Personalized education recommendation method and system based on artificial intelligence

Abstract

The invention discloses a personalized education recommendation method and system based on artificial intelligence, and relates to the technical field of artificial intelligence and education recommendation, and the method comprises the steps: collecting multi-modal data, carrying out the data fusion through a DST evidence combination formula, and carrying out the optimization through combining an adjustment factor, and obtaining a continuous vector; based on the continuous vector, calculating association strength, recommendation priority and driving score, generating an initial recommendation score in combination with the user behavior data, and then performing optimization to obtain a final recommendation score; integrating the recommendation priority, the driving score and the final recommendation score, constructing a feature vector, defining the feature vector as a main effect feature vector, calculating an interaction effect feature vector, fusing the main effect feature vector and the interaction effect feature vector, and calculating a sorting score to obtain a final recommendation list; and displaying the data through a visual interface, and storing the data through a database. The method not only enhances the robustness of emotion recognition, but also improves the accuracy and timeliness of educational resource recommendation.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and education recommendation technology, and in particular to a personalized education recommendation method and system based on artificial intelligence. Background Art

[0002] With the continuous development of artificial intelligence technology, personalized education recommendation systems have gradually become an important research direction in educational informatization. Traditional education recommendation methods mainly rely on content-based recommendations, collaborative filtering or hybrid recommendation strategies. These methods usually generate recommendation results by analyzing users' historical behavior data, knowledge point attributes and similarities between users. In recent years, with the widespread application of deep learning technology, multimodal data fusion and emotion recognition have gradually been introduced into education recommendation systems to improve the accuracy and personalization of recommendations. Existing studies have attempted to use visual images to recognize students' facial expressions, combine voice emotion analysis to judge their learning status, and further use this information in recommendation algorithms to improve the adaptability and responsiveness of recommendations.

[0003] However, existing personalized education recommendation methods and systems have low fusion efficiency when performing data fusion, and the data distribution of different modalities varies significantly, making it difficult to quantify their confidence conflicts, resulting in insufficient credibility of the fusion results. In addition, there is a lack of effective regulation mechanism for conflicting evidence, making it difficult to maintain stable emotion classification performance in complex learning scenarios. Secondly, the dynamic modeling ability of the relationship between user emotions and knowledge points is weak, and is mostly based on static statistics without introducing time decay factors and user group similarity analysis, resulting in delayed recommendation results and limited personalization. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an artificial intelligence-based personalized education recommendation method and system to solve the problems of insufficient credibility of fusion results, lack of effective adjustment mechanism for conflicting evidence, and delayed recommendation results and limited personalization.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a personalized education recommendation method based on artificial intelligence, which comprises:

[0008] Collect multimodal data, fuse the data using the DST evidence combination formula, and optimize it with the adjustment factor to obtain a continuous vector;

[0009] The multimodal data includes visual image data, audio data and behavioral data;

[0010] Based on the continuous vector, the association strength, recommendation priority, and driving score are calculated. The initial recommendation score is generated by combining user behavior data and then optimized to obtain the final recommendation score.

[0011] Integrate the recommendation priority, driving score, and final recommendation score to construct a feature vector, which is defined as the main effect feature vector. Then calculate the interaction effect feature vector, fuse the main effect and interaction effect feature vectors, and calculate the ranking score to obtain the final recommendation list.

[0012] The data is displayed through a visual interface and stored in a database.

[0013] As a preferred embodiment of the personalized education recommendation method based on artificial intelligence of the present invention, the steps of collecting multimodal data, fusing data using the DST evidence combination formula, and optimizing the data by combining the adjustment factors to obtain a continuous vector include the following steps:

[0014] Obtain the user's visual image data, audio data, and behavioral data from the education platform through the API interface, and record the test results after the user completes the test on the education platform;

[0015] The behavioral data refers to the user's click frequency and stay time on the platform;

[0016] Add a timestamp to each data type and use Fourier transform to convert the audio data into the frequency domain;

[0017] Use a random number generator to randomly generate an embedding vector for the corresponding data, and use a broadcast mechanism to perform element-wise addition of each data to the corresponding embedding vector to obtain the corresponding modal embedding tensor;

[0018] Use the Transformer architecture to build a probability distribution model, take the modal embedding tensor as input, and output the confidence distribution of the emotion classification corresponding to each data point;

[0019] The emotion classification confidence distribution includes categories of happy, sad, normal and undefined;

[0020] Use the DST evidence combination formula to fuse the confidence distribution of visual and audio data in categories;

[0021] We further use the DST evidence combination formula to fuse the behavioral data with the fusion results to obtain the final fusion distribution;

[0022] During the fusion process, when the visual and audio data or behavioral data conflicts with the fusion result, an adjustment factor is calculated and used to adjust the fusion confidence distribution value;

[0023] The confidence values of each category in the final fusion distribution result are arranged in ascending order, and the largest confidence is selected as the current emotion category. Then, the confidence values of each category in the final fusion distribution result are combined to obtain a continuous vector.

[0024] As a preferred embodiment of the AI-based personalized education recommendation method of the present invention, the calculation of association strength, recommendation priority, and driving score based on continuous vectors, and the optimization of the generated initial recommendation score based on user behavior data to obtain the final recommendation score include the following steps:

[0025] Use the mean formula to calculate the average confidence value of the same category in the continuous vector for all users;

[0026] The mean formula is further used to calculate the average of all users’ test results for the same knowledge point;

[0027] Use the covariance function combined with the mean to calculate the strength of association between each confidence value and the test result;

[0028] After calculating the standard deviation of the mean using the standard deviation formula, the Pearson correlation coefficient formula is used to combine the association strength and standard deviation to calculate the correlation coefficient between the knowledge point and each emotion category;

[0029] Use empirical rules to set the judgment threshold, compare the correlation coefficient with the threshold, and retain the correlation coefficient, emotion, and knowledge point when the correlation coefficient is greater than or equal to the threshold; otherwise, remove them;

[0030] The retained correlation coefficients are integrated to obtain the correlation vector;

[0031] Use the vector dot product method to calculate the driving score of the continuous vector and the correlation vector;

[0032] We then retrieve a set of knowledge points from the education platform through an API. Using the expert domain method, we assign initial weights to each pair of knowledge points in the set. We then use the knowledge points in the set as nodes, draw a directed graph between the nodes, and adjust the initial weights between all nodes based on the driving scores.

[0033] Use the updated weights to calculate the recommendation priority of each knowledge point;

[0034] Use statistical methods to calculate the accuracy of each user's test results, and calculate the user's initial recommendation score for the knowledge point based on the accuracy rate;

[0035] Use empirical rules to set the decay rate, and further obtain the timestamp of the user's current behavior data through the API interface. The difference between the timestamp of the user's current behavior data and the timestamp of the historical behavior data is defined as the time difference, and then combined with the exponential function to calculate the time decay weight;

[0036] Use the multiplication formula to calculate the product of the time decay weight and the initial recommendation score to obtain the optimized recommendation score;

[0037] Use the cosine similarity formula combined with the optimized recommendation score to calculate the similarity score between the current user and all users;

[0038] Set the selection threshold according to the empirical rule, compare the similarity score with the selection threshold, and when the similarity score is greater than or equal to the selection threshold, retain the similarity score and construct a similarity set; otherwise, remove it;

[0039] The optimized recommendation score and similarity score are combined to calculate the final recommendation score for each knowledge point.

[0040] As a preferred embodiment of the personalized education recommendation method based on artificial intelligence of the present invention, the steps of integrating the recommendation priority, driving score, and final recommendation score, constructing a feature vector, defining it as a main effect feature vector, calculating the interaction effect feature vector, fusing the main effect and interaction effect feature vectors, and calculating the ranking score to obtain the final recommendation list include the following steps:

[0041] Integrate the driving score, recommendation priority, and final recommendation score to construct a feature vector, which is defined as the main effect feature vector. Then, combine the features in the main effect feature vector in pairs.

[0042] Use the dot product operation to calculate the interaction effect strength between feature pairs, which is defined as combining the interaction effect features to obtain the interaction effect feature vector;

[0043] Build a hybrid model using a feedforward neural network and LambdaMART network architecture, including a front-end feature extraction layer and a back-end sorting layer;

[0044] The front-end feature extraction layer uses the ReLU activation function to fuse the main effect and interaction effect feature vectors to obtain a fused feature vector;

[0045] The backend sorting layer uses a random search method to set the learning rate and a domain method to set the initial total number of trees. The fused feature vector is input into each tree to obtain a prediction score vector. The learning rate, total number of trees, fused feature vector, and prediction score are then combined to calculate the sorting score of each knowledge point.

[0046] Use the addition formula to calculate the difference in knowledge point ranking scores;

[0047] After defining the objective function, minimizing the objective function value, and performing iterative optimization, the feature vector is input into the final hybrid model, and the sorted recommendation list is output as the final recommendation result.

[0048] As a preferred solution of the personalized education recommendation method based on artificial intelligence described in the present invention, the display of data through a visual interface refers to using the PowerBI tool to import the sorted recommendation list into a visualization platform, displaying the priority ranking of knowledge points through a bar chart, presenting the distribution of users' interest in knowledge points through a heat map, and tracking changes in recommendation trends through a line chart.

[0049] As a preferred solution of the artificial intelligence-based personalized education recommendation method described in the present invention, the data storage through the database refers to using a MongoDB distributed database to partition the recommendation list by user ID, and attach a timestamp, knowledge point ID and recommendation score field.

[0050] As a preferred solution of the artificial intelligence-based personalized education recommendation method described in the present invention, the multimodal data is collected and preprocessed first, including denoising, cleaning, time alignment and normalization operations.

[0051] In a second aspect, the present invention provides a personalized education recommendation system based on artificial intelligence, comprising:

[0052] The acquisition and optimization module is used to collect multimodal data for data fusion and optimize it with the adjustment factor to obtain a continuous vector;

[0053] The calculation and generation module is used to calculate the association strength, recommendation priority, and driving score based on the continuous vector, and then generate the initial recommendation score based on user behavior data and optimize it to obtain the final recommendation score;

[0054] The construction and ranking module is used to integrate the recommendation priority, driving score and final recommendation score, construct the feature vector, define it as the main effect feature vector, calculate the interaction effect feature vector, fuse the main effect and interaction effect feature vectors, and calculate the ranking score to obtain the final recommendation list;

[0055] The display and storage module is used to display and store data through a visual interface and database.

[0056] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the personalized education recommendation method based on artificial intelligence as described in the first aspect of the present invention is implemented.

[0057] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the personalized education recommendation method based on artificial intelligence as described in the first aspect of the present invention.

[0058] The beneficial effects of the present invention are as follows: the present invention significantly improves the fusion effect of visual, audio and behavioral data in emotion classification by introducing the DST evidence combination formula and combining the adjustment factor to dynamically correct the conflicting evidence, and by introducing the driving score and recommendation priority mechanism, combined with the time decay weight and similarity score of user behavior data, the dynamic generation and personalized optimization of recommendation scores are realized, and the relevance and adaptability of recommended content are improved. Secondly, by combining the main effect and interaction effect feature vectors, and by fusing the feedforward neural network and the LambdaMART sorting model, the refined modeling of knowledge point sorting scores is realized, thereby further improving the intelligence level and sorting performance of the present invention, so that the present invention not only enhances the robustness of emotion recognition, but also improves the accuracy and timeliness of educational resource recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0060] Figure 1 This is a flowchart of the personalized education recommendation method based on artificial intelligence in Example 1.

[0061] Figure 2 This is a structural diagram of the personalized education recommendation system based on artificial intelligence in Example 1.

[0062] Figure 3 This is a flowchart of generating a recommendation list in Example 1. DETAILED DESCRIPTION

[0063] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0064] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0065] Secondly, the term "one embodiment" or "embodiment" herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in various places throughout this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive of other embodiments.

[0066] Example 1, with reference to Figures 1-3 , which is the first embodiment of the present invention, provides a personalized education recommendation method based on artificial intelligence, comprising the following steps:

[0067] S1. Collect multimodal data, perform data fusion using the DST evidence combination formula, and optimize it with an adjustment factor to obtain a continuous vector; the multimodal data includes visual image data, audio data, and behavioral data;

[0068] Specifically, the following steps are involved in collecting multimodal data, fusing the data using the DST evidence combination formula, and optimizing the continuous vectors by combining the adjustment factors:

[0069] Obtain the user's visual image data, audio data, and behavioral data from the education platform through the API interface, and record the test results after the user completes the test on the education platform;

[0070] The behavioral data refers to the user's click frequency and stay time on the platform;

[0071] Add a timestamp to each data type and use Fourier transform to convert the audio data into the frequency domain;

[0072] Use a random number generator to randomly generate an embedding vector for the corresponding data, and use a broadcast mechanism to perform element-wise addition of each data to the corresponding embedding vector to obtain the corresponding modal embedding tensor;

[0073] Use the Transformer architecture to build a probability distribution model, including the input layer, Transformer encoding layer, branching layer, and output layer;

[0074] The Transformer encoding layer includes a self-attention mechanism;

[0075] The branch layer includes a convolutional layer, a GRU layer and a fully connected layer;

[0076] Use the cross entropy loss function as the objective function, and use the gradient descent method to update the model parameters and iterative optimization. During the iterative process, when the value of the objective function no longer decreases significantly, stop the iteration and output the final probability distribution model.

[0077] The modality embedding tensor is used as input, and the self-attention mechanism of the Transformer encoding layer is gradually used to extract the corresponding features and generate the encoding tensor of the corresponding modality. The visual encoding tensor is optimized using the convolutional layer in the branching layer, the audio encoding tensor is optimized using the GRU layer in the branching layer, and the behavior encoding tensor is optimized using the fully connected layer. All tensors are transformed using linear transformation to obtain the emotion classification probability value, which is normalized using the Softmax function, and the emotion classification confidence distribution corresponding to each data is output;

[0078] The emotion classification confidence distribution includes categories of happy, sad, normal and undefined;

[0079] The confidence distribution of visual and audio data in categories is fused using the DST evidence combination formula, which is:

[0080]

[0081] Where, Represents the confidence distribution value after the fusion of vision and audio, Indicates the emotion category, Represents a set of visual emotion categories, Represents the set of emotional categories of the audio, represents the intersection operation, represents the empty set;

[0082] The intersection represents the selection of the intersection that exactly matches the current target emotion category Z, while the denominator represents the selection of all combinations with "partial intersection (not empty)" as the normalization basis;

[0083] We further use the DST evidence combination formula to fuse the behavioral data with the fusion results to obtain the final fusion distribution;

[0084] During the fusion process, when the visual and audio data or behavioral data conflicts with the fusion result, the adjustment factor is calculated and used to adjust the fusion confidence distribution value. The formula is:

[0085]

[0086]

[0087] Where, Represents the adjusted fusion confidence distribution value of vision and audio, represents the adjustment factor;

[0088] Conflict refers to when the intersection is "empty";

[0089] The confidence values of each category in the final fusion distribution result are arranged in ascending order, and the largest confidence is selected as the current emotion category. Then, the confidence values of each category in the final fusion distribution result are combined to obtain a continuous vector.

[0090] By introducing timestamps and frequency domain conversion (Fourier transform), the present invention ensures temporal consistency and frequency resolution in the data fusion process. Dedicated optimization layers within the branched structure (such as convolutional layers, GRU layers, and fully connected layers) provide differentiated modeling for different modalities. The convolutional layers enhance the ability to capture visual edges and local features, the GRU layers enhance the temporal memory of audio signals, and the fully connected layers achieve a compressed representation of high-dimensional information about user behavior. This ensures that the present invention possesses higher expressive power when processing features from different modalities. Furthermore, an adjustment factor, C, is introduced as a basis for credibility adjustment to proactively reduce the weight of conflicting evidence in the visual and audio modalities, imposing dynamic penalties when conflicts or high inconsistencies occur. This not only enhances the present invention's robustness against anomalous data but also prevents the propagation of emotional misjudgments, making it particularly suitable for online learning scenarios subject to frequent noise interference. Furthermore, the final output emotion continuous vector retains all probability distribution characteristics of the multimodal fusion process, making it suitable not only for current emotion judgment but also for embedding historical emotional states.

[0091] S2. Based on the continuous vector, calculate the association strength, recommendation priority, and driving score, combine it with user behavior data to generate an initial recommendation score, and then optimize it to obtain the final recommendation score;

[0092] Specifically, based on the continuous vector, the association strength, recommendation priority, and driving score are calculated. The initial recommendation score is generated by combining user behavior data and then optimized to obtain the final recommendation score. The following steps are involved:

[0093] Use the mean formula to calculate the average confidence value of the same category in the continuous vector for all users;

[0094] The mean formula is further used to calculate the average of all users’ test results for the same knowledge point;

[0095] The covariance function is combined with the mean to calculate the strength of association between each confidence value and the test result. The formula is:

[0096]

[0097] Where, Express emotions Confidence value and knowledge points The strength of the association between the test results, Indicates the total number of users, Indicates the User sentiment The confidence value of Express all emotions The average of the confidence values of Indicates the Users in Knowledge Point The test results, Represents all knowledge points The average value of the test results;

[0098] Using the standard deviation formula, after calculating the standard deviation of the mean, the Pearson correlation coefficient formula is used to combine the association strength and standard deviation to calculate the correlation coefficient between the knowledge point and each emotion category. The formula is:

[0099]

[0100] Where, Representing knowledge points and emotions The correlation coefficient of Express emotions The standard deviation of the confidence values, Representing knowledge points The standard deviation of the test results;

[0101] Use empirical rules to set the judgment threshold, compare the correlation coefficient with the threshold, and retain the correlation coefficient, emotion, and knowledge point when the correlation coefficient is greater than or equal to the threshold; otherwise, remove them;

[0102] The retained correlation coefficients are integrated to obtain the correlation vector;

[0103] Use the vector dot product method to calculate the driving score of the continuous vector and the associated vector. The formula is:

[0104]

[0105] Where, Representing knowledge points The drive score, Indicates the total number of emotions;

[0106] We then obtained a set of knowledge points from the education platform through the API interface. After assigning initial weights to each pair of knowledge points in the knowledge point set using the expert domain method, we used the knowledge points in the knowledge point set as nodes and drew a directed graph between the nodes. The initial weights between all nodes were adjusted based on the driving scores. The formula is:

[0107]

[0108] Where, Indicates the next moment after update Time Knowledge Points arrive The weight of Indicates the current time Time Knowledge Points arrive The initial weight of represents the impact factor, which can be set through grid search. Representing knowledge points The driving score;

[0109] Use statistical methods to calculate the accuracy and error rate of each user's test results, and integrate the knowledge points in the error rate to define it as the set of unmastered knowledge points;

[0110] Use the updated weights combined with the set of unmastered knowledge points to calculate the recommendation priority of each knowledge point. The formula is:

[0111]

[0112] Where, Representing knowledge points The recommended priority, Represents a set of unmastered knowledge points, Representing a collection The number of knowledge points not mastered;

[0113] Based on the accuracy and recommendation priority, the user's initial recommendation score for the knowledge point is calculated using the formula:

[0114]

[0115] Where, Indicates the Users on knowledge points The initial recommendation score of Indicates the The accuracy of the test results of users;

[0116] Use empirical rules to set the decay rate, and further obtain the timestamp of the user's current behavior data through the API interface. The difference between the timestamp of the user's current behavior data and the timestamp of the historical behavior data is defined as the time difference. Combined with the exponential function, the time decay weight is calculated. The formula is:

[0117]

[0118] Where, Indicates time The time decay weight when represents the base of natural logarithm, represents the decay rate, Indicates time difference;

[0119] Use the multiplication formula to calculate the product of the time decay weight and the initial recommendation score to obtain the optimized recommendation score;

[0120] Use the cosine similarity formula combined with the optimized recommendation score to calculate the similarity score between the current user and all users. The formula is:

[0121]

[0122] Where, Indicates the User and Similarity scores between users, Represents the total number of knowledge points, The optimized Users on knowledge points Optimization recommendation score, The optimized Users on knowledge points Optimized recommendation score;

[0123] Set the selection threshold according to the empirical rule, compare the similarity score with the selection threshold, and when the similarity score is greater than or equal to the selection threshold, retain the similarity score and construct a similarity set; otherwise, remove it;

[0124] Combine the optimized recommendation score and the similarity score to calculate the final recommendation score for each knowledge point. The formula is:

[0125]

[0126] Where, Indicates the Users on knowledge points The final recommendation score, Indicates the Similarity sets for users.

[0127] By performing layer-by-layer combined operations on the mean, covariance and standard deviation of the user's emotional confidence and test scores, the potential dependency pattern between emotional state and knowledge mastery is effectively captured. By jointly calculating the covariance function and the Pearson correlation coefficient, the present invention realizes dynamic correlation modeling between emotional state and knowledge point mastery. Compared with traditional static association rules (such as fixed weights or thresholds), the present invention can adaptively capture the nonlinear impact of emotional fluctuations on learning effects, significantly improve the accuracy of recommendations, and screen out significant features by setting the correlation coefficient threshold to avoid interference from noise data and improve the robustness of the model. The resulting "correlation vector" retains the important connection between emotion and knowledge, providing high-quality input for the next step of driving score calculation, and the vector dot product effectively couples the user's current emotional state with the knowledge point correlation features. In summary, the obtained driving score not only has the similarity representation function of high-dimensional space, but also provides a quantitative basis for the subsequent edge weight adjustment in the graph. The combination of the expert domain method and the driving score enables the present invention to construct a multi-dimensional weight adjustment mechanism, which not only considers the logical relationship between knowledge points (such as pre-requisite and post-requisite knowledge), but also incorporates emotional driving factors to form a dynamic knowledge graph. Secondly, the coordinated application of time-decaying weight and cosine similarity enables the present invention to have time and space perception capabilities. The time-decaying weight ensures that the recommended content is consistent with the user's latest behavior, while the cosine similarity mines implicit needs through the behavior patterns of user groups. For example, if a user recently searches for "programming basics" and the search pattern is highly matched with that of similar user groups, the present invention can recommend advanced courses that he is interested in, achieving dual optimization of personalization and collaborative filtering.

[0128] S3. Integrate the recommendation priority, driving score, and final recommendation score to construct a feature vector, which is defined as the main effect feature vector. Then, the interaction effect feature vector is calculated. The main effect and interaction effect feature vectors are combined and the ranking score is calculated to obtain the final recommendation list.

[0129] Specifically, the following steps are involved: integrating the recommendation priority, driving score, and final recommendation score to construct a feature vector, defining it as the main effect feature vector, calculating the interaction effect feature vector, fusing the main effect and interaction effect feature vectors, and calculating the ranking score to obtain the final recommendation list:

[0130] Integrate the driving score, recommendation priority, and final recommendation score to construct a feature vector, which is defined as the main effect feature vector. Then, combine the features in the main effect feature vector in pairs.

[0131] Use the dot product operation to calculate the interaction effect strength between feature pairs, which is defined as combining the interaction effect features to obtain the interaction effect feature vector;

[0132] Build a hybrid model using a feedforward neural network and LambdaMART network architecture, including a front-end feature extraction layer and a back-end sorting layer;

[0133] The front-end feature extraction layer uses the ReLU activation function to fuse the main effect and interaction effect feature vectors to obtain a fused feature vector. The formula is:

[0134]

[0135] Where, represents the fused feature vector, represents the ReLU activation function, and represents the mapping matrix, Represents vector concatenation operation, represents the main effect eigenvector, represents the interaction effect eigenvector, and represents the bias term;

[0136] The backend sorting layer uses the random search method to set the learning rate and the domain method to set the initial total number of trees. The fused feature vector is input into each tree to obtain the prediction score vector. The learning rate, total number of trees, fused feature vector and prediction score are combined to calculate the sorting score of each knowledge point. The formula is:

[0137]

[0138] Where, Indicates the Users on knowledge points The ranking score of represents the total number of trees, Indicates the The learning rate of the tree, Indicates the The prediction score vector of the tree, Represents an aggregate function;

[0139] Use the addition formula to calculate the difference in knowledge point ranking scores;

[0140] Define the objective function and minimize the objective function value. The formula is:

[0141]

[0142] Where, represents the objective function value, Represents a user collection, Indicates the The knowledge point collection of users, represents the natural logarithm function, represents a natural constant, Indicates the Knowledge points of users and the difference in ranking scores;

[0143] The Adam optimizer is used to optimize and iterate the parameter combination of the hybrid model, and the back-propagation method is used to calculate the gradient. During the iteration process, when the decrease value of the objective function value no longer decreases significantly, the iteration is stopped and the final hybrid model is output;

[0144] The feature vector is input into the final hybrid model, and the sorted recommendation list is output as the final recommendation result.

[0145] By constructing main effect feature vectors and combining them pairwise, and then calculating the interaction effect strength by dot product, the present invention not only enhances its ability to capture high-order nonlinear features, but also realizes the modeling of complex relationships between user behavior features and content structure, thereby improving the personalization and accuracy of recommendations. By jointly modeling main effect features and interaction effect features, the present invention realizes multi-scale analysis of user-knowledge point relationships. The main effect features reflect the importance of a single dimension, and the interaction effect features capture the synergistic effects of multiple dimensions. For example, if a knowledge point has a high driving score but a low recommendation priority, the interaction effect analysis can reveal its potential correlation with the user's emotional state, thereby optimizing the recommendation strategy. This dual-path modeling method breaks through the limitations of traditional single feature dependence and significantly improves the generalization ability of the present invention. The combination of feedforward neural network and LambdaMART enables the present invention to have both feature representation and recommendation accuracy. In terms of performance and sorting efficiency, the feedforward neural network extracts implicit features through the ReLU activation function, and LambdaMART optimizes the sorting score through the gradient boosting mechanism. For example, in the process of knowledge point sorting, the present invention can dynamically adjust the prediction weights of different tree structures to avoid the sensitivity of a single tree model to noise. In addition, the combined use of the random search method and the domain method ensures the global optimal configuration of hyperparameters (such as learning rate and total number of trees), further improving the stability of the present invention. Secondly, the present invention implements a modular design of the model through a layered architecture (front-end feature extraction layer and back-end sorting layer), which is convenient for subsequent expansion and tuning. The structured expression of the front-end feature fusion formula makes the feature interaction process traceable, and the tree structure of LambdaMART provides local interpretability, which is convenient for verifying the rationality of the recommendation logic, so that the present invention meets the requirements of transparency and controllability.

[0146] S4. Display the data through a visual interface and store the data in a database;

[0147] Specifically, displaying data through a visual interface means using the PowerBI tool to import the sorted recommendation list into the visualization platform, displaying the priority ranking of knowledge points through a bar chart, presenting the distribution of users' interest in knowledge points through a heat map, and tracking changes in recommendation trends through a line chart.

[0148] The sorted recommendation list is presented in a visual form through the PowerBI tool, and the built-in bar chart, heat map and line chart modules of the PowerBI tool are used to dynamically display the priority of knowledge points, user interest distribution and recommendation trends, providing users with intuitive data insights.

[0149] Furthermore, storing data through a database refers to using a MongoDB distributed database to partition and store the recommendation list by user ID, and appending a timestamp, a knowledge point ID, and a recommendation score field.

[0150] The data is stored in the MongoDB distributed database, and the recommendation list is partitioned by user ID, with timestamp, knowledge point ID and recommendation score fields attached, making it easier for users to query the data and avoid single-point storage overload.

[0151] This embodiment also provides a personalized education recommendation system based on artificial intelligence, including:

[0152] The acquisition and optimization module is used to collect multimodal data for data fusion and optimize it with the adjustment factor to obtain a continuous vector;

[0153] The calculation and generation module is used to calculate the association strength, recommendation priority, and driving score based on the continuous vector, and then generate the initial recommendation score based on user behavior data and optimize it to obtain the final recommendation score;

[0154] The construction and ranking module is used to integrate the recommendation priority, driving score and final recommendation score, construct the feature vector, define it as the main effect feature vector, calculate the interaction effect feature vector, fuse the main effect and interaction effect feature vectors, and calculate the ranking score to obtain the final recommendation list;

[0155] The display and storage module is used to display and store data through a visual interface and database.

[0156] This embodiment also provides a computer device suitable for the case of a personalized education recommendation method based on artificial intelligence, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute computer-executable instructions to implement the personalized education recommendation method based on artificial intelligence proposed in the above embodiment.

[0157] The computer device may be a terminal, comprising a processor, memory, a communication interface, a display, and an input device connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system and computer programs. The internal memory provides an environment for the operating system and computer programs stored in the non-volatile storage media. The communication interface of the computer device is used to communicate with external terminals via wired or wireless communication. Wireless communication may be achieved via Wi-Fi, a carrier network, NFC (near-field communication), or other technologies. The display of the computer device may be a liquid crystal display or an electronic ink display. The input device may be a touchscreen overlay on the display, buttons, a trackball, or a touchpad on the computer device housing, or an external keyboard, touchpad, or mouse.

[0158] This embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the personalized education recommendation method based on artificial intelligence proposed in the above embodiment; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic memory, flash memory, magnetic disk or optical disk.

[0159] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A personalized education recommendation method based on artificial intelligence, characterized by: include, Collect multimodal data, fuse the data using the DST evidence combination formula, and optimize it with the adjustment factor to obtain a continuous vector; The multimodal data includes visual image data, audio data and behavioral data; Based on the continuous vector, the association strength, recommendation priority, and driving score are calculated. The initial recommendation score is generated by combining user behavior data and then optimized to obtain the final recommendation score. Integrate the recommendation priority, driving score, and final recommendation score to construct a feature vector, which is defined as the main effect feature vector. Then calculate the interaction effect feature vector, fuse the main effect and interaction effect feature vectors, and calculate the ranking score to obtain the final recommendation list. The data is displayed through a visual interface and stored in a database.

2. The artificial intelligence-based personalized education recommendation method according to claim 1, wherein: The method of collecting multimodal data, fusing the data using the DST evidence combination formula, and optimizing the adjustment factor to obtain a continuous vector includes the following steps: Obtain the user's visual image data, audio data, and behavioral data from the education platform through the API interface, and record the test results after the user completes the test on the education platform; The behavioral data refers to the user's click frequency and stay time on the platform; Add a timestamp to each data type and use Fourier transform to convert the audio data into the frequency domain; Use a random number generator to randomly generate an embedding vector for the corresponding data, and use a broadcast mechanism to perform element-wise addition of each data to the corresponding embedding vector to obtain the corresponding modal embedding tensor; Use the Transformer architecture to build a probability distribution model, take the modal embedding tensor as input, and output the confidence distribution of the emotion classification corresponding to each data point; The emotion classification confidence distribution includes categories of happy, sad, normal and undefined; Use the DST evidence combination formula to fuse the confidence distribution of visual and audio data in categories; We further use the DST evidence combination formula to fuse the behavioral data with the fusion results to obtain the final fusion distribution; During the fusion process, when the visual and audio data or behavioral data conflicts with the fusion result, an adjustment factor is calculated and used to adjust the fusion confidence distribution value; The confidence values of each category in the final fusion distribution result are arranged in ascending order, and the largest confidence is selected as the current emotion category. Then, the confidence values of each category in the final fusion distribution result are combined to obtain a continuous vector.

3. The artificial intelligence-based personalized education recommendation method according to claim 2, characterized in that: The calculation of association strength, recommendation priority, and driving score based on the continuous vector, combined with user behavior data to generate an initial recommendation score and then optimize it to obtain the final recommendation score includes the following steps: Use the mean formula to calculate the average confidence value of the same category in the continuous vector for all users; The mean formula is further used to calculate the average of all users’ test results for the same knowledge point; Calculate the strength of association between each confidence value and the test result using the covariance function combined with the mean; After calculating the standard deviation of the mean using the standard deviation formula, the Pearson correlation coefficient formula is used to combine the association strength and standard deviation to calculate the correlation coefficient between the knowledge point and each emotion category; Use empirical rules to set the judgment threshold, compare the correlation coefficient with the threshold, and retain the correlation coefficient, emotion, and knowledge point when the correlation coefficient is greater than or equal to the threshold; otherwise, remove them; The retained correlation coefficients are integrated to obtain the correlation vector; Use the vector dot product method to calculate the driving score of the continuous vector and the correlation vector; We then retrieve a set of knowledge points from the education platform through an API. Using the expert domain method, we assign initial weights to each pair of knowledge points in the set. We then use the knowledge points in the set as nodes, draw a directed graph between the nodes, and adjust the initial weights between all nodes based on the driving scores. Use the updated weights to calculate the recommendation priority of each knowledge point; Use statistical methods to calculate the accuracy of each user's test results, and calculate the user's initial recommendation score for the knowledge point based on the accuracy rate; Use empirical rules to set the decay rate, and further obtain the timestamp of the user's current behavior data through the API interface. The difference between the timestamp of the user's current behavior data and the timestamp of the historical behavior data is defined as the time difference, and then combined with the exponential function to calculate the time decay weight; Use the multiplication formula to calculate the product of the time decay weight and the initial recommendation score to obtain the optimized recommendation score; Use the cosine similarity formula combined with the optimized recommendation score to calculate the similarity score between the current user and all users; Set the selection threshold according to the empirical rule, compare the similarity score with the selection threshold, and when the similarity score is greater than or equal to the selection threshold, retain the similarity score and construct a similarity set; otherwise, remove it; The optimized recommendation score and similarity score are combined to calculate the final recommendation score for each knowledge point.

4. The artificial intelligence-based personalized education recommendation method according to claim 3, wherein: The process of integrating the recommendation priority, driving score, and final recommendation score to construct a feature vector and define it as a main effect feature vector, calculating the interaction effect feature vector, fusing the main effect and interaction effect feature vectors, and calculating the ranking score to obtain the final recommendation list includes the following steps: Integrate the driving score, recommendation priority, and final recommendation score to construct a feature vector, which is defined as the main effect feature vector. Then, combine the features in the main effect feature vector in pairs. Use the dot product operation to calculate the interaction effect strength between feature pairs, which is defined as combining the interaction effect features to obtain the interaction effect feature vector; Build a hybrid model using a feedforward neural network and LambdaMART network architecture, including a front-end feature extraction layer and a back-end sorting layer; The front-end feature extraction layer uses the ReLU activation function to fuse the main effect and interaction effect feature vectors to obtain a fused feature vector; The backend sorting layer uses a random search method to set the learning rate and a domain method to set the initial total number of trees. The fused feature vector is input into each tree to obtain a prediction score vector. The learning rate, total number of trees, fused feature vector, and prediction score are then combined to calculate the sorting score of each knowledge point. Use the addition formula to calculate the difference in knowledge point ranking scores; After defining the objective function, minimizing the objective function value, and performing iterative optimization, the feature vector is input into the final hybrid model, and the sorted recommendation list is output as the final recommendation result.

5. The method for personalized education recommendation based on artificial intelligence according to claim 4, characterized in that: The display of data through a visual interface refers to using the PowerBI tool to import the sorted recommendation list into the visualization platform, displaying the priority ranking of knowledge points through a bar chart, presenting the distribution of users' interest in knowledge points through a heat map, and tracking changes in recommendation trends through a line chart.

6. The artificial intelligence-based personalized education recommendation method according to claim 5, characterized in that: The data storage through the database refers to using the MongoDB distributed database to partition and store the recommendation list by user ID, and append a timestamp, knowledge point ID and recommendation score field.

7. The artificial intelligence-based personalized education recommendation method according to claim 1, characterized in that: The multimodal data are collected and preprocessed, including denoising, cleaning, time alignment, and normalization.

8. An artificial intelligence-based personalized education recommendation system, based on the artificial intelligence-based personalized education recommendation method according to any one of claims 1 to 7, characterized in that: include, The acquisition and optimization module is used to collect multimodal data for data fusion and optimize it with the adjustment factor to obtain a continuous vector; The calculation and generation module is used to calculate the association strength, recommendation priority, and driving score based on the continuous vector, and then generate the initial recommendation score based on user behavior data and optimize it to obtain the final recommendation score; The construction and ranking module is used to integrate the recommendation priority, driving score and final recommendation score, construct the feature vector, define it as the main effect feature vector, calculate the interaction effect feature vector, fuse the main effect and interaction effect feature vectors, and calculate the ranking score to obtain the final recommendation list; The display and storage module is used to display and store data through a visual interface and database.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the artificial intelligence-based personalized education recommendation method described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the artificial intelligence-based personalized education recommendation method according to any one of claims 1 to 7 are implemented.

Images