Personalized art learning path generation system and implementation method thereof

Abstract

The invention relates to the technical field of computers, and discloses a personalized art learning path generation system and an implementation method thereof. The system comprises a real-time behavior acquisition module, a standard feature library module, a behavior feature extraction module and the like. The method comprises the steps of collecting data such as student creation behaviors and work progress in real time, extracting creation style feature vectors, evaluating learning stage coefficients, generating an initial recommendation path, combining various parameters and feedback information, dynamically adjusting path node weights by utilizing a reinforcement learning algorithm, and generating a personalized learning path. Meanwhile, a media presentation form of a learning path is optimized according to multi-modal interaction data, and path parameters are iteratively updated according to learning feedback information. According to the method, the art learning efficiency can be improved, the personalized learning requirements of students are met, the learning experience is optimized, teaching resources are reasonably utilized, and an efficient and personalized learning path generation scheme is provided for art education.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and particularly to a personalized art learning path generation system and its implementation method. Background Art

[0002] Under the traditional art education mode, students usually follow a unified teaching syllabus and curriculum arrangement for learning. This "one-size-fits-all" teaching method fails to fully consider the individual differences of each student, such as learning styles, learning progress, interest preferences, etc. For art learning, the impact of individual differences on learning effects is particularly significant. Each student has a unique style and potential in art creation, and a unified teaching path is difficult to stimulate the maximum potential of students.

[0003] In terms of learning styles, some students tend to learn through visual observation and are more sensitive to elements such as colors and compositions; while some students are better at understanding and mastering knowledge through practical operations and achieve better learning effects during the hands-on creation process. However, the traditional teaching mode is difficult to provide personalized teaching content and methods for these different learning styles. Taking painting teaching as an example, teachers may explain painting techniques in a fixed order, from basic sketching to color application. For visual students, they may be able to quickly absorb knowledge in the color part, but the teaching rhythm of the basic sketching part may be too slow; for practical students, this theory-first teaching method may make them feel bored in the early stage and difficult to truly understand the knowledge, thus affecting their learning enthusiasm and effects.

[0004] Differences in learning progress are also a major problem faced by traditional art education. Different students have different speeds of mastering art knowledge and skills. Some students may be able to master a certain painting technique proficiently in a short time, while others need to spend more time and effort. In traditional class teaching, in order to ensure the overall teaching progress, teachers often have to teach according to the progress of students with medium levels. This leads to students with fast learning progress may feel that the course content is too simple to meet their learning needs, and their learning enthusiasm gradually decreases; while students with slow learning progress may not understand the knowledge well because they cannot keep up with the teaching rhythm, and thus have a sense of frustration, affecting their learning confidence.

[0005] From the perspective of interest preferences, the art field covers multiple branches such as painting, music, dance, sculpture, etc., and each student has different degrees of interest in different art forms. Even in the same art form, such as painting, students also have differences in their preferences for different painting styles (such as realism, impressionism, abstractism, etc.). The traditional teaching mode is difficult to adjust teaching content according to students' interest preferences, and students may be forced to learn content they are not interested in, which not only reduces the learning pleasure but also is not conducive to the in-depth development of students in the art field.

[0006] With the rapid development of artificial intelligence technology, its applications in the field of education are becoming increasingly widespread. However, in the field of art learning, there are still many challenges in using artificial intelligence to generate personalized learning paths. Art learning involves a large amount of multi-modal data, such as the creative behavior data (brushstrokes, lines, etc.) during the student's painting process, the work completion progress data, the voice feedback during the learning process, the visual attention distribution, etc. How to effectively collect, integrate and analyze this multi-modal data and extract valuable information from it is one of the key issues in realizing the generation of personalized learning paths. In addition, art knowledge is complex and diverse, and it is difficult to construct an accurate and comprehensive standard knowledge graph for the art discipline. At the same time, how to effectively match the student's learning data with the standard knowledge graph to accurately evaluate the student's learning progress and level is also an urgent problem to be solved.

[0007] Currently, although there are some online art learning platforms in the market, most of these platforms only provide a large amount of learning resources and lack personalized learning path planning. When faced with numerous learning resources, students often feel confused, not knowing where to start learning and finding it difficult to find learning content and a learning rhythm suitable for themselves. In summary, the traditional art education model and the existing online learning platforms cannot meet the needs of students' personalized art learning. Developing a system and method that can generate personalized art learning paths according to students' individual differences has important practical significance. Summary of the Invention

[0008] The purpose of the present invention is to provide a personalized art learning path generation system and its implementation method to solve the problems raised in the above background technology.

[0009] To achieve the above purpose, the present invention provides the following technical solutions: A personalized art learning path generation system, the system includes:

[0010] A real-time behavior collection module for obtaining in real time the creative behavior data, work completion progress data, multi-modal interaction data and learning feedback information of students during the learning cycle;

[0011] A standard feature library module for storing the standard knowledge graph of the art discipline and the historical excellent work feature data set;

[0012] A behavior feature extraction module for extracting creative style feature vectors and operation coherence parameters according to the creative behavior data;

[0013] A learning progress evaluation module for calculating the learning stage coefficient δ based on the matching degree between the work completion progress data and the standard knowledge graph;

[0014] A collaborative recommendation module for performing cross-modal similarity matching on the creative style feature vector and the historical excellent work feature dataset to generate an initial recommendation path;

[0015] A dynamic path generation module for dynamically adjusting the node weights of the initial recommendation path through a reinforcement learning algorithm by combining the learning stage coefficient δ, the operation coherence parameter, and the real-time feedback information to generate a personalized learning path;

[0016] A multi-modal adaptation module for optimizing the media presentation form of the learning path according to the visual attention distribution and the operation delay parameter in the multi-modal interaction data;

[0017] A feedback correction module for iteratively updating the priority parameters of the recommendation path through the gradient descent method according to the path deviation θ and the content acceptance η in the learning feedback information;

[0018] A path output module for synchronously pushing the optimized personalized learning path to the teaching terminal in a preset format.

[0019] Preferably, the behavior feature extraction module includes a style clustering unit, and the style clustering unit performs density clustering on the creative behavior data by using an improved DBSCAN algorithm to generate the creative style feature vector.

[0020] Preferably, the learning progress evaluation module includes a knowledge graph matching unit, and the knowledge graph matching unit calculates the semantic association degree between the current work node and the standard knowledge graph through a graph convolutional network to determine the learning stage coefficient δ.

[0021] Preferably, the collaborative recommendation module includes a transfer learning unit, and the transfer learning unit maps the creative style feature vector x c and the historical excellent work feature dataset {x h} through cross-domain feature mapping to perform adversarial generative network fusion to generate an initial recommendation path; during the adversarial generative network fusion process, the objective functions of the generator G and the discriminator D are respectively:

[0022]

[0023] where L G is the objective function of the generator G, represents the mathematical expectation, and p h (x h ) is the probability distribution function of the historical excellent work feature dataset; L D is the objective function of the discriminator D, p c (x c ) is the probability distribution function of the creative style feature vector x c , and G(x c)For the generator to create the style feature vector x c The feature vector generated for the input.

[0024] Preferably, the dynamic path generation module includes a node weight optimization unit, and the node weight optimization unit jointly optimizes the teaching duration, difficulty gradient, and resource type of the path nodes by using the double-delay deep deterministic policy gradient algorithm.

[0025] Preferably, the multimodal adaptation module is provided with an attention allocation unit, and the attention allocation unit analyzes the temporal characteristics of the visual attention distribution through a temporal convolutional network to dynamically adjust the presentation ratio ρ of the visual media in the learning path; let the output of the temporal convolutional network be o t then where w i is the weight.

[0026] Preferably, the feedback correction module includes a deviation calculation unit, and the deviation calculation unit constructs a feedback loss function L based on the weighted harmonic mean of the path deviation θ and the content acceptance η to drive the parameter update of the gradient descent method; where where α is the weight coefficient.

[0027] Preferably, the dynamic path generation module includes a path conflict detection unit, and the path conflict detection unit performs real-time verification on the logical conflicts and resource overload states in the learning path through a constraint satisfaction problem model.

[0028] Preferably, the standard feature library module includes an incremental update unit, and the incremental update unit screens high-information samples in the historical excellent work feature dataset through an online active learning mechanism to update the standard knowledge graph.

[0029] Preferably, the present invention further includes a method for implementing a personalized art learning path generation system, characterized in that the method includes the following steps:

[0030] S1: Obtain multimodal interaction data and learning feedback information during the learning period through the real-time behavior acquisition module;

[0031] S2: Use the behavior feature extraction module to perform clustering analysis on the creative behavior data to generate a creative style feature vector;

[0032] S3: Calculate the learning stage coefficient δ through the knowledge graph matching unit, and generate an initial recommended path in combination with the collaborative recommendation module;

[0033] S4: Dynamically optimize the node weights of the initial recommended path by using a reinforcement learning algorithm to generate a personalized learning path;

[0034] S5: Adjust the media presentation form of the path based on the multimodal adaptation module, and push it to the teaching terminal through the path output module;

[0035] S6: Iteratively update the path parameters according to the feedback correction module, and improve the standard feature library through the incremental update unit.

[0036] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0037] In terms of improving learning efficiency, the system obtains multi-dimensional information such as students' creative behavior data and work completion progress data through the real-time behavior acquisition module. The behavior feature extraction module and the learning progress evaluation module can accurately analyze the learning status of students. Taking painting learning as an example, the system can analyze the creative style feature vector based on the creative behavior data such as the brushstrokes and line applications when students are painting. At the same time, the learning stage coefficient is calculated through the matching degree between the work completion progress data and the standard knowledge graph, and then the learning stage where the student is located can be accurately judged. In this way, the collaborative recommendation module can generate an initial recommended path that suits the student's current level based on these accurate analysis results, avoiding students from learning overly simple or difficult content, greatly saving learning time and improving learning efficiency. For example, for students who have mastered basic sketching skills, the system will directly recommend more advanced sketching courses or related color painting courses instead of letting students repeat learning basic content.

[0038] From the perspective of meeting personalized learning needs, the present invention fully considers the individual differences of students. The dynamic path generation module combines the learning stage coefficient, the operation coherence parameter, and the real-time feedback information, and dynamically adjusts the node weights of the initial recommended path through the reinforcement learning algorithm. The learning path of each student is unique and can adapt to the learning styles, interest preferences, etc. of different students. For example, for students interested in the realistic painting style, when the system recommends a learning path, it will give priority to recommending learning resources such as courses related to realistic painting and appreciation of excellent works, and continuously optimize the learning path according to the students' learning progress and feedback. During the learning process, if a student has difficulty understanding a certain knowledge point, the system will increase the weight of the learning resources related to this knowledge point and provide more explanations and practice opportunities; if a student's learning progress is fast, the system will adjust the path in a timely manner and recommend more challenging content, truly achieving individualized teaching.

[0039] In terms of optimizing the learning experience, the multimodal adaptation module plays an important role. This module optimizes the media presentation form of the learning path according to the visual attention distribution and operation delay parameters in the multimodal interaction data. In art learning, different media forms (such as videos, pictures, texts, etc.) have different impacts on students' learning effects. The system analyzes the visual attention distribution of students to understand the degree of attention of students to different media forms during the learning process, and dynamically adjusts the presentation ratio of visual media in the learning path. If a student pays more attention to the demonstration operation part in a painting teaching video, the system will appropriately increase the playback duration of this part or provide more similar demonstration videos; if a student is more interested in the text description, the system will increase the display of relevant text materials. This can improve students' learning enthusiasm and participation, make the learning process more vivid and interesting, and enhance the overall learning experience.

[0040] From the perspective of the rational utilization of teaching resources, the path conflict detection unit and the incremental update unit are of great significance. The path conflict detection unit uses a constraint satisfaction problem model to perform real-time verification on the logical conflicts and resource overload states in the learning path, avoiding problems such as inconsistent learning content or excessive learning burden caused by over-concentration of learning resources during the learning process of students. For example, the system will not recommend advanced painting creation courses when students have not yet mastered basic painting skills, ensuring the logic and rationality of the learning path. The incremental update unit selects high-information samples from the historical excellent work feature dataset through an online active learning mechanism to update the standard knowledge graph. This enables the system to continuously absorb new art knowledge and excellent work experience, maintain the timeliness and advancement of teaching resources, provide students with higher-quality and more cutting-edge learning content, and at the same time promote the optimal allocation of teaching resources.

[0041] The feedback correction module iteratively updates the priority parameters of the recommended path according to the path deviation degree and content acceptance degree in the learning feedback information, and can continuously optimize the learning path to make it more in line with the actual learning situation of students. As the learning process of students progresses, the system will continuously adjust the learning path according to the feedback of students, so that the recommended learning content and learning methods are always consistent with the needs of students, ensuring that students can continuously progress on the most suitable learning path for themselves, improving the quality and effect of learning, and providing strong support for students on the art learning path. Brief Description of the Drawings

[0042] Figure 1 It is the working principle diagram of the personalized art learning path generation system described in the present invention;

[0043] Figure 2 It is the working flow chart of style clustering in the behavior feature extraction module;

[0044] Figure 3 It is a flowchart for optimizing node weights in the dynamic path generation module;

[0045] Figure 4 It is a flowchart for attention allocation in the multi-modal adaptation module. Specific implementation manners

[0046] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. 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.

[0047] Please refer to Figures 1-4 , the present invention provides a technical solution: a personalized art learning path generation system, and the system includes:

[0048] Real-time behavior collection module: During the learning cycle of students, it collects students' creative behavior data, work completion progress data, multi-modal interaction data, and learning feedback information in real time. For example, by embedding a monitoring program in the creative software used by students, it records every operation of students during the creative process such as painting and design, including the movement trajectory of the paintbrush, color selection, and the order of graphic drawing, etc., as creative behavior data; at the same time, it records the completion degree of students' works, such as the completion ratio of paintings and the stage progress of design projects, etc., as work completion progress data. It uses devices such as cameras and microphones to collect students' visual, auditory, and other multi-modal interaction data during the learning process, such as the position of students' gaze focus when watching teaching videos and voice questions during the learning process. In addition, it obtains students' feedback information on learning content and learning paths through methods such as online questionnaires and feedback functions of learning platforms.

[0049] Standard feature library module: Stores the standard knowledge graph of art disciplines and the historical excellent work feature data set. The standard knowledge graph covers all knowledge points of art disciplines and their interrelationships. For example, in the field of painting, it includes knowledge nodes such as color theory, composition principles, and characteristics of different painting styles, as well as the logical connections between them. The historical excellent work feature data set contains a large amount of feature information of previous excellent art works, such as color features, brushstroke style features, and theme expression features of works.

[0050] Behavior feature extraction module: Extracts creative style feature vectors and operation coherence parameters according to the collected creative behavior data. The creative style feature vectors are used to characterize the unique creative style characteristics of students, and the operation coherence parameters reflect the smoothness and coherence of students' operations during the creative process.

[0051] Learning Progress Evaluation Module: Calculate the learning stage coefficient δ based on the matching degree between the work completion progress data and the standard knowledge graph. By comparing the knowledge points involved in the student's current work with those in the standard knowledge graph, evaluate the stage at which the student is in the art learning process. The larger the δ value, the closer the student is to completing the learning goals of the current stage in terms of learning progress.

[0052] Collaborative Recommendation Module: Perform cross-modal similarity matching on the creative style feature vector and the historical excellent work feature dataset to generate an initial recommendation path. Compare the student's creative style with the styles of historical excellent works, find works with similar styles, and based on these works, combine with the general laws of art learning to recommend a series of learning contents and learning steps suitable for the student to form an initial recommendation path.

[0053] Dynamic Path Generation Module: Combine the learning stage coefficient δ, the operation coherence parameter, and the real-time feedback information, and dynamically adjust the node weights of the initial recommendation path through the reinforcement learning algorithm to generate a personalized learning path. Re-evaluate and adjust the importance of each learning content node and learning step node in the initial recommendation path according to the student's learning progress, operation coherence, and real-time feedback, so that the learning path better meets the actual learning situation and needs of the student.

[0054] Multi-modal Adaptation Module: Optimize the media presentation form of the learning path according to the visual attention distribution and operation delay parameters in the multi-modal interaction data. For example, if it is found that the student pays more attention to a certain area of the video during the learning process, it indicates that the content in this area is more important to the student. Then, in the subsequent learning path, the presentation ratio of this type of content can be appropriately increased or the presentation method can be optimized; if the operation delay is high, it indicates that there may be problems such as slow loading in the current media form, and the media format needs to be adjusted or the loading method needs to be changed.

[0055] Feedback Correction Module: Iteratively update the priority parameters of the recommendation path through the gradient descent method according to the path deviation degree θ and content acceptance degree η in the learning feedback information. The path deviation degree θ reflects the difference between the student's actual learning path and the recommendation path, and the content acceptance degree η represents the degree of recognition and acceptance of the learning content by the student. Construct a feedback loss function through these two indicators, and continuously adjust the priority of each node in the recommendation path using the gradient descent method to make the recommendation path more in line with the student's needs.

[0056] Path Output Module: Synchronously push the optimized personalized learning path to the teaching terminal in a preset format. The teaching terminal can be devices such as computers and tablets used by students. Students can view the learning path customized for them on these devices and conduct art learning according to the guidance in the path.

[0057] The following elaborates on the related technical solutions of the present invention through 7 embodiments:

[0058] Embodiment 1:

[0059] In this embodiment, the working process of the style clustering unit in the behavior feature extraction module is mainly elaborated. The style clustering unit in the behavior feature extraction module uses an improved DBSCAN algorithm to perform density clustering on the creative behavior data, thereby generating creative style feature vectors. Suppose the creative behavior data can be represented as a set of data points X = {x1, x2,..., x n}, where x i represents the i-th creative behavior data point, and these data points may contain various parameter information during creation, such as the stroke pressure and line length in painting. The improved DBSCAN algorithm first sets two key parameters: the neighborhood radius and the minimum number of points MinPts. For each point x i in the dataset, calculate the number of data points within its neighborhood with a radius of If then mark x i as a core point. Starting from the core point, continuously expand the cluster through the density reachability relationship. For example, if point x j is within the neighborhood of point x i and x i is a core point, then x j and x i belong to the same cluster. After such a clustering process, different creative behavior data points are divided into different clusters, and each cluster represents a creative style. Extract a vector that can represent the characteristics of the cluster from each cluster, that is, the creative style feature vector. For example, for a clustering of painting creative behavior data, its creative style feature vector may include the color preference eigenvalue C, the composition complexity eigenvalue S, etc. In this way, the style clustering unit effectively extracts the creative style feature vector from the creative behavior data, providing an important basis for the subsequent generation of learning paths.

[0060] Embodiment 2:

[0061] This embodiment details the working principle of the knowledge graph matching unit in the learning progress evaluation module. The knowledge graph matching unit of the learning progress evaluation module calculates the semantic correlation degree between the current work node and the standard knowledge graph through a graph convolutional network, thereby determining the learning stage coefficient δ. Let the standard knowledge graph be G = (V, E), where V is the set of nodes in the knowledge graph, representing various art knowledge concepts, and E is the set of edges, representing the relationships between nodes. The set of knowledge nodes involved in the current student work constitutes P = {p1, p2,..., p m}。The graph convolutional network learns the semantic representation of each node by performing convolutional operations on the node features and structure of the knowledge graph. Assume that the output of the l-th layer of the graph convolutional network is H l , for node v i ∈V, its feature representation is Through the operations of the multi-layer graph convolutional network, the final semantic representation of the node is obtained. Then, calculate the semantic similarity Sim(p j with the nodes in the standard knowledge graph j ,v i ), for example, cosine similarity can be used for calculation, and the formula is where p j ·v i represents the dot product of vectors, |p j | and |v i | respectively represent the norms of vectors p j and v i . According to these semantic similarities, calculate the overall semantic correlation Degree between the current work and the standard knowledge graph, and then through a specific mapping function f, obtain the learning stage coefficient δ = f(Degree). In this way, the knowledge graph matching unit can accurately determine the learning stage coefficient of the student, providing a quantitative basis for the adjustment of the subsequent learning path.

[0062] Example 3:

[0063] The transfer learning unit of the collaborative recommendation module fuses the creative style feature vector x c with the historical excellent work feature dataset x h through cross-domain feature mapping to generate an initial recommendation path. During the adversarial generative network fusion process, the objective functions of the generator G and the discriminator D are respectively:

[0064]

[0065] where L G is the objective function of the generator G, E represents the mathematical expectation, p h (x h ) is the probability distribution function of the historical excellent work feature dataset; L D is the objective function of the discriminator D, p c (x c ) is the probability distribution function of the creative style feature vector x c , and G(x c ) is the feature vector generated by the generator with the creative style feature vector x c as the input.

[0066] The transfer learning unit first takes the creative style feature vector xc Input into the generator G, and the generator G generates a new feature vector G(x c ) according to its internal neural network structure and training parameters. The role of the discriminator D is to distinguish whether the input feature vector comes from the historical excellent work feature dataset x h or the G(x c ) generated by the generator. During the training process, the generator G tries to generate vectors closer to the features of historical excellent works to minimize L G , that is, to make it difficult for the discriminator D to distinguish the generated vectors from the real historical excellent work feature vectors; while the discriminator D tries to improve its discrimination ability to maximize L D . By continuously iteratively training the generator G and the discriminator D, when the two reach a balance, the feature vectors generated by the generator G have a high similarity to the features of historical excellent works. According to these similar feature vectors, relevant parts are screened out from the learning paths of historical excellent works and combined to generate an initial recommendation path. For example, if the generated feature vector is similar to the features of a certain type of historical excellent painting works, then learning resources and learning steps suitable for the current student are selected from the learning paths of this type of works to form an initial recommendation path.

[0067] Example 4:

[0068] The node weight optimization unit of the dynamic path generation module uses the Twin Delayed Deep Deterministic Policy Gradient algorithm (TD3) to jointly optimize the teaching duration, difficulty gradient, and resource type of path nodes. Assume the path node set is N = {n1, n2, …, n k}, for each node n i , its teaching duration is represented as t i , the difficulty gradient is represented as d i , and the resource type is represented by the vector r i . The TD3 algorithm contains two policy networks (the main policy network μ and the target policy network μ′) and two Q networks (the main Q networks Q1, Q2 and the target Q networks Q′1, Q′2). The main policy network μ outputs an action a = μ(s) according to the current learning state s. Here, the action can be understood as an operation to adjust the weights of path nodes, such as the increment Δt of the teaching duration adjustment and the adjustment coefficient β of the difficulty gradient. The Q network is used to evaluate the value of the action. Q1(s, a) and Q2(s, a) respectively represent the value evaluations of the two main Q networks for executing the action a in the state s. The target network is used to stabilize the training process. The update formula of the target Q network is:

[0069] Q′1(s′, a′) = r + γmin i＝1,2 Q′ i (s′, μ′(s′))

[0070] Q′2(s′, a′) = r + γ min i＝1,2 Q′ i (s′, μ′(s′))

[0071] where s′ is the next learning state, a′ is the next action, r is the reward value, and γ is the discount factor. The update of the main policy network is achieved by maximizing the Q value, i.e., μ = arg max a Q1(s, μ(s)). By continuously performing actions in different learning states and updating network parameters, the TD3 algorithm can jointly optimize the teaching duration, difficulty gradient, and resource type of path nodes, making the learning path more in line with the learning needs of students. For example, for students with stronger learning abilities, appropriately shorten the teaching duration of simple knowledge points, increase the difficulty gradient, and match more challenging learning resources; for students with weaker learning abilities, increase the teaching duration of basic knowledge points, reduce the difficulty gradient, and provide more basic learning resources.

[0072] Example 5:

[0073] The multimodal adaptation module is provided with an attention allocation unit, which analyzes the temporal characteristics of the visual attention distribution through a temporal convolutional network to dynamically adjust the presentation ratio ρ of visual media in the learning path. Let the input of the temporal convolutional network be the visual attention distribution data sequence A = {a1, a2, …, a T}, where a t represents the visual attention distribution vector at time t, containing information such as the fixation area and fixation duration. The temporal convolutional network performs a convolution operation on the input data through the convolution kernel w. Suppose the output of the convolutional layer is o t , and its calculation formula is where k is the size of the convolution kernel. The output of the temporal convolutional network is o t , then the presentation ratio of visual media where w i is the weight. The attention allocation unit dynamically adjusts the presentation ratio of visual media in the learning path according to the calculated ρ value. For example, if within a certain period of time, the student's visual attention is concentrated on the facial area of the person in the video for a long time, and a relatively high ρ value is obtained through the analysis of the temporal convolutional network, then in the subsequent learning path, increase the display ratio of pictures or video clips containing close-ups of the person's face; conversely, if the student's visual attention to the overall scene is relatively high and the ρ value is low, then correspondingly increase the display ratio of panoramic pictures, thereby optimizing the media presentation form of the learning path and improving the learning effect of students.

[0074] Example 6:

[0075] The deviation calculation unit of the feedback correction module constructs a feedback loss function \(L\) based on the weighted harmonic mean of the path deviation \(\theta\) and the content acceptance \(\eta\) to drive the parameter update of the gradient descent method. Among them \(\alpha\) is the weight coefficient. The path deviation \(\theta\) is used to measure the degree of difference between the actual learning path of the student and the recommended learning path. For example, it can be obtained by calculating the number of differences in learning nodes, order differences, etc. between the two paths. The content acceptance \(\eta\) reflects the student's acceptance of the learning content and can be evaluated by the student's learning duration, interaction behaviors during the learning process (such as whether to frequently click pause, whether to actively participate in discussions, etc.). The weight coefficient \(\alpha\) is used to adjust the relative importance of the path deviation and the content acceptance in the feedback loss function. When the feedback correction module receives the learning feedback information, the deviation calculation unit first calculates the values of the path deviation \(\theta\) and the content acceptance \(\eta\). Then, the feedback loss function \(L\) is calculated according to the above formula. The gradient descent method gradually reduces the feedback loss function \(L\) by continuously adjusting the priority parameters of the recommended path. For example, assuming that the priority parameter of the recommended path is the vector \(p = \{p_1, p_2, \ldots, p m \}\), the update formula of the gradient descent method is where \(\lambda\) is the learning rate, is the parameter value before the update, is the parameter value after the update. By continuously iteratively updating the priority parameters, the recommended path can better meet the actual learning situation and needs of the students.

[0076] Embodiment 7:

[0077] This embodiment is used to illustrate the working process of the path conflict detection unit in the dynamic path generation module and the incremental update unit in the standard feature library module. The path conflict detection unit of the dynamic path generation module performs real-time verification on the logical conflicts and resource overload status in the learning path through a constraint satisfaction problem model. Assume that the learning path can be represented as a set \(R=\{r_1, r_2, \ldots, r m \}\) containing multiple learning tasks and resource allocations. Each learning task \(r i \) has its precondition \(Pre(r i )\), postcondition \(Post(r i )\) and required resources \(Resource(r i )\). The constraint satisfaction problem model first defines a series of constraint conditions, such as the task execution order constraint: if the precondition of task \(r j \) contains the postcondition of task \(r i \), then \(r i \) must be before \(r jExecution before; Resource constraint: The total resources required by all tasks cannot exceed the upper limit of resources provided by the system. The path conflict detection unit continuously checks whether the tasks in the learning path meet these constraints. For example, if in the learning path, a task of learning advanced painting skills is arranged before the task of learning basic painting knowledge, this violates the task execution order constraint; if multiple tasks request a large amount of video resources simultaneously, resulting in insufficient system resources, this is a state of resource overload. Once a conflict is detected, the path conflict detection unit will issue an alarm and trigger the path adjustment mechanism.

[0078] The incremental update unit of the standard feature library module screens high-information samples in the historical excellent work feature dataset through an online active learning mechanism to update the standard knowledge graph. Let the historical excellent work feature dataset be D = {d1, d2, …, d n}, for each sample d i , its information content is evaluated by calculating its uncertainty index U(d i ), for example, entropy can be used as the uncertainty index, and the formula is where p ij is the probability that the sample d i belongs to the j-th category, and k is the total number of categories. The incremental update unit sorts the samples in descending order according to the uncertainty index and selects the top m high-information samples. Then these samples are fused with the standard knowledge graph to update the nodes and edges in the knowledge graph. For example, if the new high-information sample contains a new painting style element, the incremental update unit will add a corresponding node in the standard knowledge graph to represent this style element and establish its connection with other relevant knowledge nodes, thus continuously improving the standard knowledge graph and providing more accurate and richer knowledge support for the generation of personalized art learning paths.

[0079] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the term "comprising", "including" or any other variant thereof is intended to cover non-exclusive inclusion, so that a process, method, article or device comprising a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0080] Although embodiments of the present invention have been shown and described, those of ordinary skill in the art will appreciate that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A personalized art learning path generation system, characterized in that, Including: A real-time behavior collection module for obtaining in real time the creative behavior data, work completion progress data, multi-modal interaction data, and learning feedback information of students within a learning cycle; A standard feature library module for storing the standard knowledge graph of the art discipline and the historical excellent work feature data set; A behavior feature extraction module for extracting a creative style feature vector and an operation coherence parameter according to the creative behavior data; A learning progress evaluation module for calculating a learning stage coefficient δ based on the matching degree between the work completion progress data and the standard knowledge graph; A collaborative recommendation module for performing cross-modal similarity matching between the creative style feature vector and the historical excellent work feature data set to generate an initial recommendation path; A dynamic path generation module for dynamically adjusting the node weights of the initial recommendation path through a reinforcement learning algorithm in combination with the learning stage coefficient δ, the operation coherence parameter, and the real-time feedback information to generate a personalized learning path; A multi-modal adaptation module for optimizing the media presentation form of the learning path according to the visual attention distribution and operation delay parameters in the multi-modal interaction data; A feedback correction module for iteratively updating the priority parameter of the recommendation path through the gradient descent method according to the path deviation degree θ and content acceptance degree η in the learning feedback information; A path output module for synchronously pushing the optimized personalized learning path to the teaching terminal in a preset format.

2. The personalized art learning path generation system according to claim 1, wherein The behavior feature extraction module includes a style clustering unit, and the style clustering unit uses an improved DBSCAN algorithm to perform density clustering on the creative behavior data to generate the creative style feature vector.

3. The personalized art learning path generation system according to claim 2, characterized in that, The learning progress evaluation module includes a knowledge graph matching unit, and the knowledge graph matching unit calculates the semantic association degree between the current work node and the standard knowledge graph through a graph convolutional network to determine the learning stage coefficient δ.

4. The personalized art learning path generation system according to claim 3, wherein The collaborative recommendation module includes a transfer learning unit, and the transfer learning unit maps the creative style feature vector x through cross-domain feature mapping c and the historical excellent work feature dataset {x h} for adversarial generative network fusion to generate an initial recommendation path; during the adversarial generative network fusion process, the objective functions of the generator G and the discriminator D are respectively: Among them, L G is the objective function of the generator G, represents the mathematical expectation, and p h (x h ) is the probability distribution function of the historical excellent work feature dataset; L D is the objective function of the discriminator D, and p c (x c ) is the probability distribution function of the creative style feature vector x c , and G(x c ) is the feature vector generated by the generator with the creative style feature vector x c as the input.

5. The personalized art learning path generation system according to claim 4, wherein The dynamic path generation module includes a node weight optimization unit, and the node weight optimization unit jointly optimizes the teaching duration, difficulty gradient, and resource type of the path nodes by using a double-delayed deep deterministic policy gradient algorithm.

6. The personalized art learning path generation system according to claim 5, wherein The multimodal adaptation module is provided with an attention allocation unit, and the attention allocation unit analyzes the temporal characteristics of the visual attention distribution through a temporal convolutional network to dynamically adjust the presentation ratio ρ of the visual media in the learning path; let the output of the temporal convolutional network be o t , then where w i is the weight.

7. The personalized art learning path generation system according to claim 6, wherein The feedback correction module includes a deviation calculation unit, and the deviation calculation unit constructs a feedback loss function L based on the weighted harmonic mean of the path deviation θ and the content acceptance η to drive the parameter update of the gradient descent method; where where α is a weight coefficient.

8. The personalized art learning path generation system according to claim 7, wherein The dynamic path generation module includes a path conflict detection unit, and the path conflict detection unit performs real-time verification on the logical conflicts and resource overload states in the learning path through a constraint satisfaction problem model.

9. The personalized art learning path generation system according to claim 8, wherein The standard feature library module includes an incremental update unit, and the incremental update unit screens high-information samples in the historical excellent work feature data set through an online active learning mechanism to update the standard knowledge graph.

10. An implementation method of a personalized art learning path generation system, characterized in that, Including the following steps: S1: Obtain the multi-modal interaction data and learning feedback information within a learning cycle through the real-time behavior collection module; S2: Use the behavior feature extraction module to perform clustering analysis on the creative behavior data to generate a creative style feature vector; S3: Calculate the learning stage coefficient δ through the knowledge graph matching unit and generate an initial recommendation path in combination with the collaborative recommendation module; S4: Dynamically optimize the node weights of the initial recommendation path by using a reinforcement learning algorithm to generate a personalized learning path; S5: Adjust the media presentation form of the path based on the multi-modal adaptation module and push it to the teaching terminal through the path output module; S6: Iteratively update the path parameters according to the feedback correction module, and improve the standard feature library through the incremental update unit.

Images