Multimodal learning resource intelligent recommendation method and system based on AI big model

Abstract

The present invention relates to the field of computer technology and provides a method and system for intelligently recommending multimodal learning resources based on an AI large model. The method comprises: performing intent analysis on interactive behavior data and user information based on the AI ​​large model to obtain a user feature vector; acquiring multimodal resource data based on the user feature vector, and obtaining a resource representation vector based on the integration of the multimodal resource data; mining the implicit associations between resources and knowledge points and the mapping relationship between user needs and target skills based on a knowledge graph combined with the user feature vector and the resource representation vector to obtain a knowledge network; generating candidate recommended learning resources based on contextual perception information and the knowledge network; and ranking and optimizing the candidate recommended learning resources based on the user feature vector and the resource representation vector combined with the current stage indicators of the multimodal resource data to generate an initial resource recommendation list. The embodiments of the present invention achieve comprehensive capture of the content features of learning resources and improve the accuracy of resource recommendations.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to a method and system for intelligently recommending multimodal learning resources based on an AI large model. Background Art

[0002] In the resource recommendation scenario of education, existing collaborative filtering learning resource recommendation methods primarily construct a user-resource rating matrix based on user interaction records (such as clicks, favorites, and study time). By calculating the similarity between users, other users with similar interests as the target user are identified, and learning resources favored by similar users are then recommended. However, existing collaborative filtering learning resource recommendation methods suffer from a cold start problem. When new users register, the system lacks historical interaction data, making it difficult to determine their interests and preferences, making it difficult to make accurate recommendations. Newly uploaded resources also lack interaction records, making it difficult to accurately push them to the right users. Second, this method relies solely on a single interaction data modality, ignoring the multimodal information contained in the learning resources themselves, such as text, images, and videos. This method fails to fully capture the characteristics of resource content and potential user needs, resulting in a poor match between recommended learning resources and actual user needs. Summary of the Invention

[0003] The present invention provides a multimodal learning resource intelligent recommendation method and system based on an AI large model, which is used to comprehensively capture the content characteristics of learning resources and improve the accuracy of resource recommendations.

[0004] In a first aspect, the present invention provides a multimodal learning resource intelligent recommendation method based on an AI large model, comprising:

[0005] Based on the AI ​​big model, the captured interactive behavior data and user information on the user learning platform are analyzed for intent to obtain multi-dimensional user feature vectors;

[0006] Acquiring multimodal resource data from a learning resource library based on the user feature vector, and integrating resources based on the multimodal resource data to obtain a resource representation vector;

[0007] Based on the knowledge graph, the user feature vector and the resource representation vector are combined to mine the implicit association between resources and knowledge points and the mapping relationship between user needs and target skills to obtain a knowledge network;

[0008] generating candidate recommended learning resources based on the user's context-aware information and the knowledge network;

[0009] The candidate recommended learning resources are sorted and optimized based on the user feature vector and the resource representation vector in combination with the current stage index of the multimodal resource data to generate an initial resource recommendation list.

[0010] In a second aspect, the present invention further provides a multimodal learning resource intelligent recommendation system based on an AI large model, which is applied to the multimodal learning resource intelligent recommendation method based on an AI large model as described in the first aspect; the multimodal learning resource intelligent recommendation system based on an AI large model includes:

[0011] The intent analysis module is used to analyze the interaction behavior data and user information captured on the user learning platform based on the AI ​​model to obtain multi-dimensional user feature vectors;

[0012] A resource integration module, configured to obtain multimodal resource data from a learning resource library based on the user feature vector, and perform resource integration based on the multimodal resource data to obtain a resource representation vector;

[0013] A knowledge network construction module is used to mine the implicit associations between resources and knowledge points and the mapping relationship between user needs and target skills based on the knowledge graph in combination with the user feature vector and the resource representation vector to obtain a knowledge network;

[0014] a learning resource generation module, configured to generate candidate recommended learning resources based on the user's contextual awareness information and the knowledge network;

[0015] The learning resource recommendation module is used to optimize the ranking of the candidate recommended learning resources based on the semantic similarity between the user feature vector and the resource representation vector and the current stage index of the multimodal resource data to generate an initial resource recommendation list.

[0016] In a third aspect, the present invention also provides an electronic device comprising: a memory for storing a computer software program; and a processor for reading and executing the computer software program, thereby implementing any of the above-mentioned multimodal learning resource intelligent recommendation methods based on the AI ​​large model.

[0017] In a fourth aspect, the present invention also provides a non-transitory computer-readable storage medium, in which a computer software program is stored. When the computer software program is executed by a processor, it implements any of the multimodal learning resource intelligent recommendation methods based on the AI ​​large model as described above.

[0018] In a fifth aspect, the present invention also provides a computer program product, comprising a computer program, which, when executed by a processor, implements any of the multimodal learning resource intelligent recommendation methods based on the AI ​​large model as described above.

[0019] The multimodal learning resource intelligent recommendation method based on the AI ​​large model provided by the embodiment of the present invention solves the limitation of traditional methods that rely only on single interactive data through multimodal information fusion, and realizes the comprehensive capture of the content characteristics of learning resources. On the other hand, by utilizing knowledge graphs and context-aware technology, even new users or new resources can generate accurate resource recommendations based on semantic associations and contextual information, so that the recommended learning resources are highly matched with the actual needs of users, effectively alleviating the cold start problem and improving the accuracy of resource recommendations. Therefore, the present invention has significantly improved both resource recommendation accuracy and user satisfaction, and has shown greater adaptability and intelligence when processing multimodal data and dynamic learning scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 This is a flowchart of a multimodal learning resource intelligent recommendation method based on an AI large model provided by an embodiment of the present invention;

[0021] Figure 2 Schematic diagram of the structure of a multimodal learning resource intelligent recommendation system based on an AI large model provided by an embodiment of the present invention;

[0022] Figure 3 An embodiment diagram of an electronic device provided by an embodiment of the present invention;

[0023] Figure 4 An embodiment diagram of a computer-readable storage medium provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0024] 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 those skilled in the art without making any creative efforts shall fall within the scope of protection of the present invention.

[0025] In the description of the present invention, the terms "first" and "second" are used for descriptive purposes only and should not be understood to indicate or imply relative importance or implicitly specify the number of the technical features indicated. Therefore, a feature specified as "first" or "second" may explicitly or implicitly include one or more of the specified features. In the description of the present invention, "plurality" means two or more, unless otherwise specifically defined.

[0026] In the description of the present invention, the term "for example" is used to mean "used as an example, illustration or illustration". Any embodiment of the present invention described as "for example" is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is given to enable any person skilled in the art to implement and use the present invention. In the following description, details are listed for the purpose of explanation. It should be understood that a person of ordinary skill in the art can recognize that the present invention can be implemented without using these specific details. In other examples, well-known structures and processes are not elaborated in detail to avoid obscuring the description of the present invention with unnecessary details. Therefore, the present invention is not intended to be limited to the embodiments shown, but is consistent with the widest scope consistent with the principles and features disclosed herein.

[0027] Optional, see Figure 1 , Figure 1 This is a flow chart of the multimodal learning resource intelligent recommendation method based on the AI ​​big model provided by the present invention. In the embodiment of the present invention, the execution subject of the multimodal learning resource intelligent recommendation method based on the AI ​​big model is the resource intelligent recommendation system. Therefore, the multimodal learning resource intelligent recommendation method based on the AI ​​big model includes:

[0028] Step 10: Based on the AI ​​big model, perform intent analysis on the captured interactive behavior data and user information on the user learning platform to obtain a multi-dimensional user feature vector.

[0029] Optionally, in the resource intelligent recommendation system, various data of users on the learning platform are first collected, including but not limited to course viewing time, number of video pauses, homework completion status, search keywords, collection records, and personal information filled in by users (such as education, occupation, learning goals, etc.).

[0030] Furthermore, the intelligent resource recommendation system feeds user interaction data and user information from the learning platform into a trained AI model (such as Transformer or BERT). Through in-depth data analysis and semantic understanding, the AI ​​model identifies the user's learning intent, such as whether the user wants to delve deeper into a specific area of ​​expertise, review for an exam, or expand their learning interests. Finally, the user's learning intent is abstracted into multi-dimensional features, such as learning motivation, knowledge mastery, and learning preferences. Each dimension corresponds to a numerical value, and the combined values ​​form a user feature vector, as described in steps 101 through 104.

[0031] In one embodiment, user A searches for courses related to "Python Data Analysis" on a learning platform and watches multiple Python Data Analysis course videos totaling 10 hours. During the video playback, he pauses several times to watch the data visualization explanation segments. He completes the simple data analysis homework that comes with the course. He also enters his occupation as "Market Analyst" and his learning goal as "Improve Data Analysis Skills" in his personal information. This data is input into a Transformer-based AI model. The model analyzes and concludes that user A's learning intention is to master practical Python application skills in the field of data analysis. The model then extracts a value of 0.8 for user A's learning motivation dimension (strong motivation to improve professional skills), a value of 0.3 for Python Data Analysis Knowledge Mastery dimension (basic understanding), and a value of 0.7 for Data Visualization Learning Preference dimension (preferring this direction). This ultimately results in a multi-dimensional user feature vector for user A: [0.8, 0.3, 0.7].

[0032] Step 20: Acquire multimodal resource data from a learning resource library based on the user feature vector, and perform resource integration based on the multimodal resource data to obtain a resource representation vector.

[0033] Furthermore, the intelligent resource recommendation system searches the learning resource library based on the user's feature vector. The learning resource library stores learning resources in multiple modalities, such as text-based courseware and PDF documents, video-based tutorials, and audio-based lectures. Therefore, the intelligent resource recommendation system selects matching multimodal resource data based on dimensions such as learning preferences and knowledge needs in the user's feature vector. This multimodal resource data is then integrated and processed, using technologies such as natural language processing, computer vision, and audio processing to extract key features from each modality. These key features are then fused into a unified vector, the resource representation vector, which characterizes the core content and characteristics of the resource.

[0034] Continuing with the above example, user A, whose user feature vector is [0.8, 0.3, 0.7], searches the learning resource library to find multimodal resources such as "Python Data Visualization Practical Video Course," "Python Data Analysis Case Study PDF Document," and "Data Visualization Audio Explanation." Computer vision techniques are used to extract key visual features of the data visualization demonstrations in the video courses. Natural language processing techniques are used to extract the core knowledge points and steps of the data analysis cases in the PDF documents. Audio processing techniques are used to extract key content about data visualization principles in the audio explanations. These extracted key features are combined to generate a resource representation vector, such as [0.6 (data visualization technology explanation), 0.7 (practical operation demonstration), 0.5 (case analysis)], which represents the characteristics and value of the resource in different aspects.

[0035] Step 30: Based on the knowledge graph, combined with the user feature vector and the resource representation vector, the implicit association between resources and knowledge points and the mapping relationship between user needs and target skills are mined to obtain a knowledge network.

[0036] Furthermore, a knowledge graph is a structured semantic network that graphically displays the relationships between entities such as knowledge points, resources, and skills.

[0037] Therefore, the resource intelligent recommendation system combines the user feature vector and resource representation vector with the knowledge graph for analysis. By calculating the similarity between the vectors and the existing relationships in the knowledge graph, it mines the potential, unmarked associations between resources and knowledge points. For example, although a certain video resource does not directly mark a certain knowledge point, by analyzing the semantic relationship between its content and the knowledge points in the knowledge graph, it can be found that the two have an implicit connection. At the same time, the user needs reflected in the user feature vector are analyzed, the corresponding target skills are found in the knowledge graph, and the mapping relationship between the two is clarified. These mined implicit associations and mapping relationships are integrated to construct a comprehensive knowledge network, as shown in steps 301 to 305, which clearly show the complex relationship between resources, knowledge points, user needs and target skills.

[0038] In one embodiment, in the knowledge graph, "Python Data Visualization" is a knowledge point, associated with sub-knowledge points such as "Using the Seaborn Library" and "Using the Matplotlib Library," as well as corresponding resources such as instructional videos and documents. For user A, their user feature vector reflects their need to learn Python data visualization, and the resource representation vector corresponds to related multimodal resources. Analysis revealed that although the "Python Data Visualization Practical Video Course" that user A is interested in is not directly associated with the knowledge point "Advanced Drawing Techniques for the Seaborn Library," semantic analysis of the video content revealed that the video contains an explanation of this knowledge point, thereby uncovering an implicit association between the two. At the same time, the mapping relationship between user A's need to improve data analysis capabilities and the target skills corresponding to mastering "Python Data Visualization" and its related sub-knowledge points is clarified. These associations and mapping relationships are added to the knowledge graph, forming a knowledge network containing user A's relevant information, clearly showing the relationship between resources, knowledge points, user needs, and target skills, such as user A's needs → Python data visualization → related resources → Advanced Drawing Techniques for the Seaborn Library, etc.

[0039] Step 40: Generate candidate recommended learning resources based on the user's context-aware information and knowledge network.

[0040] Furthermore, the intelligent resource recommendation system obtains contextual information about the user, including but not limited to the current time, device type (mobile phone, computer, etc.), network environment, and the user's current learning stage. Based on the user's contextual information and the relationship between resources in the knowledge network and user needs and knowledge points, the system selects learning resources that meet the current context and user needs, and generates a list of candidate recommended learning resources, as described in steps 401 to 404. For example, if the user is using a mobile device and has a poor network environment, smaller audio or text resources suitable for mobile learning will be recommended.

[0041] Step 50 : Based on the user feature vector and the resource representation vector combined with the current stage index of the multimodal resource data, the candidate recommended learning resources are sorted and optimized to generate an initial resource recommendation list.

[0042] Furthermore, the intelligent resource recommendation system calculates the similarity between the user feature vector and the resource representation vector to assess the degree of match between the resource and the user's needs. Simultaneously, it considers the current-stage indicators of the multimodal resource data, such as the playback completion rate and user ratings of video resources, and the number of collections and downloads of text resources. These factors are comprehensively considered to optimize the ranking of candidate recommended learning resources. Resources with a high degree of match and good current-stage indicator performance are given a higher ranking priority; otherwise, the ranking priority is lowered, ultimately generating an initial resource recommendation list, as described in steps 501 to 504.

[0043] Continuing with the "Python Data Visualization Basic Concepts Audio Explanation" and "Python Data Visualization Basics Short Document" in User A's list of candidate recommended learning resources, the system calculates the similarity between User A's user feature vector [0.8, 0.3, 0.7] and the resource representation vectors of these two resources. For example, the resource representation vector for "Python Data Visualization Basic Concepts Audio Explanation" is [0.7, 0.2, 0.6], indicating a high similarity. The resource representation vector for "Python Data Visualization Basic Concepts Short Document" is [0.6, 0.3, 0.5], indicating a slightly lower similarity. Referring to the current stage indicators, "Python Data Visualization Basic Concepts Audio Explanation" has an 80% play completion rate and high user ratings, while "Python Data Visualization Basics Short Document" has fewer saves. Taking all these factors into consideration, the system ranks "Python Data Visualization Basic Concepts Audio Explanation" ahead of "Python Data Visualization Basics Short Document" to generate the initial resource recommendation list, prioritizing it for User A.

[0044] The embodiments of the present invention overcome the limitations of traditional methods that rely solely on single interaction data through multimodal information fusion, achieving comprehensive capture of the content characteristics of learning resources. Furthermore, by utilizing knowledge graphs and context-aware technology, even new users or new resources can generate accurate resource recommendations based on semantic associations and contextual information, ensuring that the recommended learning resources are highly aligned with the user's actual needs, thereby improving the accuracy of resource recommendations. Therefore, the present invention significantly improves both resource recommendation accuracy and user satisfaction, demonstrating greater adaptability and intelligence when processing multimodal data and dynamic learning scenarios.

[0045] In one embodiment, steps 101 to 104 are described as follows:

[0046] Step 101 : construct a topological network with interaction behavior data and user information as entity nodes and logical connections between interaction behaviors and user information as edges.

[0047] Optionally, the intelligent resource recommendation system abstracts collected user interaction behavior data (such as course viewing history, assignment submission status, and search keywords) and user information (age, education level, occupation, learning goals, etc.) on the learning platform into entity nodes. Edges are then constructed based on the logical connections between the data. For example, if there is a connection between a user's search keyword and the course topic being viewed, an edge can be established; if there is a potential connection between a user's occupation information and the course type selected, a corresponding edge can also be constructed, thus constructing a topological network.

[0048] Continuing with the example of user A, their interaction data includes searching for the keyword "Python Data Analysis," watching the "Python Data Visualization in Action" course video, and submitting a data analysis assignment. Their occupation is market analyst, and their learning goal is to improve their data analysis skills. The intelligent resource recommendation system treats the "Python Data Analysis" keyword search history, "Python Data Visualization in Action" course video viewing history, data analysis assignment submission history, market analyst professional information, and the learning goal of improving data analysis skills as entity nodes. Because the user searches for "Python Data Analysis" and then watches related courses, an edge is established between the "Python Data Analysis" keyword search history node and the "Python Data Visualization in Action" course video viewing history node. Considering the market analyst profession's demand for data analysis skills, an edge is established between the market analyst professional information node and the learning goal node of improving data analysis skills. This process continues, ultimately constructing a topological network containing user A's relevant data.

[0049] Step 102: Perform a dynamic temporal convolution embedding operation based on the interactive behavior data with time series characteristics to embed the time dimension information into the interactive behavior data, capture the local patterns and dependencies of user behavior that change over time, and obtain the data after dynamic temporal convolution embedding.

[0050] Furthermore, we use a dynamic temporal convolutional neural network (DTCN) to process interactive behavior data with time series characteristics (such as the length of time users watch courses at different time points and the time it takes to complete learning tasks). DTCN uses multi-layer convolutional kernels to perform convolution operations on data at different time scales, capturing both local patterns of user behavior in short periods of time (such as the pattern of frequently pausing course videos within a single day) and long-term dependencies (such as the persistence of studying the same type of course for a week). Through convolution operations, temporal dimension information is embedded into the interactive behavior data, generating dynamic temporal convolution-embedded data containing time series features, enabling the data to better reflect the temporal dynamics of user behavior.

[0051] Continuing with the above example, user A's learning behavior data over a week exhibits time series characteristics. For example, they searched for "Python Data Analysis" on Monday, began watching the "Python Data Visualization Practice" course video on Tuesday, with varying viewing times each day, and submitted an assignment on Wednesday. The intelligent resource recommendation system uses this chronologically recorded data as input and performs dynamic time series convolutional embedding using DTCN. By operating multi-layer convolutional kernels at different time scales, the system captures the temporal dynamics of user A's behavior, from searching to starting the course and completing the assignment. For example, it discovers a local pattern of gradually increasing viewing time each day as user A watches the course, as well as a time interval dependency between searching and starting to learn. Ultimately, the system generates data after dynamic time series convolutional embedding that incorporates this temporal dimension information.

[0052] Step 103: Based on the AI ​​big model, diffusion propagation is performed in the topological network with semantic concepts as units to obtain the results of semantic diffusion propagation.

[0053] Furthermore, the intelligent resource recommendation system utilizes a large AI model based on graph neural networks (GNNs), using semantic concepts as the basic unit for diffusion propagation within a topological network. The GNN propagates the semantic information of each node along edges to adjacent nodes through inter-node message passing. During this propagation process, the node's semantic representation is updated based on the semantic similarity between nodes and the edge weights (which can be set based on the strength of the logical connection). Through multiple iterations of diffusion propagation, each node integrates semantic information from other relevant nodes in the network, resulting in a more comprehensive and accurate semantic representation. This is the result of semantic diffusion propagation, enabling deep semantic fusion of data within the topological network.

[0054] Continuing with the above example, for user A's topological network, taking the semantic concept of "Python Data Visualization" as an example, in the graph neural network, nodes related to "Python Data Visualization" (such as the "Python Data Visualization in Action" course video viewing record node and related document collection record nodes) propagate their semantic information along edges to other related nodes through a message-passing mechanism. For example, the "Python Data Visualization in Action" course video viewing record node transmits semantic information such as the course content and key knowledge points to the "Python Data Visualization" concept node, while also receiving semantic information from other related nodes (such as information about improving data visualization skills from the user's learning goal node). After multiple iterations of diffusion propagation, the "Python Data Visualization" concept node incorporates the semantic information of all related nodes in the network, resulting in a richer and more accurate semantic representation. This is the result of semantic diffusion propagation, leading to deep semantic integration of the entire topological network.

[0055] Step 104: perform intent analysis based on the semantic diffusion propagation results and the dynamic temporal convolution embedded data to obtain a multi-dimensional user feature vector.

[0056] Furthermore, the resource intelligent recommendation system performs intent analysis based on the results of semantic diffusion propagation and the data after dynamic temporal convolution embedding to obtain a user feature vector, as specifically shown in steps 1041 to 1044.

[0057] The embodiment of the present invention mines user data from multiple dimensions such as structure, time, and semantics, so it can understand user learning intentions and characteristics more comprehensively and accurately, providing a data basis for subsequent accurate learning resource recommendations, thereby improving the accuracy of resource recommendations.

[0058] In one embodiment, steps 1041 to 1044 are described as follows:

[0059] Step 1041, based on the results of semantic diffusion propagation and the data after dynamic temporal convolution embedding, cluster analysis is performed in the spatiotemporal dimension to aggregate data with similar learning intention patterns to obtain different learning intention pattern clusters.

[0060] Optionally, the resource intelligent recommendation system fuses the data after dynamic time series convolution embedding (including time dimension information) with the results after semantic diffusion propagation (including semantic information) to form a data set with both spatiotemporal semantic features. In this embodiment of the present invention, a density-based DBSCAN clustering algorithm is used to analyze these data in the spatiotemporal dimensions. Using the semantic similarity of data points as the spatial distance metric and combining the continuity of the time series, data points that are similar in spatiotemporal semantics are aggregated into one category to form different learning intention pattern clusters. Each cluster represents a similar learning intention pattern, such as a pattern that focuses on basic learning in a certain field, a learning pattern that expands cross-domain knowledge, and so on.

[0061] Optionally, for the DBSCAN algorithm in the embodiment of the present invention, the concepts of core point, density direct, and density reachable are defined. Let the data point set be D. For a data point p, if its If the neighborhood contains at least MinPts data points, then p is a core point. The neighborhood is centered at data point p and has a radius of The set of data points in the area is calculated as follows: ,in, is the semantic similarity distance measure between data points p and q (which can be calculated using methods such as cosine similarity). In the neighborhood, q is said to be directly accessible from p density; if there is a data point chain , making from The density is directly from Density reachable. Through these definitions, density reachable data points are aggregated into a learning intent pattern cluster.

[0062] Continuing with user A, their fused learning data includes behavioral information and semantic information such as searches for "Python Data Analysis"-related content and viewing of practical Python Data Visualization courses at different time points. The intelligent resource recommendation system clusters this data with similar fused data from other users using the DBSCAN algorithm. During the clustering process, the semantic similarity of data points (such as the semantic association of search keywords and the semantic matching of course content) is used as a distance metric, combined with the chronological order of the time series. The system discovered that user A and some other users had focused on learning the basics of Python data visualization for a period of time, and their data points were similar in spatiotemporal semantics. Therefore, these data points were aggregated into a learning intention pattern cluster, representing the "Python Data Visualization Basics Learning Pattern."

[0063] Step 1042 , between different learning intention pattern clusters, cross-domain intention association deduction is performed based on semantic information and data association relationships to obtain potential connections between different learning intention pattern clusters in different fields or topics.

[0064] Furthermore, the intelligent resource recommendation system analyzes different learning intent pattern clusters using a graph reasoning algorithm based on the knowledge graph. This algorithm treats each learning intent pattern cluster as a node in the knowledge graph. By analyzing the semantic information of the data within each cluster (such as the knowledge points and skill areas involved) and the relationships between data between clusters (such as the content corresponding to different pattern clusters learned by users at different stages), it conducts path search and reasoning within the knowledge graph. By mining the implicit paths and semantic connections between nodes in the knowledge graph, it deduces the potential connections between different learning intent pattern clusters in different fields or topics. For example, it can deduce the connection between the programming language learning pattern cluster and the software development project practice pattern cluster, indicating that mastering programming languages ​​is the foundation for practicing software development projects.

[0065] Optionally, the learning intention pattern cluster set in the embodiment of the present invention is , in the knowledge graph, for two learning intention pattern clusters and , and measure the potential connections by calculating the semantic path weights between them. Using the random walk-based path weight calculation method, we can get the clusters Starting from the node, it wanders along the edge of the knowledge graph with a certain probability and reaches the cluster Node. Path weight The calculation formula is: ,in, For arrive The number of paths, is the attenuation factor , used to control the effect of path length on weight, is the length of the kth path. The higher the path weight, the closer the potential connection between the two learning intention pattern clusters.

[0066] Continuing with the above embodiment, we obtain the "Python Data Visualization Basic Learning Mode" cluster and the "Data Report Production Learning Mode" cluster. In the knowledge graph, the "Python Data Visualization Basic Learning Mode" cluster node contains knowledge points related to Python data visualization, and the data report production learning mode cluster node contains knowledge points such as data report design and the use of production tools. Through the graph reasoning algorithm based on the knowledge graph, the system finds that learning from Python data visualization can be extended to the application of visualization elements in data reports, and there is a semantic path from the "Python Data Visualization Basic Learning Mode" cluster to the "Data Report Production Learning Mode" cluster in the knowledge graph. By calculating the path weight, it is determined that there is a potential connection between the two learning intention pattern clusters, that is, users may have the need to produce data reports after learning Python data visualization.

[0067] Step 1043 , based on the probability generation model, combining the user information and different learning intention pattern clusters and their corresponding potential connections, determines the demand probability distribution of the user's potential needs under different learning intention pattern clusters.

[0068] Furthermore, the intelligent resource recommendation system uses a variational autoencoder (VAE) as a probabilistic generative model, taking as input user information (such as age, occupation, and learning goals), different learning intention pattern clusters, and the potential connections between them. The VAE uses the encoder to map the input data to a latent variable space, learns the probability distribution of the data in this latent variable space, and then uses the decoder to generate reconstructed data from the latent variable space. During training, the model parameters are optimized by maximizing the evidence lower bound (ELBO), enabling the model to learn the probability distribution of users' potential needs within different learning intention pattern clusters. For example, within the "Python Data Visualization Basic Learning Model" cluster, the model learns user information and potential connections to determine the probability distribution of users' potential needs for further data visualization learning, in-depth use of related tools, and other related needs.

[0069] Optionally, for the variational autoencoder in the embodiment of the present invention, given the input data x (including user information, learning intention pattern clusters and potential connections), the model learns an approximate posterior distribution To approximate the true posterior distribution , where z is a latent variable, and is the model parameter. The formula for calculating the evidence lower bound (ELBO) is: ,in, For the reconstruction loss, we measure how similar the data generated by the model from the latent variables is to the original data: for Divergence, a measure of the approximate posterior distribution and the prior distribution By maximizing ELBO and optimizing model parameters, we can obtain the probability distribution of users’ potential needs under different learning intention pattern clusters. (Needs|Learning Intention Pattern Cluster).

[0070] Continuing with User A, whose user information is that of a market analyst and whose learning goal is to improve data analysis skills, they are in the "Python Data Visualization Basics Learning Mode" cluster and have a potential connection with the "Data Report Creation Learning Mode" cluster. This information is input into a variational autoencoder. The model uses the encoder to map the input data to a latent variable space. During training, the parameters are continuously adjusted to maximize the lower bound of the evidence. Ultimately, the model determines that within the "Python Data Visualization Basics Learning Mode" cluster, User A has a probability of 0.7 of subsequently wanting to learn advanced data visualization techniques, a probability of 0.6 of wanting to learn professional data report creation tools, and so on. This yields the demand probability distribution of the user's potential needs within this learning intention pattern cluster.

[0071] Step 1044 , based on the different learning intention pattern clusters and their corresponding potential connections, as well as the demand probability distributions under the different learning intention pattern clusters, a multi-dimensional user feature vector is obtained by integration.

[0072] Furthermore, the intelligent resource recommendation system comprehensively analyzes different learning intention pattern clusters, potential connections, and demand probability distributions. With learning intention pattern clusters as the core, the system converts the potential connections and demand probability distributions corresponding to each cluster into feature dimensions. For example, a learning intention pattern cluster itself is used as one dimension, the strength of its potential connections with other clusters is used as an additional dimension, and the probability of different potential demands is used as another dimension. In this way, various aspects of information are integrated into a multidimensional vector, comprehensively characterizing the user's learning characteristics and potential demands, forming the final multidimensional user feature vector.

[0073] Continuing with user A, there are a "Python Data Visualization Basics Learning Mode" cluster and a "Data Reporting Learning Mode" cluster, with a potential connection weight of 0.6. Within the "Python Data Visualization Basics Learning Mode" cluster, the probability of "learning advanced data visualization techniques" is 0.7, and the probability of "learning professional data reporting tools" is 0.6. The intelligent resource recommendation system uses the "Python Data Visualization Basics Learning Mode" cluster as a dimension, the potential connection weight of 0.6 between this cluster and the "Data Reporting Learning Mode" cluster as a dimension, the probability of learning advanced data visualization techniques (0.7) and the probability of learning professional data reporting tools (0.6) as separate dimensions, and combines this with other relevant information (such as the user's learning motivation) to form a multi-dimensional user feature vector, such as [0.8 (learning motivation), 0.3 (current knowledge mastery), 0.7 (Python Data Visualization Basics Learning Mode cluster identifier), 0.6 (connection weight with the Data Reporting Mode cluster), 0.7 (probability of needing advanced data visualization techniques), 0.6 (probability of needing professional reporting tools), ...].

[0074] The embodiments of the present invention can accurately capture the user's potential needs and learning characteristics, and provide a more comprehensive and accurate basis for subsequent learning resource recommendations, thereby improving the accuracy of resource recommendations.

[0075] In one embodiment, steps 301 to 305 are described as follows:

[0076] Step 301: Anchor and match the semantic information in the user feature vector and the resource representation vector with the corresponding semantic nodes and relationship paths in the preset knowledge graph to obtain a semantic anchoring result.

[0077] Optionally, the intelligent resource recommendation system extracts semantic information from user feature vectors and resource representation vectors, such as learning need keywords reflected in the user feature vectors and knowledge point names contained in the resource representation vectors, and compares this semantic information with a preset knowledge graph. The knowledge graph stores a large number of semantic nodes (such as various knowledge points, skills, resource types, etc.) and relationship paths between nodes (such as "mastering a certain knowledge point is a prerequisite for learning another knowledge point" and "a certain resource contains a specific knowledge point"). A matching algorithm based on semantic embedding is used to calculate the similarity between the vector semantic information and the semantics of the nodes and relationship paths in the knowledge graph. Nodes and relationship paths with similarities above a threshold are anchored to form a semantic anchoring result, which determines the corresponding positions of user features and resources in the knowledge graph.

[0078] Continuing with the above, user A's user feature vector contains the semantic information "Python Data Visualization Learning Needs," and the resource representation vector corresponds to the resource "Python Data Visualization Practical Video Course," whose semantic information includes knowledge points such as "Using the Seaborn Library" and "Using the Matplotlib Library." The resource intelligent recommendation system compares this semantic information with the preset knowledge graph. Using cosine similarity calculations, the system finds that the similarity between "Python Data Visualization" and the "Python Data Visualization" node in the knowledge graph is 0.9, and the similarity between "Using the Seaborn Library" and the corresponding knowledge point node in the knowledge graph is 0.85, both exceeding the threshold of 0.8. These nodes are then anchored. Furthermore, the system finds that the similarity between "Python Data Visualization Practical Video Course" and the path connecting the "Python Data Visualization" node under the "Contains" relationship path in the knowledge graph is 0.88, and also anchors them, resulting in a semantic anchoring result.

[0079] Step 302, starting from the nodes and relationships corresponding to the semantic anchoring results and taking the path extension length as the step length, explore the potential association paths between resources and knowledge points in the preset knowledge graph, and trace back the mapping connection paths between user needs and target skills to obtain the semantic association path.

[0080] Furthermore, the intelligent resource recommendation system uses the node and relationship path corresponding to the semantic anchoring result as the starting point, sets the path expansion length as the exploration step (for example, setting the step length to 3 means expanding up to three nodes along the relationship path in the knowledge graph). Using an algorithm based on depth-first search (DFS) combined with heuristic search, it explores potential association paths of knowledge points related to the resource forward in the knowledge graph. Starting from the resource node, it searches for other related knowledge points along the relationship path. It also traces the mapping connection path between user needs and target skills backward. Starting from the node corresponding to the user need, it searches for the relationship path between it and the target skill. During the search process, path selection is made based on the semantic relevance of nodes and relationships, as well as the importance of the path (measured by indicators such as the node's out-degree and in-degree). Ultimately, the semantic association paths between resources and knowledge points, and between user needs and target skills, are obtained.

[0081] Optionally, the starting node of the semantic anchoring result in the embodiment of the present invention is , the relationship path is , the path extension length is k. Using the depth-first search algorithm, starting from the starting node Start by searching in the knowledge graph. For the current node , its neighbor node set is Select the next node Probability The calculation formula is:

[0082] .

[0083] in, For nodes and The semantic similarity of For nodes The importance of (can be calculated based on the node's out-degree, in-degree and other indicators). When the search depth reaches k, stop searching and get the semantic association path.

[0084] Continuing with the semantic anchoring results for user A, the starting nodes are "Python Data Visualization" and "Python Data Visualization Practical Video Course," with a relationship path of "Contains." Setting the path extension length to 3, a depth-first search combined with a heuristic search algorithm is used. Starting from the "Python Data Visualization" node, it is found to be connected to the "Data Visualization Principles" node via the "Basic Knowledge Point" relationship, with high semantic similarity and high importance. Continuing to explore from the "Data Visualization Principles" node, an association with the "Visualization Chart Type" node is found, ultimately resulting in a potential resource-knowledge point association path: "Python Data Visualization" → "Data Visualization Principles" → "Visualization Chart Type." Starting from the user need node "Python Data Visualization Learning Need," tracing back, it is found to be connected to the target skill "Improving Data Analysis Skills" via the "Support" relationship. Further, "Proficiency in Using Visualization Tools" is found as an intermediate association node, resulting in a user need-target skill mapping path: "Python Data Visualization Learning Need" → "Proficiency in Using Visualization Tools" → "Improving Data Analysis Skills," thus obtaining a semantic association path.

[0085] Step 303: Dynamically cluster the pre-set knowledge graph based on the semantic association path to generate subgraph clusters. Each subgraph cluster represents the resource-knowledge point association or the user demand-target skill relationship.

[0086] Furthermore, the resource intelligent recommendation system extracts semantically associated paths from a preset knowledge graph to form multiple path subgraphs. The embodiment of the present invention uses a dynamic clustering algorithm based on a graph neural network to cluster these path subgraphs. The graph neural network learns the feature representations of nodes and relationships in each path subgraph through a message passing mechanism between nodes, and then calculates the similarity between subgraphs based on these feature representations, and aggregates subgraphs with high similarity into a subgraph cluster. Each subgraph cluster represents a specific resource-knowledge point association or user demand-target skill relationship. For example, a subgraph cluster may contain multiple path subgraphs associated with Python data visualization resources and related knowledge points, and another subgraph cluster may contain a mapping relationship path subgraph between user data analysis needs and related target skills.

[0087] Continuing with the semantic association paths obtained by user A, such as the path subgraphs "Python Data Visualization Practical Video Course" → "Using the Seaborn Library" → "Advanced Charting," and "Python Data Visualization Learning Requirements" → "Proficient in Using Visualization Tools" → "Improving Data Analysis Skills," these path subgraphs are fed into a dynamic clustering algorithm based on a graph neural network. The graph neural network learns the feature representation of each subgraph and calculates the similarity between subgraphs. It finds that the path subgraphs associated with Python data visualization resources and related knowledge points have a high degree of similarity, and these are clustered into a subgraph cluster C1, which represents the association between Python data visualization resources and knowledge points. The path subgraphs related to user needs and target skills are clustered into another subgraph cluster C2, which represents the relationship between the user's data analysis needs and target skills.

[0088] Step 304 : determining the implicit association between resources and knowledge points based on the logical relationships between different sub-graph clusters, and determining the mapping relationship between user needs and target skills based on the association levels between different sub-graph clusters.

[0089] Furthermore, the source intelligent recommendation system analyzes the logical relationships between different subgraph clusters, such as inclusion relationships, causal relationships, and parallel relationships. For subgraph clusters related to resources and knowledge points, the implicit associations between resources and knowledge points are determined by analyzing the semantics of nodes and relationships within the clusters and combining the logical relationships between different subgraph clusters. For example, if a resource node in a subgraph cluster has an indirect relationship with multiple knowledge point nodes, the coverage and association depth of the resource for these knowledge points can be inferred through the logical relationships between subgraph clusters. For subgraph clusters related to user needs and target skills, the degree of association between different subgraph clusters is calculated (this can be done by calculating the similarity of the subgraph cluster feature vectors, etc.), and the mapping relationship between user needs and target skills is determined based on the degree of association. The higher the degree of association, the closer the connection between user needs and target skills.

[0090] Continuing with user A, sub-graph cluster C1 contains association paths between the "Python Data Visualization Practical Video Course" and knowledge points such as "Using the Seaborn Library" and "Using the Matplotlib Library." Sub-graph cluster C3 contains association paths between the "Python Data Visualization Advanced Course" and the "Complex Data Visualization Project Practice" knowledge point. System analysis reveals a parallel relationship between C1 and C3. Logical reasoning indicates that the implicit association between the "Python Data Visualization Practical Video Course" and the "Seaborn Library" and "Matplotlib Library" knowledge points is a basic learning association, while the implicit association between the "Python Data Visualization Advanced Course" and the "Complex Data Visualization Project Practice" knowledge point is an advanced application association. The cosine similarity of the eigenvectors of sub-graph cluster C2 (containing the relationship path between the user's data analysis needs and target skills) and other related sub-graph clusters was calculated. The similarity between C2 and the sub-graph cluster containing the target skill "Improving Data Visualization Skills" is 0.85, indicating a strong mapping relationship between the user's data analysis needs and the target skill "Improving Data Visualization Skills."

[0091] Step 305 : construct a knowledge network with resources, knowledge points, user needs and target skills as nodes and implicit associations and mapping relationships as edges.

[0092] Furthermore, the intelligent resource recommendation system constructs a knowledge network by treating resources, knowledge points, user needs, and target skills as nodes in the knowledge network, and the implicit associations between resources and knowledge points, and the mapping relationships between user needs and target skills, as edges between nodes. During the construction process, edges are labeled based on the type and strength of the associations and mappings. For example, the weight of the labeled edges indicates the closeness of the association or mapping, and the type of the labeled edges (such as "includes," "supports," and "prerequisite") clarifies the nature of the relationship. This enables the knowledge network to clearly and intuitively display the complex relationships between resources, knowledge points, user needs, and target skills. For user A, resources such as "Python Data Visualization Practical Video Course" and "Python Data Visualization Advanced Course" are used as nodes, along with knowledge points such as "Using the Seaborn Library," "Using the Matplotlib Library," and "Practical Complex Data Visualization Projects," and user needs and target skills such as "Python Data Visualization Learning Needs" and "Improving Data Analysis Skills." A knowledge network is constructed using the "inclusion" relationship (implicit association strength of 0.7) between "Python Data Visualization Practical Video Course" and "Using the Seaborn Library" and the "support" relationship (mapping closeness of 0.85) between "Python Data Visualization Learning Needs" and "Improving Data Analysis Skills" as edges. The edges are labeled in the network, for example, labeling the line connecting "Python Data Visualization Practical Video Course" and "Using the Seaborn Library" with "inclusion, weight 0.7" and the line connecting "Python Data Visualization Learning Needs" and "Improving Data Analysis Skills" with "support, weight 0.85," forming a complete and information-rich knowledge network.

[0093] The embodiments of the present invention can accurately capture the implicit association between resources and knowledge points and the mapping relationship between user needs and target skills, providing an accurate data basis for the intelligent recommendation of learning resources, thereby improving the accuracy of resource recommendations.

[0094] In one embodiment, steps 401 to 404 are described as follows:

[0095] Step 401 : Generate a context feature vector based on the learning progress information and the learning stage information, and match the context feature vector with the user demand nodes in the knowledge network to obtain a network subgraph.

[0096] Optionally, the intelligent resource recommendation system obtains the user's learning progress information (such as quantitative data such as the number of completed course chapters and homework completion rate) and learning stage information (such as qualitative descriptions such as the introductory stage, advanced stage, and mastery stage), digitally encodes this information, and uses a deep learning model (such as a multi-layer perceptron (MLP)) to extract and fuse features to generate a contextual feature vector that contains learning progress and stage characteristics. Furthermore, the intelligent resource recommendation system calculates cosine similarity between this contextual feature vector and the vector corresponding to the user demand node in the knowledge network, screening out user demand nodes and their connected subgraph structures with similarity above a set threshold, thereby obtaining a network subgraph related to the current user context.

[0097] Continuing with user A's example, their learning progress information indicates that they have completed three chapters of the "Python Data Visualization Practice" course, with a 60% completion rate, encoded as a vector p = [0.3, 0.6]. Their learning stage is in the introductory phase, encoded as s = [1, 0, 0] (e.g., introductory, advanced, and proficient correspond to [1, 0, 0], [0, 1, 0], and [0, 0, 1], respectively). The intelligent resource recommendation system uses multi-layer perceptron fusion to generate a contextual feature vector c = MLP([0.3, 0.6, 1, 0, 0]) = [0.2, 0.4, 0.6]. In the knowledge network, the vector corresponding to the user requirement node "Python Data Visualization Basics" is u1 = [0.1, 0.3, 0.7]. The calculated similarity is 0.85, which exceeds the threshold of 0.8. This node and its connected subgraph structure are then included in the network subgraph.

[0098] Step 402 : determining the initial recommended learning resources based on the learning resource data associated with each subgraph node in the network subgraph, and determining the resource association degree between each resource and the context feature vector based on the correlation between each resource in the initial recommended learning resources and the context feature vector.

[0099] Furthermore, the intelligent resource recommendation system extracts learning resource data associated with each subgraph node (including knowledge points, user need nodes, and other nodes) from the network subgraph, and aggregates these resources into an initial set of recommended learning resources. For each learning resource in the set, its resource representation vector (such as the vector containing semantic and type features generated in the previous step) is extracted. Using a self-attention mechanism, the correlation between the resource representation vector and the contextual feature vector is calculated. This correlation value is used as the resource association degree between each resource and the contextual feature vector, reflecting the degree of fit between the resource and the user's current learning context.

[0100] Continuing in the network subgraph, the learning resources associated with the "Python Data Visualization Basics Learning" node are "Python Data Visualization Getting Started Tutorial Video" and "Basic Chart Drawing Guide Documentation," with resource representation vectors r1 = [0.3, 0.5, 0.2] and r2 = [0.4, 0.3, 0.3]. User A's context feature vector c = [0.2, 0.4, 0.6]. Using the self-attention mechanism, we can calculate the resource relevance R1 of "Python Data Visualization Getting Started Tutorial Video" to be 0.7, and the resource relevance R2 of "Basic Chart Drawing Guide Documentation" to be 0.6.

[0101] Step 403 , performing a fusion operation based on element-wise multiplication of the context feature vector and the vector corresponding to the target skill node in the knowledge network to obtain a resource matching degree.

[0102] Furthermore, the intelligent resource recommendation system obtains the vector corresponding to the target skill node associated with the user's need in the knowledge network and fuses this vector with the context feature vector by element-wise multiplication, resulting in a new vector. The fully connected neural network (FCNN) then extracts and transforms the new vector, outputting a numerical value as the resource matching degree. This value reflects the degree of match between the user's current context and the target skill, and is used to measure the potential of the recommended resource in helping the user achieve the target skill. Continuing with the example of user A, the target skill node associated with the need for "Learning Basic Python Data Visualization" in the knowledge network is "Mastering Basic Visualization Tools," and its corresponding vector is t = [0.4, 0.5, 0.1]. User A's context feature vector c = [0.2, 0.4, 0.6] is element-wise multiplied to produce the fused vector f = [0.2*0.4, 0.4*0.5, 0.6*0.1] = [0.08, 0.2, 0.06]. The fully connected neural network calculates the resource matching degree M = 0.75.

[0103] Step 404 : Screen the initially recommended learning resources based on the resource matching degree and the resource association degree between each resource and the context feature vector to obtain candidate recommended learning resources.

[0104] Furthermore, the intelligent resource recommendation system builds a decision model (such as a support vector machine (SVM)-based classification model) that uses resource matching and each resource's relevance as input features. It then uses the decision boundary learned through model training to classify and filter the initial recommended learning resources. Resources that meet the model's classification criteria (e.g., those with high matching and high relevance) are retained to form a candidate set of recommended learning resources, while resources that don't match the user's current context and target skills are excluded.

[0105] Continuing with the "Python Data Visualization Tutorial Video", its resource matching degree M=0.75, resource correlation R1=0.7, and the feature vector x1=[0.75, 0.7]. Calculated by the support vector machine decision function , the resource is retained; for other resources, if the calculation result is -1, they are excluded, and finally a set of candidate recommended learning resources is obtained.

[0106] The embodiments of the present invention utilize knowledge graphs and context-aware technologies to generate accurate resource recommendations based on semantic associations and contextual information even for new users or new resources, so that the recommended learning resources are highly matched with the user's actual needs, effectively alleviating the cold start problem and improving the accuracy of resource recommendations.

[0107] In one embodiment, steps 501 to 504 are described as follows:

[0108] Step 501 : Determine a semantic similarity difference based on a difference between a first similarity score between a user feature vector and a reference semantic vector and a second similarity score between a resource representation vector and the reference semantic vector.

[0109] Optionally, the resource intelligent recommendation system sets a baseline semantic vector, which can be generated based on common knowledge concepts in the field or semantic content that is of general concern to the user group. For each candidate recommended learning resource, the system calculates the first similarity score between the user feature vector and the baseline semantic vector, and the second similarity score between the resource representation vector and the baseline semantic vector. Optionally, the similarity calculation in an embodiment of the present invention adopts a Transformer-based semantic similarity measurement method, which captures the semantic association between vectors through a multi-head attention mechanism. Finally, the two similarity scores are subtracted to obtain a semantic similarity difference, which reflects the degree of matching difference between user needs and resources at the semantic level.

[0110] Continuing with user A as an example, their user feature vector u = [0.8, 0.3, 0.7] (representing learning motivation, knowledge mastery, and learning preferences), for example, the baseline semantic vector b = [0.7, 0.2, 0.6], and the resource representation vector r = [0.6, 0.4, 0.5] of the candidate recommended learning resource "Python Data Visualization Advanced Skills Video Course". Using the Transformer-based similarity calculation function, we get the first similarity score S u-b =0.75, the second similarity score S r-b =0.68, then the semantic similarity difference D=0.75-0.68=0.07.

[0111] Step 502 : Classify each resource in the candidate recommended learning resources into stages based on the current stage indicator to obtain a resource stage category of each resource in the candidate recommended learning resources.

[0112] Furthermore, the intelligent resource recommendation system collects current-stage metrics for candidate learning resources, such as the completion rate of video resources, the number and sentiment of user comments, and the download and collection count of text resources. Using a clustering method based on the Gaussian Mixture Model (GMM), resources are divided into different stages based on these metrics, such as introductory, advanced, and advanced. GMM clusters resource data by estimating the parameters of multiple Gaussian distributions, each representing a resource stage. The stage to which the resource belongs is determined based on the degree of probability matching between the resource data and each Gaussian distribution.

[0113] The current stage indicator data of the candidate recommended learning resources in the embodiment of the present invention is , the Gaussian mixture model consists of K Gaussian distributions, and the probability density function of the kth Gaussian distribution is:

[0114] .

[0115] in, are the mean and variance of the kth Gaussian distribution.

[0116] resource The probability of belonging to the k-th stage category is:

[0117] in, is the weight of the kth Gaussian distribution, satisfying Parameters are estimated using the expectation maximization (EM) algorithm. and , finally based on Identify resources The resource phase category.

[0118] Continuing with the candidate learning resource "Python Data Visualization Basics Tutorial Document," its current metrics are 200 downloads and 50 favorites. The "Python Data Visualization Advanced Techniques Video Course" has 80 downloads and a 30% completion rate, with user comments mostly discussing complex issues. The intelligent resource recommendation system uses these metrics as input and clusters them using a Gaussian mixture model. Using the EM algorithm to estimate parameters, the calculated probability that "Python Data Visualization Basics Tutorial Document" belongs to the introductory level category is 0.85, confirming it as an introductory resource. The probability that "Python Data Visualization Advanced Techniques Video Course" belongs to the advanced level category is 0.78, confirming it as an advanced resource.

[0119] Step 503 : Associating each resource in the candidate recommended learning resources based on the resource stage category and semantic similarity to construct an association matrix.

[0120] Furthermore, the resource intelligent recommendation system integrates semantic similarity differences and source stage categories. Assuming the number of resource stage categories is m, the number of candidate recommended learning resources is n, and constructing a The correlation matrix A. The elements in the matrix It represents the degree of association between the i-th resource stage category and the j-th resource. The association degree is obtained by jointly calculating the semantic similarity difference and the resource stage category encoding. The association calculation method based on the graph convolutional network (GCN) is adopted. The resources and stage categories are regarded as nodes in the graph, and the association degree is updated through information transmission between nodes.

[0121] In the embodiment of the present invention, the resource stage category coding vector is , the semantic similarity difference of the jth resource is . Calculate the correlation matrix elements through graph convolutional network :

[0122] in, Add self-loop adjacency matrix for The diagonal node degree matrix For the The weight matrix of the layer, is the activation function, is the number of layers in the graph convolutional network.

[0123] In one embodiment, resource stages are divided into three categories: entry-level, advanced, and advanced, and there are five candidate recommended learning resources. Taking "Python Data Visualization Basic Tutorial Document" (entry-level, semantic similarity difference 0.05) and "Python Data Visualization Advanced Skills Video Course" (advanced level, semantic similarity difference 0.07) as examples, the correlation matrix elements between the entry-level and "Python Data Visualization Basic Tutorial Document" are calculated through graph convolutional networks. The correlation matrix element is 0.82, the advanced stage and Python data visualization advanced skills video course is 0.79, and the complete Associated Torch Array.

[0124] Step 504 : rank and optimize each of the candidate recommended learning resources based on the association matrix to generate an initial resource recommendation list.

[0125] Furthermore, the intelligent resource recommendation system optimizes the ranking of each of the candidate recommended learning resources according to the association matrix and generates an initial resource recommendation list, as shown in steps 5041 to 5043 .

[0126] The embodiment of the present invention takes into comprehensive consideration multiple dimensions such as semantics, stages, and user feedback. When processing complex resource data and user personalized needs, it can significantly improve the rationality of the recommended resource list and user satisfaction, so that the recommended learning resources are highly matched with the user's actual needs, effectively alleviating the cold start problem and improving the accuracy of resource recommendations.

[0127] In one embodiment, steps 5041 to 5043 are described as follows:

[0128] Step 5041 : For each candidate recommended learning resource, sort it according to the semantic similarity difference based on the matrix element values ​​in the association matrix, and generate a semantic similarity difference sorting sequence under the resource stage category of each resource.

[0129] Optionally, the resource intelligent recommendation system extracts the matrix element values ​​corresponding to each resource from the association matrix. These element values ​​reflect the degree of association between the resource and each resource stage category, and the degree of association is related to the semantic similarity difference. For each resource, the resource intelligent recommendation system sorts the resources according to the size of the semantic similarity difference corresponding to different resource stage categories. The embodiment of the present invention adopts a multi-keyword sorting algorithm improved based on Quick Sort, with the resource stage category as the first keyword and the semantic similarity difference as the second keyword, to sort the relevant data of each resource under different stage categories, thereby generating a semantic similarity difference sorting sequence for each resource under the resource stage category to which it belongs.

[0130] In one embodiment, the resource stage categories are entry-level, advanced, and advanced. The recommended learning resources are "Python Data Visualization Basic Tutorial Document", "Python Data Visualization Intermediate Case Collection", and "Python Data Visualization Advanced Skills Video Course". , the association matrix A is as follows:

[0131] .

[0132] Taking the "Python Data Visualization Basics Tutorial Document" as an example, the semantic similarity differences for the introductory, advanced, and advanced levels correspond to 0.8, 0.2, and 0.1 in the first column of the matrix, respectively. The intelligent resource recommendation system uses an improved quick sort algorithm, first grouping resources by level and then sorting them within the group by semantic similarity difference. The resulting semantic similarity difference ranking sequence for this resource is: introductory level (0.8), advanced level (0.2), and advanced level (0.1). Similarly, the ranking sequence for the other two resources can be obtained.

[0133] Step 5042: perform a first sorting of each resource according to the priority order corresponding to the resource stage category of each resource, and recommend learning resources after obtaining the first sorting.

[0134] Furthermore, the intelligent resource recommendation system pre-defines a priority order for resource stage categories, for example, entry-level < advanced stage < advanced stage. The system ranks candidate recommended learning resources in order of priority based on the resource stage category to which each resource belongs. Optionally, embodiments of the present invention employ a multi-stage sorting algorithm based on radix sort, using resource stage category as the key basis for sorting. This algorithm performs a first round of sorting on all candidate recommended learning resources, arranging the resources in order of priority according to their stage categories, resulting in a list of recommended learning resources after the first sorting.

[0135] In this embodiment of the present invention, the resource stage category set is , whose priority order is For candidate recommended learning resource collections , each resource The resource stage category is The radix sort algorithm starts sorting from the least significant bit (resource stage category). Its core steps are as follows: 1. Initialize m buckets (corresponding to m resource stage categories). 2. Traverse the resource collection. , the resources Put it into the resource stage category 3. Take out resources from the bucket in order of priority and get the recommended learning resource list after the first ranking .

[0136] Continuing with the three candidate learning resources recommended above, "Python Data Visualization Basics Tutorial Documentation" belongs to the introductory stage, "Python Data Visualization Intermediate Case Study Collection" belongs to the advanced stage, and "Python Data Visualization Advanced Techniques Video Course" belongs to the advanced stage. Following the priority order of introductory stage < advanced stage < advanced stage, using a radix sort algorithm, first place "Python Data Visualization Basics Tutorial Documentation" in the introductory stage bucket, "Python Data Visualization Intermediate Case Study Collection" in the advanced stage bucket, and "Python Data Visualization Advanced Techniques Video Course" in the advanced stage bucket. These resources are then removed in order, resulting in the following recommended learning resource list after the first sort: "Python Data Visualization Basics Tutorial Documentation," "Python Data Visualization Intermediate Case Study Collection," and "Python Data Visualization Advanced Techniques Video Course."

[0137] Step 5043 : For the resources with the same priority in the recommended learning resources after the first sorting, perform a second sorting according to the semantic similarity difference sorting sequence under the corresponding resource stage category to obtain an initial resource recommendation list.

[0138] Furthermore, based on the resulting list of recommended learning resources after the initial ranking, the intelligent resource recommendation system re-ranks resources within the same priority level (i.e., the same resource stage category) according to the semantic similarity difference ranking sequence within each resource's resource stage category. This embodiment of the present invention utilizes a stable sorting algorithm based on Merge Sort, using semantic similarity difference as the sorting key. This fine-tunes resources within the same priority level, ensuring that resources with greater semantic similarity differences (i.e., better matching user needs) are ranked first, ultimately resulting in an initial resource recommendation list.

[0139] Continuing with the recommended learning resource list after the first sort, for example, there are multiple resources at the advanced level, such as "Python Data Visualization Intermediate Case Study" and "Python Data Visualization Advanced Theory Explanation." In the semantic similarity difference ranking sequence for the advanced level, the semantic similarity difference for "Python Data Visualization Intermediate Case Study" is 0.6, while the semantic similarity difference for "Python Data Visualization Advanced Theory Explanation" is 0.5. Using the merge sort algorithm, we sort these two resources based on semantic similarity difference, placing "Python Data Visualization Intermediate Case Study" ahead of "Python Data Visualization Advanced Theory Explanation." A similar operation is performed for all resources within the same priority level to obtain the initial resource recommendation list.

[0140] The embodiment of the present invention sorts resources hierarchically and logically based on two key dimensions: resource stage category and semantic matching degree. When dealing with diverse learning resources and complex user needs, it can significantly improve the rationality of the recommended resource list and user satisfaction, so that the recommended learning resources are highly matched with the actual needs of users, effectively alleviating the cold start problem and improving the accuracy of resource recommendations.

[0141] Furthermore, the multimodal learning resource intelligent recommendation system based on the AI ​​big model provided by the present invention is described below. The multimodal learning resource intelligent recommendation system based on the AI ​​big model described below and the multimodal learning resource intelligent recommendation method based on the AI ​​big model described above can be referenced to each other.

[0142] Reference Figure 2 , Figure 2 This is a schematic diagram of the structure of the multimodal learning resource intelligent recommendation system based on the AI ​​big model provided by the present invention. The multimodal learning resource intelligent recommendation system based on the AI ​​big model includes:

[0143] Intent analysis module 210, for performing intent analysis on the captured interactive behavior data and user information on the user learning platform based on the AI ​​big model to obtain a multi-dimensional user feature vector;

[0144] The resource integration module 220 is used to obtain multimodal resource data from the learning resource library based on the user feature vector, perform resource integration based on the multimodal resource data, and obtain a resource representation vector;

[0145] The knowledge network construction module 230 is used to mine the implicit associations between resources and knowledge points and the mapping relationship between user needs and target skills based on the knowledge graph combined with user feature vectors and resource representation vectors to obtain a knowledge network;

[0146] A learning resource generation module 240 is configured to generate candidate recommended learning resources based on the user's context-aware information and knowledge network;

[0147] The learning resource recommendation module 250 is used to optimize the ranking of candidate recommended learning resources based on the semantic similarity between the user feature vector and the resource representation vector and the current stage indicators of the multimodal resource data to generate an initial resource recommendation list.

[0148] The embodiments of the present invention overcome the limitations of traditional methods that rely solely on single interaction data through multimodal information fusion, achieving comprehensive capture of the content characteristics of learning resources. Furthermore, by utilizing knowledge graphs and context-aware technology, even new users or new resources can generate accurate resource recommendations based on semantic associations and contextual information, ensuring that the recommended learning resources are highly aligned with the user's actual needs, thereby improving the accuracy of resource recommendations. Therefore, the present invention significantly improves both resource recommendation accuracy and user satisfaction, demonstrating greater adaptability and intelligence when processing multimodal data and dynamic learning scenarios.

[0149] See also Figure 3 , Figure 3 This is a diagram of an embodiment of an electronic device provided by an embodiment of the present invention. Figure 3 As shown, an embodiment of the present invention provides an electronic device 300, including a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor 320. When the processor 320 executes the computer program 311, the following steps are implemented:

[0150] Based on the AI ​​big model, the captured interactive behavior data and user information on the user learning platform are analyzed for intent to obtain multi-dimensional user feature vectors;

[0151] Acquire multimodal resource data from the learning resource library based on the user feature vector, and integrate resources based on the multimodal resource data to obtain a resource representation vector;

[0152] Based on the knowledge graph, combined with user feature vectors and resource representation vectors, the implicit associations between resources and knowledge points, as well as the mapping relationship between user needs and target skills, are mined to obtain a knowledge network.

[0153] Generate candidate recommended learning resources based on the user's context-aware information and knowledge network;

[0154] Based on the user feature vector and resource representation vector combined with the current stage indicators of multimodal resource data, the candidate recommended learning resources are ranked and optimized to generate an initial resource recommendation list.

[0155] See also Figure 4 , Figure 4 Detailed description of an embodiment of a computer-readable storage medium provided by an embodiment of the present invention. Figure 4 As shown, this embodiment provides a computer-readable storage medium 400 on which a computer program 311 is stored. When the computer program 311 is executed by a processor, the following steps are implemented:

[0156] Based on the AI ​​big model, the captured interactive behavior data and user information on the user learning platform are analyzed for intent to obtain multi-dimensional user feature vectors;

[0157] Acquire multimodal resource data from the learning resource library based on the user feature vector, and integrate resources based on the multimodal resource data to obtain a resource representation vector;

[0158] Based on the knowledge graph, combined with user feature vectors and resource representation vectors, the implicit associations between resources and knowledge points, as well as the mapping relationship between user needs and target skills, are mined to obtain a knowledge network.

[0159] Generate candidate recommended learning resources based on the user's context-aware information and knowledge network;

[0160] Based on the user feature vector and resource representation vector combined with the current stage indicators of multimodal resource data, the candidate recommended learning resources are ranked and optimized to generate an initial resource recommendation list.

[0161] On the other hand, the present invention further provides a computer program product, which includes a computer program. The computer program can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer can execute the multimodal learning resource intelligent recommendation method based on the AI ​​large model provided by the above methods, which includes:

[0162] Based on the AI ​​big model, the captured interactive behavior data and user information on the user learning platform are analyzed for intent to obtain multi-dimensional user feature vectors;

[0163] Acquire multimodal resource data from the learning resource library based on the user feature vector, and integrate resources based on the multimodal resource data to obtain a resource representation vector;

[0164] Based on the knowledge graph, combined with user feature vectors and resource representation vectors, the implicit associations between resources and knowledge points, as well as the mapping relationship between user needs and target skills, are mined to obtain a knowledge network.

[0165] Generate candidate recommended learning resources based on the user's context-aware information and knowledge network;

[0166] Based on the user feature vector and resource representation vector combined with the current stage indicators of multimodal resource data, the candidate recommended learning resources are ranked and optimized to generate an initial resource recommendation list.

[0167] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0168] Through the above description of the embodiments, those skilled in the art will clearly understand that each embodiment can be implemented using software plus a necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the essence of the above technical solution, or the portion 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 ROM / RAM, a magnetic disk, or an optical disk, and includes a number of instructions for causing a computer device (such as a personal computer, server, or network device) to execute the methods described in each embodiment or certain portions of the embodiments.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A multimodal learning resource intelligent recommendation method based on AI large model, characterized by: include: Based on the AI ​​big model, the captured interactive behavior data and user information on the user learning platform are analyzed for intent to obtain multi-dimensional user feature vectors; Acquiring multimodal resource data from a learning resource library based on the user feature vector, and integrating resources based on the multimodal resource data to obtain a resource representation vector; Based on the knowledge graph, the user feature vector and the resource representation vector are combined to mine the implicit association between resources and knowledge points and the mapping relationship between user needs and target skills to obtain a knowledge network; generating candidate recommended learning resources based on the user's context-aware information and the knowledge network; Optimize the ranking of the candidate recommended learning resources based on the user feature vector and the resource representation vector combined with the current stage index of the multimodal resource data to generate an initial resource recommendation list; The method of mining the implicit associations between resources and knowledge points and the mapping relationship between user needs and target skills based on the knowledge graph in combination with the user feature vector and the resource representation vector to obtain a knowledge network includes: Anchor matching is performed on the semantic information in the user feature vector and the resource representation vector with the corresponding semantic nodes and relationship paths in the preset knowledge graph to obtain a semantic anchoring result; Taking the nodes and relationships corresponding to the semantic anchoring results as the starting point and the path extension length as the step length, the potential association paths between resources and knowledge points are explored forward in the preset knowledge graph, and the mapping connection paths between user needs and target skills are traced backward to obtain the semantic association path; Dynamically clustering the preset knowledge graph based on the semantic association path to generate subgraph clusters; each subgraph cluster represents a resource-knowledge point association or a user demand-target skill relationship; Determine the implicit association between resources and knowledge points based on the logical relationship between different sub-graph clusters, and determine the mapping relationship between user needs and target skills based on the degree of association between different sub-graph clusters; The knowledge network is constructed with resources, knowledge points, user needs and target skills as nodes and implicit associations and mapping relationships as edges.

2. The multimodal learning resource intelligent recommendation method based on AI large model according to claim 1 is characterized in that: The step of sorting and optimizing the candidate recommended learning resources based on the user feature vector and the resource representation vector in combination with the current stage indicator of the multimodal resource data to generate an initial resource recommendation list includes: determining a semantic similarity difference based on a first similarity score between the user feature vector and a benchmark semantic vector and a difference between a second similarity score between the resource representation vector and the benchmark semantic vector; Dividing each resource in the candidate recommended learning resources into stages based on the current stage indicator to obtain a resource stage category of each resource in the candidate recommended learning resources; Associating the resource stage category of each resource in the candidate recommended learning resources with the semantic similarity to construct an association matrix; Each resource in the candidate recommended learning resources is sorted and optimized based on the association matrix to generate the initial resource recommendation list.

3. The multimodal learning resource intelligent recommendation method based on AI large model according to claim 2 is characterized in that: The step of optimizing the ranking of each of the candidate recommended learning resources based on the association matrix to generate the initial resource recommendation list includes: For each of the candidate recommended learning resources, sort them according to semantic similarity differences based on matrix element values ​​in the association matrix, and generate a semantic similarity difference sorting sequence under the resource stage category of each resource; Sort each resource for the first time according to the priority order corresponding to the resource stage category of each resource, and recommend learning resources after obtaining the first ranking; For the resources with the same priority in the recommended learning resources after the first sorting, a second sorting is performed according to the semantic similarity difference sorting sequence under the corresponding resource stage category to obtain the initial resource recommendation list.

4. The multimodal learning resource intelligent recommendation method based on AI large model according to claim 1 is characterized in that: The context-aware information includes learning progress information and learning stage information; and the generating of candidate recommended learning resources based on the user's context-aware information and the knowledge network includes: generating a context feature vector based on the learning progress information and the learning stage information, and matching the context feature vector with a user demand node in the knowledge network to obtain a network subgraph; Determining initial recommended learning resources based on the learning resource data associated with each subgraph node in the network subgraph, and determining a resource association degree between each resource in the initial recommended learning resources and the context feature vector based on a correlation between each resource and the context feature vector; A fusion operation is performed based on element-wise multiplication of the context feature vector and the vector corresponding to the target skill node in the knowledge network to obtain a resource matching degree; The initially recommended learning resources are screened based on the resource matching degree and the resource association degree between each resource and the context feature vector to obtain the candidate recommended learning resources.

5. The method for intelligent recommendation of multimodal learning resources based on AI big model according to any one of claims 1 to 4, characterized in that: The AI-based big model performs intent analysis on the captured interactive behavior data and user information on the user learning platform to obtain a multi-dimensional user feature vector, including: Constructing a topological network with the interaction behavior data and the user information as entity nodes and the logical connections between the interaction behavior and the user information as edges; Based on the interactive behavior data with time series characteristics, dynamic temporal convolution embedding operation is performed to embed the time dimension information into the interactive behavior data, capture the local patterns and dependencies of user behavior over time, and obtain the data after dynamic temporal convolution embedding; Based on the AI ​​big model, semantic concepts are used as units to perform diffusion propagation in the topological network to obtain the results of semantic diffusion propagation; Intent analysis is performed based on the results of the semantic diffusion propagation and the data after the dynamic temporal convolution embedding to obtain a multi-dimensional user feature vector.

6. The multimodal learning resource intelligent recommendation method based on AI large model according to claim 5 is characterized in that: The intention analysis is performed based on the semantic diffusion propagation results and the dynamic temporal convolution embedded data to obtain a multi-dimensional user feature vector, including: Based on the semantic diffusion propagation results and the dynamic temporal convolution embedding data, cluster analysis is performed in the spatiotemporal dimension to aggregate data with similar learning intention patterns to obtain different learning intention pattern clusters; Between different learning intention pattern clusters, cross-domain intention association deduction is performed based on semantic information and data association relationships to obtain the potential connections between different learning intention pattern clusters in different fields or topics; Determining the demand probability distribution of the user's potential needs under different learning intention pattern clusters based on the probabilistic generation model in combination with the user information and different learning intention pattern clusters and their corresponding potential connections; Based on the integration of different learning intention pattern clusters and their corresponding potential connections, as well as the demand probability distribution under different learning intention pattern clusters, a multi-dimensional user feature vector is obtained.

7. A multimodal learning resource intelligent recommendation system based on AI big model, characterized by: Applicable to the multimodal learning resource intelligent recommendation method based on AI large model as described in any one of claims 1 to 6; The multimodal learning resource intelligent recommendation system based on the AI ​​large model includes: The intent analysis module is used to analyze the interaction behavior data and user information captured on the user learning platform based on the AI ​​model to obtain multi-dimensional user feature vectors; A resource integration module, configured to obtain multimodal resource data from a learning resource library based on the user feature vector, and perform resource integration based on the multimodal resource data to obtain a resource representation vector; A knowledge network construction module is used to mine the implicit associations between resources and knowledge points and the mapping relationship between user needs and target skills based on the knowledge graph in combination with the user feature vector and the resource representation vector to obtain a knowledge network; a learning resource generation module, configured to generate candidate recommended learning resources based on the user's contextual awareness information and the knowledge network; A learning resource recommendation module, configured to optimize the ranking of the candidate recommended learning resources based on the semantic similarity between the user feature vector and the resource representation vector, combined with the current stage indicators of the multimodal resource data, and generate an initial resource recommendation list; The method of mining the implicit associations between resources and knowledge points and the mapping relationship between user needs and target skills based on the knowledge graph in combination with the user feature vector and the resource representation vector to obtain a knowledge network includes: Anchor matching is performed on the semantic information in the user feature vector and the resource representation vector with the corresponding semantic nodes and relationship paths in the preset knowledge graph to obtain a semantic anchoring result; Taking the nodes and relationships corresponding to the semantic anchoring results as the starting point and the path extension length as the step length, the potential association paths between resources and knowledge points are explored forward in the preset knowledge graph, and the mapping connection paths between user needs and target skills are traced backward to obtain the semantic association path; Dynamically clustering the preset knowledge graph based on the semantic association path to generate subgraph clusters; each subgraph cluster represents a resource-knowledge point association or a user demand-target skill relationship; Determine the implicit association between resources and knowledge points based on the logical relationship between different sub-graph clusters, and determine the mapping relationship between user needs and target skills based on the degree of association between different sub-graph clusters; The knowledge network is constructed with resources, knowledge points, user needs and target skills as nodes and implicit associations and mapping relationships as edges.

8. An electronic device comprising: Memory for storing computer software programs; A processor for reading and executing the computer software program, characterized in that when the processor executes the computer software program, it implements the multimodal learning resource intelligent recommendation method based on the AI ​​large model as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium storing a computer software program, wherein: When the computer software program is executed by a processor, it implements the multimodal learning resource intelligent recommendation method based on the AI ​​large model as described in any one of claims 1 to 6.

Images