Online education training intelligent pushing system based on AI adaptive learning

Abstract

This invention discloses an intelligent push system for online education and training based on AI adaptive learning, belonging to the field of online education and training technology. The invention includes a data acquisition module, an AI adaptive learning analysis module, a user profile construction module, a push strategy generation module, a push execution module, a feedback optimization module, a content update module, and a permission management module. By integrating basic information, real-time learning behavior, learning outcomes, and environmental preferences, it achieves comprehensive coverage and accurate capture of learning data. Relying on the dynamic iterative model of the AI ​​adaptive learning analysis module and the multi-dimensional tag system of the user profile construction module, it achieves accurate prediction of users' short-term and long-term needs, captures changes in users' learning status, knowledge gaps, and personalized preferences, and generates a comprehensive and three-dimensional user profile. This ensures that the pushed content matches the user's current learning needs, adapts to their ability level and learning habits, and achieves the scientific generation, stable execution, and continuous optimization of personalized push strategies.

Description

Technical Field

[0001] This invention relates to the field of online education and training technology, and in particular to an intelligent online education and training push system based on AI adaptive learning. Background Technology

[0002] Traditional online education recommendation systems often use fixed rules or simple collaborative filtering algorithms, which recommend content based only on basic user information or limited historical behavior. This results in problems such as limited data collection dimensions and low accuracy in identifying user needs.

[0003] These systems often fail to fully capture changes in a user's learning status, knowledge gaps, and personalized preferences. This results in insufficient alignment between the pushed content and the user's actual learning needs, skill level, and learning scenario, frequently leading to mismatched content difficulty and a disconnect between learning pace and content. Furthermore, existing systems lack dynamic adaptive optimization mechanisms, with fixed model parameters that struggle to adjust push strategies in real time to keep pace with changes in user learning progress and needs. This not only impacts user learning efficiency and experience but may also lead to user churn, failing to meet the refined and personalized development needs of the online education and training field. Summary of the Invention

[0004] The purpose of this invention is to provide an intelligent online education and training push system based on AI adaptive learning to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an intelligent online education and training push system based on AI adaptive learning, comprising:

[0006] The data acquisition module is configured to collect multi-dimensional learning data from users during the education and training process, map numerical data in the multi-dimensional learning data to the [0,1] interval, convert text data into vector form, and perform numerical processing on categorical data.

[0007] The AI ​​adaptive learning and analysis module is configured to achieve dynamic iterative learning through deep learning algorithms, and to mine user learning patterns, knowledge mastery status and potential needs based on the user's multi-dimensional learning data.

[0008] The user profile building module is configured to combine the AI ​​analysis results output by the AI ​​adaptive learning analysis module to generate personalized user profiles. The personalized user profiles include user basic attributes, learning behavior characteristics, knowledge demand tendencies, and learning ability levels.

[0009] The push strategy generation module is configured to build a multi-dimensional matching model based on user profiles and the content tag system of the education and training resource library, and generate personalized push strategies that adapt to users' real-time needs.

[0010] The push execution module is configured to accurately deliver push content to users through multiple terminal channels and monitor the entire push process;

[0011] The feedback optimization module is configured to collect user interaction feedback data on the pushed content and iteratively optimize the AI ​​learning model and push strategy.

[0012] Furthermore, the multi-dimensional learning data acquisition process of the data acquisition module includes basic information acquisition, real-time learning behavior acquisition, learning outcome data acquisition, and environmental preference data acquisition, specifically including:

[0013] When collecting basic information, users’ age, gender, education level, professional field, occupation, learning goals and available learning time range are collected through user registration forms, identity verification, and actively filled questionnaires.

[0014] When collecting real-time learning behavior data, the system uses data tracking technology, terminal sensors, and platform interaction logs to record user behavior data within the learning platform. This behavior data includes the number of course clicks, the duration of each lesson, video playback progress, courseware download and collection records, learning note editing content, online Q&A interaction times, group discussion participation frequency, and assignment submission time.

[0015] When collecting learning outcome data, the system connects to the platform's testing system, homework correction system, and skills assessment tools to collect user unit test scores, mock exam scores, homework accuracy, error distribution, knowledge module mastery scores, and periodic learning outcome reports.

[0016] When collecting environmental preference data, user learning preference data is collected by obtaining terminal device information and geographical location-related learning scenarios through user authorization, and by combining user selection records of course type, teaching style, content difficulty and learning resource format.

[0017] Furthermore, the dynamic iterative learning process of the AI ​​adaptive learning analysis module includes:

[0018] Construct a basic learning model based on deep neural networks. The input layer of the model is set to the user's multi-dimensional data vector, the hidden layer contains 3-5 fully connected layers, and the output layer is initially set to the user's knowledge need category and learning ability score.

[0019] The real-time updated multi-dimensional learning data is input into the basic learning model, and the model parameters are continuously optimized through the backpropagation algorithm. The weight allocation of the user's recent learning behavior is increased to achieve real-time capture of changes in the user's learning status.

[0020] By combining time series analysis algorithms to perform trend analysis on users' historical learning data, we can predict users' short-term and long-term learning needs.

[0021] Among them, short-term demand forecasting focuses on the immediate needs of users at their current learning stage, while long-term demand forecasting is based on the needs of users' learning goals and plans for phased learning content.

[0022] Set the model iteration cycle and adjust the model's loss function weights in real time based on user feedback data.

[0023] Furthermore, the weighting of users' recent learning behaviors is increased to enable real-time capture of changes in users' learning status, including:

[0024] Obtain the user's historical learning behavior feature vector;

[0025] Calculate the time decay weighting coefficient;

[0026] The user's historical learning behavior feature vector is weighted and aggregated to generate a weighted behavior feature vector;

[0027] The weighted behavioral feature vector is input into a deep neural network model to calculate the user's knowledge mastery vector.

[0028] Calculate the difference between the user's current learning state and the historical stable learning state based on the user's knowledge mastery vector and the historical knowledge mastery mean vector;

[0029] When the difference between a user's current learning state and their historical stable state exceeds a preset difference threshold, the user profile is updated.

[0030] Furthermore, the personalized user profile generation process of the user profile construction module includes:

[0031] Basic attribute tags are generated based on the collected user basic information. The basic attribute tags include user age, education stage, professional field, learning goals, and available time.

[0032] By combining AI analysis of user learning behavior patterns, behavioral feature tags are generated, including learning time period preferences, learning rhythm, interaction preferences, and content selection tendencies.

[0033] Based on the user's knowledge gaps and demand prediction results mined by AI, knowledge demand tags are generated. These tags include the knowledge modules that the user currently needs to supplement, the types of skills that need to be improved, and the appropriate learning difficulty level.

[0034] Based on user learning outcome data and the learning ability scores output by AI models, user ability levels will be divided into basic, intermediate, proficient, and expert levels, and corresponding ability tags will be generated.

[0035] Furthermore, the user profile building module integrates basic attribute tags, behavioral feature tags, knowledge requirement tags, and ability tags to generate personalized user profiles. When the cumulative update volume of user learning data reaches a preset threshold or the user's learning goals change, the personalized user profile is automatically updated.

[0036] Furthermore, the personalized push strategy construction process of the push strategy generation module includes:

[0037] All content in the education and training resource library is labeled with multi-dimensional content tags, including knowledge domain tags, difficulty tags, format tags, duration tags, applicable scenario tags, and applicable ability tags.

[0038] Construct a multi-dimensional matching model between user profiles and content tags, and calculate the overall matching degree between users and content in each tag dimension;

[0039] All candidate push content is initially ranked based on comprehensive matching degree. At the same time, the initial ranking results are adjusted in a second step by taking into account the user's real-time learning scenario, learning progress and historical feedback to determine the final priority of push content.

[0040] Based on the priority ranking of the pushed content, combined with the user's available learning time and push frequency rules, a complete personalized push strategy is generated, which includes a list of pushed content and push channels.

[0041] Furthermore, the push execution module pushes educational and training content to users through mobile applications, web platforms, email, and SMS channels, collects user feedback on clicks, learning completion, evaluations, and adjustment requests, and transmits the feedback data to the AI ​​learning and analysis module in real time.

[0042] Furthermore, the push execution module is also used to monitor the push process in real time. Push monitoring includes push success rate monitoring, content delivery timeliness monitoring, and user reception status monitoring. When a push abnormality is detected, the push channel is automatically switched or the push time is adjusted.

[0043] Furthermore, the push execution module also performs the following operations:

[0044] Get the type of push channel when the push content is pushed;

[0045] When the push channel type is a mobile application or a web platform, determine the preset push interface position in the operation interface;

[0046] If the user is detected not to have clicked on the push content based on the push interface location, obtain the user's first behavior stream during the push period;

[0047] Based on the first line of flow, determine the first attention distribution change of the user interface;

[0048] The first attention distribution change features are extracted based on the changes in the first attention distribution. These features include peak attention, cumulative attention, and attention duration.

[0049] Analyze the characteristics of the first attention distribution change. If the difference between the upper and lower limits of the feature value corresponding to each feature type is less than the corresponding difference threshold, the neglect type is determined to be active neglect; otherwise, the neglect type is passive neglect.

[0050] If the ignore type is "actively ignore", then no second push will be made;

[0051] If the neglect type is passive neglect, the second behavior stream will continue to be acquired during the user's future online time period;

[0052] Determine the second attention distribution change based on the second behavior flow and extract the features of the second attention distribution change;

[0053] If the difference between the upper and lower limits of the feature value corresponding to each feature type is less than the corresponding difference threshold, then a second push will be made immediately.

[0054] If the second attention distribution change feature satisfies the condition that there is a target region within the smallest local interface region containing the interface point corresponding to the peak attention feature value, then push the content again within the target region.

[0055] Among them, the lower limit of the peak attention feature value corresponding to the target area is greater than the preset lower limit threshold, and the target area can accommodate the push interface.

[0056] Furthermore, it also includes a content update module, configured to capture, filter, and update the content in the education and training resource library in real time. Based on industry development trends, knowledge update trends, and changes in user needs, it supplements new course content, teaching materials, and learning tools, and includes the newly added content in the push content pool after tagging it.

[0057] Furthermore, it also includes a permission management module, which is configured to perform hierarchical authorization management for users, restricting the access scope and usage permissions of pushed content based on user identity, payment level, and learning permissions.

[0058] Compared with the prior art, the beneficial effects of the present invention are:

[0059] 1. This invention integrates basic information, real-time learning behavior, learning outcomes, and environmental preference data through a multi-dimensional data acquisition module, and standardizes different types of data to achieve comprehensive coverage and accurate capture of learning data. It breaks through the limitations of single data acquisition, ensures the integrity, authenticity, and timeliness of the data, and provides high-quality data support for AI analysis, user profile construction, and push strategy generation. It avoids push deviations caused by incomplete data and accurately matches user needs.

[0060] 2. This invention relies on the dynamic iterative model of the AI ​​adaptive learning analysis module and the multi-dimensional tag system of the user profile construction module, combined with the time series analysis algorithm, to achieve accurate prediction of users' short-term and long-term needs. It can keenly capture changes in users' learning status, knowledge gaps, and personalized preferences, generate a comprehensive and three-dimensional user profile, and update it dynamically. This ensures that the pushed content not only matches the user's current learning needs but also adapts to their ability level and learning habits, significantly improving the adaptability and relevance of the pushed content and enhancing the user's learning experience.

[0061] 3. This invention utilizes a closed-loop mechanism consisting of a multi-dimensional matching model in the push strategy generation module, a multi-channel push execution module, and a feedback optimization module. Combined with the synergistic effect of the content update and permission management modules, it achieves the scientific generation, stable execution, and continuous optimization of personalized push strategies. This ensures the timeliness, richness, and security of the pushed content. It can also adjust the model and strategy in real time based on user feedback, effectively improving user learning efficiency and satisfaction, and enhancing the system's adaptability and core competitiveness. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the AI ​​analysis and user profile construction process of the present invention;

[0063] Figure 2 This is a schematic diagram of the push strategy generation and feedback optimization process of the present invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Please see Figures 1-2 The present invention provides the following technical solutions:

[0066] An AI-based adaptive learning-based intelligent online education and training recommendation system includes:

[0067] The data acquisition module is configured to collect multi-dimensional learning data of users in the education and training process, map numerical data (such as test scores and learning time) in the multi-dimensional learning data to the [0,1] interval, convert text data (such as learning notes and Q&A content) into vector form, and perform numerical processing on categorical data (such as course type and learning scenario).

[0068] The AI ​​adaptive learning and analysis module is configured to achieve dynamic iterative learning through deep learning algorithms, and to mine user learning patterns, knowledge mastery status and potential needs based on the user's multi-dimensional learning data.

[0069] The user profile building module is configured to combine the AI ​​analysis results output by the AI ​​adaptive learning analysis module to generate personalized user profiles. The personalized user profiles include user basic attributes, learning behavior characteristics, knowledge demand tendencies, and learning ability levels.

[0070] The push strategy generation module is configured to build a multi-dimensional matching model based on user profiles and the content tag system of the education and training resource library, and generate personalized push strategies that adapt to users' real-time needs.

[0071] The push execution module is configured to accurately deliver push content to users through multiple terminal channels and monitor the entire push process;

[0072] The feedback optimization module is configured to collect user interaction feedback data on the pushed content and iteratively optimize the AI ​​learning model and push strategy in reverse.

[0073] The content update module is configured to capture, filter, and update the content in the education and training resource library in real time. Based on industry development trends, knowledge update trends, and changes in user needs, it supplements new course content, teaching materials, and learning tools. After adding new content, it is tagged and included in the push content pool to ensure the timeliness and richness of the push content.

[0074] The access control module is configured to manage user permissions in a tiered manner. Based on user identity, payment level, and learning permissions, it restricts the scope of access and usage rights of pushed content, while ensuring the security and privacy of user learning data. Only the authorized module can access and process sensitive user data.

[0075] The multi-dimensional learning data acquisition process of the data acquisition module includes basic information acquisition, real-time learning behavior acquisition, learning outcome data acquisition, and environmental preference data acquisition, specifically including:

[0076] When collecting basic information, user age, gender, education stage (such as primary and secondary school, university, workplace training, etc.), professional field, occupation, learning goals (such as preparation for further education, skills improvement, certificate acquisition, etc.) and available learning time range are collected through user registration forms, identity verification, and actively filled questionnaires.

[0077] When collecting real-time learning behavior data, the system uses data tracking technology, terminal sensors, and platform interaction logs to record user behavior data within the learning platform. This behavior data includes the number of course clicks, the duration of each lesson, video playback progress (such as fast forward, rewind, and repeated viewing of segments), courseware download and collection records, learning note editing content, online Q&A interaction times, group discussion participation frequency, and assignment submission time.

[0078] When collecting learning outcome data, the system connects to the platform's testing system, homework correction system, and skills assessment tools to collect user unit test scores, mock exam scores, homework accuracy, error distribution, knowledge module mastery scores (such as conclusions like "weak in algebra" or "unfamiliar with programming syntax" drawn from error analysis), and periodic learning outcome report data.

[0079] When collecting environmental preference data, user learning preference data is collected by obtaining terminal device information (such as learning using a mobile phone, tablet or computer) and geographically associated learning scenarios (such as home, commuting, office, etc.) through user authorization, combined with user selection records of course type (such as live class, recorded class, micro-class), teaching style (such as humorous and interesting, rigorous academic), content difficulty (such as beginner, intermediate, expert) and learning resource format (such as video, audio, text, practical case) to collect user learning preference data.

[0080] In the above embodiments, by integrating multiple channels, the system comprehensively acquires the user's core attributes and learning goals, providing an initial decision-making basis for the system and ensuring that the pushed content aligns with the user's basic positioning. Real-time learning behavior collection utilizes advanced technologies to accurately record every detail of the user's learning operations, dynamically reflecting changes in the user's learning status and allowing the system to promptly capture the user's interests and weaknesses. Learning outcome data collection, through integration with professional assessment tools, quantifies the user's learning effectiveness, providing an objective basis for judging the user's knowledge mastery and avoiding errors caused by subjective judgment. Environmental preference data collection fully considers the user's learning scenarios and personalized choices, ensuring that the pushed content not only matches the user's knowledge needs but also aligns with the user's usage habits and scenario limitations. The multi-stage collection process complements each other, forming a complete data collection closed loop, ensuring the comprehensiveness, authenticity, and timeliness of the collected data.

[0081] The dynamic iterative learning process of the AI ​​adaptive learning analytics module includes:

[0082] Construct a basic learning model based on deep neural networks. The input layer of the model is set to the user's multi-dimensional data vector, the hidden layer contains 3-5 fully connected layers, and the output layer is initially set to the user's knowledge need category and learning ability score.

[0083] Based on industry-standard education and training datasets and historical user learning data, the model is pre-trained to enable it to have basic user feature recognition and demand judgment capabilities.

[0084] The model inputs real-time updated multi-dimensional learning data into the basic learning model and continuously optimizes the model parameters through the backpropagation algorithm. It increases the weight allocation of the user's recent learning behavior (such as course selection and homework completion in the past 7 days) to achieve real-time capture of changes in the user's learning status. For example, when a user makes multiple mistakes in the "Calculus" chapter of the "Advanced Mathematics" course, the model automatically increases the demand recognition weight of that knowledge module and determines that the user has a knowledge gap in that area.

[0085] By combining time series analysis algorithms to perform trend analysis on users' historical learning data, we can predict users' short-term and long-term learning needs.

[0086] Among them, short-term demand forecasting focuses on the immediate needs of users at the current learning stage (such as completing the accompanying exercises for the current chapter and supplementing the relevant knowledge points with extended materials), while long-term demand forecasting is based on users' learning goals (such as taking the CET-4 and CET-6 exams in 3 months) to plan the needs of phased learning content (such as vocabulary accumulation in the first month, grammar reinforcement in the second month, and practice with past exam papers in the third month).

[0087] Set the model iteration cycle and adjust the model's loss function weights in real time based on user feedback data. When the user completion rate of a certain type of pushed content is lower than a preset threshold (such as 30%), the model automatically reduces the recommendation weight of that type of content to improve the accuracy of content recognition that better meets user needs.

[0088] In the above embodiments, the reliability of the initial performance of the model is ensured by pre-training with industry-standard data and historical user data. The continuous input of real-time learning data and the application of the backpropagation algorithm enable dynamic optimization of model parameters, allowing the system to keenly capture recent changes in user learning behavior and adjust its judgment of user needs in a timely manner, avoiding push delays caused by model solidification. The introduction of time series analysis algorithms effectively achieves accurate prediction of users' short-term and long-term learning needs, meeting the immediate needs of users at the current learning stage while providing phased planning support for users' long-term learning goals. The model iteration cycle and loss function weight adjustment mechanism ensure that the model can be continuously optimized based on user feedback, constantly improving the accuracy of demand identification, and ensuring that the pushed content always maintains a high degree of consistency with users' dynamic needs, significantly enhancing the system's adaptability and push effectiveness.

[0089] Increase the weighting of users' recent learning behaviors to enable real-time capture of changes in users' learning status, including:

[0090] Obtain user's historical learning behavior feature vector ,in For timestamps, , For the current moment, Total duration of historical data;

[0091] Calculate the time decay weighting coefficient :

[0092] ;

[0093] in, This is the time decay rate parameter;

[0094] The user's historical learning behavior feature vector is weighted and aggregated to generate a weighted behavior feature vector. :

[0095] ;

[0096] The weighted behavioral feature vector is input into the deep neural network model to calculate the user's knowledge mastery vector. :

[0097] ;

[0098] Among them, the user's knowledge mastery vector It is A vector of dimensions, each dimension using express, This represents the user's opinion on the first Assessment of mastery of each knowledge module The total number of predefined knowledge modules, Indicates transpose; It is a deep neural network model;

[0099] During deep neural network model training, the following steps are taken: First, evaluation data from educational experts on evaluators and the evaluators' historical learning behavior data are acquired. The evaluation data includes the educational experts' scores on the evaluators' mastery of each knowledge module. Next, based on the construction rule of identical weighted behavioral feature vectors, a weighted behavioral feature vector for the evaluators is constructed according to their historical learning behavior data. Similarly, based on the construction rule of identical user knowledge mastery vectors, a knowledge mastery vector for the evaluators is constructed according to the evaluation data. The weighted behavioral feature vector of the evaluators is used as the input to a pre-defined CNN neural network model, and the knowledge mastery vector of the evaluators is used as the output of the model for training. After training converges, the deep neural network model is obtained.

[0100] Based on the user's knowledge mastery vector and the historical mean knowledge mastery vector, calculate the difference between the user's current learning state and their historical stable learning state. :

[0101] ;

[0102] in, This represents the mean vector of historical knowledge mastery. Represents the L2 norm;

[0103] When the difference between a user's current learning state and their historical stable state exceeds a preset difference threshold, the user profile is updated.

[0104] In the above embodiment, firstly, the user's historical learning behavior feature data is obtained, and then differentiated weights are assigned to learning behaviors in different periods through an exponential decay function, so that recent behaviors have a greater impact on the current state evaluation. Next, the weighted behavioral features are input into a deep neural network to generate an accurate user knowledge mastery vector. Finally, the difference between the current state and the historical stable state is calculated, and when it exceeds a preset threshold, a user profile update mechanism is triggered.

[0105] For example, let's take a Python course on a programming learning platform as an example:

[0106] The system continuously collects user learning behavior data over the past 30 days, including features such as video completion rate, code practice submission frequency, and error distribution. When a user suddenly increases their learning frequency for the "Data Structures" chapter between days 25 and 30, and the completion rate of related exercises significantly decreases, a time decay weighting mechanism assigns higher weights to these recent behaviors (e.g., the weight of data from the most recent day is 0.35, while the weight of data from 5 days ago drops to 0.08). The weighted aggregated feature vector is input into a deep neural network, and the L2 distance between the current user's knowledge mastery vector and the historical average knowledge mastery vector is calculated to be 0.78, exceeding the preset threshold of 0.65. The system immediately triggers a user profile update, identifying new weaknesses in the user's "Data Structures" knowledge module, and automatically adjusts the push strategy: pausing the originally planned advanced "Web Development" content, prioritizing targeted learning resources such as "Data Structures Fundamentals" and "Linked List and Tree Structure Analysis," adjusting the content difficulty to introductory level, and increasing the proportion of accompanying practice questions.

[0107] This design enables the system to keenly identify subtle changes in a user's learning status, avoiding the problem of traditional fixed-weight models being slow to react to recent changes in user behavior. This significantly improves the timeliness and accuracy of user profiles, ultimately achieving more precise and personalized delivery of educational resources.

[0108] The personalized user profile generation process of the user profile building module includes:

[0109] Basic attribute tags are generated based on the collected user basic information. The basic attribute tags include user age, education stage, professional field, learning goals and available time. For example, the basic attribute tags are "25 years old - working professional - computer major - goal to obtain Python engineer certificate - available to study for 2 hours per day".

[0110] By combining AI analysis of user learning behavior patterns, behavioral feature tags are generated. These behavioral feature tags include learning time period preferences, learning rhythm (e.g., "30 minutes of study followed by a 5-minute break" or "2 hours of continuous study followed by a break"), interaction preferences (e.g., "likes to participate in online Q&A" or "prefers to complete exercises independently"), and content selection preferences (e.g., "prioritizes practical courses" or "prefers learning materials that combine text and images").

[0111] Based on the user's knowledge gaps and demand prediction results mined by AI, knowledge demand tags are generated. These tags include the knowledge modules that the user currently needs to supplement, the types of skills that need to be improved, and the appropriate learning difficulty level.

[0112] Based on user learning outcome data (such as test scores and skills assessment results) and the learning ability scores output by the AI ​​model, user ability levels will be divided into basic, intermediate, proficient, and expert levels, and corresponding ability tags will be generated. For example, when a user scores 80 points (out of 100) in the English vocabulary test and 65 points in the listening test, the tag "English ability - vocabulary: proficient level, listening: intermediate level" will be generated.

[0113] The system integrates basic attribute tags, behavioral feature tags, knowledge need tags, and ability tags to generate personalized user profiles. When the cumulative update volume of user learning data reaches a preset threshold (such as adding more than 5 learning behavior records or more than 1 test score) or when the user's learning goals change, the personalized user profile is automatically updated to ensure that the user profile always remains consistent with the user's actual learning status.

[0114] In the above embodiments, basic attribute tags clearly define the core positioning of users, providing a basis for the initial screening of push content; behavioral characteristic tags accurately extract users' learning habits and preferences, making the push strategy more in line with the user's learning pace; knowledge need tags focus on users' knowledge weaknesses and potential needs, ensuring that the push content can directly address users' learning pain points; ability level tags quantify users' learning abilities to achieve precise matching of the difficulty of push content, avoiding a decline in learning enthusiasm caused by content that is too difficult or too easy. The integration and generation of multi-dimensional tags constructs a comprehensive and three-dimensional personalized user profile, allowing the system to understand user needs from multiple dimensions. The dynamic update mechanism ensures that the user profile can reflect changes in the user's learning status in real time, avoiding push deviations caused by outdated profiles, and ensuring that the push content is always highly adapted to the user's current learning status, ability level, and needs, effectively improving user learning efficiency and satisfaction.

[0115] The personalized push strategy construction process of the push strategy generation module includes:

[0116] Multi-dimensional content tags are used to label all content in the education and training resource library (including courses, exercises, materials, tools, etc.). The content tags include knowledge domain tags, difficulty tags, format tags, duration tags, applicable scenario tags, and applicable ability tags.

[0117] Construct a multi-dimensional matching model between user profiles and content tags, and calculate the overall matching degree between users and content in each tag dimension using the cosine similarity algorithm;

[0118] All candidate push content is initially ranked based on the overall matching degree. At the same time, the initial ranking results are adjusted in combination with the user's real-time learning scenario, learning progress and historical feedback (such as the user's previous completion rate of "case analysis" content reaching 80%, which is given priority). The final priority of the push content is then determined.

[0119] Based on the priority ranking of the push content, combined with the user's available learning time and push frequency rules, a complete personalized push strategy is generated, which includes a list of push content (such as "Pushing 'Python Exception Handling Micro-course' + 'Exception Handling Exercises' at 10:00 AM" and "Pushing 'Python Case Analysis Live Class Preview' at 6:00 PM") and push channels (such as prioritizing pushes through the mobile APP when the user usually studies on their mobile phone, while also sending SMS reminders).

[0120] In the above embodiments, the establishment of a multi-dimensional content tagging system enables refined classification and management of educational and training resources. Each piece of content possesses a clear attribute identifier, providing a foundation for accurate matching. The system calculates the degree of matching between users and content, ensuring the objectivity and accuracy of the matching results and avoiding biases caused by subjective judgment. The combination of initial sorting and secondary adjustments ensures the core principle of prioritizing matching degree. It fully considers personalized factors such as users' real-time learning scenarios, learning progress, and historical feedback, ensuring that the pushed content not only meets users' knowledge needs but also adapts to their current learning status and habits. The complete push strategy covers the content list and push channels, clearly defining the content users need to learn, ensuring that content can reach users in a timely and convenient manner, effectively improving the completeness of the push service and user experience, and enabling users to efficiently obtain learning resources tailored to their needs.

[0121] The push execution module pushes educational and training content to users through mobile applications, web platforms, email, and SMS channels. It collects user feedback on clicks, learning completion, evaluations, and adjustment requests for the pushed content, and transmits the feedback data to the AI ​​learning and analysis module in real time to provide data support for model optimization and push strategy adjustment.

[0122] The push process is monitored in real time, including push success rate monitoring, content delivery timeliness monitoring, and user reception status monitoring. When push anomalies are detected, the push channel is automatically switched or the push time is adjusted to ensure the stability and effectiveness of the push service.

[0123] In the above embodiments, the multi-terminal channel push method fully meets users' learning needs in different scenarios, allowing users to receive learning content anytime and anywhere, breaking the limitations of time and space, and significantly improving learning convenience; the real-time collection and transmission mechanism of feedback data can promptly transmit users' evaluation and needs of the pushed content to the AI ​​learning and analysis module, providing immediate feedback for model optimization and push strategy adjustment; the full-process push monitoring ensures the stability and reliability of the push service, and by monitoring the push success rate, delivery time and user reception status in real time, abnormal problems in the push process can be promptly discovered and resolved, avoiding a decline in user experience due to push failure or delay.

[0124] The push execution module also performs the following operations:

[0125] Get the type of push channel when the push content is pushed;

[0126] In this embodiment, the types of push channels include: mobile applications, web platforms, email, and SMS channels.

[0127] When the push channel type is a mobile application or a web platform, determine the preset push interface position in the operation interface;

[0128] In this embodiment, the user interface refers to the graphical user interface through which the user interacts with the system, including the main interface of the APP, the course list page, the personal center page, etc.

[0129] In this embodiment, the push interface position refers to the coordinate area of ​​the push content in the operation interface.

[0130] If the user is detected not to have clicked on the push content based on the push interface location, obtain the user's first behavior stream during the push period;

[0131] In this embodiment, "not clicked" means that the user does not click on the push content within a certain time window after the push is displayed (e.g., 5 minutes after the push).

[0132] In this embodiment, the push period is the time interval during which the push content is displayed;

[0133] In this embodiment, the first behavior flow is a sequence of interactive behaviors generated by the user during the push period. Each behavior record contains information such as behavior type, behavior object, behavior time, and behavior location.

[0134] Based on the first line of flow, determine the first attention distribution change of the user interface;

[0135] In this embodiment, attention distribution refers to the distribution of user attention on the operation interface. The first attention distribution change refers to the process of attention shifting and intensity changing between different areas of the interface during the push period, which can be represented as a time series heat map sequence or attention trajectory. When determining the first attention distribution change, the user's click, scroll, stay and other behaviors are converted into an approximate estimate of the attention distribution based on the attention inference model of interactive behavior.

[0136] The first attention distribution change features are extracted based on the changes in the first attention distribution. These features include peak attention, cumulative attention, and attention duration.

[0137] Analyze the characteristics of the first attention distribution change. If the difference between the upper and lower limits of the feature value corresponding to each feature type is less than the corresponding difference threshold, the neglect type is determined to be active neglect; otherwise, the neglect type is passive neglect.

[0138] In this embodiment, active ignoring indicates that the user is not interested in the pushed content or actively rejects it, which is characterized by the overall stability of the first attention distribution change and the user's choice not to click.

[0139] In this embodiment, passive ignoring means that the user is unaware of the push notification or is focused on other tasks on the interface (such as answering questions). Ignoring is passive and non-subjective, and is characterized by large fluctuations in the distribution of first attention, which leads the user to choose not to click.

[0140] In this embodiment, the difference between the upper and lower limits of the feature value refers to the fluctuation range of peak attention, cumulative attention, and attention duration within the observation window. When determining the neglect type, if the difference between the upper and lower limits of each feature value is less than the corresponding difference threshold, it is determined that the overall change in the first attention distribution is stable and the neglect type is determined to be active neglect; otherwise, the neglect type is determined to be passive neglect.

[0141] If the ignore type is "actively ignore", then no second push will be sent.

[0142] If the neglect type is passive neglect, the second behavior stream will continue to be acquired during the user's future online time period;

[0143] In this embodiment, when it is determined that the user is passively ignoring the message, considering the channel advantages of mobile applications or web platforms (compared to SMS or email channels, users can directly operate after receiving the push notification without any delay, and the push conversion rate is high), a suitable opportunity for a second push is sought in real time to obtain the second behavior stream; the second behavior stream refers to the new sequence of interactive behaviors generated by the user in the future online time period, with the same structure as the first behavior stream, and is used to continuously monitor changes in the user's attention state.

[0144] Determine the second attention distribution change based on the second behavior flow and extract the features of the second attention distribution change;

[0145] In this embodiment, the principle for obtaining the second attention distribution change is the same as that for the first attention distribution change; the principle for extracting the features of the second attention distribution change is the same as that for the features of the first attention distribution change.

[0146] If the difference between the upper and lower limits of the feature value corresponding to each feature type is less than the corresponding difference threshold, then a second push will be made immediately.

[0147] In this embodiment, if the difference between the upper and lower limits of the feature value corresponding to each feature type is less than the corresponding difference threshold, it indicates that the user's attention has changed from fluctuation to stability. The system determines that it is a suitable time to push the notification and immediately triggers a second push.

[0148] If the second attention distribution change feature satisfies the condition that there is a target region within the smallest local interface region containing the interface point corresponding to the peak attention feature value, then push the content again within the target region.

[0149] Among them, the lower limit of the peak attention feature value corresponding to the target area is greater than the preset lower limit threshold, and the target area can accommodate the push interface;

[0150] In this embodiment, when a user is continuously operating, they may not have time to pay attention to push notifications in fixed locations. To facilitate user viewing, a target area is detected in real time near the user's high attention area (the smallest local interface area). The smallest local interface area is the smallest continuous operation interface area that contains all attention peak interface points. The smallest local interface area changes dynamically with changes in user attention. The lower limit of the peak attention feature value corresponding to the target area is greater than a preset lower limit threshold (e.g., 60). After detecting the target area, a push notification is sent to the corresponding area. In this way, it will not delay the user's important operation, and it will also increase the probability that the push will be noticed. For example, when a user is answering a question on the operation interface, the main operation area is the rectangular answer box whose center coincides with the center point of the operation interface. The rectangular answer box is smaller than the operation interface. The preset push notification pops up from the lower left corner of the operation interface. The user needs to cross a large cursor distance to operate on the push notification. However, this invention can plan the target area to accommodate the push notification in real time based on the user's operation. The target area will not obscure the high attention area, and it is convenient to click. It can be operated by the user when there is redundant attention, and it will not force the push to obscure key information and cause user resentment.

[0151] In the above embodiments, the push execution module intelligently identifies the push channel type, accurately locates the display position of the push content for visual interface channels such as mobile applications and web platforms, and collects behavioral flow data when the user does not click, thereby analyzing the user's attention distribution changes in the interface. It distinguishes between active and passive neglect by judging the threshold of the feature value fluctuation range. For passive neglect, the system predicts the user's future online time and continuously monitors their behavioral flow. When it detects that the user's attention state has stabilized or there is suitable space near the high attention area, it intelligently triggers a secondary push and dynamically adjusts the push position. This invention breaks the limitation of traditional push systems that rely solely on explicit click feedback. By deeply analyzing the user's implicit behavioral characteristics, it accurately identifies the reasons for neglect, avoiding repeated disturbances to users who actively reject content, while effectively capturing valuable push opportunities that are being ignored. In particular, by dynamically adjusting the secondary push position to near the user's high attention area, it neither obscures key operation areas nor fails to improve the visibility and click-through rate of the push content, significantly improving push conversion efficiency and user experience, enabling educational content to accurately reach users.

[0152] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An intelligent online education and training push system based on AI adaptive learning, characterized in that: include: The data acquisition module is configured to collect multi-dimensional learning data from users during the education and training process, map numerical data in the multi-dimensional learning data to the [0,1] interval, convert text data into vector form, and perform numerical processing on categorical data. The AI ​​adaptive learning and analysis module is configured to achieve dynamic iterative learning through deep learning algorithms, and to mine user learning patterns, knowledge mastery status and potential needs based on the user's multi-dimensional learning data. The user profile building module is configured to combine the AI ​​analysis results output by the AI ​​adaptive learning analysis module to generate personalized user profiles. The personalized user profiles include user basic attributes, learning behavior characteristics, knowledge demand tendencies, and learning ability levels. The push strategy generation module is configured to build a multi-dimensional matching model based on user profiles and the content tag system of the education and training resource library, and generate personalized push strategies that adapt to users' real-time needs. The push execution module is configured to accurately deliver push content to users through multiple terminal channels and monitor the entire push process; it determines the type of content ignored based on user behavior flow and attention distribution characteristics, and intelligently adjusts the secondary push strategy and location. The push execution module also performs the following operations: Get the type of push channel when the push content is pushed; When the push channel type is a mobile application or a web platform, determine the preset push interface position in the operation interface; If the user is detected not to have clicked on the push content based on the push interface location, obtain the user's first behavior stream during the push period; Based on the first line of flow, determine the first attention distribution change of the user interface; The first attention distribution change features are extracted based on the changes in the first attention distribution. These features include peak attention, cumulative attention, and attention duration. Analyze the characteristics of the first attention distribution change. If the difference between the upper and lower limits of the feature value corresponding to each feature type is less than the corresponding difference threshold, the neglect type is determined to be active neglect; otherwise, the neglect type is passive neglect. If the ignore type is "actively ignore", then no second push will be made; If the neglect type is passive neglect, the second behavior stream will continue to be acquired during the user's future online time period; Determine the second attention distribution change based on the second behavior flow and extract the features of the second attention distribution change; If the difference between the upper and lower limits of the feature value corresponding to each feature type is less than the corresponding difference threshold, then a second push will be made immediately. If the second attention distribution change feature satisfies the condition that there is a target region within the smallest local interface region containing the interface point corresponding to the peak attention feature value, then push the content again within the target region. Among them, the lower limit of the peak attention feature value corresponding to the target area is greater than the preset lower limit threshold, and the target area can accommodate the push interface; The feedback optimization module is configured to collect user interaction feedback data on the pushed content and iteratively optimize the AI ​​learning model and push strategy in reverse. The content update module is configured to capture, filter, and update the content in the education and training resource library in real time. Based on industry development trends, knowledge update trends, and changes in user needs, it supplements new course content, teaching materials, and learning tools, and adds new content to the push content pool after tagging the new content. The access control module is configured to manage user permissions in a tiered manner, restricting the scope of access and usage rights of pushed content based on user identity, payment level, and learning permissions.

2. The intelligent online education and training push system based on AI adaptive learning as described in claim 1, characterized in that, The multi-dimensional learning data acquisition process of the data acquisition module includes basic information acquisition, real-time learning behavior acquisition, learning outcome data acquisition, and environmental preference data acquisition, specifically including: When collecting basic information, users’ age, gender, education level, professional field, occupation, learning goals and available learning time range are collected through user registration forms, identity verification, and actively filled questionnaires. When collecting real-time learning behavior data, the system uses data tracking technology, terminal sensors, and platform interaction logs to record user behavior data within the learning platform. This behavior data includes the number of course clicks, the duration of each lesson, video playback progress, courseware download and collection records, learning note editing content, online Q&A interaction times, group discussion participation frequency, and assignment submission time. When collecting learning outcome data, the system connects to the platform's testing system, homework correction system, and skills assessment tools to collect user unit test scores, mock exam scores, homework accuracy, error distribution, knowledge module mastery scores, and periodic learning outcome reports. When collecting environmental preference data, user learning preference data is collected by obtaining terminal device information and geographical location-related learning scenarios through user authorization, and by combining user selection records of course type, teaching style, content difficulty and learning resource format.

3. The intelligent online education and training push system based on AI adaptive learning as described in claim 1, characterized in that, The dynamic iterative learning process of the AI ​​adaptive learning analysis module includes: Construct a basic learning model based on deep neural networks. The input layer of the model is set to the user's multi-dimensional data vector, the hidden layer contains 3-5 fully connected layers, and the output layer is initially set to the user's knowledge need category and learning ability score. The real-time updated multi-dimensional learning data is input into the basic learning model, and the model parameters are continuously optimized through the backpropagation algorithm. The weight allocation of the user's recent learning behavior is increased to achieve real-time capture of changes in the user's learning status. By combining time series analysis algorithms to perform trend analysis on users' historical learning data, we can predict users' short-term and long-term learning needs. Among them, short-term demand forecasting focuses on the immediate needs of users at their current learning stage, while long-term demand forecasting is based on the needs of users' learning goals and plans for phased learning content. Set the model iteration cycle and adjust the model's loss function weights in real time based on user feedback data.

4. The intelligent online education and training push system based on AI adaptive learning as described in claim 3, characterized in that, Increase the weighting of users' recent learning behaviors to enable real-time capture of changes in users' learning status, including: Obtain the user's historical learning behavior feature vector; Calculate the time decay weighting coefficient; The user's historical learning behavior feature vector is weighted and aggregated to generate a weighted behavior feature vector; The weighted behavioral feature vector is input into a deep neural network model to calculate the user's knowledge mastery vector. Calculate the difference between the user's current learning state and the historical stable learning state based on the user's knowledge mastery vector and the historical knowledge mastery mean vector; When the difference between a user's current learning state and their historical stable state exceeds a preset difference threshold, the user profile is updated.

5. The intelligent online education and training push system based on AI adaptive learning as described in claim 1, characterized in that, The personalized user profile generation process of the user profile construction module includes: Basic attribute tags are generated based on the collected user basic information. The basic attribute tags include user age, education stage, professional field, learning goals, and available time. By combining AI analysis of user learning behavior patterns, behavioral feature tags are generated, including learning time period preferences, learning rhythm, interaction preferences, and content selection tendencies. Based on the user's knowledge gaps and demand prediction results mined by AI, knowledge demand tags are generated. These tags include the knowledge modules that the user currently needs to supplement, the types of skills that need to be improved, and the appropriate learning difficulty level. Based on user learning outcome data and the learning ability scores output by AI models, user ability levels will be divided into basic, intermediate, proficient, and expert levels, and corresponding ability tags will be generated.

6. The intelligent online education and training push system based on AI adaptive learning as described in claim 5, characterized in that, The user profile building module integrates basic attribute tags, behavioral feature tags, knowledge need tags, and ability tags to generate personalized user profiles. When the cumulative update volume of user learning data reaches a preset threshold or the user's learning goals change, the personalized user profile is automatically updated.

7. The intelligent online education and training push system based on AI adaptive learning as described in claim 1, characterized in that, The personalized push strategy construction process of the push strategy generation module includes: All content in the education and training resource library is labeled with multi-dimensional content tags, including knowledge domain tags, difficulty tags, format tags, duration tags, applicable scenario tags, and applicable ability tags. Construct a multi-dimensional matching model between user profiles and content tags, and calculate the overall matching degree between users and content in each tag dimension; All candidate push content is initially ranked based on comprehensive matching degree. At the same time, the initial ranking results are adjusted in a second step by taking into account the user's real-time learning scenario, learning progress and historical feedback to determine the final priority of push content. Based on the priority ranking of the pushed content, combined with the user's available learning time and push frequency rules, a complete personalized push strategy is generated, which includes a list of pushed content and push channels.

8. The intelligent online education and training push system based on AI adaptive learning as described in claim 1, characterized in that, The push execution module pushes educational and training content to users through mobile applications, web platforms, email, and SMS channels, collects user feedback on clicks, learning completion, evaluations, and adjustment requests, and transmits the feedback data to the AI ​​learning and analysis module in real time.

9. The intelligent online education and training push system based on AI adaptive learning as described in claim 8, characterized in that, The push execution module is also used to monitor the push process in real time. Push monitoring includes push success rate monitoring, content delivery timeliness monitoring, and user reception status monitoring. When a push abnormality is detected, the push channel is automatically switched or the push time is adjusted.

Images