An intelligent learning behavior analysis method and system

Abstract

The present invention discloses an intelligent learning behavior analysis method and system, which relates to the field of intelligent learning technology in education, including constructing a knowledge graph and collecting multi-source data of users for preprocessing; calculating the final emotional fluctuation value of the target user at a future time point according to the emotional score and recommending the corresponding level of the target user in the knowledge graph; defining the action space according to the recommended level and dynamically optimizing the recommendation strategy through the SAC algorithm; constructing a user-behavior two-dimensional matrix to calculate the target user's satisfaction score for the node and generate a recommendation list according to the satisfaction score. Collect the user's emotional score and behavior data, combine the emotional fluctuation prediction model, accurately calculate the user's emotional fluctuation value, use the SAC algorithm to optimize the recommendation strategy, further improve the personalization and real-time nature of the recommendation effect, construct a user-behavior two-dimensional matrix, generate a recommendation list based on the satisfaction score, ensure that the recommended content is highly matched with the user's needs, thereby improving learning efficiency and user experience.

Description

Technical Field

[0001] The present invention relates to the field of educational intelligent learning technology, and in particular to an intelligent learning behavior analysis method and system. Background Art

[0002] In the current education field, the research and application of intelligent learning behavior analysis systems have gradually become an important part of the development of educational informatization. With the development of advanced technologies such as big data, artificial intelligence and emotional computing, the design of personalized learning systems has gradually evolved towards intelligence and contextualization. Early learning analysis systems mainly relied on students' static data, such as course completion, learning time and performance records. Recommendation algorithms are often limited to simple content push based on historical behavior, lacking dynamic analysis and prediction of user emotional states, resulting in insufficient accuracy and adaptability of recommendations. In recent years, the construction of knowledge graphs has brought new dimensions to learning behavior analysis, organizing learning content in the form of knowledge nodes and using the relationship between knowledge points for association analysis, so that personalized recommendation systems can achieve more intelligent navigation on users' learning paths. However, when constructing personalized recommendation paths, existing learning behavior analysis systems often ignore the impact of user emotional fluctuations on learning outcomes, lack deep integration and real-time analysis of multi-source data, and are difficult to meet dynamic and accurate learning needs.

[0003] The existing technology still has many shortcomings in practical applications. The existing learning behavior analysis system is difficult to effectively track and analyze the user's emotional state, and lacks a real-time prediction and feedback mechanism for emotional fluctuations. This makes it difficult for the system to dynamically adjust the difficulty and rhythm of the learning content according to the user's emotional state when generating recommended content, which can easily lead to poor learning effects due to emotional fluctuations. The data management and recommendation mechanisms of the existing systems are limited in real-time and security. Summary of the invention

[0004] The purpose of the present invention is to provide an intelligent learning behavior analysis method and system to solve the above problems.

[0005] The present invention is achieved through the following technical solutions:

[0006] An intelligent learning behavior analysis method, comprising:

[0007] Build a knowledge graph and collect multi-source data from users for preprocessing;

[0008] Calculate the final emotional fluctuation value of the target user at a future time point based on the emotional score and recommend the corresponding level of the target user in the knowledge graph;

[0009] Define the action space according to the recommendation level and dynamically optimize the recommendation strategy through the SAC algorithm;

[0010] A user-behavior two-dimensional matrix is ​​constructed to calculate the target user's satisfaction score for the node and generate a recommendation list based on the satisfaction score.

[0011] Compared with the prior art, the present invention has the following advantages and beneficial effects: collecting users' emotion scores and behavior data, combining with the emotion fluctuation prediction model, accurately calculating users' emotion fluctuation values, using the SAC algorithm to optimize the recommendation strategy, further improving the personalization and real-time performance of the recommendation effect, constructing a user-behavior two-dimensional matrix, generating a recommendation list based on the satisfaction score, ensuring that the recommended content is highly matched with user needs, thereby improving learning efficiency and user experience.

[0012] As a preferred solution of the intelligent learning behavior analysis method of the present invention, wherein: the construction of the knowledge graph includes extracting all relevant knowledge points from the user's learning goals and course requirements, and summarizing all knowledge points into a knowledge point list according to subject classification and logical relationship;

[0013] According to the knowledge point list, the logical relationship between knowledge points is used to construct a knowledge graph, set node segmentation rules, define the upstream and downstream nodes of each knowledge point, map all knowledge points to nodes of the knowledge graph, and connect the upstream and downstream nodes with directed edges;

[0014] The knowledge points are divided into three levels: basic layer, application layer and extension layer according to logical relationships.

[0015] As a preferred solution of the intelligent learning behavior analysis method of the present invention, wherein: the multi-source data collected from users includes:

[0016] By setting the collection frequency, collect multi-source data of users on each node;

[0017] The multi-source data includes behavioral data and sentiment scores;

[0018] The behavior data includes user ID, learning time, learning completion and browsing frequency;

[0019] The emotion score refers to setting the acquisition frequency of the camera, the user uses the camera to collect the user's facial image in the node learning, and transmits the collected facial image to the API interface of the Microsoft Azure Face API pre-trained expression recognition model. The API automatically analyzes the expression features of the user's facial image and returns the emotion data;

[0020] The emotion data refers to the probability of each emotion;

[0021] According to the emotional data, all emotions are divided into positive emotions and negative emotions, and the difference between the sum of the probabilities of all positive emotions and the sum of the probabilities of all negative emotions is calculated as the emotional score.

[0022] As a preferred solution of the intelligent learning behavior analysis method of the present invention, the preprocessing includes:

[0023] De-duplicate multi-source data, detect outliers and fill in missing values, unify timestamps into UTC format, and standardize and normalize multi-source data.

[0024] The intelligent learning behavior analysis method of the present invention is characterized in that: the calculation of the final emotion fluctuation value of the target user at a future time point according to the emotion score includes:

[0025] Arrange the sentiment scores in multi-source data in chronological order to form sentiment time series data;

[0026] Perform discrete Fourier transform on the emotional time series data, transform the emotional time series data in the time domain into the frequency domain, and obtain the Fourier coefficients on the frequency components;

[0027] The amplitude of the frequency component is obtained by taking the modulus of the Fourier coefficient, and the phase of the frequency component is obtained by calculating the angle of the Fourier coefficient;

[0028] The amplitude of each frequency component is combined into an amplitude spectrum, and the high-frequency amplitude threshold and the low-frequency amplitude threshold are set. If the frequency component in the amplitude spectrum is less than the low-frequency amplitude threshold, it is determined to be a low-frequency component. If the frequency component in the amplitude spectrum is greater than the high-frequency amplitude threshold, it is determined to be a high-frequency component. Otherwise, no operation is performed.

[0029] The frequency, amplitude and phase of the filtered high-frequency components and low-frequency components are combined to form a characteristic vector of emotional fluctuations;

[0030] The feature vector is converted into emotional time series data through Fourier series to obtain the emotional fluctuation value at each time point;

[0031] Select the autoregressive model to use the feature vector to predict the emotional fluctuations of future target users;

[0032] Automatically select the model order through AIC to obtain the optimal order;

[0033] The regression coefficient of the autoregressive model is calculated using the least squares method according to the optimal order. After the regression coefficient is obtained, the emotional fluctuation value of the target user at a future time point is calculated according to the initial prediction formula of the autoregressive model;

[0034] According to the emotional fluctuation value at the future time point, the probability density function of the normal distribution is set, and the final prediction formula is constructed to calculate the final emotional fluctuation value of the target user at the future time point. The formula is:

[0035]

[0036] in, represents the predicted value of the final emotional fluctuation at the future time point t+1, represents the standard deviation of the sentiment fluctuation value, c represents the constant term of the autoregressive model, and p represents the optimal order in the autoregressive model. represents the regression coefficient in the autoregressive model at the past i-th time point, represents the emotional fluctuation value at the i-th time point in the past, represents the emotional fluctuation value at the future time point t+1, u represents the mean of the emotional fluctuation value, represents the random error term.

[0037] As the intelligent learning behavior analysis method of the present invention, it is characterized in that: the levels corresponding to the recommended target users in the knowledge graph include:

[0038] Based on the final emotional fluctuation value obtained, emotional fluctuation thresholds Q1 and Q2 are set, and the corresponding level of the target user in the knowledge graph is recommended, including:

[0039] like , it means that the user's emotional fluctuation is low, and the content of the extended layer is recommended;

[0040] like , it means that the user's emotional fluctuation is medium, and the application layer content is recommended;

[0041] like , it means that the user’s emotional fluctuation is high, and basic layer content is recommended.

[0042] As the intelligent learning behavior analysis method of the present invention, it is characterized in that: the action space adjustment node defined according to the recommendation level includes:

[0043] When the target user selects a node for learning in the recommended hierarchy, new multi-source data of the target user in the current node is collected and preprocessed;

[0044] Extract features from new multi-source data and integrate them into the current state vector;

[0045] The current state vector includes sentiment score, learning time, learning completion and browsing frequency;

[0046] Define the action space and adjust the recommended nodes through the actions in the action space;

[0047] The action space includes:

[0048] A1: Adjust the recommended nodes to all downstream nodes of the current node and proceed to the next step;

[0049] A2: Adjust the recommended nodes to all upstream nodes of the current node and proceed to the next step;

[0050] A3: Keep the existing node unchanged, recommend the peer node of the current node, and do not proceed to the next step.

[0051] The intelligent learning behavior analysis method of the present invention is characterized in that: the dynamic optimization recommendation strategy through the SAC algorithm includes:

[0052] Select the strategy network and value network and initialize the model parameters;

[0053] The initialization model parameters include the weight matrix, bias term, discount factor and entropy adjustment coefficient of the initialization strategy network and value network;

[0054] In each round of training, the corresponding action is calculated based on the current state vector and the policy network, and the feedback data of the target user after performing the action is observed;

[0055] The feedback data includes a next state vector and an immediate reward;

[0056] By comparing the emotion scores and learning completion in the next state vector and the current state vector, the reward mechanism is set, and the emotion improvement reward value and learning effect reward value are defined to dynamically adjust the recommended content. If the emotion score in the next state vector is greater than the emotion score in the current state vector, the emotion improvement reward value is 1, otherwise it is 0. If the learning completion in the next state vector is greater than the learning completion in the current state vector, the learning effect reward value is 1, otherwise it is 0.

[0057] Add the emotional improvement reward value and the learning effect reward value to get the immediate reward;

[0058] Create an experience replay pool and store each state variable, selected action, immediate reward, and next state vector as a state transition sample in the experience pool;

[0059] Define the loss function of the policy network and the value network, use the value network to calculate the value of executing an action at the current state variable, and generate a probability distribution of a set of actions through the policy network;

[0060] Collect K state transition samples from the experience replay pool, and use the target value formula in the value network to calculate the target value of all collected state transition samples;

[0061] Use the target value to update the parameters of the value network, minimize the loss function of the value network, and use the loss function of the policy network to update the parameters of the policy network;

[0062] Calculate the corresponding gradients of the loss functions of the value network and the policy network respectively, and use gradient descent to update the weight parameters of the network;

[0063] Use the Adam optimizer and set the learning rate, repeat the above training steps, continue sampling, calculate the target value, and update the parameters until the loss functions of the policy network and the value network converge;

[0064] After each target user’s state vector is updated, the policy network outputs actions in real time to determine the adjustment direction of the recommended content, and updates the experience replay pool based on the target user’s state transfer samples.

[0065] As the intelligent learning behavior analysis method of the present invention, it is characterized in that: the construction of the user-behavior two-dimensional matrix to calculate the target user's satisfaction score for the node and generate a recommendation list according to the satisfaction score includes:

[0066] According to the action corresponding to the current state variable, all nodes that meet the action are screened out and an initial recommended content list H is generated;

[0067] Collect the learning time, completion rate and browsing frequency of users' multi-source data from all the selected nodes and arrange them in chronological order to form behavioral time series data;

[0068] According to the user ID, the behavior time series data is converted into a user-behavior two-dimensional matrix, where the row represents the user ID and the column represents the behavior time series data corresponding to the user;

[0069] According to the obtained user-behavior two-dimensional matrix, the target user's satisfaction score for the current node is calculated. The formula is:

[0070]

[0071] in, represents the satisfaction score of the o-th target user to the current node s, represents the learning time of the o-th user at the current node s, Indicates the maximum learning time among all users. represents the completion ratio of the o-th user at the current node s, represents the browsing frequency of the o-th user at the current node s, Indicates the maximum browsing frequency among all users;

[0072] According to the satisfaction score, the similarity between the target user and other users is calculated by the Pearson correlation coefficient to generate a user similarity matrix, in which each element represents the similarity between two different users;

[0073] According to the user similarity matrix, select the top W users with high similarity to the target user and integrate them into the neighbor user set N;

[0074] Calculate the predicted satisfaction score of the target user at the node in the initial recommended content list H;

[0075] According to the obtained predicted satisfaction scores, the nodes in the initial recommendation list H are sorted from high to low according to the predicted satisfaction scores, and the first A nodes with high predicted satisfaction scores are selected to generate the final recommendation list;

[0076] The target user's future sentiment fluctuation value is calculated regularly and the corresponding recommendation level is assigned. The predicted satisfaction score of the target user's unlearned nodes in the recommendation level is calculated and the recommendation list is regenerated.

[0077] Another object of the present invention is to provide an intelligent learning behavior analysis system, which comprises:

[0078] Graph construction module, used to build knowledge graphs and collect multi-source data of users for preprocessing;

[0079] The sentiment calculation module is used to calculate the final sentiment fluctuation value of the target user at a future time point based on the sentiment score and recommend the corresponding level of the target user in the knowledge graph;

[0080] The strategy optimization module is used to define the action space according to the recommendation level and dynamically optimize the recommendation strategy through the SAC algorithm;

[0081] The recommended content module is used to construct a user-behavior two-dimensional matrix to calculate the target user's satisfaction score for the node and generate a recommendation list based on the satisfaction score. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] The drawings described herein are used to provide a further understanding of the embodiments of the present invention, constitute a part of this application, and do not constitute a limitation of the embodiments of the present invention. In the drawings:

[0083] Figure 1 It is a flowchart of the intelligent learning behavior analysis method;

[0084] Figure 2 This is a structural diagram of the intelligent learning behavior analysis system. DETAILED DESCRIPTION

[0085] In order to make the purpose, technical scheme and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and drawings. The schematic implementation modes and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention. It should be noted that the present invention is already in the actual development and use stage.

[0086] Example 1

[0087] like Figure 1 and Figure 2 As shown, this embodiment includes:

[0088] S1, graph construction module, used to construct the knowledge graph and collect multi-source data of users for preprocessing;

[0089] Specifically, building a knowledge graph includes:

[0090] Extract all relevant knowledge points from the user's learning goals and course requirements, and summarize all knowledge points into a knowledge point list according to subject classification and logical relationships;

[0091] According to the knowledge point list, the logical relationship between knowledge points is used to construct a knowledge graph, and node segmentation rules are set to define the upstream and downstream nodes of each knowledge point. If the learning of knowledge point A requires knowledge point B, then knowledge point B is set as the upstream node. If the content of knowledge point B and knowledge point C needs to complement each other, then knowledge point B and knowledge point C are set as peer nodes. All knowledge points are mapped to nodes of the knowledge graph, and the upstream and downstream nodes are connected with directed edges.

[0092] Divide knowledge points into three levels according to logical relationships: basic layer, application layer and extension layer;

[0093] The basic layer is the starting knowledge point in the knowledge graph, that is, the content that can be understood without the support of other knowledge points;

[0094] The application layer is the content that needs to be supported by the knowledge points of the basic layer;

[0095] The expansion layer is a high-level knowledge point that can only be understood after mastering the first two layers of content.

[0096] The knowledge graph extracts knowledge points from learning objectives and course requirements to construct a hierarchical and logical structure, making the upstream and downstream relationships and complementary relationships of each knowledge point clear at a glance. The hierarchical division of the basic layer, application layer and extension layer ensures the gradual progress of learning content. Users can effectively master knowledge points according to their own learning progress, gradually from basic to advanced, improve the systematicness of learning and in-depth understanding ability, thereby significantly improving learning efficiency and effectiveness.

[0097] Furthermore, the multi-source data collected from users includes:

[0098] By setting the collection frequency, collect multi-source data of users on each node;

[0099] The multi-source data includes behavioral data and sentiment scores;

[0100] The behavior data includes user ID, learning time, learning completion and browsing frequency;

[0101] The learning duration refers to the total learning duration of the user at each node, including the time from when the user starts learning to when the user finishes learning;

[0102] The learning completion degree refers to the task completion status of the user at each node, expressed as a decimal;

[0103] The browsing frequency refers to the highest number of times a user visits the content on each node;

[0104] The emotion score refers to setting the acquisition frequency of the camera, the user uses the camera to collect the user's facial image in the node learning, and transmits the collected facial image to the API interface of the Microsoft Azure Face API pre-trained expression recognition model. The API automatically analyzes the expression features of the user's facial image and returns the emotion data;

[0105] The Microsoft Azure pre-trained expression recognition model can identify multiple expression states in facial images. The API analyzes facial expressions through a deep learning model and can identify basic expressions and emotions.

[0106] The emotion data refers to the probability of each emotion;

[0107] Each emotion described includes anger, contempt, disgust, fear, happiness, neutral, sadness, and surprise;

[0108] According to the emotional data, all emotions are divided into positive emotions and negative emotions, and the difference between the sum of the probabilities of all positive emotions and the sum of the probabilities of all negative emotions is calculated as the emotional score (positive numbers are considered positive emotions, and negative numbers are considered negative emotions).

[0109] The system collects multi-source data from users, including behavioral data and emotional scores, to fully understand the user's interaction and emotional state at the learning node. The behavioral data reflects the user's learning behavior, while the emotional score dynamically captures the user's emotions through Microsoft Azure's expression recognition technology and quantifies the emotional data into positive and negative emotional values. This multi-dimensional data collection method not only improves the monitoring accuracy of the user's learning process, but also enables the system to adjust the learning content in real time according to the user's emotional state, ultimately achieving personalized, emotion-driven intelligent recommendations, which helps to improve the user's learning experience and satisfaction.

[0110] In addition, pre-processing includes,

[0111] De-duplicate multi-source data, detect outliers and fill in missing values, unify timestamps into UTC format, and standardize and normalize multi-source data.

[0112] The preprocessing step effectively improves the integrity and consistency of the data by removing duplication, detecting outliers, filling missing values, unifying timestamps, standardizing and normalizing multi-source data, making subsequent analysis more accurate. Data standardization and normalization ensure the comparability of data from different sources and avoid bias caused by data differences, thereby providing a high-quality data foundation for the system and improving the analysis effect and prediction accuracy of the model.

[0113] S2, sentiment calculation module, used to calculate the final sentiment fluctuation value of the target user at a future time point according to the sentiment score and recommend the corresponding level of the target user in the knowledge graph;

[0114] Specifically, the final emotional fluctuation value of the target user at a future time point is calculated based on the emotional score, including:

[0115] Arrange the sentiment scores in multi-source data in chronological order to form sentiment time series data;

[0116] Perform discrete Fourier transform on the emotional time series data, transform the emotional time series data in the time domain into the frequency domain, and obtain the Fourier coefficients on the frequency components;

[0117] The amplitude of the frequency component is obtained by taking the modulus of the Fourier coefficient, and the phase of the frequency component is obtained by calculating the angle of the Fourier coefficient;

[0118] The amplitude of each frequency component is combined into an amplitude spectrum M, where the amplitude represents the intensity of the emotional fluctuation at that frequency;

[0119] Set the low frequency amplitude threshold and high frequency amplitude threshold;

[0120] The low-frequency amplitude threshold is obtained by taking the average value of all amplitudes in the amplitude spectrum M, and the formula is:

[0121]

[0122] in, represents the low-frequency amplitude threshold, represents the average value of all amplitudes in the amplitude spectrum M, N represents the number of amplitudes in the amplitude spectrum M, represents the amplitude of the kth frequency component;

[0123] The high-frequency amplitude threshold is obtained by calculating the standard deviation and average value of all amplitudes in the amplitude spectrum M, and the formula is:

[0124]

[0125] in, represents the high frequency amplitude threshold, represents the standard deviation of all amplitudes in the amplitude spectrum M;

[0126] If the frequency component in the amplitude spectrum is less than the low-frequency amplitude threshold, it is determined to be a low-frequency component. If the frequency component in the amplitude spectrum is greater than the high-frequency amplitude threshold, it is determined to be a high-frequency component. Otherwise, no operation is performed. High-frequency components usually represent a sharp change in emotions, and low-frequency components represent a stable trend in emotions.

[0127] In this method, an adaptive threshold setting method is adopted. First, the low-frequency amplitude threshold and the high-frequency amplitude threshold are obtained by calculating the mean value and standard deviation of the amplitude spectrum. This dynamic threshold setting based on statistical characteristics can better adapt to the characteristics of different data sets, thus having greater flexibility and accuracy in identifying the high-frequency and low-frequency components of emotional fluctuations.

[0128] In most Fourier transform applications, the screening of frequency components is usually performed based on a fixed threshold. Such a static threshold setting is difficult to dynamically adapt to the characteristics of the data, and it is easy to cause inaccurate screening of high-frequency or low-frequency components;

[0129] The frequency, amplitude and phase of the filtered high-frequency components and low-frequency components are combined to form a characteristic vector of emotional fluctuations;

[0130] According to the characteristic vector of emotional fluctuation, it is converted into emotional time series data through inverse Fourier series transformation to obtain the emotional fluctuation value at each time point;

[0131] The frequency information is converted back to the time domain through Fourier series, so that the frequency characteristics are associated with specific time points, thereby adding frequency components to the autoregressive model to generate emotional fluctuation values ​​containing periodic fluctuation characteristics;

[0132] Select the autoregressive model to use the feature vector to predict the emotional fluctuations of future target users;

[0133] Automatically select the model order through AIC to obtain the optimal order, that is, how many past data points are used for prediction;

[0134] Traditional AIC is used to balance the goodness of fit and complexity of the model. In order to improve the accuracy of emotion fluctuation prediction, the dynamic AIC criterion can be introduced. The dynamic AIC criterion not only considers the goodness of fit and complexity of the model, but also takes into account the temporal characteristics of emotion fluctuations. Specifically, the dynamic AIC criterion can automatically adjust the optimal order according to the amplitude of emotion fluctuations. For example, when the emotion fluctuates violently, the order can be increased to capture more details, while when the emotion is relatively stable, the order can be reduced to reduce the complexity of the model.

[0135] The AIC is a statistical indicator used for model selection. AIC measures the trade-off between the goodness of fit and complexity of the model and is usually used to select the optimal model from multiple statistical models.

[0136] The regression coefficient of the autoregressive model is calculated using the least squares method according to the optimal order, that is, the impact of different historical emotional fluctuation values ​​on future emotions. After obtaining the regression coefficient, the emotional fluctuation value of the target user at the future time point is calculated according to the initial prediction formula of the autoregressive model. The formula is:

[0137]

[0138] in, represents the emotional fluctuation value at the future time point t+1, c represents the constant term of the autoregressive model, which represents the basic level of emotional fluctuation, that is, the emotional fluctuation value when it is not affected by other factors, and p represents the optimal order in the autoregressive model. represents the regression coefficient in the autoregressive model at the past i-th time point, represents the emotional fluctuation value at the i-th time point in the past, represents the random error term, which represents the random fluctuations that cannot be explained by the model;

[0139] The emotional fluctuation value generated by Fourier series is used as the input of the autoregressive model, so that the autoregressive model can use frequency characteristics for prediction, which can better capture the periodic fluctuation of emotions and improve the accuracy of prediction;

[0140] According to the emotional fluctuation value at the future time point, the probability density function of the normal distribution is set, and the final prediction formula is constructed to calculate the final emotional fluctuation value of the target user at the future time point. The formula is:

[0141]

[0142] in, represents the predicted value of the final emotional fluctuation at the future time point t+1, represents the standard deviation of the sentiment fluctuation value, c represents the constant term of the autoregressive model, and p represents the optimal order in the autoregressive model. represents the regression coefficient in the autoregressive model at the past i-th time point, represents the emotional fluctuation value at the i-th time point in the past, represents the emotional fluctuation value at the future time point t+1, u represents the mean of the emotional fluctuation value, represents the random error term;

[0143] This formula integrates Fourier series, autoregressive model and probability density function to form a multi-level prediction formula that can capture historical dependencies and reflect periodic fluctuations and emotional distribution characteristics, thereby significantly improving the complexity and accuracy of the model and making it suitable for personalized emotion-driven systems.

[0144] Through multi-level data processing and modeling, the accuracy and personalization of emotion prediction have been significantly improved. The high-frequency and low-frequency components of emotion data are obtained through discrete Fourier transform, which enables the system to capture the rapid changes and stable trends of emotions. The frequency characteristics are restored to the time domain using Fourier series, retaining the periodic characteristics of emotions and inputting them into the autoregressive model to ensure that the model can reflect short-term fluctuations in emotions and identify long-term dependencies. The introduction of the AIC criterion ensures the balance between complexity and fit of the model and enhances the stability of the prediction. By combining the output of the autoregressive model with the normal distribution, the system generates a probabilistic prediction of emotion fluctuations, making the estimation of future emotional states more accurate.

[0145] Furthermore, the corresponding levels of the recommended target users in the knowledge graph include:

[0146] Based on the final emotional fluctuation value obtained, emotional fluctuation thresholds Q1 and Q2 are set, and the corresponding level of the target user in the knowledge graph is recommended, including:

[0147] like , it means that the user's emotional fluctuation is low, and the content of the extended layer is recommended;

[0148] like , it means that the user's emotional fluctuation is medium, and the application layer content is recommended;

[0149] like , it means that the user’s emotional fluctuation is high, and basic layer content is recommended.

[0150] This method dynamically adjusts the level of recommended content through the emotion fluctuation threshold, so that the system can recommend suitable learning content according to the user's emotional state. When the emotion fluctuation is low, the extension layer content is recommended to challenge the user's learning ability; when the emotion fluctuation is medium, the application layer content is recommended to maintain a moderate learning difficulty; when the emotion fluctuation is high, the basic layer content is recommended to help users gradually restore their learning state. This mechanism improves the adaptability of recommendations, enables users to obtain the best learning experience in different emotional states, and effectively promotes the continuity and effectiveness of learning.

[0151] S3, strategy optimization module, is used to define the action space according to the recommendation level and dynamically optimize the recommendation strategy through the SAC algorithm;

[0152] Specifically, the action space adjustment nodes defined according to the recommended level include:

[0153] When the target user selects a node for learning in the recommended hierarchy, new multi-source data of the target user in the current node is collected and preprocessed;

[0154] Extract features from new multi-source data and integrate them into the current state vector;

[0155] The current state vector includes sentiment score, learning time, learning completion and browsing frequency;

[0156] Define the action space and adjust the recommended nodes through the actions in the action space;

[0157] The action space includes:

[0158] A1: Adjust the recommended nodes to all downstream nodes of the current node and proceed to the next step;

[0159] A2: Adjust the recommended nodes to all upstream nodes of the current node and proceed to the next step;

[0160] A3: Keep the existing node unchanged, recommend the peer node of the current node, and do not proceed to the next step.

[0161] By defining the action space and dynamically adjusting the recommended nodes based on the user's learning data in the nodes, the system can flexibly recommend upstream and downstream or same-level knowledge points based on the user's learning progress and emotional state, helping users to learn efficiently at the appropriate difficulty level and ensuring the personalization and targeting of the learning path. This design enhances the learning experience and makes the learning content more in line with the user's current needs and understanding level, thereby significantly improving learning efficiency.

[0162] Furthermore, the SAC algorithm dynamically optimizes the recommendation strategy including:

[0163] The SAC algorithm is a deep reinforcement learning algorithm that maximizes the expected return of the strategy and introduces an entropy adjustment mechanism to increase the exploratory nature of the strategy. SAC introduces the entropy value of the strategy as part of the reward, so that the model retains more randomness in the decision-making process, thereby avoiding premature convergence to the local optimal solution.

[0164] Select the strategy network and value network and initialize the model parameters;

[0165] The initialization model parameters include the weight matrix, bias term, discount factor and entropy adjustment coefficient of the initialization strategy network and value network;

[0166] The policy network is used to generate the best action in the action space under each state vector, and output a probability distribution indicating the possibility of selecting different recommended contents;

[0167] The value network is used to estimate the expected total reward value of actions under each state vector;

[0168] In each round of training, the corresponding action is calculated based on the current state vector and the policy network, and the feedback data of the target user after performing the action is observed;

[0169] The feedback data includes a next state vector and an immediate reward;

[0170] By comparing the emotion scores and learning completion in the next state vector and the current state vector, the reward mechanism is set, and the emotion improvement reward value and learning effect reward value are defined to dynamically adjust the recommended content. If the emotion score in the next state vector is greater than the emotion score in the current state vector, the emotion improvement reward value is 1, otherwise it is 0. If the learning completion in the next state vector is greater than the learning completion in the current state vector, the learning effect reward value is 1, otherwise it is 0.

[0171] Add the emotional improvement reward value and the learning effect reward value to get the immediate reward;

[0172] Through the instant reward mechanism, the reward value is set according to the user's emotional score and the improvement of learning completion. The instant reward enables the system to obtain user feedback immediately after each recommendation and dynamically adjust the recommended content to meet the user's emotional needs and learning progress, thereby enhancing the user's learning motivation and satisfaction;

[0173] Create an experience replay pool and store each state variable, selected action, immediate reward, and next state vector as a state transition sample in the experience pool, so that data can be randomly extracted from different samples in subsequent updates to improve the stability of training;

[0174] Define the loss function of the policy network and the value network, use the value network to calculate the value of executing an action at the current state variable, and generate a probability distribution of a set of actions through the policy network;

[0175] The value network can provide effective feedback for the recommendation strategy by evaluating the expected return of each action. By quantifying the value of different actions, it can gradually optimize the action selection, thereby ensuring that the recommended content is highly matched with the user's current state and improving the recommendation effect.

[0176] Collect K state transition samples from the experience replay pool. The size of K is determined according to actual needs. Use the target value formula in the value network to calculate the target value of all collected state transition samples for optimizing the value network.

[0177] Use the target value to update the parameters of the value network, minimize the loss function of the value network, and use the loss function of the policy network to update the parameters of the policy network;

[0178] Calculate the corresponding gradients of the loss functions of the value network and the policy network respectively, and use gradient descent to update the weight parameters of the network;

[0179] Use the Adam optimizer and set the learning rate, repeat the above training steps, continue sampling, calculate the target value, and update the parameters until the loss functions of the policy network and the value network converge;

[0180] After each target user’s state vector is updated, the policy network outputs actions in real time to determine the adjustment direction of the recommended content, and updates the experience replay pool based on the target user’s state transfer samples.

[0181] The SAC algorithm dynamically optimizes the recommendation strategy and realizes adaptive adjustment of personalized learning paths. Through the combination of strategy network and value network, the system outputs the best action in each state and increases the exploratory nature of the strategy through entropy regulation, thereby avoiding local optimal solutions. The instant reward design based on emotion improvement and learning completion enables the system to dynamically respond to user emotions and learning changes and continuously optimize recommended content. The use of the experience replay pool improves the stability of training and ensures the effective convergence of the strategy and value network through gradient updates, thereby enhancing the system's recommendation accuracy and personalization effect, and ultimately improving the user's learning experience and efficiency.

[0182] S4, a recommended content module, used to construct a user-behavior two-dimensional matrix to calculate the target user's satisfaction score for the node and generate a recommendation list based on the satisfaction score;

[0183] Specifically, constructing a user-behavior two-dimensional matrix to calculate the target user's satisfaction score for the node and generating a recommendation list based on the satisfaction score includes:

[0184] According to the action corresponding to the current state variable, all nodes that meet the action are screened out and an initial recommended content list H is generated;

[0185] Collect the learning time, completion rate and browsing frequency of users' multi-source data from all the selected nodes and arrange them in chronological order to form behavioral time series data;

[0186] According to the user ID, the behavior time series data is converted into a user-behavior two-dimensional matrix, where the row represents the user ID and the column represents the behavior time series data corresponding to the user;

[0187] According to the obtained user-behavior two-dimensional matrix, the target user's satisfaction score for the current node is calculated. The formula is:

[0188]

[0189] in, represents the satisfaction score of the o-th target user to the current node s, represents the learning time of the o-th user at the current node s, Indicates the maximum learning time among all users. represents the completion ratio of the o-th user at the current node s, represents the browsing frequency of the o-th user at the current node s, Indicates the maximum browsing frequency among all users;

[0190] According to the satisfaction score, the similarity between the target user and other users is calculated by the Pearson correlation coefficient to generate a user similarity matrix, in which each element represents the similarity between two different users;

[0191] According to the user similarity matrix, select the first W users with high similarity to the target user and integrate them into the neighbor user set N. The number of W is determined according to actual needs;

[0192] Calculate the predicted satisfaction score of the target user at the node in the initial recommended content list H, the formula is:

[0193]

[0194] in, represents the predicted satisfaction score of the o-th target user for the h-th node in the initial recommended content list H, represents the oth neighbor user in the neighbor user set, represents the similarity between the oth target user and the vth similar user, represents the satisfaction score of the vth similar user on the hth node in the initial recommended content list H;

[0195] Based on the behavioral data of similar users, it is possible to predict the target user's potential interest in the recommended content, making the generation of recommended content more personalized and in line with the user's real needs;

[0196] According to the obtained predicted satisfaction scores, the nodes in the initial recommendation list H are sorted from high to low according to the predicted satisfaction scores, and the first A nodes with high predicted satisfaction scores are selected to generate the final recommendation list;

[0197] The resulting recommendation list is dynamically adjusted based on user interests, which improves the relevance of recommended content and user satisfaction, and provides users with a more optimized and efficient learning experience;

[0198] The target user's future sentiment fluctuation value is calculated regularly and the corresponding recommendation level is assigned. The predicted satisfaction score of the target user's unlearned nodes in the recommendation level is calculated and the recommendation list is regenerated.

[0199] By constructing a two-dimensional user-behavior matrix and calculating the satisfaction score, this method can accurately quantify the target user's preferences at specific nodes, use the Pearson correlation coefficient to screen a set of similar users, and predict the target user's potential interest in the recommended content based on the behavior of similar users, making the recommended content more personalized and accurate. The final generated recommendation list is dynamically adjusted based on user interests, which improves the relevance of the recommendation and user satisfaction, thereby optimizing the learning experience.

[0200] Example 2

[0201] Reference Figure 2 , which is the second embodiment of the present invention, is different from the previous embodiment and provides an intelligent learning behavior analysis system, including:

[0202] Graph construction module, used to build knowledge graphs and collect multi-source data of users for preprocessing;

[0203] The sentiment calculation module is used to calculate the final sentiment fluctuation value of the target user at a future time point based on the sentiment score and recommend the corresponding level of the target user in the knowledge graph;

[0204] The strategy optimization module is used to define the action space according to the recommendation level and dynamically optimize the recommendation strategy through the SAC algorithm;

[0205] The recommended content module is used to construct a user-behavior two-dimensional matrix to calculate the target user's satisfaction score for the node and generate a recommendation list based on the satisfaction score.

[0206] The specific implementation methods described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific implementation method of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. An intelligent learning behavior analysis method, characterized in that: include, According to the knowledge point list, the logical relationship between knowledge points is used to construct a knowledge graph, and the knowledge graph is divided into three layers: basic layer, application layer and extension layer according to the logical relationship between knowledge points; Collect multi-source data of users for preprocessing, wherein the multi-source data includes behavioral data and sentiment scores, wherein the sentiment scores are obtained based on facial images of users during node learning; the behavioral data includes user ID, learning time, learning completion and browsing frequency; Calculate the final emotional fluctuation value of the target user at a future time point based on the emotional score and recommend the corresponding level of the target user in the knowledge graph; Define the action space according to the recommendation level and dynamically optimize the recommendation strategy through the SAC algorithm; Filter out all nodes that meet the action requirements and generate an initial recommended content list H; Collect the learning time, completion rate and browsing frequency of multi-source data of users in all the selected nodes and arrange them in chronological order to form behavioral time series data; construct a user-behavior two-dimensional matrix to calculate the target user's satisfaction score for the node and generate the final recommendation list based on the satisfaction score; The calculation of the final emotion fluctuation value of the target user at a future time point according to the emotion score includes: Arrange the sentiment scores in multi-source data in chronological order to form sentiment time series data; Perform discrete Fourier transform on the emotional time series data, transform the emotional time series data in the time domain into the frequency domain, and obtain the Fourier coefficients on the frequency components; The amplitude of the frequency component is obtained by taking the modulus of the Fourier coefficient, and the phase of the frequency component is obtained by calculating the angle of the Fourier coefficient; The amplitude of each frequency component is combined into an amplitude spectrum, and the high-frequency amplitude threshold and the low-frequency amplitude threshold are set. If the frequency component in the amplitude spectrum is less than the low-frequency amplitude threshold, it is determined to be a low-frequency component. If the frequency component in the amplitude spectrum is greater than the high-frequency amplitude threshold, it is determined to be a high-frequency component. Otherwise, no operation is performed. The frequency, amplitude and phase of the filtered high-frequency components and low-frequency components are combined to form a characteristic vector of emotional fluctuations; The feature vector is converted into emotional time series data through Fourier series to obtain the emotional fluctuation value at each time point; Select the autoregressive model to use the feature vector to predict the emotional fluctuations of future target users; Automatically select the model order through AIC to obtain the optimal order; The regression coefficient of the autoregressive model is calculated using the least squares method according to the optimal order. After the regression coefficient is obtained, the emotional fluctuation value of the target user at a future time point is calculated according to the initial prediction formula of the autoregressive model; According to the emotional fluctuation value at the future time point, the probability density function of the normal distribution is set, and the final prediction formula is constructed to calculate the final emotional fluctuation value of the target user at the future time point. The formula is: ； in, represents the predicted value of the final emotional fluctuation at the future time point t+1, Indicates the standard deviation of the sentiment fluctuation value, represents the constant term of the autoregressive model, represents the optimal order in the autoregressive model, represents the regression coefficient in the autoregressive model at the past i-th time point, represents the emotional fluctuation value at the i-th time point in the past, Indicates the emotional fluctuation value at the future time point t+1, represents the mean value of emotional fluctuations, represents the random error term.

2. The intelligent learning behavior analysis method according to claim 1, characterized in that: The constructing of the knowledge graph further includes: Extract all relevant knowledge points from the user's learning goals and course requirements, and summarize all knowledge points into a knowledge point list according to subject classification and logical relationships; Set node segmentation rules, define the upstream and downstream nodes of each knowledge point, map all knowledge points into nodes of the knowledge graph, and connect the upstream and downstream nodes with directed edges.

3. The intelligent learning behavior analysis method according to claim 2, characterized in that: The collecting of multi-source data of users also includes: By setting the collection frequency, collect multi-source data of users on each node; The emotion score refers to setting the camera acquisition frequency, the user uses the camera to collect the user's facial image in the node learning, and transmits the collected facial image to the API interface of the Microsoft Azure Face API pre-trained expression recognition model. The API automatically analyzes the expression features of the user's facial image and returns the emotion data; The emotion data refers to the probability of each emotion; According to the emotional data, all emotions are divided into positive emotions and negative emotions, and the difference between the sum of the probabilities of all positive emotions and the sum of the probabilities of all negative emotions is calculated as the emotional score.

4. The intelligent learning behavior analysis method according to claim 3, characterized in that: The pre-processing comprises: De-duplicate multi-source data, detect outliers and fill in missing values, unify timestamps into UTC format, and standardize and normalize multi-source data.

5. The intelligent learning behavior analysis method according to claim 1, characterized in that: The levels corresponding to the recommended target users in the knowledge graph include: Based on the final emotional fluctuation value obtained, emotional fluctuation thresholds Q1 and Q2 are set, and the corresponding level of the target user in the knowledge graph is recommended, including: like , it means that the user's emotional fluctuation is low, and the content of the extended layer is recommended; like , it means that the user's emotional fluctuation is medium, and the application layer content is recommended; like , it means that the user’s emotional fluctuation is high, and basic layer content is recommended.

6. The intelligent learning behavior analysis method according to claim 5, characterized in that: Defining the action space adjustment node according to the recommendation level includes: When the target user selects a node for learning in the recommended hierarchy, new multi-source data of the target user in the current node is collected and preprocessed; Extract features from new multi-source data and integrate them into the current state vector; The current state vector includes sentiment score, learning time, learning completion and browsing frequency; Define the action space and adjust the recommended nodes through the actions in the action space; The action space includes: A1: Adjust the recommended nodes to all downstream nodes of the current node and proceed to the next step; A2: Adjust the recommended nodes to all upstream nodes of the current node and proceed to the next step; A3: Keep the existing node unchanged, recommend the peer node of the current node, and do not proceed to the next step.

7. The intelligent learning behavior analysis method according to claim 6, characterized in that: The dynamic optimization recommendation strategy through the SAC algorithm includes: Select the strategy network and value network and initialize the model parameters; The initialization model parameters include the weight matrix, bias term, discount factor and entropy adjustment coefficient of the initialization strategy network and value network; In each round of training, the corresponding action is calculated based on the current state vector and the policy network, and the feedback data of the target user after performing the action is observed; The feedback data includes a next state vector and an immediate reward; By comparing the emotion scores and learning completion in the next state vector and the current state vector, the reward mechanism is set, and the emotion improvement reward value and learning effect reward value are defined to dynamically adjust the recommended content. If the emotion score in the next state vector is greater than the emotion score in the current state vector, the emotion improvement reward value is 1, otherwise it is 0. If the learning completion in the next state vector is greater than the learning completion in the current state vector, the learning effect reward value is 1, otherwise it is 0. Add the emotional improvement reward value and the learning effect reward value to get the immediate reward; Create an experience replay pool and store each state variable, selected action, immediate reward, and next state vector as a state transition sample in the experience pool; Define the loss function of the policy network and the value network, use the value network to calculate the value of executing an action at the current state variable, and generate a probability distribution of a set of actions through the policy network; Collect K state transition samples from the experience replay pool, and use the target value formula in the value network to calculate the target value of all collected state transition samples; Use the target value to update the parameters of the value network, minimize the loss function of the value network, and use the loss function of the policy network to update the parameters of the policy network; Calculate the corresponding gradients of the loss functions of the value network and the policy network respectively, and use gradient descent to update the weight parameters of the network; Use the Adam optimizer and set the learning rate, repeat the above training steps, continue sampling, calculate the target value, and update the parameters until the loss functions of the policy network and the value network converge; After each target user’s state vector is updated, the policy network outputs actions in real time to determine the adjustment direction of the recommended content, and updates the experience replay pool based on the target user’s state transfer samples.

8. The intelligent learning behavior analysis method according to claim 7, characterized in that: The method of constructing a user-behavior two-dimensional matrix, calculating the target user's satisfaction score for the node and generating a recommendation list according to the satisfaction score includes: According to the action corresponding to the current state variable, all nodes that meet the action are screened out and an initial recommended content list H is generated; Collect the learning time, completion rate and browsing frequency of users' multi-source data from all the selected nodes and arrange them in chronological order to form behavioral time series data; According to the user ID, the behavior time series data is converted into a user-behavior two-dimensional matrix, where the row represents the user ID and the column represents the behavior time series data corresponding to the user; According to the obtained user-behavior two-dimensional matrix, the target user's satisfaction score for the current node is calculated. The formula is: ； in, represents the satisfaction score of the o-th target user to the current node s, represents the learning time of the o-th user at the current node s, Indicates the maximum learning time among all users. represents the completion ratio of the o-th user at the current node s, represents the browsing frequency of the o-th user at the current node s, Indicates the maximum browsing frequency among all users; According to the satisfaction score, the similarity between the target user and other users is calculated by the Pearson correlation coefficient to generate a user similarity matrix, in which each element represents the similarity between two different users; According to the user similarity matrix, select the top W users with high similarity to the target user and integrate them into the neighbor user set N; Calculate the predicted satisfaction score of the target user at the node in the initial recommended content list H; According to the obtained predicted satisfaction scores, the nodes in the initial recommendation list H are sorted from high to low according to the predicted satisfaction scores, and the first A nodes with high predicted satisfaction scores are selected to generate the final recommendation list; The target user's future sentiment fluctuation value is calculated regularly and the corresponding recommendation level is assigned. The predicted satisfaction score of the target user's unlearned nodes in the recommendation level is calculated and the recommendation list is regenerated.

9. An intelligent learning behavior analysis system based on the intelligent learning behavior analysis method according to any one of claims 1 to 8, characterized in that: include, Graph construction module, used to build knowledge graphs and collect multi-source data of users for preprocessing; The sentiment calculation module is used to calculate the final sentiment fluctuation value of the target user at a future time point based on the sentiment score and recommend the corresponding level of the target user in the knowledge graph; The strategy optimization module is used to define the action space according to the recommendation level and dynamically optimize the recommendation strategy through the SAC algorithm; The recommended content module is used to construct a user-behavior two-dimensional matrix to calculate the target user's satisfaction score for the node and generate a recommendation list based on the satisfaction score.

Images