Accurate correction implementation method for user log-driven interactive content recommendation system

Abstract

The invention relates to a precise correction implementation method for a user log-driven interactive content recommendation system, relates to the technical field of information recommendation, and aims to solve the problem that the performance and recommendation quality of a traditional recommendation system in a dynamic environment are limited. The method comprises the following steps of capturing user operation behavior data in real time, extracting user behavior characteristics and generating a user portrait; and when the recommendation system detects that the quality is reduced, a precise correction module is triggered, and a large language model (LLM) is called to assist decision making. The behavior logs are analyzed, the problem of data sparsity is solved, and recommendation logic is dynamically adjusted. Potential interest points or current situation demands of the user are identified, and a more accurate recommendation result is generated. According to the method, cold start challenges when new users or new projects lack enough data are solved, the system is allowed to continuously optimize recommendation strategies according to latest user feedback, user interest changes are quickly responded, the problem of data sparsity is solved, recommendation logic is dynamically adjusted, and recommendation accuracy and system response speed are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of information recommendation, and in particular to a precise correction implementation method for a user log-based interactive content recommendation system. Background Art

[0002] Traditional interactive recommendation systems play an important role in the field of modern personalized recommendation. These systems generate recommendation results by analyzing user behavior data (such as click records, ratings, purchase history, etc.). Typical recommendation techniques include collaborative filtering, content-based recommendation, hybrid recommendation and other methods. Collaborative filtering methods rely on the similarity of user behavior to recommend items, while content-based methods recommend items based on the matching of item features with user interests. Currently, there are many interactive content recommendation systems on the market, which mainly rely on user behavior data to generate personalized recommendations. The following are several typical existing technologies and their shortcomings: (1) It is difficult to respond quickly to changes in user interests: Recommendation systems based on collaborative filtering rely on the similarity of user behaviors to recommend items. Such methods are difficult to quickly capture changes in user interests because they mainly rely on historical data for prediction. When a user's preference changes, it may take some time for the system to reflect this change; (2) Data sparsity problem: For new users or new items, recommendation systems based on collaborative filtering may have difficulty making accurate recommendations due to the lack of sufficient historical data. Matrix decomposition also faces data sparsity and cold start problems, and its recommendation logic is also static and cannot adapt well to the dynamic changes in user interests; (3) Fixed recommendation logic: Content-based recommendation systems make recommendations based on the matching of project features with user interests. Content matching algorithms are usually based on fixed rules and features, which are difficult to adapt to changing user needs and interests. Hybrid recommendation systems combine collaborative filtering and content-based recommendation methods. Even if the two methods are combined, the recommendation logic of the hybrid system is still static to a certain extent, and it is difficult to flexibly respond to the diversification of user needs. The recommendation logic of collaborative filtering is based on the similarity between users, which is static to a certain extent and difficult to flexibly adjust to cope with complex user needs and changes in interests; (4) Concentration of recommendation results: Content-based recommendation systems often rely too much on the characteristics of items, resulting in recommendation results that may be concentrated on certain popular or mainstream content, while ignoring some potentially unpopular or long-tail content. Many hybrid systems fail to fully utilize real-time user feedback to optimize recommendation strategies, resulting in slow optimization of recommendation results. Although deep learning can capture deeper patterns, training such models often requires a lot of computing resources, and the model itself may also have poor interpretability. In addition, such systems may also pay too much attention to data points that appear frequently, resulting in the concentration of recommendation results; These defects limit the performance and recommendation quality of traditional recommendation systems in dynamic environments, and new technical methods are urgently needed to improve the flexibility and recommendation effect of the system. Therefore, a precise correction method for an interactive content recommendation system based on user logs is proposed. This method gives full play to the multi-step decision-making advantages of the interactive recommendation system, and combines the deep interpretation of recent user behavior logs by the large language model (LLM) to assist in decision optimization. By monitoring and capturing these changes, the present invention can quickly capture the dynamic changes of user interests, effectively solve the problem of data sparsity, promote the dynamic adjustment feedback loop of the traditional interactive recommendation system, and significantly improve the accuracy of the recommendation system and user experience. Summary of the invention

[0003] In view of the shortcomings of the prior art, the present invention proposes a precise correction method for an interactive content recommendation system based on user logs. This method gives full play to the multi-step decision-making advantages of the interactive recommendation system, and combines the deep interpretation of recent user behavior logs by the large language model (LLM) to assist in decision optimization. By monitoring and capturing these changes, the dynamic changes of user interests can be quickly captured, the data sparsity problem can be effectively solved, and the dynamic adjustment feedback loop of the traditional interactive recommendation system can be promoted, which significantly improves the accuracy of the recommendation system and the user experience.

[0004] To achieve the above objectives, the present invention is implemented through the following technical solutions: A precise correction method for an interactive content recommendation system based on user logs, based on LLM technology, includes the following steps: S1: Through the log collection module integrated in the content recommendation system, the system captures the user's operation behavior data in real time. These data include click streams, content browsing paths, page dwell time, search keywords, etc. The collected user log data will be processed through denoising, cleaning and standardization to ensure the accuracy and consistency of the data. This step includes deleting duplicate records, filtering outliers, and formatting the data in a unified manner; S2: Extract user behavior features from the processed log data and generate user portraits. This process uses a variety of machine learning algorithms, cluster analysis (K-means), collaborative filtering, etc. to build a user interest preference model. In order to cope with the dynamic changes in user interests, user portraits need to be updated periodically or in real time when significant deviations in user behavior are detected. The update process achieves refined adjustment of the portrait by fusing new log data with historical data; S3: When the recommendation system detects a decline in recommendation quality, abnormal changes in user behavior, or data sparsity problems, the precision correction module is triggered. The system determines whether to perform correction operations based on set thresholds or rules. During the recommendation correction process, the system calls the LLM decision-making assistance module and uses its powerful semantic understanding capabilities to analyze the user's recent behavior logs. LLM can identify users' potential interests or current situational needs, thereby assisting traditional algorithm decisions and generating more accurate recommendation results. After LLM correction, the system combines the output of traditional recommendation algorithms to generate an optimized recommendation list. Subsequently, the system records users' responses to these recommendation results, click-through rates, interaction depth, etc., to further optimize the recommendation strategy; S4: Based on the user's recent behavior patterns, LLM can predict the areas or activities that the user may be interested in in the future. This process involves sequence prediction tasks in natural language processing. LLM uses its trained parameters to estimate the next event or preference change that is most likely to occur. When it comes to multimedia content, LLM is not limited to text analysis, but can also combine data from multiple media forms such as images and videos to make comprehensive judgments; S5: The system continuously optimizes the recommendation algorithm based on user feedback data, including adjusting the trigger conditions of the LLM module, optimizing the feature extraction method, and improving the user portrait modeling. The overall performance of the recommendation system is regularly evaluated, and key indicators include recommendation accuracy, user satisfaction, system response time, etc. Based on these evaluation results, the system further adjusts and optimizes the correction method to ensure the continuous improvement of recommendation quality.

[0005] Preferably, in said S1, it also includes: The system first captures the user's operational behavior data in real time through the integrated log collection module. These data include but are not limited to click streams, browsing paths, page dwell time, search keywords, etc. In order to ensure the accuracy and consistency of the data, the original log data needs to be denoised, cleaned, and standardized. For the text information in the log (search queries, comments, or social media activities), natural language processing (NLP) technology will be used for pre-processing such as word segmentation, stop word removal, and word form restoration, and the text will be converted into a vector representation (through TF-IDF, Word2Vec, or BERT embedding). This step is to prepare for subsequent semantic analysis.

[0006] Preferably, in said S2, further comprising: Utilize advanced feature engineering techniques, time series analysis, social network analysis, etc. to extract more valuable user behavior features for building more detailed user portraits. Implement a real-time update mechanism for user portraits to automatically adjust user portraits by monitoring changes in user behavior to ensure the timeliness and accuracy of user portraits. Combine user behavior data with other external data sources (social media activities, geographic location information, etc.) to build a more three-dimensional and comprehensive user portrait.

[0007] Preferably, in said S3, further comprising: Based on the user's historical behavior and preferences, personalized correction trigger thresholds are set for different user groups to more accurately capture changes in user interests; content analysis of multiple media forms such as text, images, and videos is combined, and a multimodal large language model (InternVL series model) is used to more comprehensively understand user needs and interests; a context-aware mechanism is introduced to consider the user's current context to optimize the recommendation results, making the recommendation more in line with the user's immediate needs.

[0008] Preferably, in said S4, further comprising: Dynamically update the user interest model based on the user interests and their changing trends identified by LLM. Adjust the weights of interests, add new interests, or delete interests that are no longer relevant. Adjust the parameters and strategies of the recommendation algorithm based on the updated user interest model. Increase the recommendation weight for the content that the user is currently most interested in, and reduce the recommendation for content that the user has lost interest in. Record the user's response to the recommendation results, such as click-through rate, interaction depth, etc., and feed this data back to the LLM model to further optimize the accuracy of the model.

[0009] Compared with the prior art, the present invention discloses a user log driven interactive content recommendation system accurate correction implementation method. The present invention has the following beneficial effects: 1. Quickly respond to changes in user interests: By integrating the Large Language Model (LLM) to deeply interpret recent user behavior logs, capture and analyze user behavior logs in real time, it can quickly identify and respond to changes in user interests and provide recommendations that better meet the user's current needs; 2. Solve the problem of data sparsity: Even when there is insufficient data for new users or new projects, LLM can provide valuable insights based on a small amount of behavioral logs to help generate meaningful recommendations. For new users who have just started using the platform or new products that have just been launched, it can more effectively overcome the cold start challenge and provide a more personalized first experience, increasing user satisfaction from 70% to 88%; 3. Dynamically adjust the recommendation logic: Allow the system to continuously optimize the recommendation strategy based on the latest user feedback. By setting the personalized correction trigger threshold, the system can more accurately capture changes in user interests and make corresponding adjustments in a timely manner. In the face of complex and changing user needs, it has stronger adaptability and can flexibly respond to recommendation requirements in various situations; 4. Diversified recommendation results: By combining user behavior and other external data sources, we ensure that recommendations are not limited to mainstream projects, but also cover niche or unpopular content that may suit the preferences of specific user groups. By encouraging diverse recommendations, different types of creators and service providers can gain more attention and development opportunities, which promotes the healthy development of the entire ecosystem; 5. Intelligence and accuracy: The system integrates intelligent benchmark models, anomaly detection algorithm matrix, and machine learning models, enabling the system to automatically analyze processed data, mine performance trends and potential anomalies, and achieve accurate predictions. The recommendation accuracy rate has increased from 65% to 82%, and the system response time has been shortened from an average of 300 milliseconds to 150 milliseconds. 6. Real-time and high efficiency: By deploying Apache Kafka cluster and Apache Flink, efficient processing of real-time data streams is achieved to ensure low latency of data processing. The system can respond to performance changes instantly and quickly capture and process performance data. A dynamic optimization strategy engine is designed to dynamically adjust website resources, resource loading priority, image compression algorithm and cache management strategy according to real-time monitoring data to ensure that the website can maintain optimal performance in different network environments and user behaviors and improve resource utilization efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those skilled in the art, other relevant drawings can be obtained based on these drawings without creative work. Figure 1 :User behavior data collection flow chart, describing the process of collecting user behavior data from data source to data storage; Figure 2 : LLM optimization flow chart, showing the steps of interest dynamic identification, data sparsity processing, content generation and vectorization processing; Figure 3 : The recommendation system architecture diagram shows the main components of the recommendation system and their interactions, including data storage, LLM interface, vectorized processing module, recommendation logic optimization module, etc. DETAILED DESCRIPTION

[0011] Step 1: Data collection and preprocessing Through the integrated log collection module, S11 can capture the user's operation behavior data in real time, including click flow, content browsing path, page dwell time, search keywords, etc. S12 preprocesses the collected data. This includes operations such as data cleaning, outlier filtering, and format standardization. The data preprocessing algorithm is as follows: def preprocess_data(user_logs): # Delete duplicate records user_logs = remove_duplicates(user_logs) # Filter outliers user_logs = filter_outliers(user_logs) # Unified format processing user_logs = format_unification(user_logs) return user_logs In order to process user behavior data in high-concurrency scenarios, S13 uses the real-time stream processing framework Apache Kafka and Apache Flink to ensure real-time and low-latency data processing. Build an Apache Kafka cluster, including Zookeeper and multiple Kafka Brokers; configure Kafka's Topic, implement data producers, and use Kafka's Producer API to ensure high data throughput and low latency; implement data consumers, which pull data from Kafka's Topic in real time and perform subsequent processing; S14 uses the data preprocessing API of TensorFlow.js to clean the original log data, including removing useless information and filling missing values. Use TensorFlow.js to build a deep learning model for extracting user behavior features and building user portraits, and deploy the trained model to the server or client for real-time user behavior analysis; S15 uses Redis or in-memory database to store abnormal events in real-time data streams; uses Redis's aggregation function to implement data aggregation logic to quickly identify abnormal patterns in data; for each user behavior event, uses a unique identifier (a combination of user ID and event ID) as a member of the Set to automatically remove duplicate events; S16 integrates modules such as Kafka, TensorFlow.js and Redis into the user interest dynamic identification system to ensure data flow and collaboration between modules. The system is fully tested, including functional testing, performance testing and stability testing, to ensure the reliability and accuracy of the system.

[0012] Step 2: User behavior modeling S21 extracts user behavior features from the processed log data, such as click frequency, browsing time, TF-IDF value of search keywords, etc. It uses machine learning algorithms such as cluster analysis and collaborative filtering to extract user behavior features and build user portraits. The feature extraction algorithm is as follows: def extract_features(user_logs): # Extract user behavior features features = extract_user_behavior_features(user_logs) # Build user profile user_profile = build_user_profile(features) return user_profile S22 uses advanced feature engineering techniques such as time series analysis and social network analysis to extract more valuable user behavior features, and combines user behavior data with other external data sources (social media activities, geographic location information, etc.) to build a more three-dimensional and comprehensive user portrait.

[0013] Step 3: Accurate calibration of recommendation algorithm S31 When the accuracy of the recommendation system falls below a certain threshold or an abnormal change in user behavior is detected, we will trigger the precision correction mechanism; The S32 system will call the LLM decision-making support module and use its powerful semantic understanding ability to analyze the user's recent behavior log; S33 selects a pre-trained LLM, BERT, GPT, etc. as the basic model. Collect and annotate a large amount of user behavior log data as training corpus; fine-tune the basic model on the user behavior log data to enable it to capture subtle changes in user interests; The LLM fine-tuning algorithm is as follows: def fine_tune_llm(pretrained_llm, user_logs): # Prepare training data training_data = prepare_training_data(user_logs) # Fine-tune LLM fine_tuned_llm = pretrained_llm.fine_tune(training_data) return fine_tuned_llm.

[0014] Step 4: Dynamic identification of user interests S41 inputs the pre-processed user behavior log into the trained LLM. LLM uses its powerful semantic understanding and context perception capabilities to deeply analyze user behavior; S42 LLM extracts users’ potential points of interest by analyzing keywords, phrases, and contextual relationships in user behavior logs. These points of interest may be presented in the form of concepts, themes, or specific items; S43 By comparing the user's behavior logs and interest points in different time periods, LLM can detect the changing trend of user interests. This includes the rise of new interests, the decline of old interests, and the transition between interests; The user interest identification algorithm is as follows: def identify_user_interest(fine_tuned_llm, user_logs): # Use LLM to analyze user logs interest_signals = fine_tuned_llm.analyze(user_logs) # Extract points of interest interests = extract_interests(interest_signals) return interests.

[0015] Step 5: Overall optimization and iteration of the method The S51 system will continue to optimize the recommendation algorithm based on user feedback data, including adjusting the trigger conditions of the LLM module, optimizing the feature extraction method, and improving the user portrait modeling; S52 introduces reinforcement learning technology to optimize the recommendation strategy, and gradually improves the accuracy of the recommendation system and user satisfaction through continuous trial and error learning; S53 regularly evaluates the overall performance of the recommendation system, with key indicators including recommendation accuracy, user satisfaction, system response time, etc. S54 uses A / B testing and rigorous experimental design methods to evaluate the performance improvements of the recommendation system, ensuring that each iteration brings substantial improvements; S55 inputs user feedback on the newly recommended content (click rate, dwell time, rating, etc.) as new behavior log data into LLM for the next round of interest identification and recommendation optimization, forming a closed loop of continuous iteration and continuous improvement.

Claims

1. A user log driven interactive content recommendation system precision correction implementation method, characterized in that: The method comprises the following steps: S1: collect and process user operation behavior data in real time, the data at least including click flow, browsing path, page dwell time and search keywords; S2: Extract user behavior features from the processed log data to generate user profiles, which are updated periodically or in real time when significant deviations are detected; S3: When the quality of recommendations decreases, user behavior changes abnormally, or data sparsity occurs, the precision correction module is triggered to use the large language model (LLM) to assist decision-making and generate more accurate recommendation results; S4: Predict the areas or activities that users may be interested in in the future based on their recent behavior patterns, and dynamically adjust the user interest model; S5: Continue to optimize the recommendation algorithm, adjust the triggering conditions of the LLM module, optimize the feature extraction method, and improve the user portrait modeling. Regularly evaluate the performance of the recommendation system and iterate and optimize.

2. The method according to claim 1, characterized in that The S1 further comprises: De-noising, cleaning and standardization of raw log data to ensure data accuracy and consistency; Natural language processing technology is used to preprocess the text information and convert the text into a vector representation.

3. The method according to claim 1, characterized in that The S2 further includes: Use advanced feature engineering techniques such as time series analysis, social network analysis, etc. Extract more valuable user behavior features and implement a real-time update mechanism for user portraits; Combine external data sources to build a more three-dimensional and comprehensive user portrait.

4. The method according to claim 1, characterized in that: The S3 further includes: Dynamically update the user interest model based on the user interest points and their changing trends identified by LLM; Adjust recommendation algorithm parameters and strategies; The user's response to the recommendation results is recorded and this data is fed back into the LLM model to further optimize the model accuracy.

5. The method according to claim 1, characterized in that The S4 further comprises: Deploy Apache Kafka cluster and Apache Flink to achieve efficient processing of real-time data streams and ensure low-latency data processing; Introduce reinforcement learning technology to optimize recommendation strategies, and evaluate the performance improvement of recommendation systems through A / B testing and experimental design methods; Design a dynamic optimization strategy engine to dynamically adjust website resources and improve resource utilization efficiency.

Images