AI-enabled data-driven art educational resource recommendation engine

Abstract

The invention relates to the technical field of art education data analysis, and discloses an AI-enabled data-driven art education resource recommendation engine, which is characterized in that user behaviors, resource metadata and context environment data are acquired through a multi-source data acquisition module; the user portrait construction module constructs a multi-dimensional user portrait by using a deep clustering algorithm and an attention mechanism; the resource feature extraction module extracts resource features by means of a graph neural network; the privacy protection processing module uses differential privacy and federated learning to protect data privacy; and the dynamic recommendation generation module generates a personalized recommendation list based on reinforcement learning. The recommendation engine can accurately recommend art education resources meeting user requirements, data privacy is guaranteed, resource association is mined, recommendation strategies are optimized, the quality and effect of art education resource recommendation are effectively improved, and personalized learning requirements of users are met.

Description

Technical Field

[0001] The present invention relates to the technical field of art education data analysis, and specifically to an AI-enabled data-driven art education resource recommendation engine. Background Art

[0002] In today's digital age, the development of art education faces numerous challenges and opportunities. With the widespread adoption of internet technology, online art education resources have exploded, providing learners with a wealth of choices but also creating the challenge of resource selection. Traditional approaches to recommending art education resources rely primarily on simple categorization and user history, lacking in-depth data analysis and personalized recommendation capabilities, making it difficult to meet the diverse needs of different learners.

[0003] From a resource perspective, art education resources encompass a wide variety of forms, such as course videos, e-textbooks, and art appreciation. These resources possess complex and diverse metadata, including resource type, difficulty level tags, and creator information. However, existing recommendation systems often fail to fully exploit the inherent connections between this metadata, resulting in a poor match between recommended resources and learners' actual needs. For example, for a learner interested in Impressionist painting and possessing a basic understanding of the subject, traditional recommendation systems may be unable to accurately recommend advanced courses that match their skill level and style, or related resources such as art tools and artist introductions.

[0004] In terms of user behavior analysis, learners' behavioral data during the art learning process is rich and diverse, including click history, dwell time, and interaction frequency. This behavioral data not only reflects users' interests and preferences, but also contains information about their learning patterns and skill levels. However, traditional recommendation methods do not fully utilize this data, and relying solely on simple statistical analysis fails to gain deep insights into user needs. For example, recommending resources based solely on the number of clicks may overlook the differences in the depth of interest reflected by the length of time users spend on different resources, and fails to account for stage-by-stage changes in the user's learning process.

[0005] Furthermore, as awareness of data privacy protection continues to grow, data security and privacy issues in the field of art education are becoming increasingly prominent. During data collection and use, ensuring that user data is not leaked while effectively utilizing this data for accurate recommendations is a pressing issue. Existing recommendation systems often lack privacy protection, potentially exposing sensitive user information to risk. This not only harms user interests but also limits the further development of recommendation systems.

[0006] At the same time, art education possesses unique disciplinary characteristics, with complex interrelationships between different art forms and styles. For example, music and dance are interconnected in rhythm and rhyme, while painting and sculpture share similarities in form and spatial expression. However, existing recommendation engines struggle to capture these cross-disciplinary connections, making it difficult to provide learners with comprehensive and systematic recommendations for art education resources, hindering their ability to build a complete artistic knowledge system. Summary of the Invention

[0007] The purpose of the present invention is to provide an AI-enabled data-driven art education resource recommendation engine to solve the problems raised in the above-mentioned background technology.

[0008] To achieve the above objectives, the present invention provides the following technical solution: an AI-enabled data-driven art education resource recommendation engine, the engine comprising:

[0009] Multi-source data collection module, used to obtain user behavior data, resource metadata and contextual environment data through multiple channels;

[0010] The user profile construction module extracts multi-dimensional features from user behavior data based on a deep clustering algorithm, generates user interest vectors, skill level vectors, and learning mode vectors, and dynamically weights and fuses these vectors through an attention mechanism to construct a multi-dimensional user profile.

[0011] The resource feature extraction module uses a graph neural network to perform topological modeling on resource metadata, constructs the relationship between resources into a heterogeneous knowledge graph, and extracts resource features based on graph embedding technology. The resource features include content feature vectors, style feature vectors, and applicable scenario vectors.

[0012] The privacy protection processing module injects noise into user profiles and resource features through a differential privacy mechanism, and combines it with a federated learning framework for distributed data training to achieve local storage and anonymization of user data.

[0013] The dynamic recommendation generation module builds a multi-objective optimization strategy based on the reinforcement learning framework, inputs multi-dimensional user portraits, resource characteristics and privacy protection constraints into the strategy network, generates a candidate resource sequence through Monte Carlo tree search, and enhances the diversity of the sequence based on the adversarial generative network to output a personalized recommendation list.

[0014] Preferably, the user portrait construction module extracts multi-dimensional features from user behavior data based on a deep clustering algorithm, including:

[0015] Construct a temporal encoder for user behavior sequences, using a bidirectional long short-term memory network to capture the temporal dependencies of behavior data and generate preliminary interest representations.

[0016] Skill levels are hierarchically clustered through self-supervised contrastive learning, and a skill evaluation matrix is constructed based on course completion, practice accuracy, and creation complexity.

[0017] The course attention mechanism is introduced to dynamically adjust the interest weight based on the user's interaction intensity on different resource types, and combined with the gated recurrent unit to model long-term learning patterns;

[0018] The interest representation, skill evaluation matrix and learning pattern vector are input into the feature fusion layer, and a multi-head attention mechanism is used to perform cross-dimensional correlation analysis to generate an implicit vector representation of the user portrait.

[0019] Preferably, the self-supervised contrastive learning for hierarchical clustering of skill levels comprises:

[0020] Construct skill triplet samples, including anchor resources, positive sample resources, and negative sample resources. The positive samples and anchor resources have the same difficulty labels, while the negative samples have different difficulty labels.

[0021] The contrast loss function is used to maximize the similarity between the anchor point and the positive sample in the embedding space, and minimize the similarity between the anchor point and the negative sample;

[0022] A curriculum learning strategy is used to dynamically adjust the difficulty of triple sampling. In the early stage, negative samples with significant differences in difficulty are selected, and in the later stage, negative samples with similar difficulty are selected to improve the model's discriminative ability.

[0023] The learned embedding representation is input into the hierarchical clustering algorithm to automatically divide the skill levels according to the density distribution of the skill evaluation matrix.

[0024] Preferably, the resource feature extraction module uses a graph neural network to perform topological modeling on resource metadata, including:

[0025] Define heterogeneous node types for resources, including courses, tools, artists, and styles, and build multi-type edge relationships between nodes, including affiliation, collaboration, and style derivative relationships;

[0026] Based on the graph attention network, the association weights between nodes are calculated, and the feature information of neighbor nodes is aggregated through multi-layer graph convolution to generate node-level embedding representation;

[0027] A meta-path-guided random walk strategy is used to sample subgraphs of the knowledge graph, and negative sampling techniques are used to optimize the discriminability of graph embeddings.

[0028] The node embedding and the resource original metadata are input into the feature splicing layer, and high-order semantic features are extracted based on the capsule network to generate a dynamic multimodal feature vector of the resource.

[0029] Preferably, the meta-path guided random walk strategy for performing subgraph sampling on the knowledge graph includes:

[0030] Predefined meta-path templates, including "Course-Tool-Artist", "Genre-Derivative Work-User", and "Resource-Related Tags-Scenario";

[0031] Generate a walk path based on the meta-path template and filter high-frequency associated subgraphs using path importance scores, where path importance is calculated based on node co-occurrence frequency and edge weights.

[0032] A heterogeneous neighborhood sampler is used to expand the nodes of the subgraph to ensure that the sampled subgraph covers multi-hop relationships;

[0033] The graph embedding model is optimized through subgraph-level contrastive learning to maximize the embedding consistency of nodes in the same subgraph.

[0034] Preferably, the privacy protection processing module injects noise into user profiles and resource features through a differential privacy mechanism, including:

[0035] Perform sensitivity analysis on the implicit vector of the user profile and add random noise to the feature dimension based on the Laplace mechanism to satisfy differential privacy constraints.

[0036] An asynchronous parameter aggregation protocol is designed within the federated learning framework. Local model gradients are encrypted and transmitted using homomorphic encryption, and the central server only receives encrypted gradient updates.

[0037] A dynamic privacy budget allocation strategy is introduced to adjust the noise injection intensity according to the usage frequency and resource relevance of user data, achieving a balance between privacy protection and recommendation accuracy.

[0038] Preferably, the introduction of a dynamic privacy budget allocation strategy to adjust the noise injection intensity according to the usage frequency and resource relevance of user data includes:

[0039] The privacy consumption coefficient is calculated based on the update frequency of the user profile. Users with high-frequency updates are assigned a lower privacy budget to reduce the risk of long-term privacy leakage.

[0040] Data sensitivity is assessed by the correlation between resource features and user profiles, and stronger noise perturbations are applied to highly correlated features.

[0041] A privacy budget redistribution mechanism is designed to reclaim unused privacy budget after the federated learning training cycle ends and use it for dynamic allocation in the next cycle.

[0042] Preferably, the dynamic recommendation generation module constructs a multi-objective optimization strategy based on a reinforcement learning framework, including:

[0043] The state space is defined as the combination vector of the current user profile, candidate resource features, and historical recommendation records, and the action space is the discretized operation of resource recommendation order and presentation method;

[0044] Construct a multi-objective reward function, including user satisfaction reward, resource diversity reward, and privacy compliance reward. The user satisfaction reward is calculated based on click-through rate and feedback score, the resource diversity reward is measured by the entropy value of the candidate sequence, and the privacy compliance reward is evaluated by the degree of data anonymization.

[0045] Adopting a dual deep Q network structure, the target network delayed update strategy is used to reduce overestimation bias, and the priority experience replay mechanism is combined to optimize the utilization of training samples.

[0046] Based on the policy gradient algorithm, the initial distribution of the recommendation strategy is generated, and the complexity of the action space is gradually increased through course learning to achieve strategy optimization from coarse-grained to fine-grained.

[0047] Preferably, the course learning to gradually increase the complexity of the action space includes:

[0048] In the initial stage, the recommended action is limited to screening of large categories of resources, and the dimension of the action space is the number of preset resource types;

[0049] In the mid-term phase, fine-grained actions are introduced, including resource sorting strategies, display style selection, and interactive guidance operations;

[0050] In the later stages, it will be expanded to multimodal action combinations, allowing for simultaneous adjustment of the order of recommendations, presentation methods, and auxiliary tool suggestions;

[0051] Through a progressive action space expansion strategy, combined with a teacher-student network architecture, historical strategies are used to generate auxiliary supervision signals to accelerate model convergence.

[0052] Preferably, the channels include online learning platforms, art communities, sensor devices and third-party API interfaces, the user behavior data include click records, stay duration and interaction frequency, the resource metadata include resource type, difficulty label and creator information, and the contextual environment data includes timestamp, device type and geographic location.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] In terms of precise recommendations, the multi-source data acquisition module collects extensive user behavior data, resource metadata, and contextual environment data, providing a rich and comprehensive information foundation for subsequent analysis. The user portrait construction module extracts multi-dimensional features from user behavior data based on a deep clustering algorithm, generates user interest vectors, skill level vectors, and learning mode vectors, and uses the attention mechanism to perform dynamic weighted fusion to construct a multi-dimensional user portrait, which can accurately portray the user's personalized characteristics. The resource feature extraction module uses a graph neural network to perform topological modeling on resource metadata, construct a heterogeneous knowledge graph, and extract resource features, including content feature vectors, style feature vectors, and applicable scenario vectors, making the resource description more comprehensive and accurate. Based on these precise user portraits and resource features, the dynamic recommendation generation module can generate personalized recommendation lists that highly meet user needs. For example, for a user who is studying an intermediate oil painting course and often pays attention to realistic oil paintings in the art community, the recommendation engine can accurately recommend advanced oil painting courses in a realistic style, as well as related oil painting tool introductions, sharing of creative techniques by realist painters and other resources, greatly improving the accuracy and relevance of recommendations, meeting the user's personalized needs in the process of art learning, and enhancing the user's learning experience and learning results.

[0055] In terms of privacy protection, the privacy-preserving processing module injects noise into user profiles and resource features through a differential privacy mechanism. This, combined with distributed data training within a federated learning framework, enables local storage and anonymization of user data. Based on data sensitivity analysis, the Laplace mechanism adds random noise to the feature dimension that satisfies differential privacy constraints, effectively protecting the privacy of user data. Furthermore, an asynchronous parameter aggregation protocol is designed within the federated learning framework. Local model gradients are encrypted and transmitted via homomorphic encryption, with the central server receiving only encrypted gradient updates, further ensuring data security during transmission. A dynamic privacy budget allocation strategy is introduced to adjust the noise injection intensity based on the frequency of user data usage and resource relevance, achieving a balance between privacy protection and recommendation accuracy. This means that while users enjoy personalized recommendation services, their personal data is fully protected, enhancing their trust in the art education platform and providing strong support for the sustainable development of art education resource recommendation services.

[0056] In terms of resource mining and integration, the resource feature extraction module clearly displays the relationships between resources by constructing a heterogeneous knowledge graph. Defining heterogeneous node types and multi-type edge relationships of resources, such as the affiliation, collaboration, and style derivation relationships between courses, tools, artists, and styles and genres, can deeply explore the intrinsic connections between art education resources. A meta-path-guided random walk strategy is used to sample subgraphs of the knowledge graph, and negative sampling technology is combined to optimize the discriminability of graph embeddings, making the extraction of resource features more comprehensive and in-depth. This full exploration of resource relationships helps to provide users with richer and more systematic art education resources. For example, when recommending a classical music course, the recommendation engine can also recommend related classical instrument introductions, works by famous musicians who play the instrument, and appreciation of art paintings related to classical music styles, etc., helping users build a more complete art knowledge system and broaden their artistic horizons.

[0057] In terms of algorithm optimization and adaptability, the user portrait construction module uses a bidirectional long short-term memory network to capture the temporal dependencies of behavioral data. Combined with advanced technologies such as self-supervised contrastive learning and course attention mechanisms, it can better adapt to the learning behavior patterns and skill improvement processes of different users. The graph neural network and capsule network technologies in the resource feature extraction module can efficiently process complex resource metadata and accurately extract high-order semantic features of resources. The dynamic recommendation generation module constructs a multi-objective optimization strategy based on the reinforcement learning framework, adopting a dual deep Q network structure, a target network delayed update strategy, and a priority experience replay mechanism to effectively improve the optimization efficiency and accuracy of the recommendation strategy. At the same time, by gradually increasing the complexity of the action space through course learning, the recommendation system can continuously optimize the recommendation strategy from coarse-grained to fine-grained, better adapting to the diverse needs of different users and the ever-changing learning scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] Figure 1 A diagram showing the working principle of the AI-enabled data-driven art education resource recommendation engine described in the present invention;

[0059] Figure 2 Workflow diagram for hierarchical clustering of skill levels;

[0060] Figure 3 A schematic diagram of the resource feature extraction module using graph neural networks to perform topological modeling on resource metadata.

[0061] Figure 4 Workflow diagram for sampling knowledge graph subgraphs. DETAILED DESCRIPTION

[0062] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0063] See also Figure 1-4 The present invention provides a technical solution: an AI-enabled data-driven art education resource recommendation engine, the engine comprising:

[0064] Multi-source data collection module: This module collects a wide range of data through various channels, including online learning platforms, art communities, sensor devices, and third-party APIs. User behavior data collected includes click history, dwell time, and interaction frequency, reflecting the specific interactions users have with art education resources. Resource metadata, including resource type, difficulty tag, and creator information, describes the resource's attributes and characteristics. Contextual data, including timestamps, device type, and geographic location, provides more comprehensive background information for recommendations.

[0065] User profile construction module: Utilizes a deep clustering algorithm to extract multi-dimensional features from collected user behavior data. First, the temporal dependencies of the behavior data are captured to generate a preliminary interest representation. Then, through self-supervised comparative learning, skill levels are hierarchically clustered and a skill evaluation matrix is constructed. A course attention mechanism is then introduced in conjunction with a gated recurrent unit to model long-term learning patterns. Finally, the interest representation, skill evaluation matrix, and learning pattern vector are input into the feature fusion layer. A multi-head attention mechanism is used for cross-dimensional correlation analysis to generate a user interest vector, a skill level vector, and a learning pattern vector. These vectors are then dynamically weighted and fused through the attention mechanism to construct a multi-dimensional user profile.

[0066] Resource feature extraction module: Graph neural networks are used to perform topological modeling of resource metadata, defining heterogeneous node types of resources, such as courses, tools, artists, and styles, and constructing multi-type edge relationships between nodes, such as affiliation, collaboration, and style derivation. The association weights between nodes are calculated based on a graph attention network, and the feature information of neighboring nodes is aggregated through multi-layer graph convolution to generate a node-level embedding representation. A meta-path-guided random walk strategy is used to sample subgraphs of the knowledge graph, and negative sampling technology is used to optimize the discriminability of the graph embedding. Finally, the node embeddings and the original resource metadata are input into the feature concatenation layer, and high-order semantic features are extracted based on the capsule network to obtain resource features such as the content feature vector, style feature vector, and applicable scenario vector.

[0067] Privacy-preserving processing module: Leveraging differential privacy mechanisms, noise is injected into user profiles and resource features. Sensitivity analysis is first performed on the implicit vector of the user profile. Based on the Laplace mechanism, random noise that satisfies differential privacy constraints is added to the feature dimension. Distributed data training is also conducted in conjunction with a federated learning framework. Within this framework, an asynchronous parameter aggregation protocol is designed. Local model gradients are encrypted and transmitted using homomorphic encryption, with the central server receiving only encrypted gradient updates. Furthermore, a dynamic privacy budget allocation strategy is introduced to adjust the noise injection intensity based on the frequency of user data usage and resource relevance, achieving a balance between privacy protection and recommendation accuracy.

[0068] Dynamic Recommendation Generation Module: A multi-objective optimization strategy is constructed based on a reinforcement learning framework, and multi-dimensional user profiles, resource characteristics, and privacy protection constraints are input into the policy network. The state space is defined as a combination vector of the current user profile, candidate resource characteristics, and historical recommendation records, and the action space is a discretized operation of the resource recommendation order and presentation method. A multi-objective reward function is constructed, including user satisfaction rewards, resource diversity rewards, and privacy compliance rewards. A dual deep Q-network structure is adopted, combined with a delayed update strategy for the target network to reduce overestimation bias, and a prioritized experience replay mechanism is used to optimize training sample utilization. The initial distribution of the recommendation strategy is generated based on the policy gradient algorithm. The complexity of the action space is gradually increased through course learning, achieving policy optimization from coarse-grained to fine-grained. A sequence of candidate resources is generated through Monte Carlo tree search, and the diversity of the sequence is enhanced using a generative adversarial network, ultimately outputting a personalized recommendation list.

[0069] The present invention will be further described below with reference to specific examples using Examples 1 to 5:

[0070] Example 1:

[0071] In actual application scenarios, taking an online art education platform as an example, the specific implementation process of multi-dimensional feature extraction based on deep clustering algorithm and hierarchical clustering of skill levels based on self-supervised comparative learning in the user portrait construction module is explained in detail.

[0072] The platform holds a large amount of behavioral data on users learning art courses. The first step is to construct a temporal encoder for user behavior sequences. A bidirectional long short-term memory (Bi-LSTM) network is used, which can simultaneously learn the temporal information of both forward and reverse behavioral data. For example, the user's clicks on different painting courses, the duration of course video viewing, and the frequency of interaction with course content (such as commenting and submitting assignments) over a period of time are all chronologically fed into the Bi-LSTM network. Through its unique gating mechanism, the Bi-LSTM network effectively captures the temporal dependencies of behavioral data and generates a preliminary interest representation that reflects the user's interest in different art courses during that period.

[0073] Next, we perform hierarchical clustering of skill levels. We construct a sample of skill triplets, where the anchor resource is the course resource the user is currently studying or recently studied, the positive sample resources are other course resources with the same difficulty label as the anchor resource, and the negative sample resources are course resources with different difficulty labels. For example, if the anchor resource is an intermediate sketching course, then the positive sample resources could be other intermediate sketching courses, and the negative sample resources could be elementary or advanced sketching courses.

[0074] The model is optimized by contrasting the loss function. The formula of the contrasting loss function is:

[0075]

[0076] Among them, n is the number of samples, m is the preset boundary value used to control the distance difference between the anchor point and the positive sample and the negative sample, and d represents the distance metric function in the embedding space. are the representations of anchor resources, positive sample resources, and negative sample resources after being embedded by the model. The purpose of this function is to maximize the similarity between anchor points and positive samples in the embedding space and minimize the similarity between anchor points and negative samples.

[0077] During training, a curriculum learning strategy is used to dynamically adjust the difficulty of triplet sampling. Initially, negative samples with significant differences in difficulty are selected, allowing the model to quickly learn the distinct characteristics of resources of varying difficulty levels. As training progresses, negative samples of similar difficulty levels are selected to improve the model's ability to discern subtle differences in difficulty.

[0078] A skill assessment matrix is constructed based on course completion, practice accuracy, and creative complexity. For example, course completion can be measured by the percentage of course chapters completed by users, practice accuracy is the percentage of exercises completed correctly, and creative complexity can be scored based on the complexity of the user's submitted artwork in terms of composition, technique, and so on. Combining these metrics creates a multi-dimensional skill assessment matrix.

[0079] A course attention mechanism is introduced to dynamically adjust interest weights based on the intensity of user interactions across different resource types (e.g., painting, music, sculpture, etc.). For example, if a user frequently clicks on a painting course and stays there for a long time, indicating a high frequency of interaction with the course, then this interest weight will be given a higher weight when calculating the user's interest in the course. This is combined with a gated recurrent unit (GRU) to model long-term learning patterns. The GRU can remember users' long-term learning behavior patterns, such as whether they prefer systematic or fragmented learning.

[0080] Finally, the interest representation, skill evaluation matrix, and learning pattern vector are fed into the feature fusion layer, where a multi-head attention mechanism is used for cross-dimensional correlation analysis. This mechanism analyzes these vectors from different subspace perspectives, exploring potential connections between them and generating an implicit vector representation of the user profile, thereby constructing a more accurate multi-dimensional user profile.

[0081] Example 2:

[0082] This embodiment takes an art education resource library as an example to describe the specific implementation process of using graph neural networks for topological modeling and meta-path-guided random walk strategy for subgraph sampling in the display resource feature extraction module.

[0083] This resource library boasts a rich and diverse collection of resources. First, we define the heterogeneous node types of these resources. Course nodes represent different art courses, such as oil painting and piano lessons; tool nodes include brushes, paints, and musical instruments; artist nodes represent the artists who created these resources; and style and genre nodes encompass various art styles and genres, including Impressionism and Classicism. We construct multi-type edge relationships between these nodes. For example, a course and a tool have an affiliation relationship; an oil painting course might be affiliated with specific brushes, paints, and other tools; an artist might collaborate with a course, participating in the creation or teaching of a course; and a style and genre have a style-derived relationship with a course; a particular style and genre might spawn courses of a corresponding style.

[0084] The association weights between nodes are calculated based on the Graph Attention Network (GAT). Using the attention mechanism, GAT calculates the degree of attention each node pays to its neighboring nodes based on their feature information, thereby obtaining the association weights between nodes. Multi-layer graph convolutions aggregate the feature information of neighboring nodes. For example, for a given course node, after multi-layer graph convolutions, the features of its associated tool nodes, artist nodes, and style and genre nodes are aggregated to generate a node-level embedding representation that contains comprehensive feature information about the course in the resource network.

[0085] A meta-path-guided random walk strategy is used to sample subgraphs of the knowledge graph. A predefined meta-path template, such as "course-tool-artist", is used to generate a walk path using a specific guitar course as an example. Starting from the guitar course node, a walk path is formed along the tool node associated with it (such as guitar), and then to the artist node associated with the tool (such as a famous guitar player). High-frequency associated subgraphs are filtered through path importance scoring, and path importance is calculated based on node co-occurrence frequency and edge weight. For example, if the courses, tools, and artists on a certain path often appear together in the resource library, and the weight of the edge between them is high, then the subgraph corresponding to this path has a high importance.

[0086] A heterogeneous neighborhood sampler is used to expand subgraph nodes. For example, if a subgraph containing a guitar course, a specific guitar tool, and a guitarist has been sampled, the heterogeneous neighborhood sampler will select appropriate nodes from other nodes adjacent to these nodes for expansion, such as other artists who have collaborated with the guitarist or other related tools involved in the guitar course, to ensure that the sampled subgraph covers multi-hop relationships.

[0087] Optimize graph embedding models through subgraph-level contrastive learning. Within a subgraph, we maximize the consistency of embeddings across nodes within the same subgraph. For example, for a subgraph containing oil painting courses, oil paints, impressionist styles, and an impressionist painter, contrastive learning can make the representations of these nodes in the embedding space more similar, thereby better extracting resource features.

[0088] Finally, the node embeddings and the resource's original metadata are fed into the feature concatenation layer, where high-level semantic features are extracted using a capsule network. Capsule networks can better capture the spatial relationships and semantic information between different features, generating dynamic multimodal feature vectors for the resource, including content, style, and applicable scenario vectors. This provides richer and more accurate resource feature information for subsequent recommendations.

[0089] Example 3:

[0090] This embodiment takes a distributed art education data platform as an example to illustrate the specific implementation of noise injection through the differential privacy mechanism and the introduction of a dynamic privacy budget allocation strategy in the privacy protection processing module.

[0091] The platform stores a large amount of user learning data and feature data on art education resources. First, a sensitivity analysis is performed on the implicit vector of the user profile. This sensitivity analysis primarily measures the impact of each feature dimension in the user profile on privacy leakage. For example, the sensitivity of feature dimensions such as a user's age and learning preferences may vary. Based on the Laplace mechanism, random noise that satisfies differential privacy constraints is added to the feature dimension. The formula for the Laplace mechanism is: Where M(x) is the output after adding noise, f(x) is the original data, Lap represents the Laplace distribution, Δf is the sensitivity of function f, and ∈ is the privacy budget parameter, which controls the strength of privacy protection. The smaller the value of ∈, the higher the degree of privacy protection, but the data accuracy may be lower.

[0092] An asynchronous parameter aggregation protocol is designed within the federated learning framework. Each participant (e.g., educational institutions in different regions) trains the model locally and transmits the local model gradients encrypted using homomorphic encryption. Homomorphic encryption allows specific computations to be performed on encrypted data, while the central server only receives encrypted gradient updates. This ensures data security during transmission and prevents data leakage.

[0093] Introduce a dynamic privacy budget allocation strategy. Calculate the privacy consumption coefficient based on the update frequency of the user profile. Assuming the user profile update frequency is f, the privacy consumption coefficient c can be calculated by the formula Calculations show that users who frequently update their data are assigned a lower privacy budget, which can reduce the risk of long-term privacy leakage. For example, if a user frequently updates their learning records, a relatively low privacy budget will be allocated to protect their privacy.

[0094] Data sensitivity is assessed by the correlation between resource features and user profiles. The correlation is evaluated by calculating the cosine similarity between the resource feature vector and the user profile vector. The formula is: in is the resource feature vector, is the user profile vector, and sim represents the correlation. Stronger noise perturbations are applied to features with high correlation to protect sensitive data.

[0095] Design a privacy budget redistribution mechanism to reclaim unused privacy budget after the federated learning training cycle. Assume that the initial privacy budget of each training cycle is B and the used privacy budget is B. u sed, then the unused privacy budget B is recovered recycle =BB u These recovered budgets can be used for dynamic allocation in the next cycle, thereby improving the recommendation accuracy as much as possible while ensuring privacy protection.

[0096] Embodiment 4:

[0097] This embodiment takes an online art learning recommendation system as an example to introduce in detail the specific implementation methods of building a multi-objective optimization strategy based on the reinforcement learning framework in the dynamic recommendation generation module and gradually increasing the complexity of the action space through course learning.

[0098] In this recommendation system, the state space is defined as a combination vector of the current user profile, candidate resource features, and historical recommendation records. The user profile includes the user's interest vector, skill level vector, and learning mode vector; the candidate resource features include the content feature vector, style feature vector, and applicable scenario vector; and the historical recommendation record records the resources previously recommended to the user and the user's feedback. The action space is a discrete operation of the resource recommendation order and presentation method. For example, the resource recommendation order can be in ascending or descending order of course difficulty, or sorted by the degree of matching with the user's interests; the presentation method can be graphic and text display, video display, or a display method combined with interactive experience, etc.

[0099] Construct a multi-objective reward function. The user satisfaction reward is calculated by the click rate and feedback score. Assume that the click rate is ctr, the feedback score is score, and the user satisfaction reward is r user =α×ctr+β×score, where α and β are weight coefficients. Adjust their values according to the actual situation to balance the importance of click-through rate and feedback score in the reward. The resource diversity reward is measured by the entropy value of the candidate sequence. The entropy value formula is: Where n is the number of candidate resources, p i is the probability of the i-th resource in the candidate sequence. The larger the entropy value, the higher the resource diversity. The privacy compliance reward is evaluated by the degree of data anonymization. For example, the privacy compliance reward r can be determined based on the intensity of the noise added to the differential privacy mechanism and the effect of data encryption in federated learning. privacy .

[0100] A dual deep Q-network (DDQN) architecture is employed, with a delayed target network update strategy to mitigate overestimation bias. In DDQN, an online network is used to select actions, and a target network is used to calculate target Q values. The target network is updated regularly, using a delayed update strategy (for example, once every N steps). This avoids overestimation caused by frequent parameter changes in the online network. A prioritized experience replay mechanism is also incorporated to optimize training sample utilization. This mechanism samples samples based on their importance (such as the size of the reward). Highly important samples are more likely to be sampled, thereby improving training efficiency.

[0101] The initial distribution of recommendation strategies is generated using a policy gradient algorithm. The complexity of the action space is gradually increased through course learning. Initially, the recommended actions are limited to filtering resources across broad categories, and the action space dimension is the number of pre-set resource types. For example, art resources are categorized into painting, music, dance, and other broad categories, and the recommendation system selects recommendations only within these broad categories.

[0102] In the mid-term, fine-grained actions will be introduced, including resource sorting strategies, display style selection, and interactive guidance. For example, when recommending painting resources, you can choose to sort by artist popularity, select different image and text display styles, and provide interactive actions to guide users in painting practice.

[0103] In the later stages, it is expanded to multimodal action combinations, allowing for simultaneous adjustment of the recommendation order, presentation method, and auxiliary tool suggestions. For example, when recommending music courses, you can simultaneously adjust the course recommendation order, select the video presentation method, and recommend suitable instruments to assist in learning. Through a progressive action space expansion strategy, combined with the teacher-student network architecture, the historical strategy is used to generate auxiliary supervision signals to accelerate model convergence. The teacher network can be a mature model that has been trained with a large amount of data. During the learning process, the student network refers to the output of the teacher network to generate auxiliary supervision signals, thereby optimizing its own recommendation strategy more quickly.

[0104] Example 5:

[0105] This embodiment takes a comprehensive art education platform as an example to explain in detail the specific process of collecting data from various channels in the multi-source data collection module and the practical application of the collected data.

[0106] As a core data collection source, online learning platforms possess a wealth of user behavior data. The platform's back-end system records every user action in detail. For example, when a user clicks on a painting course's details page, the system accurately records the click time, the unique identifier of the course clicked, and other information, creating a click log. This log not only reflects a user's immediate interest in different courses but also tracks their path through art education resources. Regarding dwell time, the system uses time monitoring points, starting when a user enters a course video playback page and ending when they leave the page, to accurately capture the length of time a user spends on that page. This data provides a visual indicator of a user's engagement and engagement with the course content. If a user spends a long time on a course video page, it likely indicates that the course content aligns with their interests or holds significant learning value. Interaction frequency encompasses various forms of interaction, such as user comments in the course comment section, submissions, and peer review of assignments. By measuring the number and content of these interactions, we can gain a deeper understanding of user engagement and feedback, as well as uncover the areas of thought and questions users harbor during their art learning journey.

[0107] Art communities are another indispensable channel for data collection. In this vibrant community, users actively share their artistic creations and exchange learning experiences. The community's data analysis system parses every piece of content posted by users, extracting keywords and key sentences. For example, when a user shares a watercolor painting they created, they will include a text description of their creative ideas, the techniques used, and their understanding of color matching. By analyzing this text content through natural language processing technology, it is possible to identify the user's interest in watercolor as an art form, the painting techniques they have mastered, and their unique insights into the use of color. At the same time, users' likes, comments, and following behaviors in the community also contain a wealth of information. If a user frequently likes and comments on the works of a well-known painter, or follows multiple creators of the same style or genre, it can be inferred that the user has a strong interest in this style or genre.

[0108] Sensor devices play a unique role in data collection in the field of art education. Professional sensor equipment is used in some offline art practice courses, such as dance and sculpture creation courses. Taking dance courses as an example, motion capture sensors can accurately collect data on students' body movements during the dance process, including the movement trajectory of joints, the amplitude and speed of movements, etc. After analysis, this data can be generated into a detailed movement evaluation report to determine whether the student's dance movements are standardized and smooth, and whether the mastery of dance techniques has met the course requirements. In sculpture creation courses, sensors can monitor the strength and movement trajectory of students' hand movements while they are shaping their works. This data helps teachers understand students' creative habits and skill levels, providing a basis for personalized teaching guidance.

[0109] Third-party APIs provide access to a wider range of data resources for art education platforms. By integrating with APIs from renowned art databases, the platform can access rich metadata. For example, for a work of art, the platform can obtain detailed information about its creator, including biography, artistic style, and representative works. It can also obtain information about the style and genre of the work, such as Impressionism, Post-Impressionism, and Cubism, as well as key information such as the year of creation. This metadata is crucial for enriching the platform's art education resource library and improving the accuracy of resource descriptions. Furthermore, by integrating with APIs from external event platforms, the platform can obtain contextual data on various art events, such as the time, location, and exhibited works of art exhibitions, as well as the time, location, and lineup of art performances. This data can be combined with users' location and time information to recommend art events tailored to their specific circumstances, such as recommending recent art exhibitions in a specific city to users, thereby enhancing their art learning experience and engagement.

[0110] This collected data is integrated into the entire recommendation engine. User behavior data is used to build user profiles and analyze user interests and learning patterns; resource metadata helps extract resource features and better describe resource attributes; and contextual data takes user usage scenarios into account during the recommendation process. For example, it can recommend nearby art exhibitions based on the user's location, or art courses suitable for the current season or festival based on timestamps. This allows for more accurate art education resource recommendations, meeting users' diverse learning needs and improving their learning experience and effectiveness on the art education platform.

[0111] It should be noted that, in this document, relational terms such as first and second, etc., are used only 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 terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0112] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An AI-enabled data-driven art education resource recommendation engine, characterized by: include: Multi-source data collection module, used to obtain user behavior data, resource metadata and contextual environment data through multiple channels; The user profile construction module extracts multi-dimensional features from user behavior data based on a deep clustering algorithm, generates user interest vectors, skill level vectors, and learning mode vectors, and dynamically weights and fuses these vectors through an attention mechanism to construct a multi-dimensional user profile. The resource feature extraction module uses a graph neural network to perform topological modeling on resource metadata, constructs the relationship between resources into a heterogeneous knowledge graph, and extracts resource features based on graph embedding technology. The resource features include content feature vectors, style feature vectors, and applicable scenario vectors. The privacy protection processing module injects noise into user profiles and resource features through a differential privacy mechanism, and combines it with a federated learning framework for distributed data training to achieve local storage and anonymization of user data. The dynamic recommendation generation module builds a multi-objective optimization strategy based on the reinforcement learning framework, inputs multi-dimensional user portraits, resource characteristics and privacy protection constraints into the strategy network, generates a candidate resource sequence through Monte Carlo tree search, and enhances the diversity of the sequence based on the adversarial generative network to output a personalized recommendation list.

2. The AI-enabled data-driven art education resource recommendation engine according to claim 1, characterized in that: The user portrait construction module extracts multi-dimensional features from user behavior data based on a deep clustering algorithm, including: Construct a temporal encoder for user behavior sequences, using a bidirectional long short-term memory network to capture the temporal dependencies of behavior data and generate preliminary interest representations. Skill levels are hierarchically clustered through self-supervised contrastive learning, and a skill evaluation matrix is constructed based on course completion, practice accuracy, and creation complexity. The course attention mechanism is introduced to dynamically adjust the interest weight based on the user's interaction intensity on different resource types, and combined with the gated recurrent unit to model long-term learning patterns; The interest representation, skill evaluation matrix and learning pattern vector are input into the feature fusion layer, and a multi-head attention mechanism is used to perform cross-dimensional correlation analysis to generate an implicit vector representation of the user portrait.

3. The AI-enabled data-driven art education resource recommendation engine according to claim 2, characterized in that: The self-supervised contrastive learning method performs hierarchical clustering of skill levels, including: Construct skill triplet samples, including anchor resources, positive sample resources, and negative sample resources. The positive samples and anchor resources have the same difficulty labels, while the negative samples have different difficulty labels. The contrast loss function is used to maximize the similarity between the anchor point and the positive sample in the embedding space, and minimize the similarity between the anchor point and the negative sample; A curriculum learning strategy is used to dynamically adjust the difficulty of triple sampling. In the early stage, negative samples with significant differences in difficulty are selected, and in the later stage, negative samples with similar difficulty are selected to improve the model's discriminative ability. The learned embedding representation is input into the hierarchical clustering algorithm to automatically divide the skill levels according to the density distribution of the skill evaluation matrix.

4. The AI-enabled data-driven art education resource recommendation engine according to claim 1, characterized in that: The resource feature extraction module uses a graph neural network to perform topological modeling on resource metadata, including: Define heterogeneous node types for resources, including courses, tools, artists, and styles, and build multi-type edge relationships between nodes, including affiliation, collaboration, and style derivative relationships; Based on the graph attention network, the association weights between nodes are calculated, and the feature information of neighbor nodes is aggregated through multi-layer graph convolution to generate node-level embedding representation; A meta-path-guided random walk strategy is used to sample subgraphs of the knowledge graph, and negative sampling techniques are used to optimize the discriminability of graph embeddings. The node embedding and the resource original metadata are input into the feature splicing layer, and high-order semantic features are extracted based on the capsule network to generate a dynamic multimodal feature vector of the resource.

5. The AI-enabled data-driven art education resource recommendation engine according to claim 4, characterized in that: The meta-path guided random walk strategy performs subgraph sampling on the knowledge graph, including: Predefined meta-path templates, including "Course-Tool-Artist", "Genre-Derivative Work-User", and "Resource-Related Tags-Scenario"; Generate a walk path based on the meta-path template and filter high-frequency associated subgraphs using path importance scores, where path importance is calculated based on node co-occurrence frequency and edge weights. A heterogeneous neighborhood sampler is used to expand the nodes of the subgraph to ensure that the sampled subgraph covers multi-hop relationships; The graph embedding model is optimized through subgraph-level contrastive learning to maximize the embedding consistency of nodes in the same subgraph.

6. The AI-enabled data-driven art education resource recommendation engine according to claim 1, characterized in that: The privacy protection processing module injects noise into user profiles and resource features through a differential privacy mechanism, including: Perform sensitivity analysis on the implicit vector of the user profile and add random noise to the feature dimension based on the Laplace mechanism to satisfy differential privacy constraints. An asynchronous parameter aggregation protocol is designed within the federated learning framework. Local model gradients are encrypted and transmitted using homomorphic encryption, and the central server only receives encrypted gradient updates. A dynamic privacy budget allocation strategy is introduced to adjust the noise injection intensity according to the usage frequency and resource relevance of user data, achieving a balance between privacy protection and recommendation accuracy.

7. The AI-enabled data-driven art education resource recommendation engine according to claim 6, characterized in that: The dynamic privacy budget allocation strategy introduced here adjusts the noise injection intensity based on the frequency of user data usage and resource relevance, including: The privacy consumption coefficient is calculated based on the update frequency of the user profile. Users with high-frequency updates are assigned a lower privacy budget to reduce the risk of long-term privacy leakage. Data sensitivity is assessed by the correlation between resource features and user profiles, and stronger noise perturbations are applied to highly correlated features. A privacy budget redistribution mechanism is designed to reclaim unused privacy budget after the federated learning training cycle ends and use it for dynamic allocation in the next cycle.

8. The AI-enabled data-driven art education resource recommendation engine according to claim 1, characterized in that: The dynamic recommendation generation module builds a multi-objective optimization strategy based on the reinforcement learning framework, including: The state space is defined as the combination vector of the current user profile, candidate resource features, and historical recommendation records, and the action space is the discretized operation of resource recommendation order and presentation method; Construct a multi-objective reward function, including user satisfaction reward, resource diversity reward, and privacy compliance reward. The user satisfaction reward is calculated based on click-through rate and feedback score, the resource diversity reward is measured by the entropy value of the candidate sequence, and the privacy compliance reward is evaluated by the degree of data anonymization. Adopting a dual deep Q network structure, the target network delayed update strategy is used to reduce overestimation bias, and the priority experience replay mechanism is combined to optimize the utilization of training samples. Based on the policy gradient algorithm, the initial distribution of the recommendation strategy is generated, and the complexity of the action space is gradually increased through course learning to achieve strategy optimization from coarse-grained to fine-grained.

9. The AI-enabled data-driven art education resource recommendation engine according to claim 8, characterized in that: The curriculum learns to gradually increase the complexity of the action space including: In the initial stage, the recommended action is limited to screening of large categories of resources, and the dimension of the action space is the number of preset resource types; In the mid-term phase, fine-grained actions are introduced, including resource sorting strategies, display style selection, and interactive guidance operations; In the later stages, it will be expanded to multimodal action combinations, allowing for simultaneous adjustment of the order of recommendations, presentation methods, and auxiliary tool suggestions; Through a progressive action space expansion strategy, combined with a teacher-student network architecture, historical strategies are used to generate auxiliary supervision signals to accelerate model convergence.

10. The AI-enabled data-driven art education resource recommendation engine according to claim 1, characterized in that: The channels include online learning platforms, art communities, sensor devices and third-party API interfaces; the user behavior data include click records, stay duration and interaction frequency; the resource metadata include resource type, difficulty label and creator information; and the contextual environment data includes timestamp, device type and geographic location.

Images