Artificial intelligence driven organizational learning recommendation system

Abstract

The application belongs to the technical field of organization learning recommendation, and discloses an organization learning recommendation system based on artificial intelligence driving, which integrates data such as learning behaviors, interactive operations and terminal device scenes of organization members through a multi-source data acquisition module, extracts organization learning special features such as proportion of system specification courses and content staying time of organization development history through a feature engineering module, and provides comprehensive data support for a recommendation model. Meanwhile, a recommendation model construction module adopts a hybrid algorithm combining collaborative filtering and deep learning, and can also dynamically adjust parameters according to real-time behaviors of organization members. The mode of only relying on fixed rules for recommendation of the existing platform is changed, and content homogenization is avoided. In actual application, organization members can quickly locate learning resources suitable for their own needs, for example, organization members who prefer text learning will preferentially receive system analysis articles, and organization members who prefer video will obtain short films of organization development history education.

Description

Technical Field

[0001] This invention belongs to the field of organizational learning recommendation technology, specifically an artificial intelligence-driven organizational learning recommendation system. Background Technology

[0002] Digital transformation of organizational learning is a crucial direction for organizational learning innovation. Organizational learning platforms, as core carriers, have widely integrated diverse learning resources such as policy documents, organizational development history courses, and case studies. As the size of organizational members expands and their learning needs become increasingly diverse, the platform's ability to provide accurate, dynamic, and personalized recommendations is gradually becoming a core factor influencing learning efficiency and engagement. However, existing platforms face the following technical challenges in the area of ​​intelligent recommendation: Existing platforms only collect basic data such as learning duration and course completion rate, without integrating multi-source data such as interactive behavior and device scenarios, and lack the extraction of features specific to organizational learning; recommendation models mostly use fixed rules (such as sorting by publication time), without introducing algorithms such as collaborative filtering and deep learning, and cannot dynamically adjust parameters according to the behavior of organizational members, resulting in homogenized recommended content that cannot match the interests and learning styles of organizational members.

[0003] Recommendations are mostly pushed out in batches at set times, and cannot be triggered instantly based on the real-time actions of organizational members (such as just completing policy learning); at the same time, there is a lack of a sound feedback mechanism, only supporting simple evaluations, and not incorporating implicit feedback such as dwell time and secondary sharing into model optimization, resulting in the inability to continuously iterate the recommendation effect and insufficient participation of organizational members. The backend management only focuses on updating basic data, lacking real-time monitoring and quantitative evaluation of recommendation model performance (such as accuracy and response time); and it does not optimize recommendation logic (such as prioritizing content) in combination with the specific characteristics of organizational learning scenarios. The technology and business needs are disconnected, system optimization lacks data support, and operation and maintenance efficiency is low. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-driven organizational learning recommendation system to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: an artificial intelligence-driven tissue learning recommendation system, the system comprising: Multi-source data acquisition module: Collects learning data from organizational members, gathers information on their areas of interest and learning styles, and outputs raw data; Data preprocessing and feature engineering module: Upon receiving the raw data from the multi-source data acquisition module, the module first initiates a cleaning process on the raw data, then extracts three types of features, and outputs the preprocessed data and the extracted features. Recommendation model building module: Receives preprocessed data and extracted features, builds a fusion model, processes course image features and learning time series data, sets reward function, optimizes recommendation strategy, encapsulates the model as an API service, and outputs recommendation weights; Real-time recommendation module: Receives recommendation weights, generates recommendations, pushes popular and related recommendations in conjunction with a timed trigger mechanism, collects opinions from organization members, and after providing feedback on the timeliness of data labeling, sends it back to the data preprocessing and feature engineering module; System operation and maintenance management module: monitors the running status of the real-time recommendation module and the performance of the recommendation model construction module. When the indicators exceed the threshold, it notifies the operation and maintenance personnel and outputs the monitoring data to the recommendation effect evaluation and iteration module. Recommendation effect evaluation and iteration module: After receiving the monitoring data and feedback data, it regularly evaluates the core indicators and organization learning-specific indicators, and verifies the effect of the new strategy through comparative testing.

[0006] Preferably, the multi-source data acquisition module includes: The learning data includes learning behaviors and interaction behaviors; it is collected synchronously via API integration with organizational members' mobile terminals and organizational learning all-in-one machines, including scenario data such as device type, common learning time periods, and network environment; simultaneously, it collects organizational members' interest areas, learning styles, and learning goals through a learning preference questionnaire. The collected data is transmitted to the data gateway in JSON format, temporarily stored in a Redis cache cluster, and then pushed to the data preprocessing and feature engineering module to ensure the integrity and timeliness of the original data. Preferably, the data preprocessing and feature engineering module includes: Upon receiving the raw data output from the multi-source data acquisition module, the cleaning process specifically involves: filling missing learning duration values ​​with the historical average of the same user and course; removing outliers, including those with extremely short learning durations, using the 3σ principle; removing duplicate values ​​based on user ID, behavior type, and timestamp to avoid data redundancy caused by multiple records of the same operation; then performing Min-Max standardization on numerical data to map it to the [0, 1] interval to eliminate the influence of units; and performing Chinese word segmentation, stop word removal, and word vector conversion on text data to transform the text into numerical features that the model can recognize. The three types of features are behavioral features, temporal features, and organizational learning-specific features. Behavioral features include statistical features such as average daily learning time, interaction frequency, course completion rate, and the percentage of clicks on preferred areas. Temporal features, based on LSTM temporal processing technology, extract dynamic features including the distribution of learning periods over the past 7 days, the growth trend of learning progress, and the frequency of switching between interest areas. Organizational learning-specific features include features that align with the organizational learning scenario, such as the percentage of learning for policy and regulatory courses, the dwell time on content related to the organization's development history, and the number of clicks on courses related to organizational affairs exams. The processed features are stored in MySQL and MongoDB and then output to the recommendation model building module.

[0007] Preferably, the recommendation model construction module includes: Upon receiving the features processed by the data preprocessing and feature engineering modules, a multi-algorithm fusion model adapted to the organizational learning scenario is established. A collaborative filtering algorithm (user-item similarity calculation) is used to address the recommendation needs in the initial cold start phase. CNN is used to process course image features, while LSTM is used to process learning time-series data. Multimodal features are fused through a fully connected layer. A reinforcement learning algorithm is set up, using the organizational member learning completion rate and the priority of standardized content as reward functions to dynamically optimize the recommendation strategy and establish a recommendation model. After training on the training set and optimizing on the validation set, the recommendation model is encapsulated as an API service through TensorFlow Serving, and the recommendation weights are output to the real-time recommendation module. Formula for user-item collaborative filtering similarity: This represents the similarity between a user and an item, used to measure the similarity between the i-th user and the i-th item. The degree of matching for each learning content ranges from [-1, 1]. Indicates user The feature vector has a dimension of . ; This represents the value of user i on the k-th feature. The k-th element of the vector; Content The feature vectors of the learning content are the feature vectors of the learning content. Content In the The values ​​that can be taken on each feature are The first vector One element; Represents the vector dot product, i.e. , used for calculation and Linear correlation; Represents the vector magnitude, respectively corresponding to and Its function is to normalize the result of the vector dot product; This represents the number of feature dimensions, which is the total number of user and content features extracted by the system, and is fixed at 12.

[0008] Reinforcement learning reward function: In the formula: This represents the reward value of the recommended strategy at time t, used to measure... The effectiveness of the time-based recommendation strategy is measured in the range of [-1, 2]. This represents the weight of the learning completion rate, which is a fixed coefficient with a value of 0.4. The reward for completing the learning at time t is based on the organizational members' performance. Assign values ​​to the learning completion status of the recommended content in real time; This represents the priority weight of the specification, which is a fixed coefficient with a value of 0.5. This indicates that priority rewards are assigned based on the type of recommended content. This indicates the feedback adjustment weight, which is a fixed coefficient with a value of 0.1, and... , The sum is 1; express The system provides real-time feedback on adjustments, assigning values ​​based on a combination of explicit and implicit feedback from organizational members regarding the recommended content.

[0009] Preferably, the real-time recommendation module includes: The system receives recommendation weights from the recommendation model construction module. When an organizational member completes a course or logs into the platform, a real-time triggering mechanism is activated to call the recommendation model API and generate instant recommendations based on the current behavioral characteristics. Popular and related recommendations are pushed in conjunction with the organizational member's frequently used time periods. Simultaneously, opinions from organizational members are collected through explicit and implicit feedback. After the feedback data is labeled with timeliness, it is sent back to the data preprocessing and feature engineering module.

[0010] Preferably, the system operation and maintenance management module includes: The system monitors the running status of the real-time recommendation module and the performance of the recommendation model construction module. Administrators can adjust recommendation model parameters and recommendation strategy parameters through the web interface, and manage organizational learning resources and organizational member account permissions. The system also monitors model response time, server CPU utilization, and throughput of the data preprocessing and feature engineering modules in real time through the ECharts dashboard. When the metric exceeds the threshold, the system notifies the operations and maintenance personnel via SMS, records the alarm log, and has the function of rolling back to historical versions. Finally, the monitoring data is output to the recommendation effect evaluation and iteration module.

[0011] Preferably, the recommendation effect evaluation and iteration module includes: Based on the monitoring data from the system operation and maintenance management module and the feedback data from the real-time recommendation module, core indicators and organization-specific indicators are designed, and evaluation cycles are set on a daily, monthly, and quarterly basis. Monthly evaluations generate recommendation effect evaluation reports; A / B testing is used to verify the effectiveness of the new strategy. If the recommendation accuracy drops by ≥10%, the recommendation model will be retrained and the data preprocessing feature extraction logic will be optimized; new recommendation formats will be added based on feedback from organization members and then rolled out to all users after a phased rollout. Recommended accuracy formula: In the formula: This represents the recommendation accuracy rate, used to measure the degree of match between recommended content and the needs of organizational members, with a value range of [value missing]. ; The exact number of recommendations refers to the number of times that organization members have made effective interactions with the recommended content; The total number of recommendations refers to the total number of content items pushed to organization members by the real-time recommendation module.

[0012] The beneficial effects of this invention are as follows: 1. This invention integrates data on organizational members' learning behavior, interactive operations, and terminal device scenarios through a multi-source data acquisition module. Then, a feature engineering module extracts organizational learning-specific features such as the proportion of courses on regulations and standards, and the duration of time spent on content related to the organization's development history, providing comprehensive data support for the recommendation model. Simultaneously, the recommendation model construction module employs a hybrid algorithm combining collaborative filtering and deep learning, and can dynamically adjust parameters based on real-time organizational member behavior. For example, if organizational members frequently click on relevant regulatory content recently, the model automatically increases the recommendation weight for that area. This changes the existing platform's model of relying solely on fixed rules for recommendations, avoiding content homogenization. In practical applications, organizational members can quickly locate learning resources that suit their needs. For example, organizational members who prefer text-based learning will receive priority access to articles analyzing regulations, while those who prefer videos will receive educational short films on the organization's development history. The matching degree between recommended content and individual learning styles and areas of interest is significantly improved, reducing ineffective browsing time.

[0013] 2. This invention supports behavior-triggered instant push notifications through a real-time recommendation module. As soon as an organizational member finishes studying a core policy document report interpretation article, the system can immediately push related case analysis content, ensuring a continuous and uninterrupted learning process. Simultaneously, the feedback mechanism not only includes a traditional 5-point rating system but also incorporates implicit feedback such as organizational member dwell time, secondary sharing, and comment depth into data processing. This information, fed back to the model, further optimizes the recommendation strategy. This real-time response and closed-loop optimization make organizational members feel that the recommendations match their needs, enhancing their willingness to actively participate in learning interactions.

[0014] 3. This invention, through its operation and maintenance management module, allows operation and maintenance personnel to monitor the accuracy and response time of the recommendation model in real time, as well as the throughput of data collection and feature extraction. Performance bottlenecks can be located without manual troubleshooting; for example, if a recommendation response delay exceeds 500ms during a certain period, device or data transmission issues can be quickly traced through logs. Simultaneously, the effectiveness evaluation module establishes a quantitative system, including organizational learning-specific indicators such as the proportion of policy-related content recommendations and the improvement rate of organizational member learning efficiency. Monthly evaluation reports can be generated to guide model iteration by prioritizing the recommendation of policy-related and organizational development history-related content. Furthermore, organizational affairs administrators can adjust model parameters through the backend, such as the policy weight in the reinforcement learning reward function, to ensure the accuracy of recommended content. This design not only improves operation and maintenance efficiency but also ensures the system remains adaptable to organizational learning scenarios, avoiding a disconnect between technology and business needs. Attached Figure Description

[0015] Figure 1 This is a flowchart of the AI-driven tissue learning recommendation system of the present invention. Detailed Implementation

[0016] 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.

[0017] like Figure 1 As shown, this embodiment of the invention provides an artificial intelligence-driven tissue learning recommendation system, which includes: Multi-source data acquisition module: Embed data tracking points in the organization's learning app and web platform to collect learning data from organizational members, gather their areas of interest and learning styles, and output raw data; Data Preprocessing and Feature Engineering Module: Upon receiving the raw data from the multi-source data acquisition module, the module first initiates a cleaning process on the raw data, then extracts three types of features, and outputs the preprocessed data and the extracted features. Recommendation model building module: Receives preprocessed data and extracted features, builds a fusion model, processes course image features and learning time series data, sets reward function, optimizes recommendation strategy, encapsulates the model as an API service, and outputs recommendation weights; Real-time recommendation module: Receives recommendation weights, generates recommendations, pushes popular and related recommendations in conjunction with a timed trigger mechanism, collects opinions from organization members, and after providing feedback on the timeliness of data labeling, sends it back to the data preprocessing and feature engineering module; System operation and maintenance management module: Monitors the running status of the real-time recommendation module and the performance of the recommendation model building module. When the indicators exceed the threshold, it notifies the operation and maintenance personnel and outputs the monitoring data to the recommendation effect evaluation and iteration module. Recommended effect evaluation and iteration module: After receiving monitoring data and feedback data, regularly evaluate core indicators and organization learning-specific indicators, and verify the effect of new strategies through comparative tests.

[0018] The multi-source data acquisition module includes: Learning data includes learning behaviors and interaction behaviors; it is collected synchronously through API connections to the mobile terminals of organizational members and organizational learning all-in-one machines, including scenario data such as device type, common learning time periods, and network environment; at the same time, it collects organizational members' interest areas, learning styles, and learning goals through learning preference questionnaires. When organizational members make clicks, comments, or other actions, the data is transmitted to the data gateway in real time via HTTP / HTTPS protocol with a latency of ≤100ms. The collected data is transmitted to the data gateway in JSON format, temporarily stored in a Redis cache cluster, and a data collection log is established to record the collection time, source, and status of each data point. Finally, the raw data is pushed to the data preprocessing and feature engineering module to ensure the integrity and timeliness of the raw data. Learning behaviors include learning time, course progress, and pause / replay actions; interactive behaviors include comment content, sharing recipients, and like / favorite actions; areas of interest include system analysis, theoretical learning, case analysis, and education on the organization's development history; learning styles include video, audio, text, and live streaming; learning objectives include preparing for organizational affairs exams and improving theoretical knowledge.

[0019] The data preprocessing and feature engineering module includes: Upon receiving the raw data from the multi-source data acquisition module, the cleaning process is as follows: missing learning duration values ​​are filled with the historical average of the same user and the same course; outliers, including extremely short learning durations, are removed using the 3σ principle; duplicate values ​​are removed based on user ID, behavior type, and timestamp to avoid data redundancy caused by multiple records of the same operation; numerical data is then standardized using Min-Max to map it to the [0, 1] interval to eliminate the influence of units; and text data is processed using Chinese word segmentation, stop word removal, and word vector conversion to transform the text into numerical features that the model can recognize. The three categories of features are behavioral features, temporal features, and organizational learning-specific features. Behavioral features include statistical features such as average daily learning time, interaction frequency, course completion rate, and click percentage of preferred areas. Temporal features are based on LSTM temporal processing technology to extract dynamic features including the distribution of learning time periods over the past 7 days, the growth trend of learning progress, and the frequency of switching between interest areas. Organizational learning-specific features include features that fit the organizational learning scenario, such as the proportion of learning courses on rules and regulations, the dwell time on content related to the organization's development history, and the click volume of courses related to organizational affairs examinations. A hybrid storage architecture of MySQL and MongoDB is adopted: MySQL stores structured features, such as behavioral features and numerical features, while MongoDB stores unstructured features, such as text word vectors and time-series feature sequences.

[0020] The recommendation model building module includes: After receiving the features processed by the data preprocessing and feature engineering modules, a multi-algorithm fusion model adapted to organizational learning scenarios is established. A collaborative filtering algorithm (for calculating user-item similarity) is used to address basic recommendation needs in the initial cold start phase. A convolutional neural network (CNN) is used to process image features such as course cover images and video frames, while LSTM is used to process sequence features such as comment text and learning time-series data. Multimodal features are fused through fully connected layers to improve recommendation accuracy. A reinforcement learning algorithm is set up, using the improvement of organizational member learning efficiency and the accuracy of standardized content as reward functions. The Q-Learning algorithm is used to dynamically adjust the recommendation strategy to establish a recommendation model. After training on the training set (70% feature data) and fine-tuning on the validation set (20%), the recommendation model is encapsulated as an API service using TensorFlow Serving, outputting recommendation weights to the real-time recommendation module. Establish a feedback and parameter tuning process. When feedback from organizational members is below a threshold, the model parameters are automatically adjusted to ensure that the recommendation strategy adapts to dynamic needs. Collaborative filtering for users: Calculate the similarity of learning behaviors among members of an organization and recommend similar courses to members of similar organizations; Collaborative filtering of items: Calculate the correlation between clicks / completes of courses and recommend content that is "often studied together with the current course"; Formula for user-item collaborative filtering similarity: This represents the similarity between a user and an item, used to measure the similarity between the i-th user and the i-th item. The degree of matching between learning content and the value ranges from [-1, 1]. The closer the value is to 1, the higher the degree of matching between the two. It is used to calculate the recommendation weight. Indicates user The feature vector has a dimension of n, where n is the total number of user features extracted by the system. It includes 12 core features such as the average daily learning time of the data preprocessing and feature engineering module and the proportion of courses on rules and regulations. These are user structured features and are stored in a MySQL database. This represents the value of user i on the k-th feature. The k-th element of the vector, for example On behalf of users The standardized value of the average daily learning time, with a value range of [0, 1], is obtained by performing Min-Max standardization on the original learning time data by the data preprocessing and feature engineering module. Content Feature vectors, dimensions and To maintain consistency, the feature vectors belonging to the learning content are generated by combining the annotations of organizational learning-specific scenarios, such as labels for organizational development history and system analysis, and are also processed and output by the data preprocessing and feature engineering module. Content In the The values ​​that can be taken on each feature are The first vector Elements, for example Representative content The weight of policy-related tags is set to 1 if the content is policy-related, and 0.5 for other types (such as case analysis), with a value range of [0, 1], to ensure that it meets the priority requirements of the organization's learning scenario; Represents the vector dot product, i.e. , used for calculation and The linear correlation, the magnitude of which directly reflects the degree of overlap between the user and the core features of the content; Represents the vector magnitude, respectively corresponding to and Its function is to normalize the vector dot product result, avoid calculation deviations caused by differences in feature dimensions, and ensure the reasonableness of the similarity value; This indicates the feature dimension number, representing the total number of user and content features extracted by the system. It is fixed at 12 and includes behavioral features, time-series features, and organizational learning-specific features set by the data preprocessing and feature engineering modules.

[0021] Reinforcement learning reward function: In the formula: This represents the reward value of the recommended strategy at time t, used to measure... The effectiveness of the recommendation strategy at any given moment, with a value ranging from [-1, 2]. A higher value indicates a better recommendation strategy. This represents the learning completion rate weight, a fixed coefficient with a value of 0.4. The purpose of setting this weight in the recommendation model building module is to adjust the contribution of the learning completion rate to the reward value, prioritizing recommended content that encourages organizational members to complete learning, such as watching a complete video or reading 80% of an article. It is used as a model hyperparameter. Adjustments can be made through the web backend of the system operation and maintenance management module; The reward for completing the learning at time t is based on the organizational members' performance. The completion status of the recommended content is assigned as follows: 1 if the member completes the learning, 0 if not, and 0.5 if they drop out. This data comes from the course completion tracking logs recorded by the real-time recommendation module. The priority weight is a fixed coefficient of 0.5. This weight design reflects the core principle of prioritizing normative content in organizational learning scenarios, ensuring that content such as the latest system analysis and education on the organization's development history has a higher priority in the recommendation strategy. This indicates the priority reward for timely updates, assigned based on the type of recommended content: 1 for latest policy analysis and organizational development history education content, 0.6 for case studies and skills training content, and 0.3 for other types. Content type tags are derived from the organization learning-specific feature annotations in the data preprocessing and feature engineering module. This indicates the feedback adjustment weight, which is a fixed coefficient with a value of 0.1, and... , The sum is 1; its function is to integrate the immediate feedback from organizational members and balance the impact of learning completion rate and the subjective opinions of organizational members on reward value. express The feedback adjustment item is assigned a value based on a combination of explicit and implicit feedback from organizational members to the recommended content: for explicit feedback, 5 points are assigned 1, 4 points are assigned 0.8, 3 points are assigned 0.5, and 2 points and below are assigned -1; when there is no explicit feedback, the value is assigned based on implicit behavior, with sharing assigned 0.3, commenting assigned 0.2, and closing directly assigned -0.5. The feedback data comes from the real-time recommendation module.

[0022] The real-time recommendation module includes: The system receives recommendation weights from the recommendation model building module. When an organizational member performs a specific action, such as completing a course, posting a comment, or logging into the platform, it calls the recommendation model API to generate instant recommendations based on the current behavioral characteristics. For example, after completing a specialized development standard course, it recommends case analysis videos; if the member prefers text-based learning, it recommends policy articles. It also pushes popular and related recommendations based on the time periods frequently used by organizational members. Hot Recommendations: Based on the course clicks, completion rates, and positive review rates of all members across the platform in the past 7 days, a weighted ranking (clicks weighted at 0.4, completion rate at 0.3, and positive review rate at 0.3) is used to generate a hot list, which is then recommended to newly registered members or members with unclear interests. Related Recommendations: Based on the item relevance of the collaborative filtering algorithm, the "You May Also Like" section is displayed on the course playback page and article reading page. For example, after learning about the interpretation of the core institutional documents and reports of the learning organization, the audio of the keyword analysis of the core institutional documents and reports is recommended. Simultaneously, opinions from organizational members are collected through explicit feedback (5-point rating, text evaluation) and implicit feedback (stay time, sharing action). After the feedback data is labeled with timeliness, it is sent back to the data preprocessing and feature engineering module. At the same time, the system operation and maintenance management module is linked to recommend response time, click rate and other data.

[0023] The system operation and maintenance management module includes: The system monitors the operational status of the real-time recommendation module and the performance of the recommendation model building module. Administrators can adjust recommendation model parameters (such as collaborative filtering similarity threshold and reinforcement learning reward coefficient) and recommendation strategy parameters (such as the update cycle of the popular recommendation list and the accuracy matching threshold) through the web interface. They can also manage organizational learning resources and the account permissions of organizational members. The system uses the ECharts dashboard to monitor model response time, recommendation accuracy, recall, server CPU utilization, and the throughput of the data preprocessing and feature engineering modules in real time, such as the amount of data collected per second and the number of features extracted per second. When a metric exceeds a threshold, an SMS alert mechanism is used to notify operations and maintenance personnel, and an alert log is recorded, including the alert time, metric name, current value, and threshold. The log also has the function of rolling back to historical versions. Finally, the monitoring data is output to the recommendation effect evaluation and iteration module.

[0024] The recommendation effectiveness evaluation and iteration module includes: Based on monitoring data from the system operation and maintenance management module and feedback data from the real-time recommendation module, core indicators and organization learning-specific indicators are designed. Evaluation cycles are set on a daily basis (monitoring of core indicators), a monthly basis (comprehensive effect evaluation), and a quarterly basis (system iteration planning). The monthly evaluation generates a "Recommendation Effect Evaluation Report", which includes indicator data, trend analysis, and problem identification. A / B testing is used to verify the effectiveness of the new strategy. Key metrics include recommendation accuracy, recall, click-through rate, and learning completion rate. The specific indicators for organizational learning include the percentage of policy-related content recommended, the rate of improvement in organizational members' learning efficiency, and the indicators of organizational members' sense of participation. A / B testing: Randomly divide organizational members into experimental and control groups. The experimental group uses the new strategy, and the control group uses the old strategy. If the recommendation accuracy drops by ≥10%, the recommendation model will be retrained and the data preprocessing feature extraction logic will be optimized; new recommendation formats will be added based on feedback from organizational members and then rolled out to all users after a phased rollout. Recommended accuracy formula: In the formula: This represents the recommendation accuracy rate, used to measure the degree of match between recommended content and the needs of organizational members, with a value range of [value missing]. Based on the recommendation effectiveness evaluation requirements of the recommendation effectiveness evaluation and iteration module, the system sets the target threshold as follows: This metric is mainly used for daily monitoring and monthly comprehensive evaluation of the recommendation effect evaluation and iteration module. The accurate recommendation count refers to the number of times that organizational members have made effective interactions with the recommended content. Effective interaction must meet the following criteria: "After the recommended content is displayed, organizational members click to enter the learning page within 24 hours and stay on it for a certain period of time." The condition of "seconds" is based on data from the organization member behavior logs recorded by the real-time recommendation module, which ultimately serves as the molecule input for the recommendation effect evaluation and iteration module to calculate the recommendation accuracy. The total number of recommendations refers to the total number of content items pushed to organization members by the real-time recommendation module, covering the sum of the three types of pushes: precise recommendations, popular recommendations, and related recommendations. When calculating, recommendation data from test accounts and invalid accounts that have not logged in for 30 consecutive days must be excluded. This type of data filtering is completed by the system operation and maintenance management module.

[0025] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0026] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-driven tissue learning recommendation system, characterized in that: The system includes: Multi-source data acquisition module: Collects learning data from organizational members, gathers information on their areas of interest and learning styles, and outputs raw data; Data preprocessing and feature engineering module: Upon receiving the raw data from the multi-source data acquisition module, the module first initiates a cleaning process on the raw data, then extracts three types of features, and outputs the preprocessed data and the extracted features. Recommendation model building module: Receives preprocessed data and extracted features, builds a fusion model, processes course image features and learning time series data, sets reward function, optimizes recommendation strategy, encapsulates the model as an API service, and outputs recommendation weights; Real-time recommendation module: Receives recommendation weights, generates recommendations, pushes popular and related recommendations in conjunction with a timed trigger mechanism, collects opinions from organization members, and after providing feedback on the timeliness of data labeling, sends it back to the data preprocessing and feature engineering module; System operation and maintenance management module: monitors the running status of the real-time recommendation module and the performance of the recommendation model construction module. When the indicators exceed the threshold, it notifies the operation and maintenance personnel and outputs the monitoring data to the recommendation effect evaluation and iteration module. Recommendation effect evaluation and iteration module: After receiving the monitoring data and feedback data, it regularly evaluates the core indicators and organization learning-specific indicators, and verifies the effect of the new strategy through comparative testing.

2. The AI-driven tissue learning recommendation system according to claim 1, characterized in that: The multi-source data acquisition module includes: The learning data includes learning behaviors and interaction behaviors; it is collected synchronously through API connections to the mobile terminals of organizational members and the organizational learning all-in-one machine, including scenario data such as device type, common learning time periods and network environment; at the same time, it collects the organizational members' interest areas, learning styles and learning goals through learning preference questionnaires. The collected data is transmitted to the data gateway in JSON format, and after being temporarily stored in the Redis cache cluster, the raw data is pushed to the data preprocessing and feature engineering module.

3. The AI-driven tissue learning recommendation system according to claim 1, characterized in that: The data preprocessing and feature engineering module includes: Upon receiving the raw data output from the multi-source data acquisition module, the cleaning process specifically involves: filling missing learning duration values ​​with the historical average of the same user and the same course; removing outliers, including those with extremely short learning durations, using the 3σ principle; and removing duplicate values ​​based on user ID, behavior type, and timestamp. Then, the numerical data undergoes Min-Max standardization, and the text data is processed using Chinese word segmentation, stop word removal, and word vector conversion to transform the text into numerical features recognizable by the model. The three types of features are behavioral features, temporal features, and organizational learning-specific features. Behavioral features include statistical features such as average daily learning time, interaction frequency, course completion rate, and the percentage of clicks on preferred areas. Temporal features, based on LSTM temporal processing technology, extract dynamic features including the distribution of learning periods over the past 7 days, the growth trend of learning progress, and the frequency of switching between interest areas. Organizational learning-specific features include features that align with the organizational learning scenario, such as the percentage of learning for policy and regulatory courses, the dwell time on content related to the organization's development history, and the number of clicks on courses related to organizational affairs exams. The processed features are stored in MySQL and MongoDB and then output to the recommendation model building module.

4. The AI-driven tissue learning recommendation system according to claim 1, characterized in that: The recommendation model construction module includes: Upon receiving the features processed by the data preprocessing and feature engineering module, a multi-algorithm fusion model adapted to the organizational learning scenario is established; CNN is used to process course image features, while LSTM is used to process learning time series data, and multimodal features are fused through a fully connected layer; a reinforcement learning algorithm is set up, with the improvement of organizational members' learning efficiency and the accuracy of standardized content as reward functions, to dynamically optimize the recommendation strategy and establish a recommendation model; After training on the training set and optimizing on the validation set, the recommendation model is encapsulated as an API service through TensorFlow Serving, outputting recommendation weights to the real-time recommendation module.

5. The AI-driven tissue learning recommendation system according to claim 1, characterized in that: The real-time recommendation module includes: The system receives recommendation weights from the recommendation model construction module. When an organizational member completes a course or logs into the platform, a real-time triggering mechanism is activated to call the recommendation model API and generate instant recommendations based on the current behavioral characteristics. Popular and related recommendations are pushed in conjunction with the organizational member's frequently used time periods. Simultaneously, opinions from organizational members are collected through explicit and implicit feedback. After the feedback data is labeled with timeliness, it is sent back to the data preprocessing and feature engineering module.

6. The AI-driven tissue learning recommendation system according to claim 1, characterized in that: The system operation and maintenance management module includes: The system monitors the running status of the real-time recommendation module and the performance of the recommendation model construction module. Administrators can adjust recommendation model parameters and recommendation strategy parameters through the web interface, and manage organizational learning resources and organizational member account permissions. The system also monitors model response time, server CPU utilization, and throughput of the data preprocessing and feature engineering modules in real time through the ECharts dashboard. When the metric exceeds the threshold, the system notifies the operations and maintenance personnel via SMS, records the alarm log, and has the function of rolling back to historical versions. Finally, the monitoring data is output to the recommendation effect evaluation and iteration module.

7. The AI-driven tissue learning recommendation system according to claim 1, characterized in that: The recommendation effectiveness evaluation and iteration module includes: Based on the monitoring data from the system operation and maintenance management module and the feedback data from the real-time recommendation module, core indicators and organization-specific indicators are designed, and evaluation cycles are set on a daily, monthly, and quarterly basis. Monthly evaluations generate recommendation effect evaluation reports; A / B testing is used to verify the effectiveness of the new strategy. If the recommendation accuracy drops by ≥10%, the recommendation model will be retrained and the data preprocessing feature extraction logic will be optimized; new recommendation formats will be added based on feedback from organization members and then rolled out to the full network after a phased rollout.

Images