Personalized training resource recommendation method and system based on graph fusion neural network

Abstract

The invention discloses a personalized training resource recommendation method based on a graph fusion neural network, and the method comprises the steps: obtaining text data information, capturing the global and local information of a text through a convolutional neural network, adjusting a mapping function according to a constraint condition, and obtaining a multi-modal data model; optimizing the performance of the multi-modal data model through cross compression, improving the multi-modal data model by using a loss function, extracting data information of the multi-modal data model, and obtaining a training resource knowledge graph through data fusion; the method comprises the following steps: acquiring user online behavior data, performing modeling and group clustering on user online behaviors, and performing deep knowledge tracking by utilizing CRU to obtain a learner knowledge evaluation and deep learning model; through a graph fusion neural network recommendation algorithm, the training resource knowledge graph is matched with personalized demands of the user, personalized resource recommendation is obtained, richer semantic information is provided, and the recommendation algorithm can more accurately understand the demands and interests of the user.

Description

Technical Field

[0001] This invention relates to the field of personalized training resource recommendation technology, specifically to a graph fusion neural network-based personalized training resource recommendation method and system. Background Technology

[0002] In the current field of personalized training resource recommendation, various existing technologies and algorithms exist; however, they generally face a series of limitations and challenges. Traditional collaborative filtering algorithms are widely used in recommendation systems, but they rely on users' historical behavioral data and perform poorly for new users or cold-start problems. Furthermore, collaborative filtering algorithms struggle to handle multimodal data and cannot effectively capture the relationships between different data types such as text and images, resulting in a lack of comprehensiveness in recommendation results. On the other hand, content-based recommendation methods make recommendations by analyzing the attributes of training resources and learners' interest characteristics, but they are insufficient in deeply understanding user behavior. Meanwhile, traditional knowledge tracing methods such as BKT and DKT are overly simplistic in capturing learners' long-term dependencies, failing to accurately reflect learners' knowledge status. Behavioral analysis methods have limited effectiveness in mining learners' behavioral characteristics and lack an interpretation of the deeper reasons behind these behaviors.

[0003] However, this invention overcomes the limitations of traditional algorithms by innovatively integrating deep learning and knowledge graph technologies. By creatively introducing a Text CNN model, this invention can efficiently extract multimodal features, addressing the shortcomings of traditional algorithms in handling multimodal data. Employing a unique GRU for deep knowledge tracing, this invention can more accurately simulate learners' knowledge states, compensating for the deficiencies of traditional knowledge tracing methods. Combining behavioral science theory, this invention quantifies learners' behavioral characteristics, deeply exploring their learning habits and preferences, providing more targeted guidance for recommendation algorithms. Through a graph fusion neural network recommendation algorithm, this invention can more comprehensively understand the relationships between training resources, achieving more accurate matching of users' personalized needs, while also exhibiting better adaptability to the cold start problem. Therefore, this invention optimizes and improves in many aspects, overcoming the shortcomings of existing technologies and bringing significant technological progress to the field of personalized training resource recommendation. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by this invention is that traditional recommendation algorithms have limitations when processing multimodal data, making it difficult to fully explore the correlation between text and image information, affecting the accurate understanding of user interests, resulting in a lack of comprehensiveness in recommendation results, and failing to truly meet the diverse learning needs of users. Existing knowledge tracking methods are too simplistic in capturing learners' knowledge states and cannot accurately reflect the long-term dependencies of learners in the learning process, thus limiting the accuracy of personalized learning paths. Traditional behavior analysis methods do not have a deep enough understanding of learners' behavioral characteristics and cannot provide accurate personalized recommendation guidance. Existing recommendation systems do not handle the cold start problem of new users or new training resources well, making it difficult to provide accurate recommendation services when historical data is lacking.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a method for recommending personalized training resources using a graph fusion neural network, comprising:

[0008] By acquiring text data information, using convolutional neural networks to capture global and local information of the text, and adjusting the mapping function according to constraints, a multimodal data model is obtained.

[0009] The performance of the multimodal data model is optimized by cross-compression, and the data information of the multimodal data model is extracted by using a loss function. The training resource knowledge graph is obtained through data fusion.

[0010] Acquire user online behavior data, model and cluster the user online behavior, use CRU for deep knowledge tracing, and obtain learner knowledge evaluation and deep learning models;

[0011] By using a graph fusion neural network recommendation algorithm, the training resource knowledge graph is matched with the user's personalized needs to obtain personalized resource recommendations.

[0012] As a preferred embodiment of the graph fusion neural network personalized training resource recommendation method of the present invention, wherein: the mapping function is adjusted according to the constraints to obtain a multimodal data model, including, the multimodal data model, expressed as:

[0013] CLos(T1,I1,T2,I2)=max(0,Δ(T1,I1,I2)+a)+max(0,Δ(I1,T1,T2)+a)

[0014] Where T1 and T2 are the first and second inputs of the text feature vector, I1 and I2 are the first and second inputs of the image feature vector, Δ(T1,I1,I2) is used to calculate the difference between the first input of the text feature vector and the first and second inputs of the image feature vector, Δ(I1,T1,T2) represents the difference between the first input of the image feature vector and the first and second inputs of the text feature vector, and a is a constant;

[0015] The overall characteristics of the project are represented as follows:

[0016] V1 = w T ·T1+w I ·I1

[0017] Among them, w T and w I V1 represents the weights of text features and image features.

[0018] As a preferred embodiment of the graph fusion neural network personalized training resource recommendation method of the present invention, the multimodal data model is optimized for performance through cross-compression, including:

[0019] By fusing multimodal features, a feature vector is obtained. This feature vector is then subjected to cross-compression to produce the output, which is represented as follows:

[0020] V out？ =Vh T w VV +hV T w HV +b V

[0021] h out？ =Vh T w VH +hV T w HH +b H

[0022] Where w and b are the weights and bias parameters of the cross-compression unit, V is the feature vector of the item, and h is the feature vector of the entity. T h T H is the transpose of the eigenvectors, where H is the transpose of the entity's eigenvectors.

[0023] As a preferred embodiment of the graph fusion neural network personalized training resource recommendation method of the present invention, the method includes: improving the multimodal data model using a loss function to extract data information from the multimodal data model, including:

[0024] Using a comprehensive loss function, the multimodal data model is trained and optimized, as follows:

[0025]

[0026] Where λ1 and λ2 are balance factors, To constrain the square norm of the model parameter W, L KGE To embed loss terms into knowledge graphs, To reconstruct the loss term.

[0027] As a preferred embodiment of the graph fusion neural network personalized training resource recommendation method described in this invention, the method utilizes CRU for deep knowledge tracing to obtain learner knowledge evaluation and a deep learning model, including:

[0028] Learner knowledge assessment and deep learning models are represented as follows:

[0029]

[0030]

[0031] Where σ is the sigmoid function, tanh is the hyperbolic tangent function, ⊙ is element-wish multiplication, and W r U r b r W h U h b h W z U z b z For model parameters, r t To reset the door, For the candidate hidden state, z t To update the gate, h t-1 For the past state, h t This represents the current hidden state.

[0032] As a preferred embodiment of the graph fusion neural network-based personalized training resource recommendation method of the present invention, wherein: the training resource knowledge graph is matched with the user's personalized needs through a graph fusion neural network recommendation algorithm to obtain personalized resource recommendations, including:

[0033] The training resource knowledge graph is combined with training resource entities, where each node represents an entity. When an entity has neighboring entities, a method is used to limit the size of each entity's domain, represented as follows:

[0034]

[0035] Where S(v) is the set of all entities e that are adjacent to entity v, and N(v) is the entity that is adjacent to entity v.

[0036] As a preferred embodiment of the graph fusion neural network personalized training resource recommendation method of the present invention, the graph fusion neural network recommendation algorithm further includes,

[0037] The graph fusion neural network recommendation algorithm is optimized using a negative sampling strategy, as follows:

[0038]

[0039] Where λ is the equilibrium parameter, For the contribution of L2 regularizer, y uv For the relationship between entities u and v, Let p(v) be the predicted value for entities u and v. i Let be the probability distribution of entity v.

[0040] Secondly, the present invention provides a system for personalized training resource recommendation using graph fusion neural networks, comprising:

[0041] The multimodal feature extraction module is used to acquire text data information, capture global and local information of the text using a convolutional neural network, and adjust the mapping function according to constraints to obtain a multimodal data model.

[0042] The cross-compression module is used to optimize the performance of the multimodal data model through cross-compression, improve the extraction of data information from the multimodal data model using a loss function, and obtain a training resource knowledge graph through data fusion.

[0043] The user profiling module is used to acquire user online behavior data, model the user online behavior and cluster the groups, use CRU to perform deep knowledge tracking, and obtain learner knowledge evaluation and deep learning models.

[0044] The algorithm optimization module is used to match the training resource knowledge graph with the user's personalized needs through a graph fusion neural network recommendation algorithm to obtain personalized resource recommendations.

[0045] Thirdly, the present invention provides a computing device, comprising:

[0046] Memory and processor;

[0047] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the graph fusion neural network personalized training resource recommendation method.

[0048] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the graph fusion neural network personalized training resource recommendation method.

[0049] The beneficial effects of this invention are as follows: The graph fusion neural network personalized training resource recommendation method and system proposed in this invention not only improves computational efficiency, but also enables the model to simultaneously focus on local details and global features of the text through the combination of convolutional layers and pooling layers, so as to more accurately represent training resources, provide richer semantic information, and improve the system's ability to understand the correlation of training resources. By incorporating the learner's response sequence in the question-answering process into deep knowledge tracking, this method further improves the accuracy of evaluation, enabling the model to more comprehensively understand the learner's knowledge state. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0051] Figure 1 The flowchart of a personalized training resource recommendation method using a graph fusion neural network provided by the present invention is shown. Detailed Implementation

[0052] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0053] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0054] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0055] This invention is described in detail with reference to the schematic diagrams. When detailing the embodiments of this invention, for ease of explanation, the cross-sectional views illustrating the device structure may be partially enlarged, not adhering to the usual scale. Furthermore, the schematic diagrams are merely examples and should not be construed as limiting the scope of protection of this invention. In actual fabrication, the three-dimensional spatial dimensions of length, width, and depth should be included.

[0056] Furthermore, in the description of this invention, it should be noted that the terms "upper," "lower," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. These terms are used solely for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. In addition, the terms "first," "second," or "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0057] Unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" in this invention should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; similarly, they can refer to mechanical connections, electrical connections, or direct connections, or indirect connections through an intermediate medium, or internal connections between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0058] Example 1

[0059] Reference Figure 1 As an embodiment of the present invention, a method for recommending personalized training resources using a graph fusion neural network is provided, comprising:

[0060] S1: Obtain text data information, use convolutional neural networks to capture global and local information of the text, adjust the mapping function according to constraints, and obtain a multimodal data model;

[0061] Furthermore, the multimodal data model is represented as:

[0062] CLos(T1,I1,T2,I2)=max(0,Δ(T1,I1,I2)+a)+max(0,Δ(I1,T1,T2)+a)

[0063] Where T1 and T2 are the first and second inputs of the text feature vector, I1 and I2 are the first and second inputs of the image feature vector, Δ(T1,I1,I2) is used to calculate the difference between the first input of the text feature vector and the first and second inputs of the image feature vector, Δ(I1,T1,T2) represents the difference between the first input of the image feature vector and the first and second inputs of the text feature vector, and a is a constant;

[0064] The overall characteristics of the project are represented as follows:

[0065] V1 = w T ·T1+w I ·I1

[0066] Among them, w T and w I V1 represents the weights of text features and image features.

[0067] It should be noted that the Text CNN model was chosen for efficient feature extraction based on the excellent feature extraction capabilities of Convolutional Neural Networks (CNNs), which can quickly and accurately capture key information from text. Through the meticulous processing of convolutional layers, the model can focus on local details of the text, while pooling layers are responsible for extracting globally important features. This unique design structure not only reduces the number of parameters that need to be updated in each iteration, thereby improving computational power, but also offers superior processing speed compared to other traditional methods.

[0068] In a multimodal data model, max(0,a) ensures that the output value will not exceed 0, thus acting as a constraint. Adjusting the parameter 'a' affects the constraint strength of the loss function, thereby making different items more distinct in the feature space. As the value of 'a' gradually increases, the constraint effect also strengthens, which helps the model to more accurately identify and distinguish different items.

[0069] After multimodal feature extraction and fusion, the text feature vector T1 and image feature vector I1 of item V1 were successfully extracted. To obtain a comprehensive feature representation of the item, this invention employs an averaging method, adding T1 and I1 and taking half of the sum to obtain the feature vector of item V1. This approach ensures that the feature vector of the item incorporates both textual and image information, thus providing a more comprehensive description of the item, where w... T +w I =1.

[0070] It should also be noted that, in order to ensure that the mapped feature vectors have ideal distribution characteristics in the common feature space, two important constraints are introduced: First, for image and text information describing the same training resource, they are required to be as close as possible after being mapped to the common feature space. This constraint ensures that different modal information of the same training resource has consistency and similarity in the feature space, thereby enhancing the complementarity and credibility of the information.

[0071] Secondly, for image and text information describing different training resources, it is required that they be as far apart as possible after being mapped to a common feature space. This constraint helps to clearly distinguish the multimodal information of different training resources in the feature space, avoiding information confusion and overlap.

[0072] To accurately evaluate the performance of the multimodal feature fusion module, a triplet loss function is used to define the loss function. This loss function can flexibly adjust the mapping function according to the set constraints, ensuring that the mapped feature vectors meet the aforementioned constraints. By continuously optimizing this loss function, this invention can gradually improve the performance of the multimodal feature fusion module, ultimately resulting in an excellent model capable of efficiently processing multimodal data containing text and image information.

[0073] S2: Optimize the performance of the multimodal data model through cross-compression, improve the extraction of data information from the multimodal data model using the loss function, and obtain the training resource knowledge graph through data fusion;

[0074] Furthermore, the performance of multimodal data models is optimized through cross-compression, including...

[0075] By fusing multimodal features, a feature vector is obtained. This feature vector is then subjected to cross-compression to produce the output, which is represented as follows:

[0076] V out？ =Vh T w VV +hV T w HV +b V

[0077] h out？ =Vh T w VH +hV T w HH +b H

[0078] Where w and b are the weights and bias parameters of the cross-compression unit, V is the feature vector of the item, and h is the feature vector of the entity. T h T H is the transpose of the eigenvectors, where H is the transpose of the entity's eigenvectors.

[0079] It should be noted that, in order to explore the relationship between project V and its related entities h in the knowledge graph more deeply, a cross matrix C is constructed, where C∈R d×d The dimension is represented as:

[0080] C = V·h T

[0081] In the output expression, the weights and bias parameters of the cross-compression unit are adjusted. By adjusting these parameters, the model can learn and perform recommendation tasks and knowledge graph tasks simultaneously, enabling the model to achieve better performance when handling these two tasks. This allows the model to be more adaptable to different task requirements and achieve better learning results.

[0082] Furthermore, by utilizing a comprehensive loss function, the multimodal data model is trained and optimized, as follows:

[0083]

[0084] Where λ1 and λ2 are balance factors, To constrain the square norm of the model parameter W, L KGE To embed loss terms into knowledge graphs, To reconstruct the loss term.

[0085] It should be noted that balancing factors adjust and weigh the relative importance of different tasks or loss functions. Adjusting the values ​​of these parameters allows us to control how the model handles multiple tasks, ensuring ideal performance on each task. These balancing factors help improve the model training process, thereby achieving optimal balance and trade-offs across multiple tasks.

[0086] S3: Acquire user online behavior data, model user online behavior and cluster groups, use CRU for deep knowledge tracking, and obtain learner knowledge evaluation and deep learning models;

[0087] Furthermore, the learner knowledge assessment and deep learning model can be represented as follows:

[0088]

[0089]

[0090] Where σ is the sigmoid function, tanh is the hyperbolic tangent function, ⊙ is element-wish multiplication, and W r U r b r W h U h b h W z U z b z For model parameters, r t To reset the door, For the candidate hidden state, z t To update the gate, h t-1 For the past state, h t This represents the current hidden state.

[0091] It should be noted that the construction of a training resource knowledge graph includes entity extraction, attribute extraction, relation extraction, and triple extraction. Entity extraction is a key step, aiming to accurately extract entity information related to training resources from text data. Attribute extraction is the process of identifying and extracting attribute information related to entities from unstructured data. Relation extraction is the core link in knowledge graph construction, aiming to identify and extract semantic relationships between entities from unstructured data.

[0092] In a knowledge graph G, its core structure is called a triple, which consists of two entities and the relationship between them, represented as:

[0093]

[0094] Among them, T R Let R be the set of training resource entities, R be the set of relations in the knowledge system, h,t be an entity selected from the set of training resource entities, and r be a relation selected from the set of relations.

[0095] It should also be noted that GRU optimizes the computation of hidden states and reduces the risk of overfitting through a unique gating mechanism. Integrating the forget gate and input gate into a single update gate simplifies the model structure and improves computational efficiency. Simultaneously, the introduction of a reset gate and transient states achieves a deep fusion of cell states and hidden states. Weighting factors and error terms play a crucial role in model computation, helping to adjust the importance of information and fit the actual data distribution. This carefully designed gating mechanism and parameter tuning enable GRU to more effectively handle long-term dependencies in sequence data, improving model performance.

[0096] Deep Knowledge Tracking (DKT) based on GRU innovatively displays the learner's knowledge state through the hidden layer outputs of the network, significantly improving the accuracy of evaluation and effectively reducing the cost of training and validation for traditional RNN networks. By incorporating the learner's response sequence during the question-answering process into the DKT input, it can more comprehensively assess the learner's mastery of prior knowledge. This improved DKT has higher accuracy and can more meticulously track the learner's knowledge evolution, thus providing more targeted guidance for personalized learning.

[0097] This invention, based on online learning behavior analysis technology, designs a learning behavior feature labeling system. Using an index scoring method, corresponding labels are designed for each learning behavior, and a model is built using actual online learning behavior data. To accurately identify users with similar learning behaviors, a fuzzy C-means (FCM) clustering algorithm is used to group users with similar behaviors into the same group. Building upon this, a graph fusion neural network recommendation algorithm is further employed to recommend highly personalized learning paths based on user group characteristics and clustering results. This method aims to improve learning effectiveness, enhance personalized learning experiences, and provide users with truly tailored learning solutions, thereby more effectively promoting the improvement of knowledge and skills.

[0098] S4: By using a graph fusion neural network recommendation algorithm, the training resource knowledge graph is matched with the user's personalized needs to obtain personalized resource recommendations;

[0099] Furthermore, the training resource knowledge graph is combined with training resource entities, where each node represents an entity. When an entity has neighboring entities, a method is used to limit the domain size of each entity, represented as follows:

[0100]

[0101] Where S(v) is the set of all entities e that are adjacent to entity v, and N(v) is the entity that is adjacent to entity v.

[0102] It should be noted that many recommendation algorithms suffer from various problems. For example, item-based and content-based recommendations only consider single attributes of users or items, neglecting the complex relationships between them. Hybrid recommendations, while combining multiple recommendation strategies, may face issues such as data sparsity and cold start when integrating information from different sources. The LightGCN model, while utilizing high-order neighborhood information, does not leverage additional auxiliary information between items. This invention optimizes and proposes a graph fusion neural network recommendation algorithm that can represent and store a large number of entities and their relationships, and can learn complex patterns of these relationships, providing richer semantic information. This allows the recommendation algorithm to more accurately understand user needs and interests, as well as the associations and attributes between training resources. Compared to traditional algorithms that only consider first-order relationships—those directly connected to users or items—the graph fusion neural network recommendation algorithm can capture broader and deeper associations, including high-order neighborhood information. By capturing the different impacts of different training resource relationships on specific users, it can provide recommendations that better suit their needs based on user preferences, historical behavior, and similarity to other users or items. Similarly, because additional auxiliary information, such as entity attributes and relationship types, can be utilized, the impact of data sparsity is less affected when processing new users or new training resources, thereby alleviating the cold start problem in recommendation algorithms.

[0103] It should also be noted that by aggregating and merging information from neighboring entities, a prediction function can be learned. This function can be learned by aggregating and merging neighboring entities based on the characteristics of users and training resources, and is expressed as:

[0104]

[0105] in, Let Θ be the probability that user u will interact with training resource v, Θ be the model parameters of function F, and Y be the user interaction matrix, where Y∈R. M*N ,if A value of 1 indicates that user u interacts with training resource v, while a value of 0 indicates the opposite.

[0106] Graph fusion neural networks can more comprehensively consider the relationships between training resources and their neighboring entities, including relationship types and their impact on specific dimensions, thereby providing users with more accurate and personalized recommendations, represented as:

[0107]

[0108] Where u and r are the feature vectors of the user and training resources, respectively. T Let u be the transpose of u, and · be the dot product of the vectors, u∈R d ,r∈R d ;

[0109] Get the user's interest in the relationship The domain linear combination of training resources can then be represented as:

[0110]

[0111] Where e is the vector representation of entity e, It is normalization The results can be used to distinguish the importance of each neighbor;

[0112]

[0113] Furthermore, a negative sampling strategy is adopted to optimize the graph fusion neural network recommendation algorithm, as follows:

[0114]

[0115] Where λ is the equilibrium parameter, For the contribution of L2 regularizer, y uv For the relationship between entities u and v, Let p(v) be the predicted value for entities u and v. i Let be the probability distribution of entity v.

[0116] It should be noted that, in order to merge the representation of entity u with the representation of its neighborhood into a unified representation, the two vectors are first added, and then a linear transformation is performed. Through this process, a single vector that comprehensively reflects the information of entity v and its neighborhood is obtained, represented as:

[0117]

[0118] Where W is the transformation weight, b is the bias, and σ is a nonlinear function.

[0119] To improve computational efficiency, a negative sampling strategy was adopted, optimizing the original loss function. The new loss function comprehensively considers the interaction between positive and negative samples. By comparing the actual interactions between users and training resources with the non-interactive samples obtained through negative sampling, the model can be trained more effectively, improving learning performance.

[0120] The above is an illustrative scheme of a graph fusion neural network personalized training resource recommendation method according to this embodiment. It should be noted that the technical solution of this graph fusion neural network personalized training resource recommendation device belongs to the same concept as the above-described graph fusion neural network personalized training resource recommendation method. Details not described in detail in the technical solution of the graph fusion neural network personalized training resource recommendation device in this embodiment can be found in the description of the above-described graph fusion neural network personalized training resource recommendation method.

[0121] The apparatus for personalized training resource recommendation using a graph fusion neural network in this embodiment includes:

[0122] The multimodal feature extraction module is used to acquire text data information, capture global and local information of the text using a convolutional neural network, and adjust the mapping function according to constraints to obtain a multimodal data model.

[0123] The cross-compression module is used to optimize the performance of the multimodal data model through cross-compression, improve the extraction of data information from the multimodal data model using a loss function, and obtain a training resource knowledge graph through data fusion.

[0124] The user profiling module is used to acquire user online behavior data, model the user online behavior and cluster the groups, use CRU to perform deep knowledge tracking, and obtain learner knowledge evaluation and deep learning models.

[0125] The algorithm optimization module is used to match the training resource knowledge graph with the user's personalized needs through a graph fusion neural network recommendation algorithm to obtain personalized resource recommendations.

[0126] This embodiment also provides a computing device suitable for personalized training resource recommendation using graph fusion neural networks, including:

[0127] The system includes a memory and a processor. The memory stores computer-executable instructions, and the processor executes these instructions to implement the method for recommending personalized training resources for graph fusion neural networks as proposed in the above embodiments.

[0128] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the method for recommending personalized training resources for graph fusion neural networks as proposed in the above embodiments.

[0129] The storage medium proposed in this embodiment and the method for recommending personalized training resources for graph fusion neural networks proposed in the above embodiments belong to the same inventive concept. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0130] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0131] Example 2

[0132] Referring to Table 1, which illustrates an embodiment of the present invention, a method for recommending personalized training resources using a graph fusion neural network is provided to verify and explain the technical effects of the method.

[0133] Table 1 Comparison of the present invention with related technologies

[0134]

[0135]

[0136] As shown in Table 1, for multimodal processing, the technology of this invention uses a graph fusion neural network to process multimodal information, including text and images. Compared with traditional recommendation systems, this technology takes user interests more comprehensively.

[0137] This invention proposes a complete method for constructing a knowledge graph for training resources, including steps such as entity extraction, attribute extraction, and relation extraction. Compared to graph database-based methods, this technique offers better scalability and flexibility.

[0138] Learner behavior modeling: This invention utilizes behavioral science theory and online learning behavior modeling to gain a more comprehensive understanding of user behavior patterns through group clustering and user profiling. Compared to traditional recommendation systems, this technology offers a more comprehensive understanding of personalization.

[0139] This invention introduces a graph fusion neural network, which can capture the complex relationships between training resources more deeply, including high-order neighborhood information. Compared with traditional recommendation algorithms, this technology has stronger recommendation accuracy.

[0140] Loss function optimization: This invention optimizes the loss function, taking into account the interaction between positive and negative samples and introducing a balancing parameter, which helps improve the stability and performance of the model during training. Compared to traditional recommendation systems, this technique provides a more detailed consideration of the loss function.

[0141] To alleviate the cold start problem, this invention employs a graph fusion neural network method when handling new users or new training resources. This method is less affected by data sparsity, thus mitigating the cold start problem in recommendation algorithms.

[0142] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for personalized training resource recommendation using a graph fusion neural network, characterized in that, include: By acquiring text data information, using convolutional neural networks to capture global and local information of the text, and adjusting the mapping function according to constraints, a multimodal data model is obtained. The performance of the multimodal data model is optimized by cross-compression, and the data information of the multimodal data model is extracted by using a loss function. The training resource knowledge graph is obtained through data fusion. Acquire user online behavior data, model and cluster the user online behavior, use CRU for deep knowledge tracing, and obtain learner knowledge evaluation and deep learning models; By using a graph fusion neural network recommendation algorithm, the training resource knowledge graph is matched with the user's personalized needs to obtain personalized resource recommendations.

2. The method for personalized training resource recommendation using a graph fusion neural network as described in claim 1, characterized in that, Adjusting the mapping function according to the constraints yields the multimodal data model, which is represented as follows: CLos(T1,I1,T2,I2)=max(0,Δ(T1,I1,I2)+a)+max(0,Δ(I1,T1,T2)+a) Where T1 and T2 are the first and second inputs of the text feature vector, I1 and I2 are the first and second inputs of the image feature vector, Δ(T1, I1, I2) is used to calculate the difference between the first input of the text feature vector and the first input of the image feature vector and the second input of the image feature vector, Δ(I1, T1, T2) represents the difference between the first input of the image feature vector and the first input of the text feature vector and the second input of the text feature vector, and a is a constant; The overall characteristics of the project are represented as follows: V1=w T ·T1+w I ·I1 Among them, w T and w I V1 represents the weights of text features and image features.

3. The method for personalized training resource recommendation using a graph fusion neural network as described in claim 2, characterized in that, The performance of the multimodal data model is optimized through cross-compression, including: By fusing multimodal features, a feature vector is obtained. This feature vector is then subjected to cross-compression to produce the output, which is represented as follows: V out？ =Vh T w VV +hV T w HV +b V h out？ =Vh T w VH +hV T w HH +b H Where w and b are the weights and bias parameters of the cross-compression unit, V is the feature vector of the item, and h is the feature vector of the entity. T h T H is the transpose of the eigenvectors, and H is the transpose of the eigenvectors of the entity.

4. The method for recommending personalized training resources using a graph fusion neural network as described in claim 2 or 3, characterized in that, Improving the multimodal data model using a loss function to extract data information from the multimodal data model includes: Using a comprehensive loss function, the multimodal data model is trained and optimized, as follows: Where λ1 and λ2 are balance factors, To constrain the square norm of the model parameter W, L KGE To embed loss terms into knowledge graphs, To reconstruct the loss term.

5. The method for personalized training resource recommendation using a graph fusion neural network as described in claim 4, characterized in that, Using CRU for deep knowledge tracing, learner knowledge assessment and deep learning models are obtained, including... Learner knowledge assessment and deep learning models are represented as follows: Where σ is the sigmoid function, tanh is the hyperbolic tangent function, ⊙ is element-wish multiplication, and W r U r b r W h U h b h W z U z b z For model parameters, r t To reset the door, For the candidate hidden state, z t To update the gate, h t-1 For the past state, h t This represents the current hidden state.

6. The method for personalized training resource recommendation using a graph fusion neural network as described in claim 5, characterized in that, By using a graph fusion neural network recommendation algorithm, the training resource knowledge graph is matched with the user's personalized needs to obtain personalized resource recommendations, including... The training resource knowledge graph is combined with training resource entities, where each node represents an entity. When an entity has neighboring entities, a method is used to limit the size of each entity's domain, represented as follows: Where S(v) is the set of all entities e that are adjacent to entity v, and N(v) is the entity that is adjacent to entity v.

7. The method for personalized training resource recommendation using a graph fusion neural network as described in claim 6, characterized in that, The graph fusion neural network recommendation algorithm also includes, The graph fusion neural network recommendation algorithm is optimized using a negative sampling strategy, as follows: Where λ is the equilibrium parameter, For the contribution of L2 regularizer, y uv For the relationship between entities u and v, Let p(v) be the predicted value for entities u and v. i Let be the probability distribution of entity v.

8. A system for a personalized training resource recommendation method using graph fusion neural networks, characterized in that, include: The multimodal feature extraction module is used to acquire text data information, capture global and local information of the text using a convolutional neural network, and adjust the mapping function according to constraints to obtain a multimodal data model. The cross-compression module is used to optimize the performance of the multimodal data model through cross-compression, improve the extraction of data information from the multimodal data model using a loss function, and obtain a training resource knowledge graph through data fusion. The user profiling module is used to acquire user online behavior data, model the user online behavior and cluster the groups, use CRU to perform deep knowledge tracking, and obtain learner knowledge evaluation and deep learning models. The algorithm optimization module is used to match the training resource knowledge graph with the user's personalized needs through a graph fusion neural network recommendation algorithm to obtain personalized resource recommendations.

9. An electronic device, characterized in that, The device includes: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method described in any one of claims 1 to 7.

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

Images