer calculates the interaction intensity between the student features and the video features using an attention mechanism, and the rating prediction layer generates an initial rating based on the interaction features; Perform learnable bias compensation on the initial rating, including: calculating the mean bias and variance bias of the historical ratings, constructing a compensation vector based on the mean bias and variance bias, and adaptively fusing the compensation vector with the initial rating to obtain the compensated rating; Construct a two-branch calibration network. The first branch calculates the rating bias based on the historical rating distribution, and the second branch calculates the viewing satisfaction based on the viewing completion rate. Input the rating bias and the viewing satisfaction into a self-attention layer to obtain a calibration weight, and adaptively adjust the compensated rating according to the calibration weight to generate the final video rating; Dynamically adjust the recommendation strategy based on the final video rating, including updating the sorting weights of the recommendation sequence and adjusting the display probability of the recommended videos.

8. A video content on-demand system applied to school education, for implementing the method described in any one of the preceding claims 1-7, characterized in that Including: The first unit is used to establish a video source library, perform content analysis using natural language processing technology, and generate video tags in combination with educational value; The second unit is used to process the student's personalized attributes and learning behavior characteristics through the main path and competitive path networks, calculate the competitive signal, perform dynamic feature suppression and fusion, generate a student feature representation, and generate a video applicability prediction score for personalized recommendation in combination with the video tags; Cluster the student feature data through a decision tree and adaptively adjust the splitting threshold, calculate the group features in combination with the decision path weights, and introduce a time-series decay mechanism for dynamic update and optimization to generate a group feature vector for group-level recommendation; Use feature fusion and multi-task learning methods to fuse the video applicability prediction score and the group feature vector to obtain the final video recommendation list; The third unit is used to dynamically optimize the generated video recommendation list, including: constructing a time-series preference representation based on the student's viewing history data, calculating an initial rating based on the adaptive scoring network, generating a final rating through a bias compensation and two-branch calibration mechanism, and dynamically adjusting the video recommendation strategy.

9. An electronic device, characterized in that, Including: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 7 is implemented.

Images