An educational course recommendation system, method and application based on big data in the education industry

Abstract

An educational course recommendation system, method and application based on big data in the education industry, belonging to the field of education and training, by building a full-scenario label user portrait model L‑KPR‑AstralEngine based on a machine learning algorithm, and outputting user portraits; by building a DeepFM model with a gating mechanism, and using the user portraits output by the L‑KPR‑AstralEngine user portrait model to adjust parameters and train the model, and finally output a processed course recommendation list. The present invention solves the limitations of existing course recommendation systems and methods in terms of accuracy, personalization and data processing complexity, improves the accuracy and personalization of course recommendations, improves the ability to process large-scale data sets, improves system scalability, and reduces system maintenance costs and enterprise operating costs.

Description

Technical Field

[0001] The present invention belongs to the field of education and training technology, and specifically relates to an education course recommendation system, method and application based on big data in the education industry. Background Art

[0002] With the booming development of the training market, various training institutions have emerged and provided diversified course services. When purchasing courses, users need to consider factors such as service model (one-year, repeat, refund), teaching method (on-demand, live broadcast, face-to-face), business type (boutique, high-end, private school), course price, class address, and teaching teacher. Therefore, users need to conduct a detailed comparative analysis when making a choice. In addition, the large number of users and the different characteristics of course purchases have forced training institutions to recruit a large number of course consultants to answer questions one-on-one. Although this provides personalized services to users, it greatly reduces the business efficiency of the enterprise and increases operating costs.

[0003] With the increase of online education and digital learning resources, how to effectively recommend suitable courses to suitable users has become a challenge. At present, in the field of education and training, course recommendation systems usually rely on simple keyword matching or recommendation algorithms based on user historical behavior. Among them, traditional recommendation algorithms based on user historical behavior mainly include content-based recommendation, collaborative filtering and hybrid recommendation. The main principle of content-based recommendation algorithm is to make recommendations based on the attributes of items and user preferences. Although content-based recommendation algorithms can recommend personalized and interest-oriented items to users, they are suitable for new users and cold start problems. However, content-based recommendation algorithms are prone to recommendation limitations, making it difficult to explore users' potential interests and relying heavily on item labels and attributes. The main principle of collaborative filtering algorithm is to make recommendations based on user behavior data or the similarity between items by mining the association between users or items. Its advantage is that it does not rely on the content information of items and can discover users' hidden interests. However, its disadvantage is that it is easily affected by sparsity and cold start problems and requires a large amount of user behavior data. The main principle of the hybrid recommendation algorithm is that it can combine multiple recommendation algorithms to make up for the shortcomings of each algorithm and improve the overall recommendation performance. Although the hybrid recommendation algorithm can give full play to the advantages of each algorithm and improve the accuracy and coverage of the recommendation. However, the overall algorithm needs to comprehensively consider the parameter adjustment and integration of multiple algorithms, and the complexity is relatively high. The rule-based recommendation algorithm can make recommendations based on the rules defined by domain experts. The rules can be based on the user's behavior habits, the characteristics of the items, etc. The rule-based recommendation algorithm can be flexibly customized according to domain knowledge and is suitable for specific application scenarios. However, the process requires domain experts to define the rules, which may not be flexible enough for complex and large-scale data sets.

[0004] Although the existing course recommendation system can meet user needs to a certain extent, it still faces several obvious problems in terms of existing course recommendation:

[0005] 1. Insufficient data utilization: Existing course recommendation systems often rely on relatively simple algorithms such as collaborative filtering algorithms or content-based recommendation algorithms, and fail to fully utilize complex information in big data, such as multi-dimensional analysis of users' professional background, learning progress, and personal interests, and cannot fully understand users' complex needs and diverse preferences.

[0006] 2. Insufficient accuracy of course recommendations: Simple keyword matching or historical data analysis cannot accurately capture users' current learning needs and long-term educational goals. Existing course recommendation systems face performance bottlenecks when processing large-scale or high-dimensional user data, especially when the amount and complexity of data increase, resulting in inaccurate course recommendations.

[0007] 3. Poor adaptability: The existing course recommendation system has problems of poor scalability and high maintenance costs. It is difficult to adapt to the rapidly changing education and training knowledge and diverse learning needs, and it is often impossible to update the recommendation logic and content in a timely manner. When it is necessary to enhance system functions or expand the scope of services, high maintenance costs and complex system upgrades are usually required.

[0008] As the demand for personalized and precise education continues to increase, existing research technologies have failed to effectively provide a course recommendation system that can dynamically learn and adapt to the changing needs of users. Especially in the rapidly developing and information-intensive field of education and training, a smarter and more flexible course recommendation system is needed to process and analyze large amounts of complex data and provide accurate personalized course recommendations. Summary of the invention

[0009] In order to solve the problems of insufficient data utilization, insufficient course recommendation accuracy and poor adaptability in existing course recommendation systems, the present invention provides an education course recommendation system, method and application based on big data of the education industry.

[0010] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0011] The present invention provides an educational course recommendation method based on big data in the education industry, which mainly includes:

[0012] Step S1: Build a full-scenario label user portrait model L-KPR-AstralEngine based on a machine learning algorithm and output the user portrait;

[0013] The constructed L-KPR-AstralEngine user portrait model consists of a label module, a portrait module, and a prediction module, and its function formula is:

[0014] F(L,K,P,R)=F C (F P (F L (L),K,P),R)

[0015] Where L represents label data input, K represents K-means clustering algorithm, P represents principal component analysis, R represents random forest algorithm, and F L Indicates the label module, F P represents the image module, F Crepresents the prediction module;

[0016] Step S2: Build a DeepFM model with a gating mechanism, and use the user portrait output by the L-KPR-AstralEngine user portrait model to adjust parameters and train the model, and finally output the processed course recommendation list.

[0017] Furthermore, the label module is used to perform large-scale user label generation and multi-path label recovery calculation to form user basic label system data according to business needs within the company system; the formed user basic label system is divided into two categories: basic labels and extended labels.

[0018] Furthermore, the portrait module is used to construct the basic tags of the user and then use K-means clustering and principal component analysis algorithms to supplement the completeness and clearer outline of the user portrait; the function formula of the portrait module is F P (X, K, P), X represents the output from the label module, that is, the user's basic label system data; K represents K-means clustering processing, which is used for data classification and grouping; P represents principal component analysis algorithm processing, which is used for data dimension reduction and important feature extraction; principal component analysis is used to reduce the dimension of clustering results, simplify the model and improve computational efficiency.

[0019] Furthermore, the prediction module is used to calculate the matching degree between the service function representation and the user demand representation through the random forest algorithm, and predict the product or service that the user may request to call; the function formula of the prediction module is F C (Y, R), Y represents the output of the profiling module, that is, the processed user data, and R represents the random forest algorithm, which is used to make decisions and predictions based on the processed user data.

[0020] Furthermore, the constructed L-KPR-AstralEngine user portrait model adds an item-level attention mechanism and an interest-level attention mechanism; the item-level attention mechanism is used to determine which items in the system are more important to the user or task. For each item x in the system i Define a mapping function f(x i ) maps it to an attention score s i , and then these attention scores are converted into attention weights w through the softmax function j Indicates the importance of the j-th behavior s j =f(x j ):

[0021]

[0022] Finally, the attention weight a is usedi Perform weighted summation on the input items, and get the weighted item representation X weighted :

[0023]

[0024] The interest-level attention mechanism is used to determine the degree of attention paid by the user or system to different behaviors or attributes. j Define a mapping function g(b j ) maps it to an attention score s j , and then these attention scores are converted into attention weights w through the softmax function k Indicates the importance of the k-th behavior s k =g(b k ):

[0025]

[0026] Finally, the attention weight w is used j Perform weighted summation on the input items, and get the weighted item representation B weighted :

[0027]

[0028] Furthermore, the DeepFM model combines a factorization machine and a deep neural network, using the factorization machine to process the linear relationship in the multi-dimensional features of the user, while capturing complex nonlinear interactions through the deep neural network; the DeepFM model adds a long short-term memory network as a gating unit to dynamically adjust the contribution of features in the deep neural network; the calculation formula of the DeepFM model is as follows:

[0029]

[0030] Among them, w 0 and w i Respectively represent the bias term and linear weight of the model; u i and u i Both represent embedding vectors, which are used to simulate the interaction between features; DNN(p) represents the deep neural network part, which is responsible for processing high-order feature interactions; p represents the processed user portrait; the DeepFM model outputs the processed user portrait p to complete the feature interaction.

[0031] Furthermore, the parameter adjustment method is: adjusting the learning rate, optimizing the embedding dimension and adjusting the deep neural network structure; during the model training process, the loss function L of the model is defined by combining the cross entropy and the regularization term to ensure the stability and efficiency of the model during the training process:

[0032]

[0033] Among them, CrossEntropyLoss represents the cross entropy loss, which is used to evaluate the predicted label output by the model The difference between the true label y and the real label y; λ represents the regularization coefficient; All model parameters are included, and the regularization term is used to prevent overfitting.

[0034] Furthermore, the DeepFM model with gating mechanism is used to generate processed user portraits; the generation of user portraits is described as a multi-dimensional feature conversion process, in which each feature vector x is converted into a user-specific portrait through the model:

[0035] p=L-KPR(x)

[0036] Among them, L-KPR represents the mathematical model of the user portrait construction process;

[0037] The predicted label output by the model After processing the DeepFM model with a gating mechanism combined with the user portrait, the specific calculation formula is as follows:

[0038]

[0039] Where σ represents the sigmoid activation function, which is used to convert the predicted label output by the model Convert to predicted probabilities.

[0040] The present invention provides an education course recommendation system based on education industry big data, which is used to implement an education course recommendation method based on education industry big data provided by the present invention.

[0041] The present invention also provides an application of an education course recommendation system based on education industry big data in the recommendation of medical teaching course products.

[0042] The beneficial effects of the present invention are:

[0043] The present invention provides an educational course recommendation method based on big data in the education industry. The method constructs a full-scenario label user portrait model L-KPR-AstralEngine based on a machine learning algorithm and outputs a user portrait. The method constructs a DeepFM model with a gating mechanism, and uses the user portrait output by the L-KPR-AstralEngine user portrait model to perform parameter adjustment and model training, and finally outputs a processed course recommendation list. The present invention solves the limitations of existing course recommendation systems and methods in terms of accuracy, personalization, and data processing complexity. Compared with the prior art, the present invention has the following advantages:

[0044] 1. Improved the accuracy and personalization of course recommendations;

[0045] The L-KPR-AstralEngine user portrait model and the DeepFM model with a gating mechanism combine advanced machine learning technology and deep learning algorithms, enabling the course recommendation system to deeply analyze and understand the specific needs of users, significantly improve the accuracy and personalization of course recommendations, and ensure that each user receives the course recommendations that best suit their learning path.

[0046] 2. Improved the ability to process large-scale data sets;

[0047] The L-KPR-AstralEngine user portrait model effectively processes and analyzes large-scale user data sets by using PCA and random forest algorithms, reducing computational complexity while maintaining the depth and quality of data analysis. This enables the course recommendation system to not only run quickly, but also process more extensive and detailed data sets, providing more comprehensive course recommendations.

[0048] 3. Improved system scalability and reduced system maintenance costs and enterprise operating costs;

[0049] The present invention takes into account the requirements of easy maintenance and upgrading in its design. The modular design of the course recommendation method and system makes the system easy to expand and maintain, reduces long-term operating costs, and improves system scalability. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 This is the architecture diagram of the L-KPR-AstralEngine user portrait model.

[0051] Figure 2 This is the architecture diagram of the L-KPR-AstralEngine+DeepFM+LSTM model.

[0052] Figure 3 The following is a comparison chart of the ROC curves of each model.

[0053] Figure 4 Bar chart of accuracy and F1 score.

[0054] Figure 5 Confusion matrix model diagram for each model. DETAILED DESCRIPTION

[0055] The present invention is further described in detail below in conjunction with the accompanying drawings.

[0056] In a first aspect, the present invention provides an educational course recommendation method based on big data of the education industry.

[0057] The present invention provides an educational course recommendation method based on big data in the education industry, and its specific operation process is as follows:

[0058] Step S1: Build a full-scenario label user portrait model L-KPR-AstralEngine based on a machine learning algorithm and output the user portrait;

[0059] Step S1.1: The framework structure of the constructed L-KPR-AstralEngine user portrait model is as follows Figure 1 As shown in the figure, it mainly consists of three parts: label module, image module and prediction module.

[0060] The label module is mainly used to generate large-scale user labels and perform multi-channel label recovery calculations to form basic label system data for users based on business needs within the company system.

[0061] The portrait module is mainly used to construct the user's basic tags and then use K-means clustering and principal component analysis (PCA) algorithms to supplement the completeness and clearer outline of the user portrait.

[0062] The prediction module is mainly used to calculate the matching degree between the service function representation and the user demand representation through the random forest (RF) algorithm, and predict the product or service that the user may request to call.

[0063] The function formula of the constructed L-KPR-AstralEngine user portrait model is as follows:

[0064] F(L, K, P, R) = F C (F P (F L (L), K, P), R)

[0065] Where L represents the label data input, K represents the K-means clustering algorithm, P represents the principal component analysis (PCA), R represents the random forest (RF) algorithm, and F L Indicates the label module, F P represents the image module, F CRepresents the prediction module.

[0066] Step S1.2: labeling module;

[0067] Large-scale user data is used as input data, and user tags (system tags, business tags, passive tags, behavioral tags and more tags) are generated and multi-way tag recovery calculations are performed through the tag module. By extracting user behavior and preference information, a basic tag system for users is formed. The obtained user basic tag system data, namely the user data set, will be used as input data for the portrait module.

[0068] Specifically, the basic tag system of users can be divided into two categories: basic tags and extended tags. Among them, the basic tags mainly include natural attributes and location attributes. The natural attribute tags include gender, age, education level, occupation, industry, social class and income level. They provide a basic socioeconomic description for users and help the system understand the basic framework of users. Location attributes include information such as residence, work place and business district. This information is crucial for understanding the user's geographical environment and possible regional preferences, especially in the promotion of regional services. Extended tags mainly include hobbies, abilities, behaviors, system and project levels. Hobbies: Involving the user's personal interests and lifestyle, such as preferred entertainment activities, educational orientation, consumption behavior and attitude towards new technologies. Abilities: Reflecting the user's economic strength and consumption potential, including vehicle ownership, real estate and general consumption ability. Behavior: Extracted directly from the user's behavior data on the platform, such as the frequency of browsing specific courses, the number of interactions, the duration of learning, etc. System: Contains tags driven by artificial intelligence, such as smart recommendation tags and user feedback tags dynamically generated based on user behavior. Project level: Pay special attention to the user's interest in specific areas.

[0069] In order to support the above user portrait system and data feature extraction, the present invention designs a series of database tables, each of which has its specific function and purpose:

[0070] (1) User basic information table:

[0071] Function: Stores basic information of users, such as name, gender, age, educational background, etc. Significance: This information provides the necessary raw data for building basic tags and is the core component of the user portrait system.

[0072] (2) User tag table:

[0073] Function: Links users to their corresponding tags. Each tag reflects a specific attribute or behavior of the user. Significance: This table is the key to concretize abstract tags and apply them to individual users, enabling the system to dynamically classify and recommend users based on tags.

[0074] (3) Label table:

[0075] Function: Defines the detailed information of all available tags, including tag name, type, level and status. Significance: Through the management of this table, the tag system can be maintained and updated to ensure the accuracy and applicability of the tags.

[0076] (4) Source tag table:

[0077] Function: Associates labels and data sources, ensuring that each label can be traced back to its data source. Significance: This table supports the verifiability and transparency of label data, which is essential for analyzing label validity and data quality.

[0078] (5) User master table:

[0079] Function: Records user contact details and account status. Significance: Provides support to operations and customer service to ensure effective communication and service strategies can be implemented for specific users.

[0080] (6) Data source table:

[0081] Function: Manage detailed information of all data sources, such as source number and name. Significance: Support the overall architecture of data management, making the source of user data and behavior data traceable, and increasing the transparency and credibility of data processing.

[0082] Through the design and interconnection of these tables, the present invention can not only store and process a large amount of user data, but also ensure that the application of these data in the process of user portrait construction and course recommendation is accurate and efficient. Each table is an important part of the entire user portrait system, and together supports the realization of personalized medical education services.

[0083] The basic user label system formed by the present invention not only supports a fine-grained understanding of user characteristics, but also can more accurately model and predict user needs and preferences by integrating these multi-dimensional label data.

[0084] Step S1.3: Portrait module;

[0085] Based on the label module, the user's basic label system data is processed through K-means clustering and principal component analysis (PCA) algorithm, and the users are clustered based on the generated basic label system data to identify similar user groups to improve the targeting of recommendations. At the same time, the clustering results are reduced in dimension to highlight the most critical user features, optimize the data quality of subsequent processing steps, and enhance the integrity and clarity of user portraits.

[0086] The portrait module specifically includes: multiple clustering analysis layers and multiple principal component extraction layers.

[0087] Specifically, the function formula of the portrait module is F P (X, K, P), where X represents the output from the label module, that is, the user's basic label system data; K represents K-means clustering processing, which is used for data classification and grouping; P represents principal component analysis (PCA) algorithm processing, which is used for data dimensionality reduction and important feature extraction. Using principal component analysis (PCA) to reduce the dimensionality of clustering results can simplify the model and improve computational efficiency.

[0088] Specifically, the specific process of the K-means clustering algorithm is as follows:

[0089] 1) First import the original data;

[0090] 2) Setting parameters, i.e., the parameter K for how many clusters the original data is divided into, and adding clustering effectiveness indicators to determine the best K value; cycle threshold, i.e., the maximum number of cycles per clustering calculation; distance difference threshold, i.e., when the distance between the new centroid and the old centroid differs by a certain amount, clustering is stopped;

[0091] The clustering validity indexes include, but are not limited to, the silhouette coefficient (SC), the Calinski-Harabasz Index (CH), the Davies-Bouldin Index (DBI), etc.

[0092] 3) Randomly obtain K data from the original data as the initial centroid;

[0093] 4) Using Euclidean distance to calculate the distance from the original data to the identified centroid;

[0094] 5) The closest data are assigned to the same cluster;

[0095] 6) Calculate the mean of the elements in each new cluster and use this value as the new calculated centroid;

[0096] 7) When the distance between the new centroid and the old centroid is less than the distance difference threshold, the loop is stopped; otherwise, the loop threshold is continued to be judged. If the loop threshold has not been reached, the loop is started in step 4), otherwise the loop is stopped;

[0097] 8) After data processing is completed, the clustering results are output.

[0098] The principal component analysis (PCA) algorithm used by the portrait module is a feature dimension reduction method. Ideally, the principal component analysis (PCA) algorithm assumes a slope through the origin in a three-dimensional coordinate system. This slope contains all the sample points represented by three latitudes. Then it can be determined that the distribution of these samples is actually two-dimensional. Therefore, it is only necessary to rotate the three axes xyz in the three-dimensional coordinate system so that the surface composed of its two coordinate axes xy coincides with the plane where the sample points are located, and the dimensionality reduction of the samples is completed. The above principle assumes that all sample points can be contained by this slope, so no information is lost in the dimensionality reduction process. In actual situations, these sample points will more or less shake on the third latitude outside the plane. Therefore, the principal component analysis (PCA) algorithm used in the present invention considers the jitter of these sample points as the noise of the data set, and the information on the rotated xy axis is the principal component of the data. These principal components are sufficient to analyze the required results, so that the feature set can be simplified by reasonable noise processing and analyzing the principal components.

[0099] Step S1.4: prediction module;

[0100] The prediction module is mainly used to make user service predictions based on the portrait module. It uses the random forest (RF) algorithm on the reduced-dimensional data to calculate the degree of match between the service function representation and the user demand representation, and predict the products or services that the user may request to call.

[0101] Specifically, the function formula of the prediction module is F C (Y, R), where Y represents the output of the profiling module, i.e., the processed user data, and R represents the random forest (RF) algorithm, which is used to make decisions and predictions based on the processed user data.

[0102] Step S1.5: model optimization;

[0103] The item-level attention mechanism (ItemAttention) and interest-level attention mechanism (tAttention Net) are added to the constructed L-KPR-AstralEngine user portrait model.

[0104] Among them, the item-level attention mechanism aims to determine which items in the system are more important to the user or task. Through the item-level attention mechanism, the model can automatically pay attention to and weight different items to better process and understand the input data. Specifically, for each item x in the system i Define a mapping function f(x i ) maps it to an attention score s i ; These attention scores are then converted into attention weights w through the softmax function j Indicates the importance of the j-th behavior sj =f(x j ):

[0105]

[0106] Finally, using the attention weight a i Perform weighted summation on the input items, and get the weighted item representation X weighted :

[0107]

[0108] The interest-level attention mechanism aims to determine the degree of attention paid by the user or system to different behaviors or attributes. Through the interest-level attention mechanism, the model can automatically pay attention to and weight different interests or behaviors to better understand and respond to the user's interests and needs. Specifically, for each behavior b in the system j , also define a mapping function g(b j ) maps it to an attention score s j ; These attention scores are then converted into attention weights w through the softmax function k Indicates the importance of the k-th behavior s k =g(b k ):

[0109]

[0110] Finally, using the attention weight w j Perform weighted summation on the input items, and get the weighted item representation B weighted :

[0111]

[0112] Personalized processing can be achieved by adding item-level attention mechanism and interest-level attention mechanism to the constructed L-KPR-AstralEngine user portrait model: the interest-level attention mechanism enables the model to automatically weight the user's behavior or attributes according to the user's interests, thereby achieving personalized processing and prediction; by focusing on and weighting key interests or behaviors, the attention mechanism can improve the model's understanding and prediction of user behavior, thereby improving the model's performance and effectiveness; the weighting mechanism allows the model to model the user's interests and behavioral characteristics more finely, thereby improving the model's understanding and response to user needs and achieving refined modeling.

[0113] The present invention uses a combination of label generation, cluster analysis, principal component analysis (PCA) and random forest (RF) algorithms to build a powerful L-KPR-AstralEngine user portrait model. This user portrait model can not only provide detailed user portraits, but also predict user needs and behaviors, helping enterprises better understand and serve their users.

[0114] The L-KPR-AstralEngine user portrait model constructed by the present invention uses a variety of user behavior data (such as browsing, purchasing, searching, etc.) and personal information (such as age, gender, geographic location, etc.) to create multi-dimensional features of users. These features are crucial for understanding user preferences, behavior patterns, and making personalized education and training course recommendations.

[0115] Step S2: Build a DeepFM model with a gating mechanism, and use the user portrait output by the L-KPR-AstralEngine user portrait model to adjust parameters and train the model, and finally output the processed course recommendation list;

[0116] Although the L-KPR-AstralEngine user portrait model provides rich user features, relying solely on these features may not fully mine the nonlinear relationships and hidden patterns in the data. Therefore, the present invention adds a DeepFM model with a gating mechanism to better utilize these features and consider the complex interactive relationships between features to solve the feature interaction problem in the multi-level label prediction task and improve the accuracy and performance of the course recommendation method.

[0117] Step S2.1: DeepFM model;

[0118] The DeepFM model adopted combines the factorization machine (FM) and the deep neural network (DNN). The factorization machine (FM) is used to process the linear relationship in the user's multi-dimensional features, while the deep neural network (DNN) is used to capture complex nonlinear interactions.

[0119] Specifically, the calculation formula of the DeepFM model is as follows:

[0120]

[0121] Among them, w 0 and w i Respectively represent the bias term and linear weight of the model; u i and u jBoth represent embedding vectors, which are used to simulate the interaction between features; DNN(p) represents the deep neural network (DNN) part, which is responsible for processing high-order feature interactions; p represents the processed user portrait; the DeepFM model outputs the processed user portrait p to complete the feature interaction.

[0122] Among them, Factorization Machine (FM) is a machine learning model used in recommendation systems and binary classification problems. It predicts the target variable through implicit feature interactions:

[0123] Feature representation: In FM, each feature is treated as a separate entity and mapped into a low-dimensional latent vector space. This latent vector represents the potential representation of the feature, and its length is determined by the latent vector dimension of the model.

[0124] Feature Interaction: The core idea of ​​FM is to capture the interaction information between features by calculating the dot product between latent vectors. The dot product operation can quantify the similarity between two latent vectors, thereby reflecting the correlation between features. In the recommendation system, this means that the model can learn complex relationships between different features, such as the relationship between a user's age and preferences.

[0125] Model parameters: The FM model contains two parameters: latent vector and bias term. The latent vector is used to capture the interaction information between features, while the bias term captures the independent effect of each feature. By learning these parameters, FM can predict the value of the target variable.

[0126] Model training: FM models are trained by minimizing the difference between predicted values ​​and true values. A commonly used loss function is mean square error (MSE). Through optimization algorithms such as gradient descent, FM can find the optimal model parameters.

[0127] Model advantages: 1) High computational efficiency: Since only one dot product operation of the latent vector is involved, the computational complexity of FM is low and it is suitable for processing large-scale data sets. 2) Strong interpretability: The latent vector can be interpreted as a potential representation of the feature, which helps to understand the interaction between features. 3) Strong generalization ability: By capturing the interaction information between features, FM can better generalize to new data points.

[0128] Among them, the deep neural network (DNN) is an important component of the DeepFM model, which is responsible for extracting and learning high-order features in the data to capture more complex nonlinear relationships. Specifically,

[0129] 1) Deep neural network structure:

[0130] It usually consists of multiple hidden layers, each of which consists of multiple neurons (or nodes). The connection weights between these hidden layers are learned through training data, so that the deep neural network can gradually learn the abstract features and patterns in the data.

[0131] 2) Non-linear activation function:

[0132] Each hidden layer usually applies a nonlinear activation function to its output, such as ReLU (Rectified Linear Unit), Sigmoid, or tanh. The addition of these activation functions enables deep neural networks to learn more complex nonlinear relationships, thereby improving the expressiveness of the model.

[0133] 3) Feature Representation Learning:

[0134] The deep neural network structure maps the original features into a higher-dimensional space by learning the representation of the input features. In this high-dimensional space, the deep neural network can discover the complex patterns and relationships hidden in the data, thereby extracting richer and more discriminative feature representations.

[0135] 4) Feature interaction and combination:

[0136] In the deep neural network structure, the interaction and combination between features are realized through the connections and weights between hidden layers. Deep neural networks can automatically learn the complex interactive relationships between features, including the mutual influence and dependency between features, so as to better capture the important features in the data.

[0137] 5) High-level feature extraction:

[0138] Compared to the factorization machine part that mainly captures the linear relationship of low-order features, the deep neural network part focuses more on extracting and learning high-order features, that is, those features that cannot be obtained by simple linear combinations. These high-order features usually contain more semantic information and abstract representations, and can better describe the complex relationships in the data.

[0139] As a key component of the DeepFM model, Deep Neural Networks (DNN) can effectively extract and learn high-order features in the data through deep learning and feature representation learning, thereby capturing more complex nonlinear relationships. This enables the model to more accurately model and predict the complex interactive relationships between features in the data, providing more accurate and personalized recommendation results for the course recommendation method.

[0140] Step S2.2: adding a gating mechanism;

[0141] The gating mechanism is a technology that can dynamically adjust the flow of information and is often used in sequence data modeling. The core idea is to control the opening and closing of the gate to determine whether information passes through, so that the model can dynamically adjust the degree of interaction between features according to different input situations.

[0142] The long short-term memory network (LSTM) is added to the DeepFM model as a gating unit to dynamically adjust the contribution of features in the deep neural network (DNN). This enables the DeepFM model to intelligently adjust its recommendation strategy based on the temporal changes in user behavior, enhancing the model's adaptability to emerging trends and its prediction accuracy.

[0143] Specifically, the LSTM unit is used to process the dynamic changes of user behavior in time series data, and its calculation formula is as follows:

[0144]

[0145] Among them, h t and c t denote the hidden state and cell state at time step t respectively; h t-1 and C t-1 Represent the hidden state and cell state at time step t-1 respectively; p t represents the user portrait at time t; Represents the parameters of the LSTM unit.

[0146] By adding a gating mechanism, more flexibility and dynamics can be added to the model, thereby improving the performance and generalization ability of the model.

[0147] Step S2.3: model optimization;

[0148] In order to further improve the performance of the DeepFM model, the following optimization strategies are implemented. The optimization of the model mainly includes adjusting the learning rate, optimizing the Embedding dimension, and adjusting the deep neural network (DNN) structure, with the aim of improving the performance and adaptability of the overall model.

[0149] Among them, the learning rate directly affects the convergence speed and performance of the model. Too high a learning rate may cause the model to be unstable, while too low a learning rate may lead to slow convergence. Therefore, the learning rate can be adjusted through experiments. Different learning rate values ​​can be tried, and the learning rate decay strategy can be considered to find the best learning rate suitable for the data set. The Embedding layer is crucial for learning feature interactions. Adjusting the dimension of the embedding is an important aspect of optimizing model performance. By trying different embedding dimensions, a balance can be found between model complexity and computational efficiency, improving the expressiveness of the model. The DNN part of the DeepFM model can be optimized by adjusting the number of layers, activation functions, etc. You can consider increasing or decreasing the number of hidden layers, changing the activation function, or trying to add a more complex network structure.

[0150] Step S2.4: model training;

[0151] The loss function L of the model is defined by combining cross entropy and regularization terms to ensure the stability and efficiency of the model during training:

[0152]

[0153] Among them, CrossEntropyLoss represents the cross entropy loss, which is used to evaluate the predicted label output by the model The difference between the true label y and the real label y; λ represents the regularization coefficient; All model parameters are included, and the regularization term is used to prevent overfitting.

[0154] The L-KPR-AstralEngine user portrait model is integrated with the DeepFM model and the LSTM gating mechanism to obtain the L-KPR-AstralEngine+DeepFM+LSTM model, such as Figure 2 As shown in the figure, not only the personalized recommendation ability of the course recommendation method is significantly improved, but also the adaptability of the course recommendation method to changes in user behavior is enhanced. These improvements significantly improve the accuracy and user satisfaction of the course recommendation method, and provide an effective technical solution to solve the problem of personalized course recommendation.

[0155] The present invention compares the performance differences of the DeepFM model with the gating mechanism and the traditional DeepFM model under different indicators. The experimental results are shown in the table below.

[0156] Model Accuracy F1 score AUC value DeepFM 0.85 0.82 0.92 DeepFM model with gating mechanism 0.89 0.86 0.94

[0157] In the latest DeepFM model, a gating mechanism is innovatively added. This improvement is specifically for multi-level label prediction tasks. The core of the gating mechanism is that it can selectively activate or inhibit network signals at different levels, so as to more accurately process hierarchical label data. In this way, the model can focus more on the information that is most critical to the prediction task while ignoring irrelevant data noise.

[0158] Step S2.5: model output;

[0159] The DeepFM model with gating mechanism is used to generate processed user portraits; the generation of user portraits can be described as a multi-dimensional feature conversion process, where each feature vector x is converted into a user-specific portrait through the model:

[0160] p=L-KPR(x)

[0161] Among them, L-KPR represents the mathematical model of the user portrait construction process.

[0162] The predicted label output by the model After processing the DeepFM model with a gating mechanism combined with the user portrait, the specific calculation formula is as follows:

[0163]

[0164] Where σ represents the sigmoid activation function, which is used to convert the predicted label output by the model Convert to predicted probabilities.

[0165] In order to evaluate the comprehensive performance of the educational course recommendation method based on big data in the education industry of the present invention, the accuracy, F1 score and AUC value are comprehensively examined to evaluate the effectiveness of the course recommendation method in processing complex user behaviors and multi-dimensional feature interactions, with special attention paid to the responsiveness and time sensitivity of the course recommendation method to dynamic user data.

[0166] 1. Verify the dynamic adaptability of the course recommendation method of the present invention:

[0167] By comparing the capabilities of the L-KPR-AstralEngine+DeepFM+LSTM model of the present invention with the baseline model (MatrixFactorization and Neural Collaborative Filtering) in processing time series data and dynamically changing user behaviors, the advantages of the L-KPR-AstralEngine combined with DeepFM+LSTM in the present invention are demonstrated.

[0168] 2. Verify the interactive processing capability of the course recommendation method of the present invention:

[0169] The efficiency of the L-KPR-AstralEngine+DeepFM+LSTM model of the present invention in processing high-order and low-order interactions through the DeepFM component is deeply analyzed and demonstrated, and its performance improvement in complex data environments is highlighted compared with traditional recommendation system models.

[0170] 3. Verify the generalization ability of the course recommendation method of the present invention:

[0171] By performing K-fold cross validation across multiple data sets, the generalization ability of the L-KPR-AstralEngine+DeepFM+LSTM model of the present invention is evaluated to ensure that the model not only performs well on specific data sets, but can also adapt to new or unseen data environments.

[0172] Evaluation test method:

[0173] 1. Dataset composition

[0174] 1) Student data: covers the basic information of students, learning history and course participation. Through these data, we can gain insights into students’ learning habits, knowledge mastery level and personalized training needs. The analysis of student data also includes the examination of students’ demographic characteristics, such as age, gender, educational background, etc. These factors may affect learning outcomes.

[0175] 2) Product data: Contains detailed content of the training course, teaching materials and auxiliary multimedia resources. Analyzing this data helps to evaluate the market acceptance, difficulty level and comprehensiveness of teaching points of the course, so as to guide the continuous improvement of the course content. In addition, the analysis of product data also involves the evaluation of the effectiveness of course structure, teaching methods and learning tools, which are important factors affecting the training effect.

[0176] 3) Consultant data: records the consultant’s feedback, suggestions and interaction with trainees during the training process. This information is crucial to understanding the consultant’s role in training, improving service quality and enhancing trainee satisfaction. The analysis of consultant data also includes an assessment of the consultant’s professional ability, communication skills and teaching experience, all of which have a direct or indirect impact on the training effect.

[0177] The dataset fields include user ratings, timestamps, and multidimensional features of users and course products or course advisors.

[0178] 2. Experimental environment

[0179] 2.1 Hardware Configuration

[0180] CPU: Intel Xeon Processor E5-2640 v4 (2.4GHz, 10cores); Memory: 64GB DDR4; Hard disk: 1TB SSD; GPU: NVIDIA GeForce RTX 3080 (10GB GDDR6X); CUDA version: 11.4; CUDA core: 8704; Tensor Cores: 272; Video memory bandwidth: 760GB / s.

[0181] 2.2 Network Configuration

[0182] Network bandwidth: 1Gbps Ethernet; IP address: static IP configuration.

[0183] 2.3 Software Configuration

[0184] Operating system: Ubuntu 22.04LTS; Programming language: Python 3.11.4; Integrated development environment: PyCharm2023.2.5 (Professional Edition); Main third-party libraries: scikit-learn: 1.3.0, Numpy: 1.24.4, Pandas: 2.0.3, PyTorch: 2.0.1 (supports GPU acceleration), PyMysql: 1.0.2; Compilation method: pyinstaller6.2.0 cross-compilation; Version control: Git 2.37.1.

[0185] 3. Baseline model setting

[0186] Matrix Factorization (MF): A classic recommendation system model that is suitable for processing sparse data and predicts unknown ratings by decomposing the rating matrix. Neural Collaborative Filtering (NCF): A modern recommendation model that uses multi-layer perceptrons to learn feature representations of users and items.

[0187] 4. Results Analysis

[0188]

[0189] The comparison results of receiver operating characteristic (ROC) curves of different prediction models are shown in Figure 2. Figure 3 As shown in Figure 2, the ROC curve is used to demonstrate the diagnostic ability of the binary classification prediction model at different discrimination thresholds. Figure 3In the figure, the true positive rate (y-axis) indicates the proportion of positive samples that are correctly identified as positive; the false positive rate (x-axis) indicates the proportion of negative samples that are incorrectly identified as positive; the blue curve represents the "L-KPR-AstralEngine+DeepFM+LSTM" model, with an area under the curve (AUC) of 0.89, which performs best among the three models. The orange curve represents the "matrix factorization" model, with an AUC of 0.82. The green curve represents the "neural network CF (collaborative filtering)" model, with an AUC of 0.78. The dotted line represents the chance level classifier, with an AUC of 0.50. This line indicates that there is no discrimination between positive and negative categories. AUC (area under the curve) is an indicator used to measure the overall performance of a prediction model. The higher the AUC value, the better the model's performance, 1.0 is perfect, and 0.5 is no different from random guessing. Compare the performance of different models in distinguishing two categories (such as positive and negative). The model with the highest AUC value (blue curve) is considered to be the best one among the models listed here.

[0190] The accuracy and F1 score of different prediction models are as follows: Figure 4 As shown in the figure. The L-KPR-AstralEngine+DeepFM+LSTM model combines multiple technologies to provide high-performance predictive recommendations. Matrix Factorization is a common recommendation system algorithm used to discover the potential relationship between users and items. Neural CF (neural collaborative filtering) uses neural networks to simulate the interaction between users and items to improve the accuracy of recommendations. Figure 4 In the graph, the blue column represents the accuracy, which is the ratio of the total number of correct predictions of the model to the total number of samples. The orange column represents the F1 score, which is the harmonic mean of precision and recall and is often used to evaluate the effectiveness of the model in an imbalanced data set. Figure 4 It can be seen that the accuracy and F1 scores of all models are very close. In terms of accuracy and F1, L-KPR-AstralEngine+DeepFM+LSTM and Matrix Factorization are slightly higher than Neural CF.

[0191] The confusion matrix of different models is as follows Figure 5 As shown. The confusion matrix is ​​mainly used to evaluate the prediction performance of the model. The confusion matrix is ​​a particularly effective tool that can show the number of correct predictions and incorrect predictions of the model on each category. Figure 5In the L-KPR-AstralEngine+DeepFM+LSTM model, True label 0, Predicted label 0: The model correctly predicts a large number of category 0, and the darker color indicates a large number. True label 0, Predicted label 1: The number of category 0 mistakenly predicted as category 1 is small. True label 1, Predicted label 0: The number of category 1 mistakenly predicted as category 0 is also small. True label 1, Predicted label 1: The model correctly predicts a large number of category 1. In the Matrix Factorization model, True label 0, Predicted label 0: The number of category 0 correctly predicted is moderate, and the lighter color indicates a moderate number. True label 0, Predicted label 1: The number of category 0 mistakenly predicted as category 1 is large. True label 1, Predicted label 0: The number of category 1 mistakenly predicted as category 0 is also large. True label 1, Predicted label 1: The number of category 1 correctly predicted is moderate. In the Neural CF model, True label 0, Predicted label 0: The number of correctly predicted class 0 is moderate. True label 0, Predicted label 1: The number of class 0 mistakenly predicted as class 1 is small. True label 1, Predicted label0: The number of class 1 mistakenly predicted as class 0 is large. True label 1, Predicted label 1: The number of correctly predicted class 1 is moderate.

[0192] From the confusion matrix, we can see that the L-KPR-AstralEngine+DeepFM+LSTM model performs best in correctly classifying two categories with fewer misclassifications. The Matrix Factorization model seems to have more misclassifications when classifying, especially misclassifying data with the true label of 0 as 1, and misclassifying data with the true label of 1 as 0. The Neural CF model performs poorly in misclassifying data with the true label of 1 as 0.

[0193] L-KPR-AstralEngine performs best, thanks to its ability to comprehensively utilize DeepFM for feature interaction and LSTM for time series processing, providing a more sensitive response to user dynamic behavior.

[0194] Although Matrix Factorization performs well in handling implicit feature interactions, it has poor adaptability in dynamic environments.

[0195] Neural Collaborative Filtering shows the potential to improve recommendation accuracy through deep learning, but it is still slightly inferior to L-KPR-AstralEngine, especially in its ability to handle complex interactions and dynamic user behaviors.

[0196] In a second aspect, the present invention provides an education course recommendation system based on education industry big data. The system is mainly used to implement an education course recommendation method based on education industry big data provided in the first aspect of the present invention.

[0197] The present invention provides an educational course recommendation system based on big data in the education industry, mainly comprising:

[0198] Model building module: mainly used to build the L-KPR-AstralEngine user portrait model, the DeepFM model with gating mechanism, and the L-KPR-AstralEngine+DeepFM+LSTM model;

[0199] Label module: mainly used to generate large-scale user labels and calculate multi-channel label recycling based on business needs within the company system to form basic label system data for users;

[0200] Portrait module: It is mainly used to build basic user tags and then use K-means clustering and principal component analysis (PCA) algorithms to supplement the completeness and clearer outline of user portraits;

[0201] Prediction module: mainly used to calculate the matching degree between the service function representation and the user demand representation through the random forest (RF) algorithm, and predict the product or service that the user may request to call;

[0202] Model optimization module: mainly used to optimize the model's learning rate, embedding dimension, and deep neural network (DNN) structure to further improve the model's performance;

[0203] Model training module: Perform model training by setting a suitable loss function L;

[0204] Model output module: used to output the processed course recommendation list and obtain the course recommendation results.

[0205] In a third aspect, the present invention provides an application of an education course recommendation system based on big data of the education industry in a medical education system.

[0206] 1. Data collection: Collect students’ learning data from the medical school’s education platform, including but not limited to students’ test scores, forum activity records, and course feedback;

[0207] 2. Data preprocessing: clean the data, remove incomplete or abnormal records, and encode the classified data for subsequent analysis;

[0208] 3. User portrait construction: Apply label generation technology to generate preliminary user labels based on students' behaviors and grades; use K-means clustering algorithm to group student groups and identify student groups with similar learning characteristics;

[0209] 4. Implementation of the recommendation system: Use principal component analysis (PCA) to reduce the dimension of the clustering results to simplify the model and improve the computational efficiency; use the random forest (RF) algorithm to predict the medical education courses that students may be interested in based on the reduced-dimensional data and generate a personalized medical education course recommendation list;

[0210] 5. System integration and deployment: Integrate the course recommendation system of the present invention into the online education platform of the medical school;

[0211] 6. A user-friendly interface can be further designed to allow students to easily access personalized medical education course recommendations.

[0212] An example of the recommended results for medical education course products is shown in the following table.

[0213]

[0214]

[0215] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for recommending educational courses based on big data in the education industry, characterized in that: include: Step S1: Build a full-scenario label user portrait model L-KPR-AstralEngine based on a machine learning algorithm and output the user portrait; The constructed L-KPR-AstralEngine user portrait model consists of a label module, a portrait module, and a prediction module, and its function formula is: F(L,K,P,R)=F C (F P (F L (L),K,P),R) Where L represents label data input, K represents K-means clustering algorithm, P represents principal component analysis, R represents random forest algorithm, and F L Indicates the label module, F P represents the image module, F C represents the prediction module; Step S2: Build a DeepFM model with a gating mechanism, and use the user portrait output by the L-KPR-AstralEngine user portrait model to adjust parameters and train the model, and finally output the processed course recommendation list; The DeepFM model combines factorization machines and deep neural networks, using factorization machines to process linear relationships in multi-dimensional features of users, while capturing complex nonlinear interactions through deep neural networks; The DeepFM model adds a long short-term memory network as a gating unit to dynamically adjust the contribution of features in the deep neural network. The calculation formula of the DeepFM model is as follows: Among them, w0 and w i Respectively represent the bias term and linear weight of the model; u i and u j All represent embedding vectors, which are used to simulate the interaction between features; DNN(p) represents the deep neural network part, which is responsible for processing high-order feature interactions; p represents the processed user portrait; the DeepFM model outputs the processed user portrait p to complete the feature interaction; The parameter adjustment method is: adjusting the learning rate, optimizing the Embedding dimension and adjusting the deep neural network structure; During the model training process, the model loss function L is defined by combining cross entropy and regularization terms to ensure the stability and efficiency of the model during the training process: Among them, CrossEntropyLoss represents the cross entropy loss, which is used to evaluate the predicted label output by the model The difference between the true label y and the real label y; λ represents the regularization coefficient; Includes all model parameters, and regularization is used to prevent overfitting; The DeepFM model with gating mechanism is used to generate processed user portraits; the generation of user portraits is described as a multi-dimensional feature conversion process, in which each feature vector x is converted into a user-specific portrait through the model: p=L-KPR(x) Among them, L-KPR represents the mathematical model of the user portrait construction process; The predicted label output by the model After processing the DeepFM model with a gating mechanism combined with the user portrait, the specific calculation formula is as follows: Where σ represents the sigmoid activation function, which is used to convert the predicted label output by the model Convert to predicted probabilities.

2. According to claim 1, a method for recommending educational courses based on big data in the education industry is characterized in that: The label module is used to perform large-scale user label generation and multi-channel label recovery calculation to form user basic label system data based on business needs within the company system; the formed user basic label system is divided into two categories: basic labels and extended labels.

3. The method for recommending educational courses based on big data in the education industry according to claim 1, characterized in that: The portrait module is used to construct the basic tags of the user and then use K-means clustering and principal component analysis algorithms to supplement the completeness and clearer outline of the user portrait; the function formula of the portrait module is F P (X, K, P), X represents the output from the label module, that is, the user's basic label system data; K represents K-means clustering processing, which is used for data classification and grouping; P represents principal component analysis algorithm processing, which is used for data dimensionality reduction and important feature extraction; principal component analysis is used to reduce the dimensionality of clustering results, simplify the model and improve computational efficiency.

4. The method for recommending educational courses based on big data in the education industry according to claim 1, characterized in that: The prediction module is used to calculate the matching degree between the service function representation and the user demand representation through the random forest algorithm, and predict the product or service that the user may request to call; the function formula of the prediction module is F C (Y, R), where Y represents the output of the profiling module, i.e., the processed user data, and R represents the random forest algorithm, which is used to make decisions and predictions based on the processed user data.

5. The method for recommending educational courses based on big data in the education industry according to claim 1, characterized in that: The constructed L-KPR-AstralEngine user portrait model adds item-level attention mechanism and interest-level attention mechanism; the item-level attention mechanism is used to determine which items in the system are more important to the user or task. For each item x in the system i Define a mapping function f(x i ) maps it to an attention score s i , and then these attention scores are converted into attention weights w through the softmax function j Indicates the importance of the j-th behavior s j =f(x j ): Finally, the attention weight a is used i Perform weighted summation on the input items, and get the weighted item representation X weighted : The interest-level attention mechanism is used to determine the degree of attention paid by the user or system to different behaviors or attributes. j Define a mapping function g(b j ) maps it to an attention score s j , and then these attention scores are converted into attention weights w through the softmax function k Indicates the importance of the k-th behavior s k =g(b k ): Finally, the attention weight w is used j Perform weighted summation on the input items, and get the weighted item representation B weighted :

6. An educational course recommendation system based on big data in the education industry, characterized in that: Used to implement an educational course recommendation method based on big data in the education industry as described in any one of claims 1-5.

7. Application of an education course recommendation system based on education industry big data as described in claim 6 in recommending medical teaching course products.

Images