Personalized teaching resource recommendation method and system for large-scale users

Abstract

The present application provides a personalized teaching resource recommendation method and system for large-scale users. The method includes: constructing a dynamic interactive network; extracting teaching resources that users are already interested in and teaching resources that users are about to be interested in from the dynamic interactive network; establishing a multi-layer interest graph for each user based on the teaching resources that users are already interested in and the teaching resources that users are about to be interested in; determining a personalized teaching resource recommendation list for each user based on the multi-layer interest graph of each user, performing a comparative analysis of the local teaching resources of interest and the global teaching resources of interest in the personalized teaching resource recommendation list, and optimizing the multi-layer interest graph based on the comparative results; and co-training the dynamic interactive network on multiple edge devices through federated learning to optimize the network structure of the dynamic interactive network. The technical solution provided by the present application improves the accuracy of personalized teaching resource recommendations for large-scale users.

Description

Technical Field

[0001] The embodiments of the present application relate to the technical field of personalized recommendation of teaching resources for smart campuses, and in particular, to a personalized teaching resource recommendation method and system for large-scale users. Background Art

[0002] With the rapid development of online education, the demand for personalized teaching resource recommendation systems is increasing. The system can provide customized learning materials and course recommendations based on each student's learning habits, interests, and historical interaction behaviors, thereby improving learning efficiency and personal satisfaction. Especially in a large-scale user environment, how to effectively provide each user with teaching resources that match their interests has become a key challenge. The application scenarios are not limited to school education, but also cover multiple fields such as vocational training and self-improvement, meeting the diverse learning needs of users of different ages and backgrounds.

[0003] The existing personalized teaching resource recommendation method includes the following steps: obtaining user interaction behavior data, performing data preprocessing on the user interaction behavior data, and obtaining a user resource rating matrix; performing feature dimension reduction on the user resource rating matrix, and obtaining a user teaching resource feature matrix; clustering the teaching resource feature matrix, and obtaining teaching resource clusters, and sorting the teaching resources in the teaching resource clusters; obtaining the user's ratings for all teaching resources, and using the teaching resource interest model in turn to calculate the user's interest in all teaching resources, and arranging all teaching resources in descending order according to the interest, and generating a teaching resource recommendation list.

[0004] Most existing personalized teaching resource recommendation methods are based on static data analysis, which makes it difficult to adapt to rapidly changing user interests and emerging teaching resources. Summary of the invention

[0005] The embodiments of the present application provide a method and system for recommending personalized teaching resources for large-scale users, so as to solve the problem of low accuracy of personalized teaching resource recommendations for large-scale users existing in the prior art.

[0006] In a first aspect, the present application embodiment provides a personalized teaching resource recommendation method for large-scale users, including: Based on the historical interaction behavior data between large-scale users and teaching resources, a dynamic interaction network is constructed using a quantum annealing algorithm, where the number of large-scale users exceeds a first preset number threshold; Extracting teaching resources that the user has been interested in and teaching resources that the user is likely to be interested in from the dynamic interactive network, and establishing a multi-layer interest graph for each user based on the teaching resources that the user has been interested in and the teaching resources that the user is likely to be interested in; Determine a personalized teaching resource recommendation list for each user based on the multi-layer interest graph of each user, compare and analyze the local teaching resources of interest and the global teaching resources of interest in the personalized teaching resource recommendation list, and optimize the multi-layer interest graph based on the comparison results; The dynamic interaction network is collaboratively trained on multiple edge devices through federated learning to optimize the network structure of the dynamic interaction network.

[0007] Optionally, the historical interaction behavior data includes: user identifier, user personal information, user preference information, interaction mode, teaching resource identifier, teaching resource metadata, teaching resource format and user feedback information; The method uses quantum annealing algorithm to construct a dynamic interaction network based on historical interaction behavior data between large-scale users and teaching resources, including: Define nodes and edges, a node is a user node or a teaching resource node, the attribute information of the user node includes: user identifier, user personal information and user preference information, the attribute information of the teaching resource node includes: teaching resource identifier, teaching resource metadata and teaching resource format, the information of an edge includes the similarity between users, the similarity between teaching resources, or the association between users and teaching resources, the association includes the interaction mode and user feedback information, the attribute information of the node and the information of the edge are mapped to the Ising model, the stable state of the Ising model is determined by using the quantum annealing algorithm, and the stable state of the Ising model is used as the initial user and teaching resource association graph; When new interactive behavior data appears, determine the content to be updated and the teaching scene to which the content to be updated belongs, determine the update strategy corresponding to the teaching scene in combination with the scene characteristics of the teaching scene and the update content, and update the initial user and teaching resource association map according to the update strategy corresponding to the teaching scene to obtain an updated user and teaching resource association map; Structural optimization is performed on the updated user-teaching resource association graph to obtain an optimized user-teaching resource association graph; According to preset evaluation indicators, the optimized user-teaching resource association graph is cross-validated, and the user-teaching resource association graph that passes the cross-validation is determined as a dynamic interaction network. The preset evaluation indicators include recommendation accuracy and recall rate.

[0008] Optionally, the teaching scenario includes at least one of the following: an exam preparation scenario, an autonomous learning scenario, and an online classroom scenario; The initial user-teaching resource association graph is updated according to the update strategy corresponding to the teaching scenario to obtain an updated user-teaching resource association graph, including at least one of the following: Filter out key update data for the test preparation scenario from the content to be updated, and based on the key update data for the test preparation scenario, in combination with a first preset annealing time and a first preset Hamiltonian strength, use a batch processing and incremental update mechanism to update the attribute information of the nodes and the edge information in the initial user-teaching resource association graph, use natural language processing technology to perform sentiment analysis on the comment content in the newly added interactive behavior data, and adjust the edge weight in the corresponding edge information according to the sentiment analysis result, so as to obtain an updated user-teaching resource association graph; Filtering non-critical update data of the autonomous learning scenario from the content to be updated, and periodically updating the attribute information of the nodes and the edge information in the initial user-teaching resource association graph according to the non-critical update data of the autonomous learning scenario, in combination with a second preset annealing time and a second preset Hamiltonian strength, to obtain an updated user-teaching resource association graph; the second preset annealing time is greater than the first preset annealing time, and the second preset Hamiltonian strength is less than the first preset Hamiltonian strength; Filter out key update data and non-key update data of the online classroom scene from the content to be updated; based on the key update data of the online classroom scene, in combination with the first preset annealing time and the first preset Hamiltonian strength, use batch processing and incremental update mechanism to update the attribute information of the nodes and edge information in the initial user-teaching resource association graph; use natural language processing technology to perform sentiment analysis on the comment content in the newly added interactive behavior data, and adjust the edge weight in the corresponding edge information according to the sentiment analysis result to obtain an adjusted association graph; based on the non-key update data of the online classroom scene, in combination with the second preset annealing time and the second preset Hamiltonian strength, periodically update the adjusted association graph to obtain an updated user-teaching resource association graph; Among them, the critical update data includes: update data whose number of included update items exceeds a second preset number threshold, and / or update data of nodes whose number of affected nodes exceeds a third preset number threshold; the non-critical update data is other update data in the content to be updated except the critical update data. In different teaching scenarios, the value of the second preset number threshold is different, and the value of the third preset number threshold is different.

[0009] Optionally, the structural optimization of the updated user-teaching resource association graph to obtain an optimized user-teaching resource association graph includes: Using density-based clustering algorithm, identify sparse areas and dense areas in the updated user-teaching resource association graph; Merge nodes in sparse areas, supplement edges in sparse areas, prune edges in dense areas, and use community detection algorithms to divide dense areas into communities to obtain an intermediate association graph. The intermediate association graph is mapped to four-dimensional space-time, and abnormal nodes and abnormal edges in the intermediate association graph are detected in the four-dimensional space-time based on the lens effect and the user life cycle, and abnormal processing is performed on the abnormal nodes and abnormal edges to obtain an optimized user and teaching resource association graph, and the abnormal processing is adjustment or deletion.

[0010] Optionally, the using of a density-based clustering algorithm to identify sparse areas and dense areas in the updated user-teaching resource association graph includes: Initializing parameters of a density-based clustering algorithm, including a neighborhood radius and a minimum number of neighbors; Traverse each node in the updated user-teaching resource association graph, calculate the number of neighbor nodes in the domain determined by the domain radius, and if the number of neighbor nodes is greater than or equal to the minimum number of neighbors, determine the node as a core node, merge the core node and its reachable neighbor nodes into a cluster, evaluate the cluster, and determine whether the cluster is a dense area based on the cluster evaluation result; Alternatively, if the number of neighbor nodes is less than a preset node number threshold, the node is determined to be a boundary point or a noise point. The noise point is not a core node and is a node that is not reachable by any core node. The sparse area is determined based on the location distribution of the noise point and the minimum number of neighbors.

[0011] Optionally, the step of establishing a multi-layer interest graph for each user based on the teaching resources that the user has been interested in and the teaching resources that the user is expected to be interested in includes: Extracting the features of the teaching resources that the user has been interested in and the teaching resources that the user will be interested in at each layer from the teaching resource metadata; According to the characteristics of the teaching resources that the user is already interested in and the teaching resources that the user is to be interested in at each layer, combined with the weight of the teaching resource type and the weight of the layer, a multi-layer interest map for each user is established, wherein the multiple layers include a macro layer, a meso layer and a micro layer. The macro layer reflects the fields and subjects that the user is interested in, the meso layer reflects the knowledge points and topics that the user is interested in, and the micro layer reflects the resource characteristics and user behavior characteristics that the user is interested in. The resource characteristics include distributors, release time and content characteristics. Different teaching resource types correspond to different weights, and different layers correspond to different weights.

[0012] Optionally, determining a personalized teaching resource recommendation list for each user according to the multi-layer interest graph of each user includes: For each user, the fields and subjects of teaching resources that the user is not interested in are converted into a first vector group, and the fields and subjects that the user is interested in are converted into a second vector group, and the cosine similarity is used to calculate the interest similarity at the macro level; The knowledge points and topics of the teaching resources that the user is not interested in are converted into a third vector group, the knowledge points and topics that the user is interested in are converted into a fourth vector group, and the cosine similarity is used to calculate the interest similarity of the meso-level; The resource features and user behavior features of the teaching resources that the user is not interested in are converted into a fifth vector group, the resource features and user behavior features that the user is interested in are converted into a sixth vector group, and the cosine similarity is used to calculate the interest similarity of the micro-level; According to the interest similarity at the macro level, the interest similarity at the meso level, and the interest similarity at the micro level, multiple candidate teaching resources are screened out, where the candidate teaching resources are teaching resources that the user is not interested in and that meet the similarity design rules corresponding to the teaching scenario; The multiple candidate teaching resources are sorted, and a personalized teaching resource recommendation list for the user is generated according to the sorting result.

[0013] In a second aspect, the embodiment of the present application provides a personalized teaching resource recommendation system for large-scale users, including: A construction module is used to construct a dynamic interaction network using a quantum annealing algorithm based on historical interaction behavior data between large-scale users and teaching resources, where the number of large-scale users exceeds a first preset number threshold; An extraction and establishment module, used for extracting teaching resources that users have been interested in and teaching resources that users are expected to be interested in from the dynamic interactive network, and establishing a multi-layer interest graph for each user based on the teaching resources that users have been interested in and the teaching resources that users are expected to be interested in; Determine a comparison module, which is used to determine a personalized teaching resource recommendation list for each user according to the multi-layer interest map of each user, compare and analyze the local teaching resources of interest and the global teaching resources of interest in the personalized teaching resource recommendation list, and optimize the multi-layer interest map based on the comparison results; A training optimization module is used to collaboratively train the dynamic interaction network on multiple edge devices through federated learning to optimize the network structure of the dynamic interaction network.

[0014] In a third aspect, an embodiment of the present application provides a computing device, comprising a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a personalized teaching resource recommendation method for large-scale users as described in any one of the first aspects.

[0015] In a fourth aspect, an embodiment of the present application provides a computer storage medium storing a computer program. When the computer program is executed by a computer, it implements a personalized teaching resource recommendation method for a large number of users as described in any one of the first aspects.

[0016] In an embodiment of the present application, a personalized teaching resource recommendation method for large-scale users is provided, and the method includes the following steps: constructing a dynamic interaction network using a quantum annealing algorithm based on historical interaction behavior data between large-scale users and teaching resources, wherein the number of large-scale users exceeds a first preset number threshold; extracting teaching resources that users are already interested in and teaching resources that users are to be interested in from the dynamic interaction network, and establishing a multi-layer interest graph for each user based on the teaching resources that users are already interested in and the teaching resources that users are to be interested in; determining a personalized teaching resource recommendation list for each user based on the multi-layer interest graph of each user, performing a comparative analysis of local teaching resources of interest and global teaching resources of interest in the personalized teaching resource recommendation list, and optimizing the multi-layer interest graph based on the comparison results; and collaboratively training the dynamic interaction network on multiple edge devices through federated learning to optimize the network structure of the dynamic interaction network.

[0017] The personalized teaching resource recommendation method for large-scale users provided in this application can reflect the latest preferences and behavior patterns of large-scale users in real time by constructing a dynamic interactive network using a quantum annealing algorithm, so that the graph structure in the dynamic interactive network can be updated and adjusted according to the latest interactive behavior data. The dynamic interactive network can also capture the complex relationships between users and between users and resources, which helps to discover potential points of interest and provide users with more personalized and timely recommendations. The design of the multi-layer interest graph allows the capture of user interests at different levels, from macroscopic fields and disciplines to microscopic specific resource characteristics and behavioral characteristics. This multi-level method can more comprehensively describe the user's interest distribution than simple feature dimensionality reduction, thereby reducing the risk of information loss. Combining a multi-layer interest graph to generate a personalized recommendation list can not only improve the relevance and accuracy of the recommendation, but also increase the diversity of the recommended content, so that the recommendation list can better meet the actual needs of the user and improve the user's satisfaction and user experience. By comparing and analyzing the local and global interest teaching resources in the personalized recommendation list, we can effectively explore the correlation between users' explicit preferences and potential learning needs, dynamically calibrate the semantic representation level of the multi-layer interest graph, and thus improve the interest coverage breadth and long-tail resource discovery capabilities of the recommendation system; at the same time, based on the distributed collaborative training mechanism of federated learning, while ensuring user privacy and security, the computing resources of edge nodes are used to realize the asynchronous aggregation and structural evolution of parameters of the dynamic interactive network, which not only enhances the response efficiency of the model to massive user concurrent requests, but also continuously optimizes the network topology through cross-device knowledge migration, so that the recommendation system has the elastic expansion capability to dynamically adapt to the update of teaching resources and the drift of user interests, and finally forms a personalized recommendation ecology with a balance between accuracy and diversity. Furthermore, for different teaching scenarios (such as exam preparation, self-study, and online classes), specific update strategies are adopted to process critical and non-critical update data, ensuring the high relevance and pertinence of the recommended content and meeting the needs of different learning situations. A density-based clustering algorithm is used to identify sparse and dense areas, and the graph structure is optimized by merging nodes, supplementing edges, pruning, and community division, further improving the quality of the graph, which helps to improve the accuracy and efficiency of the recommendation system.

[0018] These and other aspects of the present application will become more clearly understood in the description of the following embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, a brief introduction will be given below to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0020] Figure 1 A flowchart of a method for recommending personalized teaching resources for large-scale users provided in an embodiment of the present application; Figure 2 A structural diagram of a personalized teaching resource recommendation system for large-scale users provided in an embodiment of the present application; Figure 3 A schematic diagram of the structure of a computing device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0021] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application.

[0022] In some of the processes described in the specification and claims of this application and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or executed in parallel. The sequence numbers of the operations, such as 11, 12, etc., are only used to distinguish between different operations, and the sequence numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., do not represent the order of precedence, and do not limit the "first" and "second" to different types.

[0023] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of this application.

[0024] Figure 1 A flowchart of a method for recommending personalized teaching resources for large-scale users provided in an embodiment of the present application is shown in FIG. Figure 1 As shown, the method includes: S11. Based on the historical interaction behavior data between large-scale users and teaching resources, a dynamic interaction network is constructed using a quantum annealing algorithm, where the number of large-scale users exceeds a first preset number threshold. The first preset number threshold can be any integer of hundreds, thousands, tens of thousands, hundreds of thousands, millions, tens of millions, or other orders of magnitude.

[0025] Among them, the quantum annealing algorithm is used to efficiently handle the dynamic association modeling problem between ultra-large-scale users and teaching resources. In addition, the construction of the dynamic interaction network is based on the historical interaction behavior data between large-scale users and teaching resources, where the number of users exceeds a specific threshold, ensuring the scale and diversity of the dynamic interaction network. The dynamic interaction network not only records the user's interaction methods such as access, browsing, and downloading of different teaching resources, but also combines user feedback information to more accurately understand user interests and their changing trends. By analyzing these interaction patterns, potential points of interest and learning paths can be identified, providing basic data support for subsequent steps. The specific process of constructing the dynamic interaction network is described in steps 111 to 114 below, which will not be repeated here.

[0026] For example, on a large online education platform, the personalized teaching resource recommendation system collects learning activity data of millions of users and establishes an initial user-teaching resource association graph. As the new semester begins, a large number of students enter the exam preparation stage. The personalized teaching resource recommendation system identifies this key update data and applies a batch processing mechanism to quickly update the attribute information of the nodes and edge information in the association graph to ensure the real-time and accuracy of the recommendation system.

[0027] S12. Extract teaching resources that the user is already interested in and teaching resources that the user is to be interested in from the dynamic interactive network, and establish a multi-layer interest graph for each user based on the teaching resources that the user is already interested in and the teaching resources that the user is to be interested in.

[0028] Among them, the teaching resources that the user is already interested in are teaching resources that the user has interacted with and whose user rating value is greater than a preset rating threshold, the teaching resources that the user is to be interested in are teaching resources that similar users are already interested in and the user has not interacted with, and similar users are other users whose similarity is greater than the first preset similarity threshold.

[0029] Optionally, this embodiment may use a collaborative filtering algorithm to analyze user behavior, determine similar user groups, and mine potential points of interest based on this, and then combine the user's historical preferences and the preferences of similar users to make personalized resource recommendations.

[0030] For example, the personalized teaching resource recommendation system identifies a student A who is particularly interested in programming, and also finds that another student B who is highly similar to him also shows a strong interest in machine learning. Based on this, the personalized teaching resource recommendation system recommends several high-quality introductory machine learning courses to student A as teaching resources that may be of interest to him.

[0031] Optionally, the multiple layers include a macro layer, a meso layer and a micro layer, the macro layer reflects the fields and subjects that the user is interested in, the meso layer reflects the knowledge points and topics that the user is interested in, and the micro layer reflects the resource characteristics and user behavior characteristics that the user is interested in, and the resource characteristics include distributors, release time and content characteristics.

[0032] It should be understood that the multi-layer interest graph can fully reflect the user's interest level, from the macro-level subject interests to the micro-level specific resource characteristics, covering details such as knowledge points, topics, publisher background, etc. Different levels correspond to different weights, which helps to deeply understand and predict the user's interest focus.

[0033] Specifically, this embodiment can extract the features of each layer from the metadata of teaching resources, combine the resource type and the weight of the layer, and build a personalized interest map for each user. This map not only reflects the user's current interests, but also provides clues for the development of future interests.

[0034] For example, the personalized teaching resource recommendation system builds a detailed multi-layer interest map for Student A, which not only covers the field of computer science, but also includes specific skill points such as Python programming language and data analysis tools, and is further refined to courses related to the latest advances in artificial intelligence technology.

[0035] S13. Determine a personalized teaching resource recommendation list for each user based on the multi-layer interest graph of each user, perform comparative analysis on the local teaching resources of interest and the global teaching resources of interest in the personalized teaching resource recommendation list, and optimize the multi-layer interest graph based on the comparative results.

[0036] Correspondingly, this step generates a personalized teaching resource recommendation list based on the user's multi-layer interest graph, and by calculating the interest similarity at different levels, screens and sorts the candidate resources to meet the user's personalized needs. Specifically, this embodiment can convert the resources that the user is not interested in into a vector group, calculate the cosine similarity between it and the resources that the user is interested in, and comprehensively consider the macro, meso, and micro interest similarities to select the best matching resources and generate a personalized teaching resource recommendation list.

[0037] For example, for student A, the personalized teaching resource recommendation system calculates the similarity scores of a series of resources that student A has not been exposed to but may be interested in based on his interest graph, and finally generates a personalized recommendation list including advanced courses on the latest advances in artificial intelligence technology and personal development lectures.

[0038] It should be understood that local teaching resources of interest can refer to teaching resources that are highly matched with the user's current multi-layer interest graph, usually including: recent interactive resources, similar user preference resources, interest graph direct coverage resources, etc. Among them, recent interactive resources can refer to resources that the user has recently visited, rated or collected (such as in the past 7 days). Similar user preference resources can refer to resources that other users with high similarity to the user's interests (similarity> preset threshold) are recently interested in. Interest graph direct coverage resources can refer to extended resources in the same field, knowledge point or feature of the user's already interested resources.

[0039] In addition, global teaching resources of interest can refer to resources with broad universality or long-tail value on the platform, including: popular resources, emerging trend resources, cold start resources, etc. Among them, popular resources can refer to resources with high frequency access and high ratings by users across the platform. Emerging trend resources can refer to resources that have been recently released and are rapidly gaining popularity (such as a weekly visit growth rate of >50%). Cold start resources can refer to high-quality resources that have been newly released but have not yet accumulated enough interactive data (such as "potential resources" manually labeled by the platform).

[0040] For example, the local resources of user A may include the "Python Basics" he recently learned, and the global resources may include the popular "XXX Introduction" on the entire platform or the newly released "Introduction to Quantum Computing".

[0041] It should also be understood that the process of comparative analysis and optimization of multi-layer interest graphs based on comparative results is as follows: for each level (macro / meso / micro), the weight difference between local and global interests is calculated, and the level weight of the user's interest graph is corrected proportionally. For example, if the global trend shows that the macro-level weight in the field of "artificial intelligence" is 20% higher than the user's current level, the macro-weight of "artificial intelligence" in the user's interest graph is increased, and more relevant resources are recommended.

[0042] S14. The dynamic interaction network is collaboratively trained on multiple edge devices through federated learning to optimize the network structure of the dynamic interaction network.

[0043] It should be understood that federated learning is a distributed machine learning framework that allows multiple edge devices to perform model training on local data without sharing the original data.

[0044] Federated learning optimizes its network structure through the following steps: 1. Initialize the global model. Specifically, the central server is used to initialize the global model parameters of the dynamic interactive network (such as node embedding vectors and edge weight matrices). The distribution model is used to distribute the initial model parameters to each edge device (such as the regional education platform server).

[0045] 2. Local training of edge devices, specifically, input data so that each edge device uses the interaction data of local users (such as click records, learning time, and ratings) to train the model. The training objectives include node attribute updates, edge weight adjustments, and structural optimization, where node attribute updates can be understood as optimizing the embedded representation of user nodes (such as learning preferences) and resource nodes (such as content features). Edge weight adjustment can be understood as dynamically adjusting the weight of the edge according to the intensity of interaction (such as access frequency, sentiment analysis results). Structural optimization can be understood as pruning redundant edges through local graph clustering (such as community detection) to enhance the connectivity of sparse areas. In an embodiment of the present application, privacy protection technologies include differential privacy, homomorphic encryption, etc., where differential privacy means adding noise when the model parameters are updated to prevent the original data from being inferred. Homomorphic encryption means encrypting the local parameters before uploading them, and decrypting them after aggregation by the central server.

[0046] 3. Parameter aggregation and global update, including two steps: secure aggregation and dynamic interactive network optimization. In order to achieve secure aggregation, the central server uses the federated average algorithm to integrate the model updates of each device. In order to achieve dynamic interactive network optimization, the user / resource node embedding vectors of each device are weighted averaged to generate a globally consistent representation to achieve node embedding fusion. The global edge weights are adjusted according to the cross-device interaction mode (such as the weight increase of popular resources) to achieve edge weight synchronization. The low-weight edges in the global model are removed to perform structural pruning, thereby simplifying the network topology.

[0047] The embodiment of the present application can synchronize the parameters of the network structure of the optimized dynamic interaction network to all edge devices for the next round of training until the model converges or reaches a preset number of iterations.

[0048] By executing S11~S14, the dynamic interactive network constructed by the embodiment of the present application can reflect the latest preferences and behavior patterns of large-scale users in real time, so that the graph structure in the dynamic interactive network can be updated and adjusted according to the latest interactive behavior data. The dynamic interactive network can also capture the complex relationship between users and between users and resources, which helps to discover potential points of interest and provide users with more personalized and timely recommendations. The potential interests of the target user are inferred by the interests of other users whose similarity is greater than the first preset similarity threshold. This method increases the robustness of the personalized teaching resource recommendation system and reduces the deviation caused by individual abnormal scores. The design of the multi-layer interest map allows the user's interests to be captured at different levels, from macroscopic fields and disciplines to microscopic specific resource features and behavioral characteristics. This multi-level method can more comprehensively describe the user's interest distribution than simple feature dimensionality reduction, thereby reducing the risk of information loss. The macro, meso and micro levels focus on different interest dimensions, ensuring that the recommendation results can meet the user's long-term learning needs and respond to short-term changes in learning interests. Combining multi-layer interest graphs to generate personalized recommendation lists can not only improve the relevance and accuracy of recommendations, but also increase the diversity of recommended content, making the recommendation list more in line with the actual needs of users and improving user satisfaction and usage experience. By comparing and analyzing the local and global interest teaching resources in the personalized recommendation list, it is possible to effectively explore the correlation between users' explicit preferences and potential learning needs, dynamically calibrate the semantic representation level of the multi-layer interest graph, thereby improving the interest coverage breadth and long-tail resource discovery capabilities of the recommendation system; at the same time, based on the distributed collaborative training mechanism of federated learning, while ensuring user privacy and security, the computing resources of edge nodes are used to achieve asynchronous aggregation and structural evolution of parameters of dynamic interactive networks, which not only enhances the model's response efficiency to massive user concurrent requests, but also continuously optimizes the network topology through cross-device knowledge migration, so that the recommendation system has the elastic expansion capability to dynamically adapt to teaching resource updates and user interest drift, and finally forms a personalized recommendation ecosystem with a balance between accuracy and diversity.

[0049] In a possible embodiment, the historical interaction behavior data includes: user identifier, user personal information, user preference information, interaction mode, teaching resource identifier, teaching resource metadata, teaching resource format and user feedback information; the present application embodiment does not specifically limit the content of the historical interaction behavior data. An accurate dynamic interaction network is the basis for improving the accuracy of personalized teaching resource recommendations. Therefore, the present application embodiment specifically describes step S11 as follows.

[0050] S11. Based on the historical interaction behavior data between large-scale users and teaching resources, a dynamic interaction network is constructed using the quantum annealing algorithm, including: Step 111, define nodes and edges, a node is a user node or a teaching resource node, the attribute information of the user node includes: user identifier, user personal information and user preference information, the attribute information of the teaching resource node includes: teaching resource identifier, teaching resource metadata and teaching resource format, the information of an edge includes the similarity between users, the similarity between teaching resources, or the association between users and teaching resources, the association includes the interaction method and user feedback information, the attribute information of the node and the information of the edge are used to determine the stable state of the Ising model using the quantum annealing algorithm, and the stable state of the Ising model is used as the initial user and teaching resource association graph.

[0051] Among them, user personal information includes age, school, college, subject, elective courses and other information, and user preference information types include: learning field preference, resource type preference, interaction method preference, time arrangement preference, time arrangement preference, learning progress preference, learning progress preference, evaluation preference, social network preference and technology preference.

[0052] Specifically, learning field preferences refer to the subject or professional field that users are interested in, such as mathematics, physics, computer science, etc. Resource type preferences refer to the types of teaching resources that users tend to use, such as video tutorials, online courses, e-books, interactive exercises, etc. Interaction method preferences refer to the way users prefer to interact with the platform or other users, such as forum discussions, real-time Q&A, social media sharing, etc. Schedule preferences refer to the time pattern of when users tend to conduct learning activities, such as mornings, evenings, or weekends. Learning progress preferences refer to whether users want to learn at their own pace or follow a fixed course schedule. Evaluation preferences refer to understanding users' likes and dislikes of specific types of content and their standards for resource quality based on their rating history. Social network preferences include: whether they are willing to participate in study groups or prefer to study independently. Technology preferences refer to the technical devices that users prefer to use, such as mobile phones, tablets, or desktop computers.

[0053] Teaching resource metadata includes descriptive information, technical requirements, resource characteristics, etc. Resource characteristics include publisher, publishing time and content characteristics. Teaching resource formats can exist in many forms, each of which has its own characteristics and applicable scenarios. For example, text materials include documents, files, e-books, etc., which are suitable for in-depth reading and detailed study. Videos include recorded classes, live lectures, animated demonstrations, etc., which are very suitable for visual learners and situations where concepts need to be presented dynamically. Audio, such as podcasts and audiobooks, are convenient for use on the move or when it is inconvenient to watch the screen. Interactive content includes online quizzes, virtual laboratories, simulators, etc., allowing users to directly participate in the learning process. Slides / presentations are used to summarize key points or provide structured information presentation. Images and charts can be static pictures or dynamic charts, which help explain complex information or show relationships. In addition, for programming courses, code repositories on designated platforms are important learning resources. Augmented reality and virtual reality technologies provide immersive learning experiences, which are suitable for certain skill training or exploratory learning.

[0054] It should be noted that the specific process of step 111 can be replaced by the following exemplary process: When the user scale reaches millions or more, the traditional graph computing method will be difficult to capture the nonlinear interaction relationship between users and resources in real time due to the exponential growth of computational complexity. The quantum annealing algorithm transforms the construction of a dynamic interaction network into a combinatorial optimization problem, and uses the quantum tunneling effect to quickly search for the global optimal solution in the solution space: first, the user-resource historical behavior (such as clicks, learning time, test feedback) is encoded into a weighted graph structure, where nodes represent user and resource entities, and edge weights reflect the interaction intensity and time decay characteristics; second, the potential energy of different network topologies is evaluated in parallel through the super-position state of quantum bits, with the objective function of minimizing the network reconstruction error (such as the conflict of user implicit interests and the sparsity of resource associations), and quickly generating a dynamic interaction graph that meets the time evolution constraints. This process breaks through the local optimal trap of classical algorithms when dealing with high-dimensional sparse data. It is particularly suitable for capturing the interest drift of massive users (such as changes in course popularity and knowledge association migration) in real time. At the same time, it retains the weak correlation signals of long-tail resources through the probabilistic sampling mechanism of quantum annealing, providing low-noise and high-information relationship primitives for the subsequent construction of multi-layer interest graphs.

[0055] Accordingly, step 111 ensures the accuracy and richness of the initial data structure, and through detailed node attribute definition, it can more accurately capture the user's interests and preferences, as well as the specific characteristics of educational resources, thereby improving the relevance and accuracy of recommendations.

[0056] Step 112: When new interactive behavior data appears, determine the content to be updated and the teaching scenario to which the content to be updated belongs, and determine the update strategy corresponding to the teaching scenario in combination with the scenario characteristics and update content of the teaching scenario. Update the initial user and teaching resource association map according to the update strategy corresponding to the teaching scenario to obtain an updated user and teaching resource association map.

[0057] It should be understood that the teaching scenario includes at least one of the following: exam preparation scenario, autonomous learning scenario and online classroom scenario; the scenario characteristics of the exam preparation scenario include learning environment, content characteristics and evaluation methods, the scenario characteristics of the autonomous learning scenario include: path planning characteristics, interactivity characteristics and collaboration characteristics, and the scenario characteristics of the online classroom scenario include: implementation interactivity, learning progress tracking and learning feedback efficiency. Step 112 considers the different needs of different teaching scenarios (such as exam preparation, autonomous learning or online classroom), and formulates corresponding update strategies accordingly. For example, in the exam preparation scenario, key update data will be processed first so as to quickly adjust the relationship between the associated users and teaching resources.

[0058] Step 112 takes into account the changes in resource requirements for different types of learning activities, so that the personalized teaching resource recommendation system can quickly respond and adjust the recommended content, thereby enhancing the flexibility and adaptability of the personalized teaching resource recommendation system, enabling it to provide resource recommendations that are more in line with actual needs in different learning scenarios, thereby improving user experience satisfaction.

[0059] Specifically, in step 112, the initial user-teaching resource association map is updated according to the update strategy corresponding to the teaching scenario, and an updated user-teaching resource association map is obtained, including at least one of the following: (1) Filter out the key update data of the exam preparation scenario from the content to be updated, and based on the key update data of the exam preparation scenario, combine the first preset annealing time and the first preset Hamiltonian strength, use batch processing and incremental update mechanism to update the attribute information of the nodes and edge information in the initial user-teaching resource association graph, use natural language processing technology to perform sentiment analysis on the comments in the newly added interactive behavior data, and adjust the edge weight in the corresponding edge information according to the sentiment analysis results to obtain an updated user-teaching resource association graph. Among them, the key update data of the exam preparation scenario includes newly released review materials, simulation test questions, important announcements, etc.

[0060] For example, with the arrival of the new semester, many students are entering a tense exam preparation phase, and the personalized teaching resource recommendation system detects a large number of new interactions about mathematics and physics courses. In response to this trend, the system immediately adopts a targeted update strategy to efficiently update the nodes and edges related to exam preparation, ensuring that the latest interaction information can be reflected in the association graph in a timely manner. Among them, batch processing and incremental update mechanisms can improve rapid response.

[0061] (2) Non-critical update data of the autonomous learning scenario are selected from the content to be updated, and based on the non-critical update data of the autonomous learning scenario, the attribute information of the nodes and the edge information in the initial user-teaching resource association graph are periodically updated in combination with a second preset annealing time and a second preset Hamiltonian strength to obtain an updated user-teaching resource association graph. The second preset annealing time is greater than the first preset annealing time, and the second preset Hamiltonian strength is less than the first preset Hamiltonian strength.

[0062] It should be understood that non-critical update data in the autonomous learning scenario mainly includes daily learning progress records, slight changes in interests, etc. This type of data has little impact on the overall network structure. Therefore, the demand can be met by using a periodic update method without the need for an immediate response.

[0063] (3) Filter out the key update data and non-key update data of the online classroom scene from the content to be updated, and update the attribute information of the nodes and edge information in the initial user-teaching resource association graph by using the batch processing and incremental update mechanism according to the key update data of the online classroom scene, and use the natural language processing technology to perform sentiment analysis on the comment content in the newly added interactive behavior data in combination with the first preset annealing time and the first preset Hamiltonian strength, and adjust the edge weight in the corresponding edge information according to the sentiment analysis result to obtain an adjusted association graph; according to the non-key update data of the online classroom scene, in combination with the second preset annealing time and the second preset Hamiltonian strength, periodically update the adjusted association graph to obtain an updated user-teaching resource association graph. Exemplarily, the first preset annealing time is 30 minutes, the second preset annealing time is 24 hours, and it is assumed that the range of the Hamiltonian strength is 0~1, the first preset Hamiltonian strength is 0.9, and the second preset Hamiltonian strength is 0.3. Only when the amount of data or the scope of influence exceeds the threshold, a high-sensitivity fast update (30 minutes annealing and 0.9 intensity) is triggered to avoid frequent processing of small-scale data. Even if a small amount of data reaches the threshold (30 pieces / day), the cost is reduced through periodic batch updates (24 hours annealing and 0.3 intensity).

[0064] It should be understood that in the online classroom scenario, there are both critical update data that require a quick response (such as feedback from live classes and real-time Q&A) and non-critical update data that can be processed periodically (such as homework submission status). This hybrid update strategy ensures that it can quickly respond to emergency changes while maintaining the stable operation of the personalized teaching resource recommendation system.

[0065] Among them, in the above teaching scenarios, the critical update data includes: update data in which the number of included update items exceeds the second preset number threshold, and / or update data of nodes in which the number of affected nodes exceeds the third preset number threshold; non-critical update data is other update data in the content to be updated except the critical update data. In different teaching scenarios, the value of the second preset number threshold is different, and the value of the third preset number threshold is different.

[0066] Exemplarily, the second preset quantity threshold in the exam preparation scenario can be 100, and the third preset quantity threshold is 500, the second preset quantity threshold in the self-study scenario can be 50, and the third preset quantity threshold is 200, the second preset quantity threshold for key parts in the online classroom scenario can be 80, and the third preset quantity threshold is 400, and the second preset quantity threshold for non-key parts in the online classroom scenario can be 60, and the third preset quantity threshold is 300.

[0067] Updating the initial user-teaching resource association map according to the update strategy corresponding to each teaching scenario has the following advantages: maximizing the user experience while ensuring the efficient operation of the personalized teaching resource recommendation system. For the exam preparation scenario, it can quickly adapt to high-frequency major changes to ensure the timeliness and relevance of exam preparation resources; in the self-learning scenario, it supports smooth and continuous personalized recommendations; and in the online classroom scenario, it can achieve an effective combination of immediate feedback and long-term planning. Overall, this method enhances the response speed and flexibility of the personalized teaching resource recommendation system, improves the accuracy of recommendations and user satisfaction, and promotes a more personalized learning experience.

[0068] Optionally, if the number of affected nodes exceeds a fourth preset number threshold, and the fourth preset number threshold is much larger than the third preset number threshold, the number of nodes updated at a single time is limited to allow the update operation to proceed in an orderly manner and prevent omission of update data.

[0069] Step 113: structurally optimize the updated user-teaching resource association graph to obtain an optimized user-teaching resource association graph.

[0070] It should be understood that the embodiments of the present application improve the graph structure through a series of optimization techniques to make it more compact and accurately represent the relationship between users and resources. These optimization measures help improve the performance and effect of subsequent analysis. The optimized graph structure is more compact and effective, which helps to speed up the calculation process, and can also better reflect the deep relationship between users and resources, thereby improving the efficiency and accuracy of recommendations.

[0071] Exemplarily, the present application can combine multiple technologies to achieve structural optimization. Specifically, step 113, structural optimization is performed on the updated user and teaching resource association diagram to obtain an optimized user and teaching resource association diagram, including: Step a1: using a density-based clustering algorithm to identify sparse areas and dense areas in the updated user-teaching resource association graph.

[0072] Specifically, step a1, using a density-based clustering algorithm to identify sparse areas and dense areas in the updated user-teaching resource association graph, includes: Step a11: Initialize the parameters of the density-based clustering algorithm, including the community radius and the minimum number of neighbors. The community radius defines the spatial range around the node, within which neighbor nodes are found; and the minimum number of neighbors specifies the minimum number of neighbors a node must have to be considered a "core node".

[0073] For example, in a large online education platform, in order to optimize the association graph between users and teaching resources, the administrator of the personalized teaching resource recommendation system sets an initial domain radius value of 0.5 based on historical data analysis, and sets the minimum number of neighbors to 50. If a node has at least 50 neighbor nodes within its 0.5 unit range, the node is marked as a core node.

[0074] Step a12, traverse each node in the updated user and teaching resource association graph, calculate the number of neighbor nodes in the domain determined by the domain radius, if the number of neighbor nodes is greater than or equal to the minimum number of neighbors, determine the node as a core node, merge the core node and its reachable neighbor nodes into a cluster, evaluate the cluster, and determine whether the cluster is a dense area based on the cluster evaluation result; or, if the number of neighbor nodes is less than the preset node number threshold, determine the node as a boundary point or a noise point, the noise point is not a core node and is a node that is not reachable by any core node, and determine the sparse area based on the position distribution of the noise point and the minimum number of neighbors.

[0075] By executing steps a11 to a12, the embodiment of the present application can effectively identify dense areas and sparse areas in the large-scale user-teaching resource association graph, thereby achieving an in-depth understanding and optimization of the graph structure. This method not only helps to distinguish key user groups and popular educational resources, but also provides a basis for cleaning up unnecessary noise, thereby improving the performance and accuracy of the personalized recommendation system. Specifically, it enables the personalized teaching resource recommendation system to more accurately locate the user's interests and provide learning resource recommendations that are more in line with personal needs. It also helps to improve the user experience and promote continuous learning and development. In addition, by evaluating the quality of the cluster, it is ensured that the user group formed has a high internal consistency, which helps to promote the formulation of more targeted education strategies.

[0076] Step a2: merge nodes in sparse areas, perform supplementary processing on edges in sparse areas, prune edges in dense areas, and divide dense areas into communities using a community detection algorithm to obtain an intermediate association graph.

[0077] Step a3: Map the intermediate association graph to four-dimensional space-time, and detect abnormal nodes and abnormal edges in the intermediate association graph based on the lens effect and user life cycle in the four-dimensional space-time, perform abnormal processing on the abnormal nodes and abnormal edges, and obtain an optimized user and teaching resource association graph. The abnormal processing is adjustment or deletion.

[0078] Among them, four-dimensional space-time can be understood as a combination of three-dimensional space and time dimensions, which can expand the relationship between users and resources into a dynamic evolution process. Specifically, the time dimension records the timestamps of users' access to resources (such as a student frequently accessing a certain knowledge point late at night for a week), the periodic changes in resource popularity (such as a surge in visits to mock test questions before an exam), and analyzes the topological distribution of users and resources in three-dimensional space (such as a course resource gradually forming a dense subgraph in the association graph).

[0079] Moreover, the lens effect refers to focusing on data in a specific area or time period by dynamically adjusting the granularity and scope of spatiotemporal analysis to amplify hidden local anomalies or key patterns. Its core is multi-scale analysis, which is similar to an optical lens observing details at different levels by adjusting the focal length. Therefore, the lens effect can focus on key areas and amplify hidden anomalies. The embodiment of the present application can dynamically define the window size of four-dimensional space-time according to analysis requirements (such as a time range of 1 hour, 1 month, etc.; a spatial range of a single course subgraph, a full platform association graph, etc.). This embodiment can also combine coarse-grained global analysis and fine-grained local analysis to solve the problem of local anomalies being masked by global noise in traditional methods.

[0080] For example, we can focus on analyzing a specific time and space range (such as the recent learning behavior of a class) to identify small-scale anomalies (such as a student suddenly deviating from the group learning path). Another example is to focus the time lens on "72 hours before the exam" 3 days before the exam to analyze the user's access density to the simulation test questions and capture the high-frequency abnormal access of the test-taking robot. Another example is to perform spatial local magnification on the association map of a certain regional user group (such as students in a certain province) to identify regional resource demand hotspots.

[0081] For example, three days before the exam, the edge weight between a user node and the "Previous Years' Questions" resource suddenly returns to zero (normally it should be an upward trend). This may be due to the user switching the exam preparation strategy or the system recommendation failing. This embodiment can locate this period through the lens effect for repair. Alternatively, a resource node is frequently visited by users from different regions during the early morning hours (abnormal spatiotemporal distribution), which may be malicious crawler behavior, and abnormal edges need to be pruned. For example, global analysis shows that the overall visits to a course are normal, but the local lens finds that the visits to a certain knowledge point in its sub-chapter have dropped suddenly (possibly due to content errors that need to be repaired).

[0082] Another example is that if a user node establishes high-weight edges (abnormal time density) with multiple unpopular resource nodes within a short period of time (such as 1 hour), it may be a robot's behavior and needs to be marked as an abnormal node. Alternatively, the number of visits to a teaching resource node drops sharply after the end of the exam season (normal life cycle), but if its number of visits suddenly soars in the non-exam season (time and space anomaly), it may be due to external promotion that causes a sudden change in traffic, and the recommendation strategy needs to be adjusted.

[0083] The user life cycle can refer to the full-stage behavioral evolution process of the user from registration to silence in the platform, which is divided into different stages (such as new user exploration period, active learning period, and silence period). In this application, life cycle analysis is used to define differentiated abnormal judgment rules to distinguish normal behavior from potential abnormalities in different stages. Therefore, the embodiment of the present application can define differentiated abnormal judgment rules: for the new user period, frequent attempts to unpopular resources are allowed (normal behavior), but if the core knowledge points are not reached for a long time (abnormal stagnation), active guidance is given. For the silence period, the user node has no interactive edges for a long time (normal), but if an edge is suddenly established with a high-difficulty resource (abnormal activation), it may be account theft or demand mutation. Exemplarily, a user continues to access introductory courses within 3 months after registration (normal life cycle), but if the intermediate course is suddenly skipped in the fourth month to directly access expert content (leapfrog abnormality), it is necessary to detect whether it is a misoperation or demand upgrade. A resource node is visited in large quantities in the middle of user learning (normal), but if it is frequently visited during the user's silence period (life cycle mismatch), it may be due to outdated recommendation algorithms, and the weight needs to be adjusted.

[0084] By executing steps a1 to a3, the embodiment of the present application can not only accurately identify the key structural features in the large-scale user-teaching resource association graph, but also can perform targeted optimization processing, including enhancing the connectivity of sparse areas, streamlining dense areas, and improving the quality of the overall network. Ultimately, this method improves the performance of the personalized teaching resource recommendation system, making the recommendation more accurate and efficient, while also improving the user experience and promoting the development of personalized learning paths. In addition, the lens effect is a spatiotemporal dynamic analysis tool that improves the accuracy and efficiency of anomaly detection through multi-scale focusing. The user life cycle is a behavioral evolution modeling framework that enhances the scene adaptability of anomaly determination through staged rules. The combination of the two enables the system to realize intelligent exception handling and resource optimization from the dual dimensions of time and space and user behavior, and the personalized teaching resource recommendation system can continuously optimize itself to ensure long-term high-quality services.

[0085] Step 114: cross-validate the optimized user-teaching resource association graph according to preset evaluation indicators, and determine the user-teaching resource association graph that passes the cross-validation as a dynamic interaction network. The preset evaluation indicators include recommendation accuracy and recall rate.

[0086] The optimized network can improve the accuracy and coverage of recommendations, so that each student can get the most suitable teaching resource recommendation. The strictly verified dynamic interactive network has higher credibility and reliability, and can ensure the provision of high-quality personalized recommendation services to users. Therefore, step 114 can ensure that the generated dynamic interactive network has good recommendation accuracy and recall rate.

[0087] By executing steps 111 to 114, the embodiment of the present application can extract valuable information from large-scale historical interaction data, build and optimize a dynamic interaction network that accurately reflects the complex relationship between users and teaching resources. This method not only improves the quality of personalized recommendations, but also greatly enhances the user experience and promotes continuous learning and personal development. Through continuous iterative optimization, the personalized teaching resource recommendation system can adapt to different teaching scenarios and provide more accurate and personalized education resource recommendation services.

[0088] In a possible embodiment, in S12, a multi-layer interest graph of each user is established based on the teaching resources that the user has been interested in and the teaching resources that the user is expected to be interested in, including: Step 121: extract the features of the teaching resources that the user has already been interested in and the teaching resources that the user is about to be interested in at each layer from the teaching resource metadata.

[0089] Step 122: Based on the characteristics of the teaching resources that the user is interested in and the teaching resources that the user is about to be interested in at each layer, combined with the weight of the teaching resource type and the weight of the layer, a multi-layer interest map of each user is established. The multi-layer includes a macro layer, a meso layer and a micro layer. The macro layer reflects the fields and subjects that the user is interested in, the meso layer reflects the knowledge points and topics that the user is interested in, and the micro layer reflects the resource characteristics and user behavior characteristics that the user is interested in. The resource characteristics include distributors, release time and content characteristics. Different teaching resource types correspond to different weights, and different layers correspond to different weights.

[0090] By executing steps 121 to 122, the embodiment of the present application can assign corresponding weights according to the type of teaching resources and the importance of different levels, thereby constructing a map that accurately reflects the multi-level interest structure of the user. This method can not only comprehensively capture the user's interest preferences and their dynamic changes, but also provide a solid foundation for personalized recommendations, so that the personalized teaching resource recommendation system can more accurately match the user's actual needs, promote a more efficient learning experience and the development of personalized educational paths. Ultimately, through this weight-optimized interest map establishment method, the accuracy and relevance of the personalized teaching resource recommendation system are improved, the user experience is enhanced, and continuous learning and personal growth are promoted.

[0091] Based on the personalized teaching resource recommendation method for large-scale users provided in the above embodiment, in S13, according to the multi-layer interest graph of each user, a personalized teaching resource recommendation list for each user is determined, including: Step 131: for each user, convert the fields and subjects of teaching resources that the user is not interested in into a first vector group, convert the fields and subjects that the user is interested in into a second vector group, and use cosine similarity to calculate the interest similarity at the macro level.

[0092] Step 132: convert the knowledge points and themes of the teaching resources that the user is not interested in into a third vector group, convert the knowledge points and themes that the user is interested in into a fourth vector group, and use cosine similarity to calculate the interest similarity of the meso-level.

[0093] Step 133: convert resource features and user behavior features of teaching resources that the user is not interested in into a fifth vector group, convert resource features and user behavior features that the user is interested in into a sixth vector group, and use cosine similarity to calculate interest similarity at the micro level.

[0094] Among them, teaching resources that users are not interested in are the basic pool for mining potential interests. By analyzing users' behavior patterns, preference changes, and interest tendencies of similar user groups, the personalized teaching resource recommendation system can identify resources that may arouse users' interest.

[0095] Step 134 , based on the interest similarity at the macro level, the interest similarity at the meso level, and the interest similarity at the micro level, a plurality of candidate teaching resources are screened out, where the candidate teaching resources are teaching resources that the user is not interested in and that meet the similarity design rules corresponding to the teaching scenario.

[0096] Specifically, the interest similarity of the macro level, the interest similarity of the meso level, and the interest similarity of the micro level are weighted and summed. The similarity design rule indicates that different weights are corresponding to different levels in different teaching scenarios. For example, the weighted summation formula is as follows: ; in, represents the comprehensive interest similarity score, represents the macro-level weight, represents the meso-level weight, represents the micro-level weight, represents the macro level, Represents the meso-level, Represents the microscopic layer.

[0097] Step 135: sort the multiple candidate teaching resources, and generate a personalized teaching resource recommendation list for the user according to the sorting result.

[0098] It should be noted that the alternative teaching resources for the teaching resources to be of interest to the user are ranked higher than the alternative teaching resources for the teaching resources not to be of interest to the user.

[0099] By executing steps 131 to 135, the embodiment of the present application can accurately generate a personalized teaching resource recommendation list based on each user's multi-layer interest map. Specifically, this embodiment improves the accuracy and relevance of recommendations through multi-level and comprehensive similarity calculations, so that the personalized teaching resource recommendation system can not only capture the user's broad interests, but also deeply understand their specific needs, provide resource recommendations that are more in line with personal learning paths, and improve user experience and learning efficiency. In addition, through comprehensive consideration of interests at different levels, the personalized teaching resource recommendation system can continuously optimize the recommendation strategy and support the personalized development of users under diversified learning goals.

[0100] In a possible embodiment, in the above step S13, a comparative analysis is performed on the local teaching resources of interest and the global teaching resources of interest in the personalized teaching resource recommendation list, and the multi-layer interest graph is optimized based on the comparison results, which specifically includes the following steps: During the comparative analysis stage, the embodiment of the present application generates two types of key results: a difference resource list and a hierarchical deviation index, wherein the difference resource list may refer to a set of resources with significant differences in similarity between local and global interests, and the hierarchical deviation index may refer to: the weight offset of the macro, meso, and micro levels in local and global interests (such as the global weight of the macro layer is higher than the local weight).

[0101] By comparing the results, the system optimizes the interest graph from three dimensions: dynamically adjusting the hierarchical weights, updating the interest vector representation, and sparsification and pruning. Specifically: Dynamically adjusting the hierarchical weights can balance the personalized interests of users with the global trends of the platform. Specifically, (1.1) Calculate the weight offset: for each level (macro / meso / micro), calculate the weight difference between local and global interests; (1.2) If the average weight of global resources at the macro level is greater than the preset weight threshold, the hierarchical weight of the user interest graph is corrected proportionally.

[0102] Updating the interest vector representation can fuse local preferences with global trends to generate a more comprehensive interest vector. Specifically, (2.1) vector concatenation of local and global features of different resources; (2.2) using a lightweight neural network to generate an optimized interest vector; (2.3) replacing the vector of the original user node with the optimized interest vector, and adjusting the associated edge weights simultaneously. For example, the user's original interest vector for "machine learning" only contains basic algorithm features (local), and after fusing the "deep learning framework" features in the global trend, a new vector containing comprehensive information of both is generated.

[0103] Sparsification and pruning can eliminate noise in the interest graph and improve computational efficiency. Specifically, (3.1) in the comparative analysis, if the local and global interest similarities of a resource are both lower than the preset similarity threshold (e.g. 0.2), it is marked as a low-contribution node. (3.2) Remove low-contribution nodes and their associated edges to retain the core interest structure. For example, the "quantum physics" resource that the user briefly browsed was pruned to simplify the graph because it had no interaction for a long time and deviated from the global trend.

[0104] In some other possible embodiments, S14, collaboratively training the dynamic interaction network on multiple edge devices through federated learning to optimize the network structure of the dynamic interaction network, specifically includes the following steps: Sub-models of the dynamic interactive network are deployed on multiple edge devices. Each sub-model is trained based on local user data, and the initial parameters are synchronized by the central server. During local training, noise (such as Laplace noise) is added to the user interaction data to ensure that individual data cannot be traced. Each edge device calculates the model parameter gradient, uploads it to the central server for weighted averaging, and updates the global model. The dynamic interactive network structure is optimized according to the global model. For example, according to the parameters after federation aggregation, the weights of the user-resource edges are updated, and the edges with weights below the preset weight threshold are removed to reduce network complexity. For example: A regional server detects that the user group has an increased interest in the "Data Analysis" course. After updating the global model through federated learning, the central server adjusts the edge weight of the "Data Analysis" node in the association graph of the entire platform to increase its recommendation priority.

[0105] Figure 2 A structural diagram of a personalized teaching resource recommendation system for large-scale users provided in an embodiment of the present application is shown in FIG. Figure 2 As shown, the system includes: The construction module 21 is used to construct a dynamic interaction network using a quantum annealing algorithm based on historical interaction behavior data between large-scale users and teaching resources, where large-scale means that the number of users exceeds a first preset number threshold.

[0106] The extraction and establishment module 22 is used to extract the teaching resources that the user is already interested in and the teaching resources that the user is to be interested in from the dynamic interactive network, and establish a multi-layer interest graph for each user based on the teaching resources that the user is already interested in and the teaching resources that the user is to be interested in.

[0107] A determination comparison module 23 is used to determine a personalized teaching resource recommendation list for each user based on the multi-layer interest map of each user, compare and analyze the local teaching resources of interest and the global teaching resources of interest in the personalized teaching resource recommendation list, and optimize the multi-layer interest map based on the comparison results; The training optimization module 24 is used to collaboratively train the dynamic interaction network on multiple edge devices through federated learning to optimize the network structure of the dynamic interaction network.

[0108] Figure 2 The personalized teaching resource recommendation system for large-scale users can be implemented Figure 1 The implementation principle and technical effect of the personalized teaching resource recommendation method for large-scale users described in the illustrated embodiment will not be described in detail. The specific manner in which each module and unit performs operations in the personalized teaching resource recommendation system for large-scale users in the above embodiment has been described in detail in the embodiment of the method, and will not be elaborated here.

[0109] In one possible design, Figure 2 The personalized teaching resource recommendation system for large-scale users in the embodiment shown can be implemented as a computing device, such as Figure 3 As shown, the computing device may include a storage component 31 and a processing component 32. The storage component 31 stores one or more computer instructions, wherein the one or more computer instructions are called and executed by the processing component 32. The processing component 32 is used to: execute the method flow provided by steps S11 to S14.

[0110] The processing component 32 may include one or more processors to execute computer instructions to complete all or part of the steps in the above method. Of course, the processing component may also be implemented by one or more application-specific integrated circuits (ASIC), digital signal processors (DSP), digital signal processing devices (DSPD), programmable logic devices (PLD), field programmable gate arrays (FPGA), controllers, microcontrollers, microprocessors or other electronic components to perform the above method.

[0111] The storage component 31 is configured to store various types of data to support operations at the terminal. The storage component can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as random access memory (RAM), static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0112] Of course, the computing device may also include other components, such as input / output interfaces, display components, communication components, etc.

[0113] The input / output interface provides an interface between the processing component and the peripheral interface module, which may be an output device, an input device, etc.

[0114] The communication component is configured to facilitate, among other things, wired or wireless communications between the computing device and other devices.

[0115] Among them, the computing device can be a physical device or an elastic computing host provided by a cloud computing platform, etc. In this case, the computing device can refer to a cloud server, and the above-mentioned processing components, storage components, etc. can be basic server resources rented or purchased from the cloud computing platform.

[0116] The present application also provides a computer storage medium storing a computer program, wherein the computer program can achieve the above-mentioned Figure 1 The illustrated embodiment is a method for recommending personalized teaching resources for large-scale users.

[0117] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0118] The system embodiment described above is merely illustrative, wherein the units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules may be selected according to actual needs to achieve the purpose of the solution of this embodiment. Those of ordinary skill in the art may understand and implement it without creative work.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application 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 embodiments of the present application.

Claims

1. A personalized teaching resource recommendation method for large-scale users, characterized in that: include: Based on the historical interaction behavior data between large-scale users and teaching resources, a dynamic interaction network is constructed using a quantum annealing algorithm, where the number of large-scale users exceeds a first preset number threshold; Extracting teaching resources that the user has been interested in and teaching resources that the user is likely to be interested in from the dynamic interactive network, and establishing a multi-layer interest graph for each user based on the teaching resources that the user has been interested in and the teaching resources that the user is likely to be interested in; Determine a personalized teaching resource recommendation list for each user based on the multi-layer interest graph of each user, compare and analyze the local teaching resources of interest and the global teaching resources of interest in the personalized teaching resource recommendation list, and optimize the multi-layer interest graph based on the comparison results; The dynamic interaction network is collaboratively trained on multiple edge devices through federated learning to optimize the network structure of the dynamic interaction network.

2. The method according to claim 1, characterized in that The historical interaction behavior data includes: user identifier, user personal information, user preference information, interaction mode, teaching resource identifier, teaching resource metadata, teaching resource format and user feedback information; The method uses quantum annealing algorithm to construct a dynamic interaction network based on historical interaction behavior data between large-scale users and teaching resources, including: Define nodes and edges, a node is a user node or a teaching resource node, the attribute information of the user node includes: user identifier, user personal information and user preference information, the attribute information of the teaching resource node includes: teaching resource identifier, teaching resource metadata and teaching resource format, the information of an edge includes the similarity between users, the similarity between teaching resources, or the association between users and teaching resources, the association includes the interaction mode and user feedback information, the attribute information of the node and the information of the edge are mapped to the Ising model, the stable state of the Ising model is determined by using the quantum annealing algorithm, and the stable state of the Ising model is used as the initial user and teaching resource association graph; When new interactive behavior data appears, determine the content to be updated and the teaching scene to which the content to be updated belongs, determine the update strategy corresponding to the teaching scene in combination with the scene characteristics of the teaching scene and the update content, and update the initial user and teaching resource association map according to the update strategy corresponding to the teaching scene to obtain an updated user and teaching resource association map; Structural optimization is performed on the updated user-teaching resource association graph to obtain an optimized user-teaching resource association graph; According to preset evaluation indicators, the optimized user-teaching resource association graph is cross-validated, and the user-teaching resource association graph that passes the cross-validation is determined as a dynamic interaction network. The preset evaluation indicators include recommendation accuracy and recall rate.

3. The method according to claim 2, characterized in that The teaching scenario includes at least one of the following: an exam preparation scenario, an independent learning scenario, and an online classroom scenario; The initial user-teaching resource association graph is updated according to the update strategy corresponding to the teaching scenario to obtain an updated user-teaching resource association graph, including at least one of the following: Filter out key update data for the test preparation scenario from the content to be updated, and based on the key update data for the test preparation scenario, in combination with a first preset annealing time and a first preset Hamiltonian strength, use a batch processing and incremental update mechanism to update the attribute information of the nodes and the edge information in the initial user-teaching resource association graph, use natural language processing technology to perform sentiment analysis on the comment content in the newly added interactive behavior data, and adjust the edge weight in the corresponding edge information according to the sentiment analysis result, so as to obtain an updated user-teaching resource association graph; Filtering non-critical update data of the autonomous learning scenario from the content to be updated, and periodically updating the attribute information of the nodes and the edge information in the initial user-teaching resource association graph according to the non-critical update data of the autonomous learning scenario, in combination with a second preset annealing time and a second preset Hamiltonian strength, to obtain an updated user-teaching resource association graph; the second preset annealing time is greater than the first preset annealing time, and the second preset Hamiltonian strength is less than the first preset Hamiltonian strength; Filter out key update data and non-key update data of the online classroom scene from the content to be updated; based on the key update data of the online classroom scene, in combination with the first preset annealing time and the first preset Hamiltonian strength, use batch processing and incremental update mechanism to update the attribute information of the nodes and edge information in the initial user-teaching resource association graph; use natural language processing technology to perform sentiment analysis on the comment content in the newly added interactive behavior data, and adjust the edge weight in the corresponding edge information according to the sentiment analysis result to obtain an adjusted association graph; based on the non-key update data of the online classroom scene, in combination with the second preset annealing time and the second preset Hamiltonian strength, periodically update the adjusted association graph to obtain an updated user-teaching resource association graph; Among them, the critical update data includes: update data whose number of included update items exceeds a second preset number threshold, and / or update data of nodes whose number of affected nodes exceeds a third preset number threshold; the non-critical update data is other update data in the content to be updated except the critical update data. In different teaching scenarios, the value of the second preset number threshold is different, and the value of the third preset number threshold is different.

4. The method according to claim 2, characterized in that: The structural optimization of the updated user-teaching resource association graph to obtain an optimized user-teaching resource association graph includes: Using density-based clustering algorithm, identify sparse areas and dense areas in the updated user-teaching resource association graph; Merge nodes in sparse areas, supplement edges in sparse areas, prune edges in dense areas, and use community detection algorithms to divide dense areas into communities to obtain an intermediate association graph. The intermediate association graph is mapped to four-dimensional space-time, and abnormal nodes and abnormal edges in the intermediate association graph are detected in the four-dimensional space-time based on the lens effect and the user life cycle, and abnormal processing is performed on the abnormal nodes and abnormal edges to obtain an optimized user and teaching resource association graph, and the abnormal processing is adjustment or deletion.

5. The method according to claim 4, characterized in that The method of using a density-based clustering algorithm to identify sparse areas and dense areas in the updated user-teaching resource association graph includes: Initializing parameters of a density-based clustering algorithm, including a neighborhood radius and a minimum number of neighbors; Traverse each node in the updated user-teaching resource association graph, calculate the number of neighbor nodes in the domain determined by the domain radius, and if the number of neighbor nodes is greater than or equal to the minimum number of neighbors, determine the node as a core node, merge the core node and its reachable neighbor nodes into a cluster, evaluate the cluster, and determine whether the cluster is a dense area based on the cluster evaluation result; Alternatively, if the number of neighbor nodes is less than a preset node number threshold, the node is determined to be a boundary point or a noise point. The noise point is not a core node and is a node that is not reachable by any core node. The sparse area is determined based on the location distribution of the noise point and the minimum number of neighbors.

6. The method according to claim 1, characterized in that The step of establishing a multi-layer interest graph for each user based on the teaching resources that the user has been interested in and the teaching resources that the user is to be interested in includes: Extracting the features of the teaching resources that the user has been interested in and the teaching resources that the user will be interested in at each layer from the teaching resource metadata; According to the characteristics of the teaching resources that the user is already interested in and the teaching resources that the user is to be interested in at each layer, combined with the weight of the teaching resource type and the weight of the layer, a multi-layer interest map for each user is established, wherein the multiple layers include a macro layer, a meso layer and a micro layer. The macro layer reflects the fields and subjects that the user is interested in, the meso layer reflects the knowledge points and topics that the user is interested in, and the micro layer reflects the resource characteristics and user behavior characteristics that the user is interested in. The resource characteristics include distributors, release time and content characteristics. Different teaching resource types correspond to different weights, and different layers correspond to different weights.

7. The method according to claim 1, characterized in that Determining a personalized teaching resource recommendation list for each user based on the multi-layer interest graph of each user includes: For each user, the fields and subjects of teaching resources that the user is not interested in are converted into a first vector group, and the fields and subjects that the user is interested in are converted into a second vector group, and the cosine similarity is used to calculate the interest similarity at the macro level; The knowledge points and topics of the teaching resources that the user is not interested in are converted into a third vector group, the knowledge points and topics that the user is interested in are converted into a fourth vector group, and the cosine similarity is used to calculate the interest similarity of the meso-level; The resource features and user behavior features of the teaching resources that the user is not interested in are converted into a fifth vector group, the resource features and user behavior features that the user is interested in are converted into a sixth vector group, and the cosine similarity is used to calculate the interest similarity of the micro-level; According to the interest similarity at the macro level, the interest similarity at the meso level, and the interest similarity at the micro level, multiple candidate teaching resources are screened out, and the candidate teaching resources are teaching resources that the user is not interested in that meet the similarity design rules corresponding to the teaching scenario; The multiple candidate teaching resources are sorted, and a personalized teaching resource recommendation list for the user is generated according to the sorting result.

8. A personalized teaching resource recommendation system for large-scale users, characterized by: include: A construction module is used to construct a dynamic interaction network using a quantum annealing algorithm based on historical interaction behavior data between large-scale users and teaching resources, where the number of large-scale users exceeds a first preset number threshold; An extraction and establishment module, used to extract teaching resources that users have been interested in and teaching resources that users are expected to be interested in from the dynamic interactive network, and establish a multi-layer interest graph for each user based on the teaching resources that users have been interested in and the teaching resources that users are expected to be interested in; Determine a comparison module, which is used to determine a personalized teaching resource recommendation list for each user according to the multi-layer interest map of each user, compare and analyze the local teaching resources of interest and the global teaching resources of interest in the personalized teaching resource recommendation list, and optimize the multi-layer interest map based on the comparison results; A training optimization module is used to collaboratively train the dynamic interaction network on multiple edge devices through federated learning to optimize the network structure of the dynamic interaction network.

9. A computing device, characterized in that It comprises a processing component and a storage component; the storage component stores one or more computer instructions; the one or more computer instructions are used to be called and executed by the processing component to implement a personalized teaching resource recommendation method for large-scale users as described in any one of claims 1 to 7.

10. A computer storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a computer, a personalized teaching resource recommendation method for large-scale users as described in any one of claims 1 to 7 is implemented.

Images