Learning resource recommendation method and system based on user portrait

Abstract

The invention relates to the technical field of resource recommendation, in particular to a learning resource recommendation method and system based on a user portrait, and the method comprises the steps: building a user portrait matrix and a learning resource feature matrix, and constructing a latent semantic matching model of a user and a learning resource through a factorization machine and a collaborative filtering algorithm; learning parameters of the latent semantic matching model through an alternating least square method, and generating user-learning resource latent semantic association features; and taking the user-learning resource latent semantic association features as individual codes, searching a Pareto optimal recommendation solution set of the multi-target learning resource recommendation optimization model by using a non-dominated sorting genetic algorithm, and generating a learning resource recommendation scheme in combination with an optimal recommendation decision model. According to the method, the recommendation strategy is dynamically adjusted by combining technologies such as multi-objective optimization and deep reinforcement learning, the recommendation accuracy is improved, meanwhile, the diversity and personalization of recommended contents can be balanced, and the dynamic adaptability and robustness of a recommendation system are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of resource recommendation, and in particular, to a learning resource recommendation method and system based on user portraits. Background Art

[0002] In today's digital age, personalized learning has become an important trend in the education field. By analyzing users' behaviors and preferences, customized learning resources are provided for users. However, users' interests and needs are constantly changing, while existing recommendation systems often rely on static user data such as age, gender, educational level, etc. These data are difficult to capture the subtle differences of users over time. Although users' behavioral data can provide a more dynamic perspective, the collection and analysis of these data often lack depth and breadth, resulting in inaccurate construction of user portraits and being unable to truly reflect users' current learning status and needs. Especially for new users or new courses, there is often a lack of sufficient user interaction data to construct accurate user portraits, which leads to the difficulty of the recommendation system to provide effective recommendations for new users or new learning resources, thus resulting in the cold start problem.

[0003] Secondly, traditional recommendation systems mainly rely on single recommendation algorithms such as content-based recommendation or collaborative filtering. Although these algorithms can provide personalized learning resources to a certain extent, there are some significant limitations. These algorithms not only have low efficiency in processing large-scale data sets, but also tend to recommend popular or high-rated learning resources, which leads to the homogenization of recommendation results and ignores users' needs for diversity. At the same time, this trend exacerbates the uneven distribution of learning resources, making some niche or low-rated but educationally valuable learning resources not receive enough attention. In addition, users' learning needs and abilities are constantly changing, while traditional recommendation systems often lack the ability to adapt to this dynamic change. Users need different types of learning resources at different learning stages, but traditional recommendation systems often cannot capture these changes in time and are difficult to adjust recommendation strategies in real time to adapt to such changes of users.

[0004] In summary, although traditional learning resource recommendation methods have made certain progress in personalized recommendation, there are deficiencies in terms of recommendation diversity, dynamic adaptability, etc. To solve these problems, there is an urgent need to provide a learning resource recommendation method based on user portraits to improve the accuracy and diversity of recommendation results. Summary of the Invention

[0005] The purpose of the present invention is to provide a learning resource recommendation method and system based on user portraits, which dynamically adjust the recommendation strategy by introducing technologies such as latent semantic matching, multi-objective optimization, and deep reinforcement learning to improve the precision and dynamic adaptability of learning resource recommendation.

[0006] To solve the above technical problems, the present invention provides a learning resource recommendation method and system based on user portraits.

[0007] In a first aspect, the present invention provides a learning resource recommendation method based on user portraits, and the method includes the following steps:

[0008] Based on the collected user portrait information and learning resource information, establish a user portrait matrix and a learning resource feature matrix;

[0009] Based on the user portrait matrix and the learning resource feature matrix, use a factorization machine and a collaborative filtering algorithm to construct a latent semantic matching model between users and learning resources;

[0010] Learn the parameters of the latent semantic matching model through the alternating least squares method to generate user-learning resource latent semantic association features;

[0011] Taking the minimization of user learning time and the maximization of the cumulative learning resource quality as optimization objectives, establish a multi-objective learning resource recommendation optimization model;

[0012] Use the user-learning resource latent semantic association features as individual codes, and use the non-dominated sorting genetic algorithm to search for the Pareto optimal recommendation solution set of the multi-objective learning resource recommendation optimization model;

[0013] Construct a multi-dimensional learning state representation based on user portraits and learning resources, use the deep Q-network algorithm to estimate the long-term cumulative rewards of different learning resources, and autonomously learn to obtain an optimal recommendation decision model;

[0014] Generate a learning resource recommendation plan according to the Pareto optimal recommendation solution set and the optimal recommendation decision model.

[0015] In a further implementation, the step of establishing a user portrait matrix and a learning resource feature matrix according to the collected user portrait information and learning resource information includes:

[0016] Collect user portrait information and learning resource information, and use the maximum relevance minimum redundancy algorithm to extract features from the user portrait information and the learning resource information respectively to obtain corresponding user portrait feature subsets and learning resource feature subsets;

[0017] Perform principal component analysis on the user portrait feature subset and the learning resource feature subset respectively to obtain corresponding user portrait key feature vectors and learning resource key feature vectors;

[0018] According to the pre-obtained feature importance weights, perform weighted average processing on the user portrait key feature vectors to obtain user portrait reference feature values;

[0019] Standardize the user portrait key feature vectors according to the user portrait benchmark feature values to establish a user portrait matrix;

[0020] Use the moving average method to perform learning time series aggregation on the learning resource key feature vectors to obtain learning resource benchmark feature values;

[0021] Standardize the learning resource key feature vectors according to the learning resource benchmark feature values to establish a learning resource feature matrix.

[0022] In a further embodiment, the step of constructing a latent semantic matching model of users and learning resources using a factorization machine and a collaborative filtering algorithm based on the user portrait matrix and the learning resource feature matrix includes:

[0023] Map the user portrait matrix and the learning resource feature matrix to the same latent semantic space to obtain a user portrait latent vector representation and a learning resource latent vector representation;

[0024] Use a factorization machine to capture the feature interaction information between the user portrait latent vector representation and the learning resource latent vector representation;

[0025] Use a collaborative filtering algorithm to decompose the feature interaction information to obtain a user implicit feature vector and a learning resource implicit feature vector;

[0026] Fuse the feature interaction information, the user implicit feature vector, and the learning resource implicit feature vector to obtain a fused feature;

[0027] Construct a latent semantic matching model of users and learning resources using a deep autoencoder attention network based on the fused feature.

[0028] In a further embodiment, the step of learning the parameters of the latent semantic matching model by alternating least squares to generate user-learning resource latent semantic association features includes:

[0029] Randomly initialize the parameters in the latent semantic matching model; the parameters in the latent semantic matching model include the user latent semantic space representation and the learning resource latent semantic space representation;

[0030] Use alternating least squares to alternately update the user latent semantic space representation and the learning resource latent semantic space representation by minimizing the reconstruction error to obtain user latent semantic features and learning resource latent semantic features;

[0031] Capture the potential association between users and learning resources based on the user latent semantic features and the learning resource latent semantic features to generate user-learning resource latent semantic association features.

[0032] In a further embodiment, the multi-objective learning resource recommendation optimization model is specifically:

[0033] min(λ*F1+(1-λ)F2)

[0034] Wherein,

[0035]

[0036] In the formula, F1 is the user learning time function; F2 is the cumulative learning resource quality function; λ is a weight parameter between 0 and 1; S is the number of learning resources; a and b are fitting parameters; A i is the difficulty of learning resource i; x i is the recommendation variable, x i ∈{0,1}; T i is the learning time of learning resource i; f i is the user's familiarity with learning resource i; γ, α, δ, ζ, ε are weight coefficients; cover i is the knowledge point coverage breadth of learning resource i; g i is the relevance of learning resource i to the user's current learning progress requirement; h i is the contribution value of the user interest degree of learning resource i to the cumulative learning resource quality; z i is the interaction frequency between the user and learning resource i; z max is the maximum interaction frequency between the user and learning resource i; diversity(S,i) is the diversity gain after adding learning resource i to the learning resource set S; cos(i,j) is the cosine similarity between learning resource i and learning resource j, i≠j.

[0037] In a further embodiment, the constraint conditions of the multi-objective learning resource recommendation optimization model include learning time constraint conditions and learning resource difficulty constraint conditions. The learning time constraint conditions are specifically:

[0038]

[0039] The learning resource difficulty constraint conditions are specifically:

[0040]

[0041] In the formula, B is the learning saturation of the user.

[0042] In a further embodiment, the steps of constructing a multi-dimensional learning state representation based on user portraits and learning resources include:

[0043] Obtain the interaction data of the user portrait information and the learning resource information, evaluate the connection strength between the user and the learning resource according to the interaction data, and construct a user-learning resource adjacency matrix;

[0044] Define multi-dimensional attributes for each learning resource according to the learning resource information, and construct a learning resource multi-dimensional attribute matrix;

[0045] Perform feature splicing on the user-learning resource adjacency matrix and the learning resource multi-dimensional attribute matrix to form a learning feature splicing matrix;

[0046] Input the learning feature splicing matrix into a graph neural network to learn the feature representations of the user and the learning resource, and obtain a user feature vector and a learning resource feature vector;

[0047] Classify and extract multi-dimensional information entropy from the learning resource information in sequence to obtain multi-dimensional learning resource representation data;

[0048] Perform weighted fusion on the user feature vector, the learning resource feature vector, and the multi-dimensional learning resource representation data to obtain a multi-dimensional learning state representation.

[0049] In a further implementation, the deep Q-network algorithm uses the multi-dimensional learning state representation as the state space.

[0050] In a further implementation, the step of generating a learning resource recommendation scheme according to the Pareto optimal recommendation solution set and the optimal recommendation decision model includes:

[0051] Traverse each recommended candidate solution in the Pareto optimal recommendation solution set, and calculate the matching degree between the recommended candidate solution and the user's current progress requirements;

[0052] According to the matching degree, screen out the recommended candidate set with the highest matching degree with the user's current progress requirements;

[0053] According to the optimal recommendation decision model, obtain the Q value of each learning resource in the recommended candidate set;

[0054] According to the Q value of each learning resource, screen out the learning resource with the highest Q value from the recommended candidate set to generate a learning resource recommendation scheme.

[0055] In a second aspect, the present invention provides a learning resource recommendation system based on a user portrait, and the system includes:

[0056] A data analysis module for establishing a user portrait matrix and a learning resource feature matrix according to the collected user portrait information and learning resource information;

[0057] A latent semantic analysis module, used to construct a latent semantic matching model between users and learning resources by using a factorization machine and a collaborative filtering algorithm according to the user portrait matrix and the learning resource feature matrix;

[0058] A parameter learning module, used for learning the parameters of the latent semantic matching model by alternating least squares method to generate user-learning resource latent semantic association features;

[0059] A multi-objective optimization module is used to establish a multi-objective learning resource recommendation optimization model with the optimization objectives of minimizing user learning time and maximizing the quality of accumulated learning resources; and to use the user-learning resource latent semantic association features as individual codes and use a non-dominated sorting genetic algorithm to search for the Pareto optimal recommendation solution set of the multi-objective learning resource recommendation optimization model;

[0060] The recommendation decision module is used to construct a multi-dimensional learning state representation based on user portraits and learning resources, use a deep Q network algorithm to estimate the long-term cumulative returns of different learning resources, and autonomously learn to obtain the optimal recommendation decision model; and generate a learning resource recommendation plan based on the Pareto optimal recommendation solution set and the optimal recommendation decision model.

[0061] The present invention provides a method and system for recommending learning resources based on user portraits. The method establishes a user portrait matrix and a learning resource feature matrix according to collected user portrait information and learning resource information, and uses a factor decomposition machine and a collaborative filtering algorithm to build a latent semantic matching model between users and learning resources; learns the parameters of the latent semantic matching model by an alternating least squares method to generate latent semantic association features between users and learning resources; establishes a multi-objective learning resource recommendation optimization model with minimization of user learning time and maximization of cumulative learning resource quality as optimization goals; uses the latent semantic association features between users and learning resources as individual codes, and uses a non-dominated sorting genetic algorithm to search for the Pareto optimal recommendation solution set of the multi-objective learning resource recommendation optimization model; constructs a multi-dimensional learning state representation based on user portraits and learning resources, uses a deep Q network algorithm to estimate the long-term cumulative returns of different learning resources, and autonomously learns to obtain an optimal recommendation decision model; and generates a learning resource recommendation plan according to the Pareto optimal recommendation solution set and the optimal recommendation decision model. Compared with the existing technology, this method uses the latent semantic matching model to deeply explore the potential correlation between users and learning resources, combines the multi-objective optimization algorithm to balance the recommendation accuracy and diversity, and dynamically adjusts the recommendation strategy through deep reinforcement learning to meet the personalized and dynamically changing learning needs of users, thereby achieving the personalization, accuracy and dynamic adaptability of learning resource recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1It is a schematic flowchart of a learning resource recommendation method based on user portraits provided by an embodiment of the present invention;

[0063] Figure 2 It is a block diagram of a learning resource recommendation system based on user portraits provided by an embodiment of the present invention. Specific embodiments

[0064] The following specifically illustrates the implementation manner of the present invention in conjunction with the drawings. The given embodiments are only for illustrative purposes and should not be construed as a limitation of the present invention. The included drawings are for reference and illustration only and do not constitute a limitation on the protection scope of the present invention's patent, because many changes can be made to the present invention without departing from its spirit and scope.

[0065] Refer to Figure 1 , an embodiment of the present invention provides a learning resource recommendation method based on user portraits. As Figure 1 shown, the method includes the following steps:

[0066] S1. According to the collected user portrait information and learning resource information, establish a user portrait matrix and a learning resource feature matrix.

[0067] In this embodiment, the step of establishing a user portrait matrix and a learning resource feature matrix according to the collected user portrait information and learning resource information includes:

[0068] Collect user portrait information and learning resource information, and use the maximum correlation minimum redundancy algorithm to extract features from the user portrait information and the learning resource information respectively, to obtain corresponding user portrait feature subsets and learning resource feature subsets;

[0069] Perform principal component analysis on the user portrait feature subset and the learning resource feature subset respectively, to obtain corresponding user portrait key feature vectors and learning resource key feature vectors;

[0070] Perform weighted average processing on the user portrait key feature vectors according to the pre-obtained feature importance weights to obtain user portrait reference eigenvalues;

[0071] Standardize the user portrait key feature vectors according to the user portrait reference eigenvalues to establish a user portrait matrix;

[0072] Use the moving average method to perform learning time series aggregation on the learning resource key feature vectors to obtain learning resource reference eigenvalues;

[0073] Standardize the learning resource key feature vectors according to the learning resource reference eigenvalues to establish a learning resource feature matrix.

[0074] Specifically, in this embodiment, user portrait information and learning resource information are collected respectively, and data cleaning is performed on the user portrait information and learning resource information to process missing values, outliers, and duplicate data, providing a data basis for the subsequent establishment of the user portrait matrix and the learning resource feature matrix; wherein, the user portrait information includes but is not limited to user basic information (such as age, gender, educational background), user behavior data (such as browsing history, learning progress, interaction records, knowledge mastery level), and user preference data (such as preferred learning resources, learning habits); the learning resource information includes but is not limited to learning resource metadata and learning resource performance data. The learning resource metadata may include topics, difficulty levels, resource types, authors, and the scope of knowledge points applicable, etc., and the learning resource performance data may include the completion rate, positive review rate, and user ratings of the learning resources, etc.

[0075] In this embodiment, the maximum relevance minimum redundancy (mRMR) algorithm is used to select the most representative and least redundant features from the user portrait information and the learning resource information respectively, obtaining a user portrait feature subset and a learning resource feature subset. Then, principal component analysis (PCA) is used to perform dimensionality reduction on the user portrait feature subset and the learning resource feature subset respectively, and the main change information in the dataset is retained, extracting the user portrait key feature vector and the learning resource key feature vector. Next, for each user portrait key feature in the user portrait key feature vector, based on the different contributions of different features to the user portrait, statistical methods such as Pearson correlation coefficient or mutual information can be used to calculate its feature importance weight in the entire feature vector, and the user portrait key feature vector is weighted by the feature importance weight to obtain weighted eigenvalues. Importance weighting can more accurately reflect the actual impact of features. Then, the average value of the weighted eigenvalues is calculated as the user portrait benchmark eigenvalue. It should be noted that in this embodiment, considering that the key feature vector of the learning resource will be different in different time periods, such as the regular update of learning resource knowledge points or content, the impact of different times or events on the learning resource access volume, and features such as the learning resource quality will change over time. Therefore, in this embodiment, the sliding average method is used to aggregate the learning resource key feature vector, extracting the average performance or trend of the learning resource in different time periods to obtain the learning resource benchmark eigenvalue, so as to more accurately evaluate the quality and effect of the learning resource and provide support for subsequent decision-making. At the same time, considering that different features have different dimensions and numerical ranges, in this embodiment, the user portrait key feature vector and the learning resource key feature vector are standardized according to their respective benchmark eigenvalues to convert the eigenvalues to the same scale, thereby constructing the user portrait matrix and the learning resource feature matrix.

[0076] S2. According to the user portrait matrix and the learning resource feature matrix, use a factorization machine and a collaborative filtering algorithm to construct a latent semantic matching model between users and learning resources.

[0077] In this embodiment, the steps of constructing a latent semantic matching model between users and learning resources according to the user portrait matrix and the learning resource feature matrix by using a factorization machine and a collaborative filtering algorithm include:

[0078] Map the user portrait matrix and the learning resource feature matrix to the same latent semantic space to obtain a user portrait latent vector representation and a learning resource latent vector representation;

[0079] Use a factorization machine to capture the feature interaction information between the user portrait latent vector representation and the learning resource latent vector representation;

[0080] Use a collaborative filtering algorithm to decompose the feature interaction information to obtain a user implicit feature vector and a learning resource implicit feature vector;

[0081] Fuse the feature interaction information, the user implicit feature vector, and the learning resource implicit feature vector to obtain a fused feature;

[0082] According to the fused feature, use a deep auto-encoding attention network to construct a latent semantic matching model between users and learning resources.

[0083] In this embodiment, the user portrait latent vector representation and the learning resource latent vector representation are combined as the input of a factorization machine (FM). The factorization machine captures the implicit feature interactions between users and learning resources, such as clicks and learning durations. At the same time, a collaborative filtering algorithm is used to decompose the feature interaction information to obtain the implicit feature vectors of users and learning resources. Then, by combining the factorization machine and the collaborative filtering algorithm, a latent semantic matching model is constructed. In this embodiment, the feature interaction information learned by the factorization machine and the implicit feature vectors obtained by collaborative filtering are fused. The fusion process can be achieved by methods such as feature concatenation or weighted summation to obtain a fused feature, and a latent semantic matching model that can express the deep relationship between users and learning resources is constructed based on the fused feature.

[0084] Specifically, in this embodiment, according to the fusion features, a deep auto-encoding attention network is used to construct a latent semantic matching model between users and learning resources. Different from traditional recommendation methods, this embodiment combines a factorization machine, a collaborative filtering algorithm, and a deep auto-encoder to capture the deep association between users and learning resources. The deep auto-encoding attention network combines a deep auto-encoder, a multi-scale attention mechanism, and matrix factorization. The deep auto-encoding attention network includes a matrix factorization initialization layer, a deep auto-encoder, and a multi-scale attention mechanism introduced in the deep auto-encoder. Among them, the deep auto-encoder includes an encoder, a bottleneck layer, and a decoder. The encoder includes multiple neural network layers and a multi-scale attention mechanism set before each neural network layer. The matrix factorization initialization layer is used to generate preliminary implicit feature vectors for users and learning resources using matrix factorization technology, and perform weighted fusion on the fusion features and the preliminary implicit feature vectors to obtain enhanced fusion features. In this embodiment, a multi-scale attention mechanism is introduced in the encoder. The role of the multi-scale attention mechanism is to make the model more flexible and accurate in capturing the complex relationships between different users and learning resources by dynamically adjusting the weights. In this embodiment, the multi-scale attention mechanism of the first layer is used to adjust the weights of the enhanced fusion features, and perform a non-linear transformation on the enhanced fusion features weighted by the multi-scale attention mechanism through multiple neural network layers. According to the above steps, further extract and compress the feature information through the non-linear transformation layer by layer of the multiple neural network layers, so as to capture the core latent semantic information output by the multiple neural network layers through the bottleneck layer, generate a latent semantic feature vector, and then decode the latent semantic feature vector through the decoder.

[0085] S3. Learn the parameters of the latent semantic matching model by the alternating least squares method to generate user-learning resource latent semantic association features.

[0086] In this embodiment, the step of learning the parameters of the latent semantic matching model by the alternating least squares method to generate user-learning resource latent semantic association features includes:

[0087] Randomly initialize the parameters in the latent semantic matching model; the parameters in the latent semantic matching model include the user latent semantic space representation and the learning resource latent semantic space representation;

[0088] Use the alternating least squares method to alternately update the user latent semantic space representation and the learning resource latent semantic space representation by minimizing the reconstruction error, and obtain user latent semantic features and learning resource latent semantic features;

[0089] Capture the potential association between users and learning resources according to the user latent semantic features and the learning resource latent semantic features, and generate user-learning resource latent semantic association features.

[0090] In this embodiment, the Alternating Least Squares (ALS) method is used to learn the parameters of the latent semantic matching model. In the latent semantic matching model, for the user latent semantic space representation and the learning resource latent semantic space representation, the Alternating Least Squares method iteratively minimizes the reconstruction error by alternately fixing one of the latent semantic space representations (user latent semantic space representation or learning resource latent semantic space representation) and optimizing the other latent semantic space representation, and solving the optimal solution under the current fixed parameters until the preset convergence condition is reached. Then, the newly solved parameters are fixed, and the previously fixed parameters are optimized. This process is repeated alternately until the model parameters converge. In this embodiment, the preset convergence condition is set to that the change in the reconstruction error is less than the preset reconstruction error threshold or the preset number of iterations is reached. This embodiment optimizes the model parameters by the Alternating Least Squares method to capture the latent semantic association between the user and the learning resource, and obtains the user-learning resource latent semantic association feature.

[0091] S4. Taking the minimization of the user's learning time and the maximization of the cumulative learning resource quality as the optimization objectives, a multi-objective learning resource recommendation optimization model is established.

[0092] S5. Using the user-learning resource latent semantic association feature as the individual coding, the non-dominated sorting genetic algorithm is used to search for the Pareto optimal recommendation solution set of the multi-objective learning resource recommendation optimization model.

[0093] In this embodiment, taking the minimization of the user's learning time and the maximization of the cumulative learning resource quality as the optimization objectives, a multi-objective learning resource recommendation optimization model is established, and the non-dominated sorting genetic algorithm (NSGA-II) is used to handle the multi-objective optimization problem, so that the multi-objective learning resource recommendation optimization model can find the optimal balance among multiple optimization objectives, that is, simultaneously minimize the user's learning time and maximize the cumulative learning resource quality, and obtain the Pareto optimal recommendation solution set. The Pareto optimal recommendation solution set represents a comprehensive recommendation strategy that achieves a balance among multiple objectives such as the user's learning time, the cumulative learning resource quality, and diversity, providing more flexibility and adaptability for the recommendation system, and being able to meet the needs of a wider range of users and scenarios. In this embodiment, the multi-objective learning resource recommendation optimization model is specifically:

[0094] min(λ*F1+(1-λ)F2)

[0095] where,

[0096]

[0097] In the formula, F1 is the user learning time function; F2 is the cumulative learning resource quality function; λ is a weight parameter between 0 and 1; S is the number of learning resources; a and b are fitting parameters; A iis the difficulty of learning resource i; x i is a recommendation variable, x i ∈ {0, 1}, if the user is recommended to learn resource i, then x i = 1; T i is the learning time of learning resource i; f i is the familiarity of the user with learning resource i; γ, α1, δ, ζ, ε are weight coefficients; cover i is the breadth of knowledge point coverage of learning resource i; g i is the relevance of learning resource i to the user's current learning progress requirement; h i is the contribution value of the user interest degree of learning resource i to the cumulative learning resource quality; z i is the interaction frequency between the user and learning resource i; z max is the maximum interaction frequency between the user and learning resource i; diversity(S, i) is the diversity gain after adding learning resource i to the learning resource set S; cos(i, j) is the cosine similarity between learning resource i and learning resource j, i ≠ j.

[0098] Meanwhile, the constraint conditions of the multi-objective learning resource recommendation optimization model may include learning time constraint conditions, learning resource difficulty constraint conditions, etc. Among them, the learning time constraint conditions are specifically:

[0099]

[0100] The learning resource difficulty constraint conditions are specifically:

[0101]

[0102] In the formula, B is the learning saturation of the user, which reflects the user's acceptance ability of new learning resources.

[0103] Meanwhile, this embodiment introduces learning resource attributes such as learning resource importance and learning path coherence to construct an individual evaluation function. According to the individual evaluation function, the non-dominated sorting genetic algorithm is used to solve the multi-objective learning resource optimization model, and the Pareto optimal solution set is searched. Among them, the individual evaluation function is specifically:

[0104]

[0105] Among them,

[0106]

[0107] In the formula, P(S) is the individual evaluation function; R1(i) is the comprehensive benefit of learning resource i; R2(i, j) is the learning path coherence from learning resource i to learning resource j; w iis the time weight coefficient for user learning; Y i is the availability of learning resource i; u i is the importance of the learning resource; α2 is the non - linear adjustment parameter of the importance of the learning resource; k ij is the conversion degree weight coefficient; E ij is the conversion degree from learning resource i to learning resource j; P ij indicates whether learning resource j is a prerequisite for learning resource i; G ij is the difficulty gap between learning resource i and learning resource j; α3 is the non - linear adjustment parameter of the difficulty of the learning resource.

[0108] S6. Construct a multi - dimensional learning state representation based on the user profile and learning resources, use the deep Q - network algorithm to estimate the long - term cumulative return of different learning resources, and autonomously learn to obtain an optimal recommendation decision model.

[0109] In this embodiment, the steps of constructing the multi - dimensional learning state representation based on the user profile and learning resources include:

[0110] Obtain the interaction data of the user profile information and the learning resource information, evaluate the connection strength between the user and the learning resource according to the interaction data, and construct a user - learning resource adjacency matrix;

[0111] Define multi - dimensional attributes for each learning resource according to the learning resource information, and construct a learning resource multi - dimensional attribute matrix;

[0112] Perform feature splicing on the user - learning resource adjacency matrix and the learning resource multi - dimensional attribute matrix to form a learning feature splicing matrix;

[0113] Input the learning feature splicing matrix into a graph neural network to learn the feature representations of the user and the learning resources, and obtain the user feature vector and the learning resource feature vector;

[0114] Classify and extract the multi - dimensional information entropy of the learning resource information in sequence to obtain multi - dimensional learning resource representation data;

[0115] Perform weighted fusion on the user feature vector, the learning resource feature vector and the multi - dimensional learning resource representation data to obtain a multi - dimensional learning state representation.

[0116] In this embodiment, by combining the user - learning resource adjacency matrix with the learning resource multi - dimensional attribute matrix to form a learning feature splicing matrix, and inputting the learning feature splicing matrix into a graph neural network to learn the feature representations of the user and the learning resources, the internal structure and features of the data can be fully utilized to learn more comprehensive and accurate feature representations of the user and the learning resources, and obtain the user feature vector and the learning resource feature vector.

[0117] In this embodiment, by capturing and analyzing the interaction data between the user profile information and the learning resource information, the connection strength between the user profile information and the learning resource information is evaluated, and a user-learning resource adjacency matrix is constructed according to the connection strength between the user profile information and the learning resource information. In this embodiment, the user-learning resource adjacency matrix can reflect the degree of investment of the user in different learning resources. Then, in order to more comprehensively characterize the characteristics of the learning resources, this embodiment defines multi-dimensional attributes for each learning resource and constructs a learning resource multi-dimensional attribute matrix. By combining the user-learning resource adjacency matrix with the learning resource multi-dimensional attribute matrix, a learning feature splicing matrix is formed, so that the learning feature splicing matrix not only includes the degree of investment of the user in the learning resources, but also incorporates the learning resource multi-dimensional attribute matrix, providing comprehensive data support for the subsequent graph neural network. Then, this embodiment uses the graph neural network (GNN) to take the learning feature splicing matrix as the input, and learns and extracts the deep feature representations of the user and the learning resources, namely the user feature vector and the learning resource feature vector. This embodiment effectively captures the complex relationships and structural characteristics in the learning feature splicing matrix by introducing GNN, improving the accuracy and efficiency of feature learning.

[0118] This embodiment constructs a multi-dimensional learning state representation according to the user feature vector, the learning resource feature vector and the multi-dimensional learning resource characterization data, and takes the multi-dimensional learning state representation as the state space of the deep Q-network algorithm. Taking the recommendation action of recommending a certain learning resource as the action space of the deep Q-network algorithm, the deep Q-network algorithm is used to estimate the long-term cumulative rewards of different recommendation actions, and by continuously updating its policy to maximize the long-term cumulative rewards, the optimal recommendation decision strategy taken in the current state is output. In summary, the deep Q-network algorithm uses the multi-dimensional learning state representation as the state vector to interact with the environment, selects the recommendation action that can maximize the long-term cumulative rewards according to the current state, and obtains the optimal recommendation decision strategy. This embodiment combines the graph neural network and the deep Q-network algorithm to achieve accurate estimation and optimization of the long-term cumulative rewards of the user's learning resources, enabling the optimal recommendation decision model to intelligently adjust the recommendation strategy according to the dynamic changes of the learning state, and being able to adaptively optimize the decision-making learning resource recommendation scheme by maximizing the user's long-term learning benefits.

[0119] S7. Generate a learning resource recommendation scheme according to the Pareto optimal recommendation solution set and the optimal recommendation decision model.

[0120] In this embodiment, the step of generating a learning resource recommendation scheme according to the Pareto optimal recommendation solution set and the optimal recommendation decision model includes:

[0121] Traverse each recommended candidate solution in the Pareto optimal recommendation solution set, and calculate the matching degree between the recommended candidate solution and the user's current progress requirements;

[0122] According to the matching degree, screen out the recommended candidate set with the highest matching degree to the current progress requirement of the user;

[0123] According to the optimal recommendation decision model, obtain the Q value of each learning resource in the recommended candidate set;

[0124] According to the Q value of each learning resource, screen out the learning resource with the highest Q value from the recommended candidate set to generate a learning resource recommendation plan.

[0125] Specifically, the Pareto optimal recommendation solution set contains multiple recommended candidate solutions. In this embodiment, by traversing the entire Pareto optimal recommendation solution set, calculate the matching degree for each recommended candidate solution in the Pareto optimal recommendation solution set to quantify its degree of fit with the current learning progress requirement of the user. Then, according to the matching degree, further screen out the recommended candidate solution with the highest matching degree to the current progress requirement of the user to form a recommended candidate set. These recommended candidate sets can most likely meet the current actual learning needs of the user. Next, this embodiment uses the optimal recommendation decision model to calculate the Q value of each learning resource in the screened recommended candidate set. The Q value reflects the potential value of this learning resource in meeting the current needs of the user and optimizing the future learning path. Finally, sort according to the Q value of each learning resource and select the learning resource with the highest Q value from them. The learning resource with the highest Q value is the one that can best meet the current and future learning needs of the user. Generate a learning resource recommendation plan according to the screened learning resource with the highest Q value. Through the above steps, this embodiment effectively combines the optimal recommendation decision model with the Pareto optimal recommendation solution set, not only considering the balance of multiple goals, but also being able to dynamically adjust according to the real-time feedback of the user, and efficiently screen out the most suitable recommendation plan for the current progress requirement of the user from the Pareto optimal recommendation solution set.

[0126] In summary, the learning resource recommendation method based on user portrait provided in this embodiment combines latent semantic matching, multi-objective optimization, and deep reinforcement learning technologies. It uses the latent semantic matching model to deeply mine the potential association between users and learning resources, combines the multi-objective optimization algorithm to balance recommendation accuracy and diversity, and dynamically adapts to changes in the learning state through deep reinforcement learning, improving the accuracy and diversity of learning resource recommendations, enhancing the dynamic adaptation ability of the recommendation system, and realizing more accurate, diverse, and intelligent personalized learning resource recommendations.

[0127] An embodiment of the present invention provides a learning resource recommendation method based on a user profile. The method establishes a user profile matrix and a learning resource feature matrix according to the collected user profile information and learning resource information, and constructs a latent semantic matching model between users and learning resources by using a factorization machine and a collaborative filtering algorithm; learns the parameters of the latent semantic matching model by the alternating least squares method to generate latent semantic association features between users and learning resources; establishes a multi-objective learning resource recommendation optimization model with the minimization of the user learning time and the maximization of the cumulative learning resource quality as the optimization objectives; uses the latent semantic association features between users and learning resources as individual encodings, and searches for the Pareto optimal recommendation solution set of the multi-objective learning resource recommendation optimization model by using a non-dominated sorting genetic algorithm; constructs a multi-dimensional learning state representation based on the user profile and learning resources, estimates the long-term cumulative rewards of different learning resources by using a deep Q-network algorithm, and autonomously learns to obtain an optimal recommendation decision model; and generates a learning resource recommendation scheme according to the Pareto optimal recommendation solution set and the optimal recommendation decision model. Compared with the prior art, the method proposed in this embodiment solves the problems of insufficient accuracy of personalized recommendation, homogenization of recommended content, and poor adaptability to the dynamic needs of users in traditional recommendation systems by introducing technologies such as latent semantic matching, multi-objective optimization, and deep reinforcement learning. The latent semantic matching model significantly improves the accuracy and relevance of recommendations by mining the potential associations between users and learning resources. At the same time, by combining multi-objective optimization and deep reinforcement learning technologies, dynamic learning and adaptive recommendation are realized, the dynamic adaptability of the system is improved, and the diverse and dynamically changing learning needs of users are met.

[0128] It should be noted that the magnitudes of the sequence numbers of the above processes do not mean the order of execution. The order of execution of each process should be determined according to its function and internal logic, and should not constitute any limitation to the implementation process of the embodiments of the present application.

[0129] In one embodiment, as Figure 2 shown, an embodiment of the present invention provides a learning resource recommendation system based on a user profile. The system includes:

[0130] A data analysis module 101, configured to establish a user profile matrix and a learning resource feature matrix according to the collected user profile information and learning resource information;

[0131] A latent semantic analysis module 102, configured to construct a latent semantic matching model between users and learning resources according to the user profile matrix and the learning resource feature matrix by using a factorization machine and a collaborative filtering algorithm;

[0132] A parameter learning module 103, configured to learn the parameters of the latent semantic matching model by the alternating least squares method to generate latent semantic association features between users and learning resources;

[0133] The multi-objective optimization module 104 is used to establish a multi-objective learning resource recommendation optimization model with the minimization of the user's learning time and the maximization of the cumulative learning resource quality as the optimization objectives; and, using the user-learning resource latent semantic association feature as an individual encoding, searching for the Pareto optimal recommendation solution set of the multi-objective learning resource recommendation optimization model by using the non-dominated sorting genetic algorithm;

[0134] The recommendation decision module 105 is used to construct a multi-dimensional learning state representation based on the user profile and learning resources, estimate the long-term cumulative rewards of different learning resources by using the deep Q-network algorithm, and autonomously learn to obtain an optimal recommendation decision model; and, generate a learning resource recommendation scheme according to the Pareto optimal recommendation solution set and the optimal recommendation decision model.

[0135] For the specific limitations of a learning resource recommendation system based on a user profile, reference can be made to the above limitations on a learning resource recommendation method based on a user profile, which will not be elaborated here. Those of ordinary skill in the art can realize that, in combination with the various modules and steps described in the embodiments disclosed in the present application, they can be implemented by hardware, software, or a combination of both. Whether these functions are executed in a hardware or software manner depends on the specific application and design constraints of the technical solution. Professional technicians can use different methods to implement the described functions for each specific application, but such implementation should not be considered to exceed the scope of the present application.

[0136] An embodiment of the present invention provides a learning resource recommendation system based on user portraits. The system establishes a user portrait matrix and a learning resource feature matrix through a data analysis module; a latent semantic analysis module constructs a latent semantic matching model between users and learning resources by using a factorization machine and a collaborative filtering algorithm; a parameter learning module learns the parameters of the latent semantic matching model through the alternating least squares method to generate user-learning resource latent semantic association features; a multi-objective optimization module takes minimizing the user's learning time and maximizing the cumulative learning resource quality as optimization objectives, and establishes a multi-objective learning resource recommendation optimization model; uses the user-learning resource latent semantic association features as individual coding, and uses the non-dominated sorting genetic algorithm to search for the Pareto optimal recommendation solution set of the multi-objective learning resource recommendation optimization model; a recommendation decision module constructs a multi-dimensional learning state representation based on user portraits and learning resources, uses the deep Q-network algorithm to estimate the long-term cumulative rewards of different learning resources, and autonomously learns to obtain an optimal recommendation decision model; generates a learning resource recommendation plan according to the Pareto optimal recommendation solution set and the optimal recommendation decision model. Compared with the prior art, the system proposed in this embodiment solves the problems of insufficient accuracy of personalized recommendations, homogenization of recommended content, and poor adaptability to users' dynamic needs in traditional recommendation systems by introducing technologies such as latent semantic matching, multi-objective optimization, and deep reinforcement learning. The latent semantic matching model significantly improves the accuracy and relevance of recommendations by mining the potential associations between users and learning resources. At the same time, combined with multi-objective optimization and deep reinforcement learning technologies, it realizes dynamic learning and adaptive recommendation, improves the dynamic adaptability of the system, and meets the diverse and dynamically changing learning needs of users.

[0137] The above embodiments only represent several preferred embodiments of the present application. The description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the technical principle of the present invention, several improvements and substitutions can be made, and these improvements and substitutions should also be regarded as the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the protection scope of the claims.

Claims

1. A learning resource recommendation method based on user portraits, characterized in that It includes the following steps: Based on the collected user profile information and learning resource information, establish a user profile matrix and a learning resource feature matrix; According to the user profile matrix and the learning resource feature matrix, use a factorization machine and a collaborative filtering algorithm to construct a latent semantic matching model between users and learning resources; Learn the parameters of the latent semantic matching model through the alternating least squares method to generate user-learning resource latent semantic association features; Establish a multi-objective learning resource recommendation optimization model with the minimization of user learning time and the maximization of the cumulative learning resource quality as the optimization objectives; Use the user-learning resource latent semantic association features as individual encoding, and use the non-dominated sorting genetic algorithm to search for the Pareto optimal recommendation solution set of the multi-objective learning resource recommendation optimization model; Construct a multi-dimensional learning state representation based on the user profile and learning resources, use the deep Q-network algorithm to estimate the long-term cumulative rewards of different learning resources, and autonomously learn to obtain an optimal recommendation decision model; Generate a learning resource recommendation plan according to the Pareto optimal recommendation solution set and the optimal recommendation decision model.

2. The learning resource recommendation method based on user portraits according to claim 1, characterized in that, The step of establishing a user profile matrix and a learning resource feature matrix according to the collected user profile information and learning resource information includes: Collect user profile information and learning resource information, and use the maximum relevance minimum redundancy algorithm to extract features from the user profile information and the learning resource information respectively to obtain corresponding user profile feature subsets and learning resource feature subsets; Perform principal component analysis on the user profile feature subset and the learning resource feature subset respectively to obtain corresponding user profile key feature vectors and learning resource key feature vectors; Perform weighted average processing on the user profile key feature vectors according to the pre-acquired feature importance weights to obtain user profile reference feature values; Standardize the user profile key feature vectors according to the user profile reference feature values to establish a user profile matrix; Use the moving average method to perform learning time series aggregation on the learning resource key feature vectors to obtain learning resource reference feature values; Standardize the learning resource key feature vectors according to the learning resource reference feature values to establish a learning resource feature matrix.

3. The learning resource recommendation method based on user portraits according to claim 1, wherein The step of constructing a latent semantic matching model between users and learning resources using a factorization machine and a collaborative filtering algorithm according to the user profile matrix and the learning resource feature matrix includes: Map the user profile matrix and the learning resource feature matrix to the same latent semantic space to obtain user profile latent vector representations and learning resource latent vector representations; Use a factorization machine to capture the feature interaction information between the user profile latent vector representations and the learning resource latent vector representations; Use a collaborative filtering algorithm to decompose the feature interaction information to obtain user implicit feature vectors and learning resource implicit feature vectors; Fuse the feature interaction information, the user implicit feature vectors, and the learning resource implicit feature vectors to obtain fused features; According to the fused features, use a deep auto-encoding attention network to construct a latent semantic matching model between users and learning resources.

4. The learning resource recommendation method based on user portraits according to claim 1, characterized in that, The steps of learning the parameters of the latent semantic matching model by the alternating least squares method to generate user-learning resource latent semantic association features include: Randomly initialize the parameters in the latent semantic matching model; the parameters in the latent semantic matching model include the user latent semantic space representation and the learning resource latent semantic space representation; Using the alternating least squares method, alternately update the user latent semantic space representation and the learning resource latent semantic space representation by minimizing the reconstruction error to obtain user latent semantic features and learning resource latent semantic features; Capture the potential association between the user and the learning resource according to the user latent semantic features and the learning resource latent semantic features, and generate user-learning resource latent semantic association features.

5. The learning resource recommendation method based on user portraits according to claim 1, wherein The multi-objective learning resource recommendation optimization model is specifically: min(λ*F1+(1-λ)F2) Where, Wherein, F1 is the user learning time function; F2 is the cumulative learning resource quality function; λ is a weight parameter between 0 and 1; S is the number of learning resources; a and b are fitting parameters; A i is the difficulty of learning resource i; x i is the recommendation variable, x i ∈{0,1}; T i is the learning time of learning resource i; f i is the familiarity of the user with learning resource i; γ, α, δ, ζ, ε are weight coefficients; cover i is the knowledge point coverage breadth of learning resource i; g i is the relevance of learning resource i to the user's current learning progress requirement; h i is the contribution value of the user interest degree of learning resource i to the cumulative learning resource quality; z i is the interaction frequency between the user and learning resource i; z max is the maximum interaction frequency between the user and learning resource i; diversity(S,i) is the diversity gain after adding learning resource i to the learning resource set S; cos(i,j) is the cosine similarity between learning resource i and learning resource j, i≠j.

6. The learning resource recommendation method based on user portraits according to claim 5, characterized in that The constraint conditions of the multi-objective learning resource recommendation optimization model include learning time constraint conditions and learning resource difficulty constraint conditions. The learning time constraint conditions are specifically: The learning resource difficulty constraint conditions are specifically: In the formula, B is the learning saturation of the user.

7. The learning resource recommendation method based on user portraits according to claim 1, wherein, The steps of constructing a multi-dimensional learning state representation based on the user portrait and learning resources include: Obtain the interaction data of the user portrait information and the learning resource information, evaluate the connection strength between the user and the learning resource according to the interaction data, and construct a user-learning resource adjacency matrix; Define multi-dimensional attributes for each learning resource according to the learning resource information, and construct a learning resource multi-dimensional attribute matrix; Perform feature splicing on the user-learning resource adjacency matrix and the learning resource multi-dimensional attribute matrix to form a learning feature splicing matrix; Input the learning feature splicing matrix into a graph neural network to learn the feature representations of the user and the learning resource, and obtain a user feature vector and a learning resource feature vector; Classify and extract multi-dimensional information entropy from the learning resource information in turn to obtain multi-dimensional learning resource representation data; Perform weighted fusion on the user feature vector, the learning resource feature vector and the multi-dimensional learning resource representation data to obtain a multi-dimensional learning state representation.

8. The learning resource recommendation method based on user portraits according to claim 1, wherein: The deep Q-network algorithm uses the multi-dimensional learning state representation as the state space.

9. The learning resource recommendation method based on user portraits according to claim 1, wherein The steps of generating a learning resource recommendation scheme according to the Pareto optimal recommendation solution set and the optimal recommendation decision model include: Traverse each recommended candidate scheme in the Pareto optimal recommendation solution set, and calculate the matching degree between the recommended candidate scheme and the user's current progress requirements; According to the matching degree, screen out the recommended candidate set with the highest matching degree with the user's current progress requirements; According to the optimal recommendation decision model, obtain the Q value of each learning resource in the recommended candidate set; According to the Q value of each learning resource, screen out the learning resource with the highest Q value from the recommended candidate set to generate a learning resource recommendation scheme.

10. A learning resource recommendation system based on user portraits, characterized in that, The system includes: A data analysis module for establishing a user portrait matrix and a learning resource feature matrix according to the collected user portrait information and learning resource information; The latent semantic analysis module is used to construct a latent semantic matching model between users and learning resources according to the user portrait matrix and the learning resource feature matrix by using a factorization machine and a collaborative filtering algorithm; The parameter learning module is used to learn the parameters of the latent semantic matching model by the alternating least squares method and generate latent semantic association features between users and learning resources; The multi-objective optimization module is used to establish a multi-objective learning resource recommendation optimization model with the minimization of the user learning time and the maximization of the cumulative learning resource quality as the optimization objectives; and, using the latent semantic association features between users and learning resources as individual coding, a non-dominated sorting genetic algorithm is used to search for the Pareto optimal recommendation solution set of the multi-objective learning resource recommendation optimization model; The recommendation decision module is used to construct a multi-dimensional learning state representation based on the user portrait and learning resources, use the deep Q-network algorithm to estimate the long-term cumulative rewards of different learning resources, and autonomously learn to obtain an optimal recommendation decision model; and, according to the Pareto optimal recommendation solution set and the optimal recommendation decision model, generate a learning resource recommendation scheme.

Images