College student-oriented online resource personalized recommendation method and system

Abstract

The invention belongs to but is not limited to the technical field of learning resource recommendation, and particularly relates to an online resource personalized recommendation method and system for college students, and the system comprises a front-end interaction module which is used for sending a request to an algorithm engine module below and returning a learning resource personalized recommendation result, a basic interaction interface function and a personalized learning resource recommendation service display function are provided for students, and the system is divided into a front-end display sub-module and a business processing sub-module; the rear-end algorithm engine module is used for processing data of the lower-layer data storage module and providing data and algorithm support for upper-layer services, and is divided into a user portrait sub-module, a knowledge graph sub-module, a model training sub-module and a recommendation list sub-module; the data storage module is used for storing and managing data in the system; and the system parameter setting module is connected with the data storage module and is used for setting system parameters.

Description

Technical Field

[0001] The present invention belongs to, but is not limited to, the technical field of learning resource recommendation, and in particular relates to a method and system for personalized online resource recommendation for college students. Background Art

[0002] With the rapid growth of online education resources, college students have become the primary audience. However, the vast number of courses can easily lead to "knowledge wandering" and difficulty in choosing. To address this, personalized recommendation systems leverage big data to analyze student behavior, precisely matching learning resources to students and improving learning efficiency. Recommendation systems offer advantages over search engines in that they eliminate the need for users to actively enter keywords, instead intelligently recommending content based on user characteristics.

[0003] Knowledge graph technology builds semantic networks through entities and relationships, improving the understanding capabilities of recommendation systems and effectively alleviating cold start and data sparsity issues. Knowledge graph-based recommendations can be divided into three categories: connection-based methods use path relationships in the knowledge graph to calculate similarity but do not consider user interests; embedding-based methods vectorize entities and relationships in the knowledge graph to improve recommendation accuracy but tend to ignore semantic connections; and propagation-based methods combine path and embedding concepts to enhance personalized recommendations but are computationally more complex.

[0004] In recent years, the application of knowledge graphs in personalized recommendations has continued to develop. For example, the Hete-MF algorithm combines matrix factorization with knowledge graph similarity to improve user preference matching; RippleNet optimizes recommendation effectiveness through multi-hop neighborhood propagation in knowledge graphs; KGCN introduces graph convolutional neural networks (GCN) to improve high-order connectivity mining capabilities; and KGAT combines attention mechanisms to optimize user-item interaction information and enhance recommendation interpretability. However, existing technologies still face challenges such as high computational cost and strong dependence on path quality.

[0005] Current search engine recommendation models still rely on user input, making it difficult to accurately understand user needs. Personalized recommendations still have room for improvement. Therefore, knowledge graph-based recommendation systems need to further improve path connection quality, optimize embedding capabilities, and reduce computational complexity to more accurately meet the needs of online education users. Summary of the Invention

[0006] In view of the problems existing in the prior art, the present invention provides a method and system for personalized online resource recommendation for college students.

[0007] The present invention is implemented as follows: a personalized online resource recommendation system for college students, comprising:

[0008] The front-end interaction module sends requests to the algorithm engine module below and returns personalized learning resource recommendation results. It also provides students with basic interactive interface functions and the display function of personalized learning resource recommendation services. It is divided into a front-end display sub-module and a business processing sub-module.

[0009] The backend algorithm engine module processes data from the lower-level data storage module and provides data and algorithm support for upper-level businesses. It is divided into user portrait sub-module, knowledge graph sub-module, model training sub-module, and recommendation list sub-module;

[0010] Data storage module, used to store and manage data in the system;

[0011] Furthermore, the front-end interaction module specifically includes: The front-end display submodule is primarily responsible for receiving student data requests, transmitting them downward to the system business processing submodule to obtain the results, and realizing the visual transmission of the recommendation list. The business processing submodule's main functions are to receive data requests from the front-end display submodule, perform corresponding data processing and analysis, and deliver them to the front-end; receive recommendation list input from the back-end algorithm engine module, and provide the results to the front-end display submodule for display.

[0012] Furthermore, the business processing submodule is responsible for the requests and scheduling of the entire system, including basic class data requests and data requests for recommendation services, and completes tasks such as data return and database update through interfaces and front-end interfaces.

[0013] After receiving a data request from the system, basic data requests use the interface to query, verify, and update the data, returning the results to the requester. If the request fails, a failure status is returned. This includes student registration, login, user profile query, user profile query, learning resource query, and student history query. Universities generally use a unified identity authentication system to verify student identities.

[0014] The recommendation service data request refers to when the system receives a request for a student's recommended resource list, it parses the request parameters, calls the interface for generating the recommended learning resource list, obtains the recommendation results, and returns the results to the request sender in a fixed-format JSON data.

[0015] The business processing submodule is responsible for receiving and dispatching incoming system requests, calling corresponding methods to respond to front-end requests and returning results. Database operations include querying and updating the database using encapsulation classes; accessing the database and retrieving data in the Request class, as well as obtaining student history records, and then generating a personalized list of recommended learning resources. The User class contains student-related fields, including personal information; and the Resource class defines fields related to learning resources.

[0016] (1) User registration page

[0017] On the user registration page, design text boxes for "Username" and "Password." If the user enters the correct information, a "Registration Successful" message is displayed. If the user enters incorrect information, the user is provided with the specific reason for the error, such as an incorrect password or an unavailable personal account name. Additionally, a redirect link is designed to facilitate the user's login to the "Login" page. This is typically implemented in universities and colleges, where unified identity authentication systems are integrated.

[0018] (2) User login page

[0019] On the user login page, similarly design text boxes for "Username" and "Password." If the user enters the correct information, provide a "Login Successful" message and redirect to the user's profile page. If the user enters incorrect information, provide a detailed explanation of the error, such as a missing username or incorrect password. The page also requires a redirect link to facilitate the user's redirect to the "Register" page to register a personal account.

[0020] (3) Homepage

[0021] This page displays information about the platform's popular resources. When the user is logged in (guest status), the platform only displays popular learning resources. Each learning resource on the homepage is displayed by name and cover. Clicking on a resource's name or cover takes the user directly to the resource's introduction page, allowing them to learn more about the resource.

[0022] (4) Resource recommendation page

[0023] If the user is logged in, the system will use the recommendation engine to generate recommended content for the user, and then through the resource recommendation page, it will be able to show the user a personalized and diverse list of learning resources.

[0024] (5) Resource Introduction Page

[0025] The resource overview page provides users with a visual overview of the learning resource, including its description, number of participants, and ratings. Clicking "Join Learning" on this page completes course selection and leads them to the corresponding resource website to begin studying the selected course.

[0026] (6) History page

[0027] Users can enter the personal history page by clicking on their login name and see their personal historical course selection information.

[0028] Furthermore, the backend algorithm engine module specifically includes:

[0029] The user portrait submodule is responsible for data processing and modeling of user portraits, providing data for subsequent cold start processing. This includes basic data preprocessing and user portrait construction.

[0030] The knowledge graph submodule is responsible for knowledge graph data processing and modeling, preparing data for the recommendation model training submodule. This includes knowledge graph construction, knowledge graph vectorization, and student similarity calculation. Knowledge graph construction includes raw data preprocessing and triple extraction. The optimal model is used to vectorize the knowledge graph. Student similarity is calculated using knowledge graph relationship weights and entity vectors.

[0031] The model training submodule uses information such as student similarity obtained by the knowledge graph submodule to derive the optimal weight coefficient and update the user resource evaluation matrix, providing relevant data for the recommendation list submodule. This module includes recommendation model data processing and recommendation model training.

[0032] The recommendation list submodule consists of two parts: algorithm cold start processing and generation of a recommended learning resource list. Cold start processing mitigates the sparsity of the rating matrix caused by insufficient historical student behavior data by predicting student needs. For each student, this module retrieves their user profile from the lower-level data storage module, makes recommendations based on their major and interests, and then generates a recommended list after sorting. This list is then returned to the front-end interaction module for display.

[0033] Furthermore, the user profiling submodule and knowledge graph submodule within the backend algorithm engine module serve as pre-recommendations for the recommendation model training submodule and the recommendation list submodule, providing data preparation for subsequent modules. The user profiling submodule is divided into two parts: data preprocessing and profile construction; the graph data processing module is divided into three parts: knowledge graph construction, knowledge graph vectorization, and student similarity calculation.

[0034] The process of building a user profile includes data acquisition, data preprocessing, model building, and profile generation. The data acquisition phase involves acquiring basic data and user behavior data. Basic data can be obtained from the campus integrated database, while user behavior data can be obtained from the log analysis mapping URL. All this data is stored in the database, and data preprocessing and label modeling are performed as needed to ultimately form a user profile.

[0035] Knowledge graph construction primarily involves processing raw data and extracting triples. Knowledge graph vectorization primarily utilizes the triples generated during knowledge graph construction and uses TransD training to generate entity vectors. Inter-student similarity calculation primarily utilizes knowledge graph weights and entity vectors to generate a similarity matrix.

[0036] During the recommendation model training phase, the pre-trained model calculates parameters such as similarity and processes the raw data to generate the resource rating matrix and similarity data to be optimized. The user-resource rating matrix is updated based on similarity. Parameters are optimized through model training to obtain the optimal weight coefficients and the optimal model.

[0037] The learning resource recommendation list submodule's primary task is to generate personalized recommendation lists for students based on the recommendation model training submodule. The primary input is the student's historical behavior data. This submodule consists of two parts: a student cold start process and the generation of a student learning resource list. The student cold start process involves predicting the resource needs of new students based on their majors and interests, supplementing their resource needs, and improving the recommendation process. The learning resource list recommendation process loads the student's learning history and the latest recommendation model to recommend learning resources to the student.

[0038] Furthermore, the system parameter settings specifically include:

[0039] The system can be set to turn resource recommendations on or off. Since turning on the recommendation function requires collecting students' personal behavior information and requires their consent, it is turned off by default;

[0040] The number of recommended resources each time, the default is 20;

[0041] Recommended message sending methods: Message Center supports email, WeChat, SMS, and their combinations. By default, only email is sent.

[0042] Student personal information maintenance, phone and email, student status changes, automatic synchronization with the academic affairs system;

[0043] Students can customize their learning interests and receive recommended academic resources accordingly;

[0044] After completing the connection with various data sources, the learning resource recommendation service platform regularly generates labels, calculates weights and time decay every day to ensure that the platform maintains the latest learning resources and student data. Recommended resources are sent regularly every week. The automatic operation mode reduces dependence on management. College leaders and student management departments access the platform to understand students' learning situation. In addition, the platform records the entire service process, which is conducive to the school's grasp of students' independent learning situation.

[0045] Another object of the present invention is to provide a personalized recommendation method for online resources for college students using the method, comprising:

[0046] (1) Check whether the model needs to be retrained: There are currently two situations where model training is required. One is when there is no model, in which case model training is required. The other is to regularly collect student update data and resource update data, and regularly train the model on the data in the database.

[0047] (2) Pre-trained model loading: Since the system needs to meet the fluency requirement, and the computational efficiency of online real-time computing is low, which affects the user experience, we use the offline model loading method to predict learning resource recommendations. The offline model includes parameters such as user profile, vectorized parameters of the knowledge graph, similarity matrix between students, user resource rating matrix, and weight coefficient. After receiving the user's request, the model parameters will be requested from the database, and then a recommendation list will be generated.

[0048] (3) User login and request: In the recommendation system, different students use their student ID (number) to distinguish their user requests, which are passed layer by layer through the algorithm engine module and database sub-module, and finally a list of learning resource recommendations is returned to the student.

[0049] (4) Recommendation result generation: A ranked list of learning resources is obtained through a similarity-based collaborative filtering algorithm. If the number requirement is not met, the user profile parameters are requested to be supplemented and returned to the student.

[0050] Another object of the present invention is to provide a computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to execute the steps of the method for personalized online resource recommendation for college students.

[0051] Another object of the present invention is to provide an information data processing terminal, which includes the online resource personalized recommendation system for college students.

[0052] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solutions to be protected by the present invention are as follows:

[0053] Aiming at the personalized learning needs of college students, this paper constructs an online educational resource recommendation system based on user profiles and knowledge graphs. First, the system obtains students' basic information and online behavior data from university databases, performs data cleaning and label modeling, and constructs user profiles. At the same time, the profiles are dynamically updated using a fixed time window to ensure that the recommendation system can promptly adapt to students' interests and learning needs. Second, the system constructs a course domain knowledge graph, collects course data through a web crawler, analyzes entities and their relationships, establishes a course ontology model, and uses the Neo4j database to store and visualize the knowledge graph, representing knowledge points in the form of triples to ensure that the knowledge acquisition of the recommendation system is more accurate and structured.

[0054] The present invention adopts an online resource recommendation algorithm based on similarity and combines it with user portraits to optimize the recommendation effect. First, entity vectors are trained through knowledge graphs to calculate the similarity between different course resources, thereby optimizing the user rating matrix. Then, a collaborative filtering algorithm is combined with user portraits as a supplement to improve the accuracy and personalization of recommendations. The system verifies the recommendation effect of different models through experiments, optimizes the weight coefficients, and finally selects the optimal parameter α for personalized recommendations. The recommendation results not only include courses that the user may be interested in, but also make expanded recommendations based on the relevance of knowledge points, thereby improving the utilization rate and matching degree of learning resources.

[0055] In terms of system architecture, the present invention establishes a complete personalized recommendation system, realizing automated processing from data collection to recommendation. The system uses MySQL to store user structured information, including basic student information, learning behavior records, learning resources, etc., and knowledge graph information is stored in Neo4j to ensure efficient management of complex knowledge structures. At the same time, the system includes six major modules: front-end display, business data processing, user profile construction, knowledge graph management, model training, and learning resource recommendation. Ultimately, it can automatically capture Internet learning resources and accurately push them based on student interests to meet the personalized education needs of colleges and universities.

[0056] The technical solution of the present invention has broad commercial value in the education industry. At present, the scale of online education resources in my country continues to expand, and college students have become the main audience. However, the disorderly growth of massive course resources can easily lead to students getting lost in their studies and reduce their learning efficiency. The intelligent recommendation system provided by the present invention effectively solves the problem of course overload in a data-driven manner and matches personalized learning resources for college students. There are nearly 3,000 colleges and universities in my country. If the construction cost is estimated at 500,000 to 1 million yuan per college, the overall commercial value of this technology can reach more than 1 billion yuan. If the service fee is charged on an annual basis, each college will be charged 50,000 to 100,000 yuan, and the annual income can reach more than 50 million yuan, which has broad market application prospects.

[0057] This invention fills a technological gap in the current field of personalized learning resource recommendations. Currently, most recommendation systems are still limited to closed learning platforms, making it difficult to balance cross-platform open resource acquisition with personalized recommendations. However, this invention captures open internet learning resources and combines them with university management system data and student online behavior data to achieve accurate cross-platform recommendations. Core technologies such as cross-platform data fusion, dynamic user profile construction, and knowledge graph-based recommendation optimization significantly enhance the intelligent level of personalized recommendations, breaking through the technical bottlenecks of existing systems.

[0058] This invention overcomes the technical limitations of traditional recommendation systems and achieves true automation, personalization, and intelligence. While existing learning recommendation systems mostly rely on users actively inputting keywords, this system can automatically match the most appropriate course resources based on user profiles and knowledge graphs, reducing user screening time and improving learning efficiency. Furthermore, this system uses fully automated operations, requiring no human intervention, and can adaptively optimize recommendation strategies based on user data, enabling college students to quickly find the most suitable learning path for themselves within the vast online educational resources, thereby enhancing their learning experience and achieving the modern development of intelligent education. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a structural diagram of a personalized recommendation system for online learning resources for universities provided by an embodiment of the present invention;

[0060] Figure 2 This is a functional module diagram of a personalized recommendation system for online learning resources for universities provided by an embodiment of the present invention;

[0061] Figure 3 This is a functional flow chart of the recommendation system provided by an embodiment of the present invention;

[0062] Figure 4 This is a database ER diagram provided by an embodiment of the present invention;

[0063] Figure 5 This is a schematic diagram of the Flask front-end and back-end interaction provided by an embodiment of the present invention;

[0064] Figure 6 This is a timing diagram of the service processing submodule provided by an embodiment of the present invention;

[0065] Figure 7 This is a screenshot of the homepage of the recommendation system provided by an embodiment of the present invention;

[0066] Figure 8 This is a schematic diagram of a personalized recommendation page provided by an embodiment of the present invention;

[0067] Figure 9 This is a time sequence diagram for constructing a user portrait provided by an embodiment of the present invention;

[0068] Figure 10 This is a time sequence diagram for constructing a knowledge graph provided by an embodiment of the present invention;

[0069] Figure 11 This is a model training timing diagram provided by an embodiment of the present invention;

[0070] Figure 12 This is a timing diagram of a recommendation list provided by an embodiment of the present invention;

[0071] Figure 13This is a schematic diagram of the user portrait construction process provided by an embodiment of the present invention;

[0072] Figure 14 This is a schematic diagram of a data center service architecture provided by an embodiment of the present invention;

[0073] Figure 15 This is a personal portrait display diagram provided by an embodiment of the present invention;

[0074] Figure 16 It is a knowledge graph construction diagram provided by an embodiment of the present invention;

[0075] Figure 17 This is a specific idea diagram of the algorithm optimization provided by the embodiment of the present invention;

[0076] Figure 18 This is a specific flow chart of online resource recommendation provided by an embodiment of the present invention;

[0077] Figure 19 Schematic diagram of the evaluation index matrix of the recommendation system provided by an embodiment of the present invention;

[0078] Figure 20 This is a diagram showing the improvement effect of the recall rate of the recommendation algorithm provided by an embodiment of the present invention;

[0079] Figure 21 This is a diagram showing the improvement in accuracy of the recommendation algorithm provided by an embodiment of the present invention;

[0080] Figure 22 This is a diagram showing the F1-score improvement effect of the recommendation algorithm provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0081] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0082] The personalized online learning resource recommendation method and system for college students provided by the present invention include:

[0083] The front-end interaction module, the back-end algorithm engine module and the data extraction and storage module are three parts. Among them, the front-end and back-end interaction modules and the back-end algorithm engine module are functional modules, the data storage module is the foundation of the entire system, and there are auxiliary modules such as the system parameter setting module. Figure 1 shown.

[0084] Front-end interaction module: The front-end interaction basic module sends requests to the algorithm engine module below and returns the personalized recommendation results of learning resources, and provides students with basic interactive interface functions and display functions of personalized learning resource recommendation services. This module is divided into front-end display sub-module and business processing sub-module, such as Figure 2 shown.

[0085] The front-end display submodule is primarily responsible for receiving student data requests, passing them down to the system's business processing submodule to retrieve the results and visualize the recommended list. The business processing submodule's primary function is to receive data requests from the front-end display submodule, perform corresponding data processing and analysis, and deliver them to the front-end; receive recommended list input from the back-end algorithm engine module, and provide the results to the front-end display submodule for display.

[0086] Backend algorithm engine module:

[0087] The backend algorithm engine primarily processes data from the underlying data storage module, providing data and algorithm support for upper-level businesses. This backend algorithm engine module is primarily divided into a user profiling submodule, a knowledge graph submodule, a model training submodule, and a recommendation list submodule.

[0088] The user portrait submodule is responsible for data processing and modeling of user portraits, providing data for subsequent cold start processing. This includes basic data preprocessing and user portrait construction.

[0089] The knowledge graph submodule is responsible for knowledge graph data processing and modeling, preparing data for the recommendation model training submodule. This includes knowledge graph construction, knowledge graph vectorization, and student similarity calculation. Knowledge graph construction includes raw data preprocessing and triple extraction. The optimal model is used to vectorize the knowledge graph. Student similarity is calculated using knowledge graph relationship weights and entity vectors.

[0090] The model training submodule uses information such as student similarity obtained by the knowledge graph submodule to derive the optimal weight coefficient and update the user resource evaluation matrix, providing relevant data for the recommendation list submodule. This module includes recommendation model data processing and recommendation model training.

[0091] The recommendation list submodule consists of two parts: algorithm cold start processing and generation of a recommended learning resource list. Cold start processing mitigates the sparsity of the rating matrix caused by insufficient historical student behavior data by predicting student needs. For each student, this module retrieves their user profile from the lower-level data storage module, makes recommendations based on their major and interests, and then generates a recommended list after sorting. This list is then returned to the front-end interaction module for display.

[0092] 1. System function process design

[0093] The overall functional process of the personalized recommendation system for online learning resources in colleges and universities is as follows: Figure 3 shown.

[0094] (1) Check whether the model needs to be retrained. There are currently two situations where model training is required. One is when there is no model currently, in which case model training is required. The other is to regularly collect student update data and resource update data, and regularly train the model on the data in the database.

[0095] (2) Loading the pre-trained model. Since the system needs to meet the fluency requirement, and the computational efficiency of online real-time computing is low, which affects the user experience, the offline model is loaded to perform learning resource recommendation prediction. The offline model includes parameters such as user profile, vectorized parameters of the knowledge graph, similarity matrix between students, user resource rating matrix, and weight coefficient. After receiving the user's request, the model parameters will be requested from the database, and then a recommendation list will be generated.

[0096] (3) User login and request. In the recommendation system, different students use their student IDs (numbers) to distinguish their user requests. Through the layer-by-layer transmission of the algorithm engine module and database sub-module below, a list of recommended learning resources is finally returned to the student.

[0097] (4) Recommendation result generation: A ranked list of learning resources is obtained through a collaborative filtering algorithm based on similarity. If the number requirement is not met, the user profile parameters are requested to be supplemented and the list is returned to the student.

[0098] Given the frequently updated user behavior and resource data, the recommendation system needs to provide a model retraining function. In the flowchart, if the model is determined to require retraining, it first requests new data to update the user profile and knowledge graph. It then re-vectorizes the knowledge graph, recalculates student similarities, and determines the final weight coefficients, completing the training of the collaborative filtering algorithm. After evaluating the algorithm results and failing to meet expectations, the model is re-tuned. After model training is complete, the model parameters are stored offline.

[0099] 2. System database design

[0100] The content of this section mainly focuses on database and database interface design based on system organizational structure design and system functional process design.

[0101] The database is primarily based on MySQL, supplemented by the graph database Neo4j and file storage. Student registration and login information, basic information, behavioral interaction data, dynamic information, and learning resource information are stored in MySQL. Knowledge graph information is stored in Neo4j, and model-related parameters are stored in files.

[0102] (1) MySQL structured data storage design

[0103] The structured data storage based on MySQL mainly includes six tables, namely student login information table, student information table, training model information table, learning resource information table, student behavior interaction table and student dynamic information table. The database ER diagram is as follows Figure 4 shown.

[0104] The student login credential information table is shown in Table 1, which includes student ID, login credential string, expiration time, login ID and login status information, where login ID is the primary key.

[0105] Table 1 Student login information table

[0106]

[0107] The student information table, shown in Table 2, includes the student ID as the primary key, along with the student's name, password, gender, user type, college, major, grade, and participating model ID. The student ID and password are primarily used for login authentication. The participating model ID is used to determine whether a student is a new student. When the user's interaction data is used in model training, this field is marked as the training model ID.

[0108] Table 2 Student Information Table

[0109]

[0110] The training model information table includes the model ID, model training date, and model storage path. See Table 3.

[0111] Table 3 Training model information table

[0112]

[0113] The learning resource information table includes the learning resource ID, address, name, description, number of learners, number of reviewers, rating, type, and participating model IDs. The learning resource ID is the primary key, and the ID of the participating model training is used to determine the latest model ID. See Table 4.

[0114] Table 4 Learning resource information table

[0115]

[0116] The student interaction behavior table is shown in Table 5, which includes behavior ID, student ID, resource ID and interaction time, where behavior ID is the primary key.

[0117] Table 5 Student interaction behavior table

[0118]

[0119] The dynamic information table is shown in Table 6, including student ID, book ID, book title, lecture ID, lecture name, elective course status, major course ID, and major course scores.

[0120] Table 6 Dynamic Information Table

[0121]

[0122] (2) Graph database storage design

[0123] Knowledge entities and relationships are extracted from the acquired data, and a knowledge graph is constructed based on the triple data obtained from entity extraction and relationship extraction, and Neo4j is used for storage and visualization.

[0124] (3) Model file storage

[0125] Deep learning-related model files, such as the TransE model, will be stored using ckpt model files. Entity vectors, inter-student similarity matrices, and weight coefficients will be stored using JSON files.

[0126] 3. System call interface design

[0127] Carry out interface design for the relevant functional implementation of the system front-end interaction module and back-end algorithm engine module.

[0128] 3.1 Backend Algorithm Engine Module

[0129] In the backend algorithm engine module, the deep learning model is based on PyTorch 1.8.0 and deployed on a Linux system. The deep learning model is trained using an NVIDIA GeForce RTX 4090 GPU. The user profile submodule includes user profile data processing and user profile construction interfaces (see Table 7). The knowledge graph submodule includes knowledge graph construction, graph vectorization, and similarity calculation interfaces (see Table 8). The model training submodule includes recommendation model data processing and recommendation model training interfaces (see Table 9). The recommendation list submodule includes interfaces for student demand prediction, recommendation list generation, and model ID query (see Table 10).

[0130] Table 7 User portrait interface comparison table

[0131]

[0132] Table 8 Knowledge graph interface comparison table

[0133]

[0134] Table 9 Comparison table of recommended model training interfaces

[0135]

[0136] Table 10 Comparison table of interfaces related to recommendation list generation

[0137]

[0138] 3.2 System front-end interaction module

[0139] The interface design of the system front-end interactive module mainly includes the student registration interface, the student login interface, the student user portrait display interface, the student personalized recommendation list interface, the history record interface, and the learning resource details interface, as shown in Table 11.

[0140] Table 11 Comparison table of student interaction interfaces

[0141]

[0142] The recommendation results are sent; the message center is called to send the corresponding recommendation content to the students. All previous recommendation results are saved in the recommendation result table for students to review and evaluate the accuracy of the recommendations, which can be used as a reference for improving the recommendation algorithm.

[0143] Furthermore, the system parameter setting module:

[0144] The system can be set to turn resource recommendations on or off. Since turning on the recommendation function requires collecting students' personal behavior information and requires their consent, it is turned off by default;

[0145] The number of recommended resources each time, the default is 20;

[0146] Recommended message sending methods: Message Center supports email, WeChat, SMS, and their combinations. By default, only email is sent.

[0147] Student personal information maintenance, phone calls, emails, and student status changes can be automatically synchronized with the academic affairs system;

[0148] Students can customize their learning interests and receive recommended academic resources accordingly;

[0149] After completing the connection with various data sources, the learning resource recommendation service platform regularly generates labels, calculates weights and time decay every day to ensure that the platform maintains the latest learning resources and student data. Recommended resources are sent regularly every week. The automatic operation mode reduces dependence on management. College leaders and student management departments access the platform to understand students' learning situation. In addition, the platform records the entire service process, which is conducive to the school's grasp of students' independent learning situation.

[0150] Based on the system design of a personalized recommendation system for online learning resources in universities, carry out detailed design and implementation of functional modules. This includes the detailed design and implementation of front-end and back-end modules, as well as the detailed design and implementation of the user profile submodule, graph data submodule, recommendation model training submodule, and recommendation list submodule of the back-end algorithm engine module. Visualize the key components with class diagrams or sequence diagrams.

[0151] 1. Front-end interaction module

[0152] The business processing submodule in the system's front-end interaction module is responsible for responding to front-end requests and completing data returns. The front-end display submodule is mainly used to provide functions such as student registration, student login, student portrait display, student basic information display, and recommended learning resource list display.

[0153] This project is based on the Flask framework to develop the front-end interactive module of the system. When students access it through a browser, the URL distributor distributes the interface to the View module, and then calls the corresponding class method in the Model module to complete the database query and update operations and return the result data layer by layer. Finally, the View module encapsulates it into an Html file and displays it to the Web browser, such as Figure 5 shown.

[0154] The business processing submodule is responsible for the requests and scheduling of the entire system, including basic data requests and data requests for recommended services, and completes tasks such as data return and database update through interfaces and front-end interfaces.

[0155] After receiving a data request from the system, the basic data request class queries, verifies, and updates the data through the interface and returns the results to the requester. If the request fails, a failure status will be returned. This includes student registration, login, user profile query, user personal information query, learning resource query, and student history query.

[0156] The recommendation service data request refers to when the system receives a request for a student's recommended resource list, it parses the request parameters, calls the interface for generating the recommended learning resource list, obtains the recommendation results, and returns the results to the request sender in a fixed-format JSON data.

[0157] The timing diagram of the business processing submodule is as follows Figure 6As shown in the figure, it is responsible for receiving and dispatching system requests, calling corresponding methods to respond to front-end requests and returning results. Database operations include querying and updating the database using encapsulation classes; accessing the database and retrieving data in the Request class, as well as obtaining student history records, and then generating a personalized list of recommended learning resources; the User class contains student-related fields, including personal information; and the Resource class defines fields related to learning resources.

[0158] (1) User registration page

[0159] On the user registration page, design text boxes for "Username" and "Password." When the user enters the correct information, the user will receive a "Registration Successful" message. If the user enters incorrect information, the user will receive a detailed explanation of the error, such as an incorrect password or an unavailable personal account name. Additionally, design a redirect link to facilitate the user's login to the "Login" page.

[0160] (2) User login page

[0161] On the user login page, similarly design text boxes for "Username" and "Password." If the user enters the correct information, provide a "Login Successful" message and redirect to the user's profile page. If the user enters incorrect information, provide a detailed explanation of the error, such as a missing username or incorrect password. The page also requires a redirect link to facilitate the user's redirect to the "Register" page to register a personal account.

[0162] (3) Homepage

[0163] This page displays the popular resources of the platform. When the user is not logged in (guest identity), the platform only displays popular learning resources; each learning resource on the homepage is displayed in the form of name and cover. When the user clicks the name and cover of a learning resource, it will jump directly to the resource introduction page, allowing the user to further understand the detailed information of the learning resource. Figure 7 shown.

[0164] (4) Resource recommendation page

[0165] If the user is logged in, the system will use the recommendation engine to generate recommended content for the user, and then through the resource recommendation page, it can show the user a personalized and diverse list of learning resources. Figure 8 Shown:

[0166] (5) Resource Introduction Page

[0167] The resource overview page provides users with a visual overview of the learning resource, including its description, number of participants, and ratings. Clicking "Join Learning" on this page completes course selection and leads them to the corresponding resource website to begin studying the selected course.

[0168] (6) History page

[0169] Users can enter the personal history page by clicking on their login name and see their personal historical course selection information.

[0170] 2. Backend algorithm engine module

[0171] The user profiling submodule and knowledge graph submodule within the backend algorithm engine module serve as pre-recommendations for the recommendation model training submodule and the recommendation list submodule, providing data preparation for subsequent modules. The user profiling submodule is divided into two parts: data preprocessing and profile construction. The graph data processing module is divided into three parts: knowledge graph construction, knowledge graph vectorization, and student similarity calculation.

[0172] The process of building a user portrait includes data acquisition, data preprocessing, model building, and portrait generation. In the data acquisition stage, it is divided into acquiring basic data and user behavior data. Basic data can be obtained through the campus integrated database, and user behavior data can be obtained through the log easy analysis mapping URL. These data are stored in the database, and data preprocessing and label modeling are performed as needed to finally form a user portrait. The time series diagram is as follows Figure 9 shown.

[0173] The construction of knowledge graph mainly includes operations such as raw data processing and triple extraction. The vectorization of knowledge graph mainly uses the triples generated in the knowledge graph construction process and uses TransD (not necessarily) to train to obtain entity vectors. The similarity calculation between students mainly uses the weights of the knowledge graph and the entity vectors to obtain the similarity matrix. The time series diagram is as follows Figure 10 shown.

[0174] During the recommendation model training phase, the pre-model has calculated similarity and other parameters, and processed the original data to generate the resource rating matrix to be optimized and similarity and other data. The user resource rating matrix is updated based on similarity. Parameters are optimized through model training to obtain the optimal weight coefficient and the optimal model. The timing diagram is shown in the figure below. Figure 11 shown.

[0175] The main task of the learning resource recommendation list submodule is to generate a personalized recommendation list for students based on the recommendation model training submodule. The student's historical behavior data is the main input data. This submodule is mainly divided into two parts: student cold start and student learning resource list generation. Student cold start is to predict the required resources of new students based on their majors and interests as a supplement to the student's resource needs and improve the recommendation process. The learning resource list recommendation process completes the recommendation of student learning resources by loading the student's learning records and the latest recommendation model. The timing diagram is shown as follows: Figure 12 shown.

[0176] 2.1 Construction of Student User Portrait

[0177] This section describes the user profile construction process, including data acquisition, data preprocessing, model building, and profile generation. The data acquisition phase involves acquiring basic data and user behavior data. Both basic data and user behavior data can be obtained from the campus integrated database. Data preprocessing and label modeling are performed as needed. Finally, a user profile is formed and updated within a fixed time window.

[0178] User portrait construction is divided into four levels: original data layer, data preprocessing layer, label model layer and portrait application layer, such as Figure 13 As shown in the figure, the raw data layer targets college students and collects basic attribute information from information systems such as the student affairs system, personnel system, academic affairs system, library management system, and retrieval system, as well as course information and historical behavior data that can represent user interests, preferences, and needs. The data preprocessing layer cleans and transforms the data, and then uses statistical analysis and data mining techniques to model the collected information to generate a label model layer. The portrait application layer forms student user portraits, sets periodic updates, and applies the portraits to the knowledge graph algorithm.

[0179] 2.1.1 Data Acquisition

[0180] The construction of a portrait for resource recommendation requires the collection of data that can represent the user's interests and needs. User data can be divided into two categories according to the speed of data change. The first category is mainly static data that changes slowly and is relatively stable, such as the user's student ID, gender, user type, college, major, courses currently being studied, and past course grades, among which majors and course grades are the main basis for resource recommendation; the second category is user dynamic data that changes rapidly, including borrowing and returning books, searching for digital resources, participating in academic reports, participating in thematic lectures, etc. These dynamic behavior data represent the user's interests and needs and are stored in multiple information systems such as student affairs, personnel, academic affairs, clubs, activity centers, library management, OPAC retrieval, etc. Each system is relatively independent and the data are unrelated. In order to establish a connection between these data, they need to be imported into the target database for centralized storage, and then data processing is performed. Here, a method based on shared data in the school data center is adopted, such as Figure 14 As shown, ESB technology (Enterprise Service Bus) is used for shared forwarding of data messages, and structured data and unstructured data are integrated into a shared database through traditional "extract-transform-load" technology, and reliable data services are provided to other applications through interface calls.

[0181] 2.1.2 Data Preprocessing

[0182] The basic attribute data collected from various systems may include structured, semi-structured, and unstructured data. There may also be a series of problems such as missing data, formatting errors, and logical errors. Therefore, the data must be pre-processed through conversion, specification, and integration to form effective student portrait data and store the data for subsequent calculations and other operations. Constructing user portraits inevitably requires collecting a large amount of user data, which involves user privacy issues. When integrating user personal information and historical behavior data, user data should be screened and privacy data, such as ID card numbers and mobile phone numbers, should not be integrated. Personal privacy labels should also be minimized when generating user portraits. Timed-out data can be destroyed, or technical means can be used to ensure user privacy.

[0183] 2.1.3 Model Construction

[0184] The collected data is organized. Static data is a basic attribute. If there is no historical behavioral data, personalized recommendations can be made based on the majors and compulsory course scores in the static data to avoid the "cold start" problem of the recommendation system; user dynamic data can express the user's true interests through in-depth mining and is an important basis for resource recommendations.

[0185] Data modeling allows the program to automatically assign tags and weights to users based on their behavior and pre-defined calculation rules. Tags represent user attributes, behaviors, interests, needs, and preferences, while weights represent quantifiable indicators such as the degree of interest, preference, and need.

[0186] The pre-processed data is labeled to form a user tag system for the online learning resource system. Based on the different attributes of the tags, the tags are divided into three levels of indicators: the first level is divided into basic attributes and learning behavior; the second level is divided into basic information, academic performance, book borrowing, digital resources, and other hobbies; and the third level is divided into name, student ID, gender, and user type.

[0187] 2.1.4 Generate portrait

[0188] The rules for generating user profiles are stored in MySQL. Adding, deleting, modifying, and querying rules are all achieved by connecting to MySQL. Based on the user's basic attributes and learning behavior, the rules for generating user profiles are formulated in conjunction with resource recommendation applications.

[0189] (1) Generate a profile based on the selected courses. Use Python functions to slice the obtained course names to obtain keywords, use the keywords as labels, and the number of occurrences as weights.

[0190] (2) Generate a profile based on the borrowing situation of books. Slice the book title to obtain keywords, use the keywords as labels, and use the keyword frequency as weight.

[0191] (3) Generate a profile based on offline course scores. Offline compulsory courses are very important. In order to improve the understanding of previous course knowledge, users are recommended based on their scores. Scores of 86 points and above are considered "excellent"; scores of 60-85 points are considered "good"; and scores below 60 points are considered "average". Because offline courses are highly directional, the slicing method is not used for offline courses. The course name is directly used as a label, and the weight of "good" is set to 1, and the weight of "average" is set to 1.5.

[0192] (4) Generate a portrait based on the lecture. Slice the lecture name and extract keywords as labels, and use the keyword frequency as the weight.

[0193] The elective courses, book borrowing and lectures are used as interest labels, and the offline course scores are used as professional labels to form a Figure 15 The personal portrait display shown in the figure outputs the results in reverse order of the weights of interest tags and professional tags, and the final number of recommended courses is supplemented in a 1:1 ratio.

[0194] 2.2 Knowledge Graph Construction

[0195] The knowledge graph for the purpose of personalized recommendation of online learning resources is mainly based on text data, including basic information of courses, information of lecturers, etc. The interactive data generated by students during the learning process is also valuable for mining. This invention takes the online education platform of China University Mooc as an example to build a knowledge graph of some courses for recommendation. The main contents include data acquisition, entity recognition and fusion, relationship extraction, knowledge graph construction storage and embedding. The idea of knowledge graph construction is mainly divided into four levels: data layer, technology layer, storage visualization layer and graph vectorization layer. Figure 16 shown.

[0196] 2.2.1 Data Acquisition

[0197] 1. User Data

[0198] Target student data is obtained in the same way as profiled student data. However, in personalized recommendation systems, to train the algorithm on massive amounts of data for more accurate recommendations, we can also obtain behavioral data from other students on online education platforms like Chinese University Moocs. In addition to student data, we can also obtain information such as the instructor's name, school, education background, professional title, and courses taught on the course page.

[0199] 2. Resource Selection

[0200] Resource data refers to the courses that need to be recommended. There are a large number of universities on the Chinese University Mooc platform. If all of them are crawled, it will not only take a lot of time, but the quality of courses offered by different universities is uneven. The value of the data will not increase significantly with the increase in quantity. Therefore, only double first-class universities are selected as the actual processing objects. The courses offered by such universities are wide-ranging, the number of students is large, and the data is highly standardized.

[0201] 3. Course data collection

[0202] A web crawler is a type of computer script that automatically captures and analyzes data from web pages according to user-defined rules, thereby retrieving and storing relevant information. However, writing a crawler script requires extensive programming experience and considerable time. Therefore, this paper uses a scraper for crawling, which operates on a similar principle. The China University Mooc online education platform was selected as the crawling target. The platform's course information, primary users, and data formats generally meet the requirements of this experiment, requiring only standardization of the scraped data. Using the Octopus scraper, data was crawled from the website's homepage, obtaining the homepage addresses of all universities on the platform and storing them in a list of URLs to be processed. Only Double First-Class universities were selected as the actual processing targets. After obtaining the target university's homepage address and storing it in the list, the scraper automatically opens the university's details interface. Clicking on a course to scrape information, the scraper returns to the previous page to continue scraping the next course information, continuing until all courses at the university are scraped. The scraper then returns to the previous page and clicks on the next university. This method obtains information such as the course name, course description, category, number of students, number of reviewers, and ratings for all courses offered by the university.

[0203] 2.2.2 Entity Classification and Entity Attribute Definition

[0204] Recommendations are primarily based on the similarity between students and courses, calculated based on their historical interests and course types. To make the online course knowledge graph more complete and comprehensive, we abstracted five entity concepts from the acquired data: courses, students, teachers, keywords, and schools.

[0205] After conceptually classifying entities, their attributes and value ranges need to be defined. Entity attributes include data attributes and object attributes. Data attributes refer to the characteristics and features of an entity, and their ability to describe things. Object attributes represent relationships between entities.

[0206] 2.2.3 Knowledge Extraction

[0207] After analyzing and processing the student and course data captured by the collector, most entity and attribute information can be obtained, which can be directly populated into the knowledge graph. However, keywords that highlight the course's characteristics must be extracted from the course title or course description. This requires the use of word segmentation technology. Currently, tools that support Chinese word segmentation include Python's Jieba word segmentation, Snownlp, and Thulac. Jieba is used as the word segmentation tool. Jieba word segmentation constructs a prefix dictionary based on a statistical dictionary and uses the prefix dictionary to segment the input sentence.

[0208] (1) Entity extraction

[0209] When splitting data sources, there are often function words, meaningless words, or punctuation marks such as "的", "是", "啊", "。". We can use publicly available stop word lists on the Internet to filter these words. When constructing a knowledge graph, the keywords to be extracted are those included in online courses. These professional terms such as "deep learning", "learning community", "set theory", etc. are not available in the built-in word segmentation dictionary of jieba, which will reduce the quality of word segmentation. Therefore, to improve the accuracy of the word segmentation results, an expert word list related to this is imported, and jieba word segmentation with a custom dictionary can extract entities more accurately.

[0210] (2) Relationship extraction

[0211] After extracting entities, constructing a knowledge graph also requires extracting the relationships between entities. According to the definitions of the above entity attributes, the relationships of the types "learning", "teacher-student", "colleague", "professor", "include", "belong to" can be extracted.

[0212] 2.2.4 Knowledge graph storage

[0213] After the construction of the model and knowledge extraction, to complete the construction of the knowledge graph, the extracted content also needs to be stored. In traditional relational databases, multiple Join operations are required when querying multi-level nested relationships. This method greatly increases the system query burden and reduces the query efficiency. To avoid this defect, a graph database represented by Neo4j is used for storage. The graph database is more efficient when querying multi-level relationships.

[0214] Neo4j uses the property graph model to store data. Neo4j mainly includes two basic data types: nodes (Nodes) and relationships (Relation). Nodes are connected by relationships to form a relationship network. The previously obtained knowledge entities and relationships are stored in the Neo4j graph database in the form of triples (head, relation, tail). Neo4j can display the relationships between entities in the form of a visual graph, and using the Cypher query language, it can efficiently query graph data.

[0215] 2.2.5 Knowledge graph embedding

[0216] After the knowledge graph is constructed, the entities and relationships in the graph are still expressed in text form and cannot be directly used when training a recommendation model. Therefore, the representations of entities and relationships are vectorized to achieve the mapping of low-dimensional vectors. The selected knowledge embedding models include: distance translation models (TransE, TransD) and semantic matching network models (Rescal). The above methods are experimented, and the model with the best embedding performance is selected.

[0217] The 250,658 triples obtained were divided into 8:1:1, of which 200,527 were randomly selected as training data, and the remaining 50,131 triples were divided into two parts, used as test data and validation set respectively.

[0218] To construct a knowledge graph for university online learning resources, we first used a scraper to crawl the required data from Chinese university Moocs. We then extracted knowledge entities and relationships, resulting in five entity types and six relationships. We then constructed a knowledge graph based on the triples generated from entity and relationship extraction, storing and visualizing it in Neo4j. Finally, we used three models to embed the knowledge graph, comparing and analyzing the performance of the different models. The optimal model was selected as a reference for subsequent recommendation model experiments.

[0219] 2.3 Personalized Learning Resource Recommendation Algorithm

[0220] 2.3.1 Shortcomings of Collaborative Filtering Algorithms and Optimization Ideas

[0221] The present invention proposes an optimization idea based on the user collaborative filtering algorithm, and integrates user portraits and user resource knowledge graphs to update the similarity calculation weights. The collaborative filtering algorithm is insufficient. According to the different calculation objects, it is divided into a user-based collaborative filtering algorithm (UserCF) and an item-based collaborative filtering algorithm (ItemCF). In the field of education, since students of the same major tend to learn similar learning resources and have a strong clustering effect, the present invention adopts a user-based collaborative filtering algorithm. This algorithm calculates the score of the resource by counting the interactive behavior between users and resources, selects similar users and then makes recommendations, which solves the problem of information navigation to a certain extent. However, the number of online learning resources in colleges and universities is huge. Although students study many courses, the behavioral data of most students is still sparse, which leads to a sparse user resource rating matrix, which makes personalized recommendations difficult. At the same time, new users have no historical behavioral data, and there is a problem of system cold start.

[0222] To solve the above problems, in learning resource recommendation, we use the entity vectors in the user resource knowledge graph to calculate the similarity between users, update the user resource rating matrix with the help of weights, and draw the obtained basic attributes and behavioral attributes into user portraits. We supplement them with weights according to professions and interests, and finally form a recommendation list. Therefore, for users with no or only a small amount of historical behavior data, we can still make recommendations based on user portraits, which alleviates the problem of data sparsity to a certain extent and avoids cold starts. The specific ideas of the optimization algorithm are as follows: Figure 17 As shown:

[0223] 2.3.2 Generation of recommendation lists

[0224] This section focuses on the specific methods for calculating user similarity and supplementing recommendations based on user profiles. This includes two main parts: first, calculating user similarities based on a weighted fusion of user entity vectors using a user-resource rating matrix to determine neighbor selection; second, determining users' professional and interest needs based on their profiles and supplementing recommendations based on their ratings. Finally, the specific online resource recommendation process is described.

[0225] 1. Rating matrix acquisition and user similarity calculation

[0226] Generally speaking, user ratings of resources are explicit feedback of user interests. These explicit feedback data can be used for direct calculation of collaborative filtering recommendation algorithms. The more overall user rating data there is, the more accurate the recommendations will be. However, the reality is not like this. Not every user is keen on rating the resources they visit. It can be said that only a few users have the habit of rating. Therefore, the rating matrix is basically sparse, which makes it difficult to calculate similarity. Even in an online learning environment, users who have the habit of rating only rate the subjects they study or related learning resources they are interested in, and it is impossible to evaluate all resources. The rating matrix is still sparse. Therefore, the acquisition of the rating matrix needs to rely on the user's implicit feedback. The present invention extracts the user's access log file to the online resource, obtains the user's access frequency and the last access time according to the timestamp, and calculates the user similarity after comprehensive processing and combining the entity vectors between users in Chapter 4.

[0227] (1) Obtaining the scoring matrix

[0228] Log files record the user's resource learning trajectory, including the number of times a resource has been studied and previously visited resources. This information reflects the user's implicit preferences. However, due to the platform's privacy policy and user rating habits, it's difficult to fully capture the user's explicit preferences for training. Only the resources the user has previously visited are available. Since the platform's courses are arranged by learning time, with the most recently studied courses at the top, the user resource rating matrix is weighted by course resources according to their time series.

[0229] A time series is a series of observations taken at intervals in a certain time sequence. Time series model forecasting methods include arithmetic averaging, moving average, weighted moving average, exponential smoothing, autoregressive and moving average. Here, we use the weighted moving average method, giving more weight to recent data and less weight to distant data to meet data weighting requirements. The formula is as follows:

[0230]

[0231] Among them, w1+w2+…+w n =1;xk represents the predicted value; w i Represents x k-i The weight of , i = 1, 2, 3…, n.

[0232] According to the calculation, we get an M*N rating matrix, r ij is the rating of user i on resource j. It is worth noting that every user has resources that they have not learned. ij is 0, as shown in Table 12.

[0233] Table 12 Example of user-resource rating matrix

[0234]

[0235] (2) User similarity calculation

[0236] The entity vector of the knowledge graph represents the position of the user in the knowledge graph network. The closer the spatial position is, the higher the similarity between the two is. The user vector of the knowledge graph is used to calculate the user similarity t using the cosine similarity algorithm. uu , as shown in Table 13.

[0237] Table 13 User similarity calculation example

[0238]

[0239] The similarity calculation of the user resource rating matrix also uses the cosine similarity algorithm. The core of this method is to use similarity to calculate similar neighbors and recommend the top-N similar neighbor rating results to the target user. The formula is as follows:

[0240]

[0241] Among them, Sim(a,b) represents the similarity between user a and user b, R a,j and R b,j denote the ratings of user a and user b on item i respectively.

[0242] The user resource rating matrix is updated based on the entity vector similarity. The parameter weight coefficient α is set. The value of α is between 0 and 1. The weight coefficient affects the entity vector and the user rating matrix. The specific value of α is determined through experiments. The formula is as follows:

[0243] sim(u, v) = αt uu (u,v)+(1-α)sim cf (u, v)

[0244] sim(u,v) represents the final similarity between user u and user v, t uu(u,v) represents the vector similarity between user u and user v in the knowledge graph relationship network, sim cf (u,v) represents the similarity between user u and user v in the user resource rating matrix.

[0245] After similarity calculation, the nearest neighbor is found through the similarity threshold. The similarity threshold of this project is set to 0.8. Then, users with similarity greater than the similarity threshold are selected, and online resources that have been learned by the nearest neighbor but not by user u are found for recommendation.

[0246] 2. User portrait recommendation and supplement

[0247] By identifying user attributes through user profiles, new users with limited or no historical behavioral data can be recommended courses based on their majors and interests, avoiding a cold start and providing additional insights for other users. The Ministry of Education categorizes university disciplines into 12 categories: philosophy, history, engineering, economics, science, medicine, law, literature, agriculture, management, education, and art. Each category offers different majors. While courses within the same major vary across institutions, similarities are common, particularly in core courses, which are almost mandatory. Therefore, online courses corresponding to required courses for each major are identified and stored in a database for easy search. Users can search for their interests directly in the database using keywords to find corresponding courses, which are then sorted and recommended based on ratings. A set number of recommendations is set, prioritizing online courses based on similarity calculations. If the total number of recommendations is not met, a 1:1 ratio of recommendations is made based on the major and interests in the user profile.

[0248] 3. Specific process of online resource recommendation

[0249] The online learning resource recommendation algorithm based on user similarity recommends learning resources to students according to the number set by the system. The algorithm generates an entity vector matrix of user portraits and knowledge graphs based on the user's basic attributes and historical interaction data, and makes recommendations based on data such as weights and totals. The recommendation process of the algorithm in the system is as follows: Figure 18 shown.

[0250] ① When a student enters the system, it is determined whether he or she is a new user. If so, recommendations are made based on majors and interests. If not, the knowledge graph entity vector and user resource rating matrix are obtained.

[0251] ② Use the formula to calculate user similarity and update the user resource rating matrix based on the weight.

[0252] ③ Use the formula to calculate user similarity, find the nearest neighbors based on the similarity threshold, and make recommendations.

[0253] ④ If the total number of recommendations does not meet the requirements, the top-N resources will be generated based on majors and interests to supplement them.

[0254] 2.3.3 Recommendation Algorithm Experimental Design and Experimental Results

[0255] 1. Experimental Design

[0256] This experiment, written in Python, experimentally calculated the range of weight coefficients and compared the recommendation algorithms. The experimental data was derived from online learning resources and user data from Chinese university Moocs, as well as data from academic administration systems, student affairs systems, library management systems, and activity centers. The experimental results were compared between the actual resource lists viewed by users in the training sample and the recommendation lists generated by the recommendation algorithms.

[0257] (1) Experimental design of weight coefficient values

[0258] The recommendation algorithm based on user similarity involves the weight coefficient α of similarity calculation. Some scholars will choose a fixed value based on their own experience, such as 0.5. The present invention will determine the optimal value of the weight coefficient α through experiments, and take the weight coefficient α from 0-1 with a step size of 0.2, and give a fixed value to α by comparing the experimental results under different values. In addition, according to the best performance of TransE in the experiment of knowledge graph vectorization in the previous stage, the entity vector is trained by the TransE model by default in this value experiment. According to the final similarity calculation formula, the smaller the value of the weight coefficient α, the smaller the proportion of entity vector similarity, and the greater the proportion of collaborative filtering similarity, and vice versa.

[0259] (2) Design of comparative experiments on recommendation algorithms

[0260] In order to compare the similarity-based recommendation algorithm with the traditional collaborative filtering algorithm (CF) and entity vector similarity algorithm, these three algorithms are used as input for experimental comparison, and are named CF, KGE and KGE-CF respectively. The weight coefficient used in KGE-CF is determined by the optimal weight coefficient α obtained in the previous experiment.

[0261] The favorite recommendation list and the algorithm recommendation list of the training sample can form multiple evaluation indicators. The results of the above two experiments can be obtained by comparing the sizes of the evaluation indicators.

[0262] 2. Evaluation indicators

[0263] There are many evaluation metrics for recommendation algorithms. For recommendation systems that predict ratings, these metrics primarily include mean square error (MSE), root mean square error (RMSE), and mean absolute error (MAE). These metrics measure the difference between predicted and actual ratings. Of these three types of metrics, mean absolute error (MAE) and root mean square error (RMSE) are more widely used to measure recommendation accuracy, better reflecting the actual level of error in predictions. Using RMSE or MAE can measure whether the ratings predicted by a recommendation system are close to the actual values, allowing for targeted improvements to the recommendation algorithm.

[0264] For the recommendation system of this experiment, the evaluation indicators used are accuracy (Precision), recall (Recall), and comprehensive evaluation index (F-Measure) to evaluate the quality of the recommendation algorithm.

[0265] The accuracy rate is also called the precision rate, which indicates the proportion of samples recommended to the user that the user is interested in among the total recommended samples.

[0266]

[0267] The recall rate is also called the recall rate. The recall rate indicates the proportion of items recommended to users that the user is interested in among all items of interest to the user in the dataset.

[0268]

[0269] Where R(u) represents the resource set in the recommendation list, and T(u) represents the learning resource set actually liked by the samples in the test dataset.

[0270] Precision and recall can evaluate the recommendation quality of a recommendation system to a certain extent, but there may sometimes be some contradictions between these two indicators. To deal with this special situation, the comprehensive evaluation indicator F-Measure was introduced to weightedly reconcile precision and recall.

[0271]

[0272] When α in the comprehensive evaluation formula is set to 1, the comprehensive evaluation metric becomes the common F1 index, also known as the F1-Score. This metric is often used in statistics to evaluate binary classification systems. Its value range is [0, 1], with larger values indicating more accurate classification. The F1-Scores of individual and hybrid algorithms can be calculated to evaluate the algorithms, demonstrating that the overall recommendation quality of hybrid algorithms is superior to that of individual algorithms.

[0273] The following is generated based on the resource list that users actually like in the training sample and the resource list recommended by the algorithm: Figure 19 The matrix shown here completes the calculation of accuracy, recall and F1-Score.

[0274] 3. Experimental Results

[0275] This section will present the results of experiments on the value of the weight coefficient α and the comparison of the recommendation algorithm, with the aim of selecting the optimal weight coefficient α and the appropriate learning resource recommendation algorithm.

[0276] (1) Experimental design of weight coefficient values

[0277] In order to determine the optimal value of the weight coefficient α, the weight coefficient α was varied from 0 to 1 with a step size of 0.2. Ten experiments were conducted for each value, and the final evaluation index was averaged.

[0278] As α increases, the recall rate increases. When α is 0, the recall rate is 0.45. When α is 0.8, the recall rate reaches a higher level of about 0.77. Compared with α being 0, the recall rate increases by about 71%.

[0279] As α increases, the accuracy rate first increases and then decreases. When α is 0, the accuracy rate is 0.16. When α is 0.2, the accuracy rate reaches a high level of approximately 0.21. Compared to when α is 0, the accuracy rate increases by approximately 31.2%.

[0280] As α increases, the F1-score shows a trend of first increasing and then decreasing. When α is 0, the F1-score is 0.2. When α is 0.2, the F1-score reaches a high level of approximately 0.22. Compared to when α is 0, the F1-score improves by approximately 10%.

[0281] In summary, the inclusion of knowledge graph entity vectors significantly improves the quality of the recommendation algorithm. Compared to an α of 0 (i.e., no knowledge graph entity vectors), recall, precision, and F1-score all increase by over 10%. Overall, a weight coefficient α of 0.2 is optimal.

[0282] (2) Design of comparative experiments on recommendation algorithms

[0283] To compare the similarity-based recommendation algorithm with the traditional collaborative filtering algorithm (CF) and entity vector similarity algorithm, we selected different numbers of online resource recommendations and conducted an experimental comparison of the three algorithms in terms of recall rate, precision rate, and F1-score. The weight coefficient used in KGE-CF was determined by the optimal weight coefficient α obtained in the previous experiment, which is 0.2. The experimental results are as follows. In terms of recall rate, KGE-CF improved by an average of 17 percentage points compared to CF and KGE, an increase of about 29.9%. Figure 20 As shown in Figure 2, KGE-CF improves the accuracy by an average of 9 percentage points compared to CF and KGE, an increase of approximately 25.6%. Figure 21 As shown in Figure 2. In F1-score, KGE-CF improves by an average of 4 percentage points compared to CF and KGE, an increase of approximately 10.9%. Figure 22 shown.

[0284] In summary, the recommendation algorithm based on similarity calculation is superior to the traditional collaborative filtering algorithm (CF) and knowledge graph entity vector (KGE) similarity calculation in terms of the three evaluation indicators, proving that the algorithm has an improvement effect.

[0285] We designed an online resource recommendation algorithm based on similarity, supplemented by user profiles. Through experimental comparisons, we determined the optimal weighting coefficient and demonstrated the algorithm's improved performance. We first introduced the shortcomings of traditional collaborative filtering algorithms and optimized approaches based on knowledge graphs. We then presented the detailed recommendation process for obtaining a rating matrix, calculating similarity, and supplementing user profiles. Finally, through value selection experiments, we determined the optimal value of α to be 0.2. Through algorithm comparison experiments, we demonstrated that our algorithm outperformed both traditional collaborative filtering algorithms and knowledge graph entity vector similarity calculation algorithms.

[0286] 3. System testing

[0287] The system was tested primarily from the perspectives of functionality and performance, with the goal of identifying problems in the system design and code development process. The system test environment is shown in Table 14.

[0288] Table 14 System test environment

[0289]

[0290] 3.1 Functional Testing

[0291] Functional testing verifies all system functions, primarily to ensure that users can access all system features normally. This system functional testing consists of three parts: front-end interaction module testing, model training testing, and personalized recommendation testing. The system interaction module testing primarily focuses on student registration, login, personal information, and user profile querying, as shown in Table 15. Model training testing primarily tests automated model updates, as shown in Table 16. The personalized recommendation module primarily tests the ability to provide personalized resource recommendations for new and existing users, as well as for the same user over time, as shown in Table 17.

[0292] Table 15 System interaction module test

[0293]

[0294] Table 16 Model training and testing

[0295]

[0296] Table 17 Personalized recommendation module test

[0297]

[0298] There are 150 test cases in total, of which 2 cases failed, and the pass rate is 98.6%, which is higher than 98%, and the test passed.

[0299] Performance testing focuses on system performance, such as response time and high-concurrency processing. In this system, response time is used as a performance test indicator, and the response time is required to meet TP90 (Top Percentile 90) = 500ms. That is, among all requests initiated to a certain interface, at least 90% of the request response time is within 500ms. "TP90" is one of the commonly used test indicators in the Internet industry.

[0300] Testing was conducted using the browser's performance component, a performance testing tool in the browser's developer mode that measures page performance while the browser is running. DOM loading (DomContentLoaded, DCL) begins at 200ms, at which point the HTML document has been parsed. First Paint (FP), First Contentful Paint (FCP), and Largest Contentful Paint (LCP) begin around 240ms, and loading (onLoad, L) is complete around 630ms. Monitoring the overall interface performance shows excellent front-end performance, meeting the platform's requirements for front-end interfaces.

[0301] The results show that the constructed personalized recommendation system for online learning resources in colleges and universities has passed the test and can operate normally.

[0302] The present invention can be widely applied to autonomous learning assistance systems in universities, providing personalized learning resource recommendations for university students. By constructing student user profiles and combining learning behavior data with professional focus, it can intelligently screen and match high-quality course resources from online learning platforms such as China MOOC, thereby reducing the time students spend blindly searching through massive learning resources and improving learning efficiency. At the same time, the present invention can effectively promote the rational use of online educational resources, enabling students to obtain learning paths that better meet their personal needs, assisting their professional development and knowledge expansion, and promoting the upgrading of university education models towards intelligentization.

[0303] The present invention can be integrated into a variety of computer devices, including but not limited to smart learning terminals, online learning platforms, educational apps, and university teaching management systems. The computer device includes a memory and a processor. The memory stores a computer program. When the processor executes this program, it automatically analyzes students' interests and learning behaviors, calculates course associations based on a knowledge graph, and generates a personalized list of recommended learning resources. This computer device can be used in the construction of smart campuses in universities, or as a standalone smart education terminal, providing more accurate and personalized learning guidance for university teachers and students.

[0304] The computer-readable storage medium of the present invention can be applied to the recommendation system module of an online education platform to store and execute intelligent recommendation algorithms to achieve personalized push of learning resources. Based on knowledge graphs and user profiles, the storage medium can track students' learning progress over the long term and perform real-time optimization in conjunction with artificial intelligence algorithms, thereby improving the accuracy and matching of recommendations. The storage medium can be integrated into university education management systems, distance education platforms, online tutoring platforms, etc., helping to build an intelligent education ecosystem and provide students with a customized learning experience.

[0305] Furthermore, the intelligent information data processing terminal of the present invention can be widely used in university autonomous learning systems, educational cloud service platforms, intelligent courseware recommendation systems, and other applications to improve the utilization efficiency of learning resources. The terminal integrates multiple functional units, including a knowledge graph analysis module, a user behavior data analysis module, and an intelligent recommendation algorithm module. It can update learning resource information in real time, optimize recommendation strategies, and ensure that students can quickly access the highest-quality learning content. The application of this terminal not only enhances the personalization level of online education, but also provides technical support for the development of university education informatization, promoting the development of smart education.

[0306] It should be noted that the embodiments of the present invention can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in a memory and executed by an appropriate instruction execution system, such as a microprocessor or dedicated design hardware. Those skilled in the art will appreciate that the above-mentioned devices and methods can be implemented using computer-executable instructions and / or contained in processor control code, for example, such as a carrier medium such as a disk, CD or DVD-ROM, a programmable memory such as a read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field programmable gate arrays, programmable logic devices, etc., can also be implemented by software executed by various types of processors, or can be implemented by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0307] The above description is only a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions and improvements made by any technician familiar with this technical field within the technical scope disclosed by the present invention and within the spirit and principles of the present invention should be covered by the scope of protection of the present invention.

Claims

1. A personalized online resource recommendation system for college students, characterized by: include: The front-end interaction module sends requests to the algorithm engine module below and returns personalized learning resource recommendation results. It also provides students with basic interactive interface functions and the display function of personalized learning resource recommendation services. It is divided into a front-end display sub-module and a business processing sub-module. The backend algorithm engine module processes data from the lower-level data storage module and provides data and algorithm support for upper-level businesses. It is divided into user portrait sub-module, knowledge graph sub-module, model training sub-module, and recommendation list sub-module; Data storage module, used to store and manage data in the system; A system parameter setting module, connected to the data storage module, for setting system parameters; The front-end interaction module specifically includes: the front-end display sub-module is mainly responsible for receiving students' data requests, and passing them down to the system business processing sub-module to obtain the results, thereby realizing the visual transmission of the recommendation list; the main function of the business processing sub-module is to receive data requests from the front-end display sub-module, perform corresponding data processing and analysis, and deliver them to the front end; receive the recommendation list input from the back-end algorithm engine module, and provide the results to the front-end display sub-module for display.

2. The online resource personalized recommendation system for college students according to claim 1, characterized in that: The business processing submodule is responsible for the requests and scheduling of the entire system, including basic data requests and data requests for recommendation services, and completes data return and database update tasks through interfaces and front-end interfaces; After receiving the data request from the system, the basic data request completes the query, verification and update of the data through the interface and returns the result to the requester; if the request fails, a failure status will be returned; including student registration, login, user portrait query, user personal information query, learning resource query and student history information query; Recommendation service data request refers to when the system receives a request for a student's recommended resource list, it parses the request parameters, calls the interface for generating the recommended learning resource list, obtains the recommendation results, and returns the results to the request sender in a fixed-format JSON data; The business processing submodule is responsible for receiving and scheduling received system requests, and calling corresponding methods to respond to front-end requests and return results; in related operations on the database, it includes using encapsulation classes to query and update the database; in the Request class, the database is accessed and data is obtained, and at the same time, the student's historical records are obtained, and then a personalized recommended learning resource list is generated; the User class is the relevant fields for students, including the student's personal information fields; the Resource class defines the relevant fields for learning resources.

3. The online resource personalized recommendation system for college students according to claim 1, characterized in that: The backend algorithm engine module specifically includes: The user portrait submodule is responsible for data processing and modeling of user portraits, providing data for subsequent cold start processing; it includes basic data preprocessing and user portrait construction; The knowledge graph submodule is responsible for knowledge graph data processing and modeling, and prepares data for the recommendation model training submodule. This includes knowledge graph construction, knowledge graph vectorization, and student similarity calculation. Knowledge graph construction includes raw data preprocessing and triple extraction; the optimal model is used for knowledge graph vectorization representation; and student similarity is calculated using knowledge graph relationship weights and entity vectors. The model training submodule uses the student similarity information obtained by the knowledge graph submodule to obtain the optimal weight coefficient and update the user resource evaluation matrix, providing relevant data for the recommendation list submodule; this module includes recommendation model data processing and recommendation model training; The recommendation list submodule includes two parts: algorithm cold start processing and learning resource recommendation list generation. Cold start processing alleviates the sparse rating matrix problem caused by insufficient student historical behavior data by predicting student needs. For different students, this module obtains the user portrait of the student from the lower-level data storage module, makes recommendations based on their professional characteristics and interests, generates a recommendation list after sorting, and returns it to the front-end interaction module for display.

4. The online resource personalized recommendation system for college students according to claim 1, characterized in that: The user portrait submodule and knowledge graph submodule in the backend algorithm engine module are pre-modules for the recommendation model training submodule and the recommendation list submodule, and their purpose is to provide data preparation for subsequent modules. The user portrait submodule is divided into two parts: data preprocessing and portrait construction. The graph data processing module is divided into three parts: knowledge graph construction, knowledge graph vectorization, and similarity calculation between students. The process of building a user portrait includes data acquisition, data preprocessing, model building, and portrait generation; In the data acquisition stage, it is divided into acquiring basic data and user behavior data; Knowledge graph construction, mainly including raw data processing and triple extraction operations; knowledge graph vectorization; During the recommendation model training phase, the pre-model has calculated similarity parameters and processed the raw data to generate the resource rating matrix and similarity data to be optimized. The user resource rating matrix is updated based on similarity. Parameters are optimized through model training to obtain the optimal weight coefficient and the optimal model. The main task of the learning resource recommendation list submodule is to generate a personalized recommendation list for students based on the recommendation model training submodule; the students' historical behavior data is the main input data.

5. The online resource personalized recommendation system for college students according to claim 1, characterized in that: The system parameter setting module specifically includes: The system can be set to turn resource recommendations on or off. Since turning on the recommendation function requires collecting students' personal behavior information and requires their consent, it is turned off by default; The number of recommended resources each time, the default is 20; Recommended message sending methods: Message Center supports email, WeChat, SMS, and their combinations. By default, only email is sent. Student personal information maintenance, phone calls, emails, and student status changes can be automatically synchronized with the academic affairs system; Students can customize their learning interests and receive recommended academic resources accordingly; After completing the connection with various data sources, the learning resource recommendation service platform regularly generates labels, calculates weights and time decay every day to ensure that the platform maintains the latest learning resources and student data. Recommended resources are sent regularly every week. The automatic operation mode reduces dependence on management. College leaders and student management departments access the platform to understand students' learning situation. In addition, the platform records the entire service process, which is conducive to the school's grasp of students' independent learning situation.

6. A personalized online resource recommendation method for college students using the system according to any one of claims 1 to 5, characterized in that: include: (1) Check whether the model needs to be retrained: There are currently two situations where model training is required. One is when there is no model currently, model training is required; the other is to regularly collect student update data and resource update data, and regularly train the model on the data in the database; (2) Pre-trained model loading: Since the system needs to meet the fluency requirements, and the computational efficiency of online real-time computing is low, which affects the user experience, the offline model is loaded to perform learning resource recommendation prediction. The offline model includes user portraits, vectorized parameters of the knowledge graph, similarity matrix between students, user resource rating matrix, and weight coefficient parameters. After receiving the user's request, the model parameters will be requested from the database, and then a recommendation list will be generated. (3) User login and request: In the recommendation system, different students use their student IDs to distinguish their user requests, which are passed layer by layer through the algorithm engine module and database submodule below, and finally a list of recommended learning resources is returned to the student; (4) Recommendation result generation: Obtain a sorted list of learning resources through a similarity-based collaborative filtering algorithm. If the quantity requirement is not met, request the user profile parameters to supplement it and return it to the student.

7. A computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of the method for personalized online resource recommendation for college students as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the method for personalized online resource recommendation for college students as described in any one of claims 1 to 5.

9. An information data processing terminal, comprising the online resource personalized recommendation system for college students according to claim 8.

Images