Personalized recommendation method, device and storage medium based on real-time user behavior

Abstract

The application relates to the field of information technology and provides a personalized recommendation method based on real-time user behavior, equipment and a storage medium, which comprises the following steps: collecting real-time user behavior data, performing stream processing on the real-time user behavior data to obtain user behavior feature data; inputting the user behavior feature data into a stream computing architecture, combining a distributed storage and computing architecture, adopting data sharding storage and parallel computing tasks, and optimizing the user behavior feature data; in view of the fault and data loss risk in the running process of the personalized recommendation method, adopting a fault-tolerant recovery mechanism and data backup disaster recovery, realizing load balancing through dynamic resource scheduling, judging the available state of the personalized recommendation method, and generating a method high-availability evaluation report; caching the optimized personalized recommendation result to a distributed cache cluster, adopting a multi-copy storage mechanism, routing user requests to the nearest cache node through load balancing, and improving the access speed of the personalized recommendation result.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a personalized recommendation method, device, and storage medium based on real-time user behavior. Background Technology

[0002] In large-scale real-time personalized recommendation methods, the core technical challenge lies in achieving efficient, accurate, and real-time recommendations under massive user behavior data and complex business scenarios. With the rapid growth of user scale and data volume, traditional recommendation methods face multiple technical bottlenecks. First, there is a contradiction between real-time performance and accuracy. Methods need to complete complex calculations such as user behavior analysis, feature extraction, and interest modeling within milliseconds, while simultaneously ensuring the accuracy of recommendation results. Second, there is a conflict between scalability and resource utilization efficiency. Faced with sudden surges in traffic and data, methods must be able to scale flexibly to maintain service quality while avoiding resource waste. Third, there is a tension between continuous optimization and stability of recommendation algorithms. Frequent model updates may lead to instability, while fixed algorithms struggle to adapt to changes in user interests. Furthermore, how to fully utilize user behavior feature data for personalized recommendations while ensuring data security and user privacy is also a thorny issue. These technical challenges are interconnected and mutually influential, constituting the core challenges faced by large-scale real-time personalized recommendation methods. Finding a balance among these contradictions to create high-performance, highly available, and sustainably optimizable recommendation methods is a crucial issue that urgently needs to be addressed. Summary of the Invention

[0003] To achieve the objectives of this invention, in a first aspect, this invention provides a personalized recommendation method based on real-time user behavior, mainly comprising:

[0004] Real-time user behavior data is collected and streamed. A message queue is used to buffer and smooth traffic fluctuations to obtain user behavior feature data. This user behavior feature data is then input into a distributed computing architecture, where it is processed through sharded storage and parallel computing tasks to obtain optimized user behavior feature data. Flink-SQL is used to perform real-time statistics and aggregation on the optimized user behavior feature data. A distributed coordination service is used to schedule tasks across multiple computing nodes to determine popular items and user interests. Recommendations are then generated through a personalized recommendation engine. To address the risks of failure and data loss during the personalized recommendation method's operation, fault tolerance and data backup disaster recovery mechanisms are employed. Load balancing is achieved through dynamic resource scheduling. The system monitors the operation status of the personalized recommendation method and generates an availability assessment report. It acquires user growth data, analyzes the user growth rate, and if the growth rate exceeds a preset threshold, it determines a significant user increase. In response, it dynamically adds distributed storage and computing nodes, optimizes data transmission paths using the principle of data proximity for computation, reduces network overhead, and adjusts resource allocation to cope with user growth in the personalized recommendation method. It continuously collects user feedback data on the recommendation results, optimizes the results based on this feedback, and obtains optimized personalized recommendation results. Finally, it caches the optimized personalized recommendation results in a distributed cache cluster, employing a multi-replica storage mechanism and using load balancing to route user requests to the nearest cache node, improving the access speed of the optimized personalized recommendation results.

[0005] In a second aspect, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of any of the methods described above.

[0006] Thirdly, a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0007] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0008] This invention discloses a personalized recommendation method, device, and storage medium based on real-time user behavior. Targeting business scenarios involving large-scale real-time user behavior data processing and personalized recommendations, this invention employs a personalized recommendation engine and streaming processing architecture to achieve millisecond-level latency in extracting user behavior features. Through distributed storage and parallel computing, combined with real-time statistical aggregation using Flink-SQL, popular items and user interests are efficiently determined. This invention introduces a fault-tolerant recovery mechanism and dynamic resource scheduling to ensure high availability and scalability of the recommendation service. Simultaneously, user feedback is continuously collected and processed in real time, forming a closed-loop recommendation algorithm optimization process. Finally, this invention caches personalized recommendation results in a distributed cluster and employs a load balancing strategy, significantly improving the processing efficiency, reliability, and access speed of the recommendation method, providing high-quality personalized recommendation services to a massive number of users. Attached Figure Description

[0009] Figure 1 This is a flowchart of the personalized recommendation method based on real-time user behavior according to the present invention.

[0010] Figure 2 This is a schematic block diagram of the structure of a computer device according to an embodiment of the present invention. Detailed Implementation

[0011] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0012] like Figure 1 The personalized recommendation method based on real-time user behavior in this embodiment may specifically include:

[0013] S101. Collect real-time user behavior data, perform streaming processing on the real-time user behavior data, and smooth traffic fluctuations through message queue buffering to obtain user behavior feature data.

[0014] Real-time user behavior data is acquired from the data acquisition layer to obtain the raw behavior data stream. A real-time data stream processing pipeline is established using Flink. The raw behavior data stream is buffered using Kafka to smooth out instantaneous traffic fluctuations, and the buffered data stream is then input into the real-time data stream processing pipeline. In Flink, a window function is used to perform real-time aggregation processing on the input buffered data stream to obtain aggregated user behavior data. The aggregated user behavior data is then classified using the K-nearest neighbor classification algorithm to obtain classification result data representing user behavior patterns. The classification result data is then used to perform feature filtering using the random forest algorithm with a feature importance screening method to obtain user behavior feature data reflecting user behavior characteristics. The feature importance screening method refers to using the random forest algorithm to calculate feature importance scores and filter out features that have a significant impact on the target variable.

[0015] Specifically, acquiring real-time user behavior data from the data acquisition layer is the starting point of the entire process, with the core being capturing dynamic user actions. For example, in an e-commerce platform, real-time data collection of user behaviors such as browsing products, adding items to the cart, and placing orders and making payments can be achieved through event tracking technology, forming a raw behavior data stream. This behavior data typically includes fields such as timestamps, user IDs, and behavior types, reflecting the user's immediate activities on the platform. Buffering the raw behavior data stream using Kafka aims to smooth traffic fluctuations. When using Flink to build a real-time data stream processing pipeline, the focus is on its low latency and high throughput characteristics. For example, the data stream buffered by Kafka is input into the real-time data stream processing pipeline built by Flink. By defining data processing logic, such as filtering out invalid clicks or duplicate operations, the accuracy of subsequent processing is ensured. Specifically, Flink can process tens of thousands of user behavior records per second, making it suitable for scenarios with surges in traffic during peak periods. In one possible implementation, assuming an e-commerce platform generates 100,000 behavior data records per second during a promotional event, Kafka can shard and store this data, setting the buffer size to the amount of data for one minute, i.e., 6 million records. This ensures system stability by preventing a sudden surge in traffic from directly impacting Flink. In Flink, using window functions for real-time aggregation can be understood as dividing the buffered data stream into time windows. For example, a window is created every 5 seconds to count the number of times a product is viewed or the frequency of a user's purchases, resulting in aggregated user behavior data. This approach effectively reduces data redundancy and improves subsequent analysis efficiency. After acquiring the aggregated data, the K-nearest neighbor classification algorithm is used for classification to identify user behavior patterns. Preferably, users can be categorized based on their browsing and purchase records, such as "high-frequency buyers" or "primarily browsing users." In one embodiment, assuming a user browses 10 times and places 2 orders within 5 minutes, they are classified as a "potential high-value user" based on distance calculations from historical data, providing a basis for personalized recommendations. Subsequently, a random forest algorithm is used to filter features in the classification results data using feature importance filtering, further mining deeper features. For example, from the classification results data, random forests select high-weight features such as "historical low-priced product click-through rate," "brand repurchase frequency," and "price sensitivity (discount participation frequency)" by calculating feature importance. Based on these high-weight features, user behavior features are constructed in conjunction with business logic (e.g., "brand repurchase frequency / total purchase frequency" quantifies "brand preference," and "discount participation frequency / total purchase frequency" reflects "low-price preference"), resulting in user behavior feature data that reflects user behavior characteristics. The advantage of this method is that it can integrate multi-dimensional information and improve the expressive power of features. Furthermore, this application stores user behavior feature data in Redis and establishes a real-time query interface, enabling efficient access.For example, Redis can store user behavior characteristic data for each user, and its query interface supports processing 10,000 requests per second, meeting the needs of real-time recommendation systems. It should be noted that Redis's in-memory storage significantly reduces latency. This application also employs Kubernetes to dynamically adjust Flink computing resources. In one embodiment, an access frequency threshold is set to 5,000 times per second; if this threshold is exceeded, the number of Flink task slots is automatically increased, for example, from 10 to 15; if it is below 2,000 times per second, it is reduced to 8. This dynamic adjustment optimizes resource utilization and reduces costs. The user behavior characteristic data obtained through the above process not only supports real-time business decisions but also improves user experience. For example, e-commerce platforms can use this data to optimize recommendation algorithms, increasing user retention and conversion rates, demonstrating the efficiency and practicality of the technology.

[0016] S102. Input the user behavior feature data into a distributed computing architecture, and process the user behavior feature data through sharded storage and parallel computing tasks to obtain optimized user behavior feature data.

[0017] User behavior feature data is input into a streaming computing architecture. Based on a pre-defined sharding strategy, the data is sharded and stored using a distributed storage architecture, evenly distributing it across at least one storage node. This results in sharded data. For the sharded data, at least one parallel computing task is initiated to process it using the MapReduce framework. This yields MapReduce-processed data. The MapReduce-processed data is then clustered using the KMeans algorithm to obtain grouped user behavior data. Finally, Principal Component Analysis (PCA) is used to reduce the dimensionality of the grouped user behavior data, resulting in optimized user behavior feature data.

[0018] Specifically, when inputting user behavior feature data into a distributed computing architecture, the first step is to define a data sharding strategy to ensure that the data is evenly distributed across multiple storage nodes. Sharding strategies may be based on user ID, geographic location, or timestamps. For example, an e-commerce platform can shard users by province to ensure that user data from the same region is stored on nearby nodes, improving query efficiency. Assuming 10 million users distributed across 30 provinces, each shard would average approximately 330,000 user data points. The MapReduce framework processes massive amounts of data in a distributed environment. The Map phase may convert user behavior feature data into key-value pairs, such as {User ID: [List of browsed product IDs]}. The Reduce phase may aggregate user behavior feature data from the same user. This parallel processing significantly improves the processing speed of large-scale data. The KMeans clustering algorithm is used to discover user groups. In e-commerce scenarios, users may be categorized into groups such as "price-sensitive," "brand-loyal," and "new product seekers." The algorithm iteratively calculates the cluster centers until convergence. Assuming 5 clusters are set, a stable result is obtained after 50 iterations, with each cluster containing approximately 2 million users. Principal Component Analysis (PCA) reduces data dimensionality while retaining key features. For example, for user A, 10 key metrics are extracted from 100 original user behaviors, such as user A's browsing, clicking, and purchasing behavior data, resulting in optimized user behavior feature data. This not only reduces storage space but also improves the efficiency of subsequent analysis. Furthermore, the optimized user behavior feature data in this application is stored in an HBase database, and a query interface based on the user behavior key feature data is established. HBase, as a big data storage system, is suitable for storing and quickly retrieving massive amounts of user feature data. The data table design may use user IDs as rows and various features as columns. This structure supports efficient random read / write and range scanning, providing data support for applications such as personalized recommendations. This application also determines whether the access volume of the query interface exceeds a preset access threshold. If it exceeds, the number of computing nodes in the Hadoop cluster is dynamically expanded; if it is below, the number of computing nodes is reduced. Dynamically adjusting the number of computing nodes in the Hadoop cluster is crucial to ensuring system performance. In one embodiment, assuming the original cluster has 40 nodes, when access volume exceeds the threshold, the system automatically adds 20 additional nodes to the cluster via a script. After the node expansion, the total processing capacity increases from 50,000 behavioral data entries per second to 80,000 entries per second, ensuring stable system operation. Conversely, if access volume drops to 3,000 times per second, below the access threshold, the system can be reduced to 30 nodes, freeing up excess resources. This dynamic adjustment can also optimize costs, such as reducing nodes during off-peak hours at night to lower expenses. The advantage of this architecture lies in its scalability and flexibility. Through distributed storage and computing, the method can easily handle the growth of data volume.Dynamic resource adjustment ensures performance stability under varying loads. For example, an e-commerce platform might process 10,000 user behavior data points per second on a normal day, but this could surge to 100,000 during peak sales periods. The architecture described above allows for smooth scaling of processing capacity, ensuring an unaffected user experience. In summary, this big data processing architecture optimizes the processing of user behavior data through distributed storage, parallel computing, intelligent algorithms, and dynamic resource scheduling. This enables efficient processing and utilization of massive amounts of user behavior data, providing enterprises with robust data support and decision-making support.

[0019] S103. Use Flink-SQL to perform real-time statistics and aggregation on the optimized user behavior feature data, and realize task scheduling of multiple computing nodes through distributed coordination service to determine popular items and user interests and preferences, and generate recommendation results through personalized recommendation engine.

[0020] The optimized user behavior feature data is acquired, and Flink-SQL is used to perform real-time statistical processing on the optimized user behavior feature data to obtain statistical values. Real-time aggregation calculation is performed on the statistical values ​​to obtain scheduling values. Distributed computing processing is performed on multiple computing nodes according to the scheduling values ​​to obtain calculated values. Based on the calculated values, popular items are analyzed using a preset popularity analysis algorithm to obtain item values. A preset user interest analysis model is used to analyze user interests based on the item values ​​to obtain interest values. The interest values ​​are input into a preset user preference analysis model for processing to obtain user preference values. If the preference values ​​are greater than a preset preference threshold, user preference profile data is generated based on the preference values. Using this user preference profile data, a personalized recommendation engine generates targeted recommendation results in real time.

[0021] Specifically, real-time statistical processing is key to efficient user behavior analysis. Flink-SQL (a stream-batch integrated SQL query engine based on Apache Flink) can process massive data streams within milliseconds due to its low latency and high throughput. In one possible implementation, when optimized user behavior feature data undergoes real-time statistical processing via Flink-SQL, the data stream can be segmented based on time windows. Below is a complete example of the process based on the above steps. For instance, an e-commerce platform extracts and summarizes user A's browsing, click, and purchase behavior data for the past 10 days from optimized user behavior feature data, calculating the number of times user A viewed the product "sports shoes" and the purchase conversion rate. Assuming "sports shoes" were viewed 50 times and purchased 2 times in 10 days, the statistical value is calculated using the following formula: "Sports shoes" statistical value = number of purchases / number of views × 100% = 2 / 50 × 100% = 4%. Real-time aggregation calculations are performed on the statistical value, grouping and aggregating by purchase time periods, etc., to obtain the scheduling value. For example, the scheduling value is calculated as follows: Scheduling value = Click-through rate × Statistical value × Time period weight (preset weight 1.2 for 8-10 PM). To calculate the scheduling value for user A's "sneakers," the optimized user behavior feature data shows a click-through rate of 40% and a statistical value of 4% for user A's "sneakers." Therefore, the scheduling value for user A's "sneakers" is 40% × 4% × 1.2 = 1.92%. When performing distributed computing processing on multiple computing nodes based on the scheduling value, task sharding can be used. Assuming there are 10 computing nodes, product categories with higher scheduling values ​​will be allocated more node resources, resulting in higher computational values. For example, when obtaining the scheduling values ​​for all products, "sneakers," with the highest scheduling value (1.92%), is allocated to 4 nodes for parallel processing, while "books," with a lower scheduling value, is allocated only to 1 node. Therefore, the computational value for "sneakers" is 4, and the computational value for "books" is 1. Based on a pre-defined popularity analysis algorithm, such as a time-decay-based weighted scoring algorithm, and combining the calculated value with real-time data such as clicks, favorites, purchases, shares, and time decay factors, the calculated value is analyzed to obtain the item value. For example, the item value for "Sneakers" is calculated as follows: Item Value = Calculated Value × {α × (Clicks) + β × (Favorites) + γ × (Purchases) + δ × (Shares) - θ}, where the weighting coefficients (α=0.4, β=0.3, γ=0.2, δ=0.1) can be pre-determined through A / B testing, and θ (time decay factor) = log 10 (Current time - Item listing time + 1). Example input data: "Sneakers" calculated value = 4, clicks = 1000, favorites = 50, purchases = 20, shares = 10. These calculated values ​​are derived from optimized user A's behavioral characteristics data. Listing time: 7 days ago, corresponding to θ = log 10(7×24×3600+1)≈5.8, resulting in the item value for "Sports Shoes" = 4×[0.4×1000+0.3×50+0.2×20+0.1×10-5.8] = 4×[400+15+4+1-5.8] = 4×414.2 = 1656.8. When analyzing interest values ​​using a pre-defined user interest analysis model, item values ​​and user history can be combined to infer interest points. For example, interest value = total behavior score × (item value / maximum item value of similar items), where total behavior score = Σ (number of behaviors × behavior weight × time decay factor); behavior weights: browsing = 1, favorites = 3, purchases = 5. Input example data: The item value for "Sneakers" is 1656.8. User A's interest in "Sneakers": 5 views (score decays after each view = 1 × 0.9 = 0.9) → 5 × 0.9 = 4.5; 2 favorites (score decays after each view = 3 × 0.8 = 2.4) → 2 × 2.4 = 4.8; 1 purchase (score decays after each purchase = 5 × 0.7 = 3.5) → 1 × 3.5 = 3.5; Total behavior score = 4.5 + 4.8 + 3.5 = 12.8; Maximum value for similar items: Item value = 2000 ("Sneakers" category). The item value is calculated based on optimized user A's behavioral feature data, and user A's relevant data (views, favorites, purchases) are extracted from the optimized user behavior feature data. Therefore, user A's interest value for "Sneakers" is 12.8 × (1656.8 / 2000) ≈ 12.8 × 0.828 ≈ 10.60. When inputting User A's interest value for "sports shoes" into the user preference analysis model, the preference value can be determined through a weighted scoring mechanism. The preference value is calculated as: Preference Value = Σ(Behavior Type Weight × Interest Value) × Category Coefficient. For example, the behavior type weights might be: Browsing 0.3 (duration > 30 seconds), Adding to Favorites 0.5 (Added to Favorites / Wishlist), Adding to Cart 0.6 (Added to Cart but not purchased), Purchasing 0.7 (Successful payment), Reviewing 0.4 (Additional +0.1 for reviews with pictures), and the category coefficient (sports shoes) is 1.2. Example input data: User A's interest value for "sports shoes" = 10.60. Behavior types: Browsing (0.3), Adding to Favorites (0.5), Purchasing (0.7), where the interest value is calculated based on the optimized user A's behavioral characteristics data. The resulting preference value for "sports shoes" is: (0.3 + 0.5 + 0.7) × 10.60 × 1.2 = 1.5 × 10.60 × 1.2 = 19.08. If user A's preference value for "sports shoes" is greater than the preset preference threshold of 15, then user A is determined to have a strong preference for "sports shoes," and preference profile data is generated, such as "sports enthusiast." This profile clearly depicts the user's tendencies. When a personalized recommendation engine generates recommendation results based on the preference profile, the possible calculation basis is obtained by using the user preference profile data through collaborative filtering and machine learning models, including the similarity between the user and the item, the predicted click-through rate, and the real-time weight data of the item.A specific example of the actual calculation process: By analyzing users' historical behavioral data (such as browsing, purchasing, rating, etc.) and the attribute characteristics of items, collaborative filtering is used to calculate the similarity between users and items; based on user behavioral data and item characteristics, machine learning models (such as logistic regression, decision trees, neural networks, etc.) are used to predict the probability of whether a user will click on a certain item, thereby estimating the predicted click-through rate; real-time weight data of items is pre-set according to user usage time, platform usage, etc. (e.g., 7 pm, using a mobile phone). The similarity between users and items, predicted click-through rate, and real-time weight data of items are obtained and scores are calculated. Base score = similarity between users and items (0.8) × predicted click-through rate (0.9) × real-time weight of items (1.2), adjustment score = popularity (0.3) + diversity (same category penalty 0.7) + clearance bonus (0.2), final score = base score + adjustment score, final output result: the top 10 ranked products (such as running socks, sports water bottles, cushioned running shoes, etc.). Products that meet the conditions are filtered in real time (recommendation results) and pushed to users. This approach enhances the targeting of recommendations. Understandably, this architecture transforms user behavior into precise recommendations through multi-layered analysis and real-time processing. The distributed design and algorithm optimization at each stage ensure efficient integration from data input to output, providing robust support for e-commerce platforms.

[0022] S104. To address the risks of failure and data loss during the operation of personalized recommendation methods, a fault-tolerant recovery mechanism and data backup disaster recovery are adopted. Load balancing is achieved through dynamic resource scheduling, the operating status of personalized recommendation methods is determined, and an availability assessment report is generated.

[0023] The process involves acquiring the runtime log data of the personalized recommendation method, which refers to the collection of events and data automatically recorded during the method's operation. A pre-established fault-tolerance and recovery mechanism is used to analyze the runtime log data, detect the running status of the personalized recommendation method, and obtain runtime status detection results. Based on a preset data backup strategy, optimized user behavior feature data in the personalized recommendation method is periodically backed up to full. Combined with a preset disaster recovery strategy, full backup data is deployed in a remote data center. Based on the runtime status detection results, if at least one fault point is detected, the preset data backup disaster recovery mechanism is triggered, switching user requests to the backup data and recovering the data lost due to the at least one fault point, obtaining recovered data. Based on the recovered data, a pre-established dynamic resource scheduling algorithm is used to analyze the load situation during the personalized recommendation method's operation, obtaining the current load value. For the current load value, computing resources for the personalized recommendation method are dynamically allocated, obtaining computing resource allocation results. A pre-established load balancing algorithm is used to dynamically adjust various resources of the personalized recommendation method based on the computing resource allocation results, obtaining resource balancing results. Based on the resource balancing results, the availability index of the personalized recommendation method is calculated to obtain a method availability value. It is then determined whether the method availability value is greater than a preset availability threshold. If the method availability value is greater than the preset availability threshold, the personalized recommendation method is determined to be in a high-availability state, and a availability assessment report is generated.

[0024] Specifically, ensuring the high availability of personalized recommendation methods is crucial for guaranteeing user experience and business continuity. By acquiring operational log data of personalized recommendation methods, various metrics of the personalized recommendation engine can be monitored in real time. For example, in e-commerce platforms, operational log data can record request and response data, performance metrics, user behavior feedback, system status, errors, and alarms to assess the smooth operation of personalized recommendation methods. Establishing fault-tolerant recovery mechanisms can improve the stability of personalized recommendation methods. For instance, in e-commerce platform recommendation systems, multi-level fault detection thresholds can be set. When the error rate of a recommendation module exceeds a preset value, a backup algorithm is automatically triggered to ensure uninterrupted service. This mechanism can detect and respond to anomalies within milliseconds, significantly reducing the risk of method failures. Regular full backups are the cornerstone of data security. For e-commerce platform recommendation systems, incremental backups can be performed daily at midnight, and full backups can be performed weekly. This not only protects users' historical purchase information and other data but also provides a guarantee for emergency recovery. Deploying off-site data centers further enhances the method's resilience. For example, setting up the main data center in Beijing and a mirror center in Guangzhou can effectively cope with regional disasters. When a fault is detected, the data backup and disaster recovery mechanism plays a crucial role. In an e-commerce platform recommendation system, if the primary database (fault point) becomes unavailable due to hardware failure, the system can switch to a backup data source within seconds and recover lost data using backup data, ensuring that users' investment decisions are not affected. Alternatively, if the operational status detection results show that at least one off-site data center (fault point) has failed, the system can also switch to other backup or primary data sources within seconds and recover lost data using backup data, ensuring that users' investment decisions are not affected. Dynamic resource scheduling algorithms can optimize method performance. When dynamically allocating computing resources based on the current load value, computing resources can be increased for nodes with high loads. For example, allocating two additional CPU cores to a node with a load of 1200 operations / minute results in a balanced adjustment of the computing resources across all nodes to a processing capacity of approximately 1000 operations / minute. When adjusting resources using load balancing algorithms, the adjustable resources for personalized recommendation methods include computing resources (such as CPU, GPU), memory resources, service instances (number of containers, virtual machines), storage resources (such as database access), network bandwidth, task parallelism (number of shards in a distributed task), Kafka partitions, etc.For example, a recommendation system faces a surge in load during peak evening hours: Computational resources: scaling from 50 to 100 nodes to handle real-time clickstreams; Memory resources: adding 3 nodes to the Redis cluster to cache embedding vectors for popular products; Service instances: scaling the Kubernetes recommendation API Pods from 20 to 50; Network bandwidth resources: enabling dedicated network channels for cross-availability zone communication to reduce latency; Load balancing: adjusting Nginx weights to direct 70% of traffic to the newly added Pods. Result: Recommendation response time decreased from 800ms to 200ms, and system throughput increased by 3 times. Availability metrics are a set of indicators such as request response time, fault tolerance and recovery time, and request success rate (SLA achievement rate). Calculating availability metrics is an important means of evaluating service quality. For example, availability percentage can be calculated as "(Total service time - downtime) / Total service time". For critical businesses, availability of 99.99% or higher is typically required, meaning no more than 52.56 minutes of downtime per year. Through these measures, the personalized recommendation method can maintain high availability in the face of various challenges, providing users with continuous, stable, and high-quality services. The availability assessment report not only reflects the current operational status but also provides direction for future optimization, enabling the recommendation service to continuously evolve and meet users' growing personalized needs.

[0025] S105. Obtain user growth data and analyze the user growth rate. If the user growth rate is greater than the preset growth rate threshold, it is determined that the user scale has grown significantly. Dynamically increase distributed storage and computing nodes, optimize data transmission paths, reduce network transmission overhead, and adjust resource allocation results to cope with the user growth of personalized recommendation methods.

[0026] User growth data is retrieved from the log database. Linear regression analysis is used to analyze the user growth rate. If the growth rate is less than or equal to a preset threshold, the user base is considered stable and no action is taken. If the growth rate exceeds the threshold, a significant increase in user size is identified. If a significant increase is identified, storage capacity is dynamically increased based on the growth rate. Computing nodes are deployed in densely populated user areas. Computing tasks are allocated based on user geographic location information, and the shortest path algorithm is used to optimize data transmission paths, reducing data transmission distance and network overhead. Load data is retrieved from a pre-set monitoring system, and a dynamic resource scheduling algorithm is used to allocate idle computing resources, improving computational efficiency. Based on the load data, a round-robin algorithm is used to adjust resource allocation results, ensuring the stability and efficiency of the personalized recommendation method.

[0027] Specifically, analyzing user growth in personalized recommendation methods is a crucial starting point for optimizing these methods. By extracting user activity and registration data from log databases, linear regression models can be used to analyze user growth rates. For example, an e-commerce platform analyzed user data from the past 12 months and found a stable month-on-month growth in user numbers, predicting a 15% increase in users in the next quarter. This growth rate directly impacts storage requirements. To cope with the surge in data brought about by user growth, the number of storage nodes needs to be dynamically adjusted. If a significant increase in user numbers is anticipated, capacity can be expanded in advance. For instance, based on user growth analysis, an e-commerce platform decided to add 20 nodes to its existing 100 storage nodes, increasing the total storage capacity from 1PB to 1.2PB to ensure data storage needs for the next six months. User geographic distribution information is critical for optimizing method performance. By analyzing user IP addresses or GPS data, densely populated user areas can be identified. For example, an e-commerce platform found that 80% of its active users were concentrated in 10 major cities. Based on this finding, the platform deployed additional computing nodes in these cities, enabling most user requests to be processed within a 100-kilometer radius, significantly reducing network latency. Shortest path algorithms play a crucial role in optimizing data transmission. Taking product recommendation methods as an example, when a user requests to view a product, the system calculates the optimal path from the nearest storage node to the user's device. This not only reduces data transmission distance but also lowers the likelihood of network congestion. Practice shows that this optimization can shorten the loading time of product details by more than 20%, significantly improving the user experience. Load data is a vital basis for resource scheduling. Monitoring systems collect metrics such as CPU utilization, memory usage, and network throughput in real time, providing a decision-making basis for dynamic resource scheduling. For example, when an e-commerce platform detects a surge in load between 7 PM and 9 PM, it automatically allocates idle computing resources to the recommendation engine to handle peak user access. This elastic scheduling strategy ensures that the method's response time during peak periods only increases by 5% compared to normal times, guaranteeing a consistent user experience. Round-robin algorithms are an effective means of ensuring service stability. In personalized recommendation methods, round-robin algorithms can evenly distribute user requests across multiple servers. For instance, during the Double 11 shopping festival, an e-commerce platform's recommendation system adopted a weighted round-robin strategy, dynamically adjusting weights based on the processing capacity of each server to ensure load balance. This not only improves the method's concurrent processing capabilities but also reduces the risk of single points of failure, ensuring that even during peak traffic periods, the average response time remains below 200 milliseconds. Through the combined application of these technologies, the personalized recommendation method can maintain efficient and stable operation even with rapid user growth, providing users with a smooth and personalized service experience and effectively addressing the challenges of user growth.

[0028] S106. Continuously collect user feedback data on the recommendation results, optimize the recommendation results based on the feedback data, and obtain optimized personalized recommendation results.

[0029] Click-through rate, dwell time, and conversion rate data of recommendation results are extracted from user behavior logs. The data is cleaned and features extracted using a streaming computing architecture. The user behavior logs record detailed user operations and interactions. A sliding window mechanism is used to aggregate the click-through rate, dwell time, and conversion rate data of recommendation results to obtain temporal features of user behavior. If the feature value of the temporal feature exceeds a preset feature value threshold, an engine update mechanism is triggered, and the temporal features exceeding the threshold are input into the personalized recommendation engine. The parameters of the personalized recommendation engine are optimized using a gradient descent algorithm, updating the user preference profile data in the personalized recommendation engine to obtain updated user preference profile data. The personalized recommendation engine uses the updated user preference profile data to generate new recommendation results. A collaborative filtering algorithm is used to calculate the similarity between the updated user preference profile data and other user preference profile data. Target user preference profile data with a similarity greater than a preset similarity threshold is selected, and the target recommendation result corresponding to the target user preference profile data is obtained. The new recommendation result and the target recommendation result are combined to update the recommendation result, resulting in the optimized personalized recommendation result.

[0030] Specifically, personalized recommendation methods optimize recommendation performance by analyzing user behavior logs. User behavior logs are detailed user actions and interactions within the application (such as clicks, dwell time, and purchases), containing information such as time, object, and context. These logs are used to calculate key metrics like click-through rate (CTR), dwell time, and conversion rate in real time. Taking an e-commerce platform as an example, CTR, dwell time, and conversion rate data are extracted from user behavior logs. CTR reflects a user's initial interest in the recommended content, dwell time reflects the content's attractiveness, and conversion rate shows whether the user completed the expected action, such as adding to the cart or making a purchase. A streaming computing architecture can process massive amounts of data in real time. This platform uses Apache Flink for data cleaning and feature extraction. During cleaning, the method filters out abnormal data, such as records with excessively short dwell times. Feature extraction includes calculating user viewing preferences and active time periods. A sliding window mechanism is used to capture dynamic changes in user behavior. Assuming the platform sets a 5-minute sliding window, updating every minute, the method calculates time-series features such as average dwell time and CTR within this window. If a user's dwell time suddenly increases from an average of 10 minutes to 30 minutes, exceeding the preset 20-minute threshold, the recommendation engine's update mechanism will be triggered. The time-series features exceeding the feature value threshold will be input into the personalized recommendation engine, resulting in updated user preference profile data. Specifically, when optimizing the personalized recommendation engine parameters using the gradient descent algorithm, the weights related to user preference profile data in the personalized recommendation engine can be adjusted based on the newly input time-series features. For example, if system analysis reveals that the proportion of time a user spends browsing fitness equipment products increases from 40% to 60%, the gradient descent algorithm will gradually adjust the weight of fitness equipment preference in the user preference profile data, increasing it from 0.5 to 0.7. The core of this user preference profile data optimization method lies in using the difference between historical and new data to gradually approach the optimal solution, ensuring that the user preference profile data can reflect the latest interests in a timely manner. Preferably, this method can also avoid instability in recommendation results due to excessively rapid parameter adjustments. After obtaining the updated user preference profile data, the personalized recommendation engine uses the updated user preference profile data to generate new recommendation results. For example, a collaborative filtering algorithm is used to calculate the similarity between the updated user preference profile data and other user preference profile data based on different users' interaction data with items (such as click-through rate, dwell time, and conversion rate data), using cosine similarity or Pearson correlation coefficient. Other user preference profile data with similarity higher than a preset similarity threshold can be used as target user preference profile data. Then, the target recommendation result corresponding to the target user preference profile data is obtained. Finally, the recommendation result is updated using the target recommendation result and the new recommendation result to obtain the optimized personalized recommendation result.For example, with a preset similarity threshold of 80%, if a user's preference profile data shows a strong interest in athletic shoes and fitness equipment after updates, the system finds target user preference profile data with a similarity greater than 80%, discovering that these users frequently browse yoga mat recommendations, meaning yoga mats appear in the target recommendation results. Combining these target recommendation results with the new recommendation results, the system merges them, removing duplicate data, to obtain personalized recommendations. In the yoga mat example above, if neither the target nor the target recommendation results contain yoga mat recommendations, then yoga mats will be added to the optimized personalized recommendations. This method can uncover users' potential interests and improve the diversity of recommendations. Understandably, by inputting personalized recommendation results into a streaming computing architecture, the personalized recommendation results can be continuously optimized. For example, when a user browses a page, the system updates the recommendation results every 5 minutes based on the recommendation list generated in the aforementioned steps using a streaming computing architecture. Specifically, when a user clicks on a recommended pair of sneakers, the system immediately feeds this action back to the engine, triggering a new round of feature extraction and recommendation list generation, updating the personalized recommendation results. The core of this closed-loop iterative process lies in rapidly responding to changes in user behavior, ensuring that the recommendations always closely match the user's current needs. Through the comprehensive application of these technologies, the system can provide users with more accurate and personalized content recommendations. This not only improves the user experience but also increases user stickiness on the platform. Simultaneously, this data-driven approach provides valuable insights for the platform's content creation and operational strategies.

[0031] S107. Cache the optimized personalized recommendation results to a distributed cache cluster, adopt a multi-replica storage mechanism, and route user requests to the nearest cache node through load balancing to improve the access speed of the optimized personalized recommendation results.

[0032] A distributed caching cluster retrieves the optimized personalized recommendation results and uses a multi-replica storage mechanism to store these results across multiple cache nodes. A load balancing algorithm analyzes the source of user requests, determining the optimal cache node based on geographical location and network latency information. User requests are then routed to the optimal cache node, from which the optimized personalized recommendation results are retrieved, thus improving the access speed of the optimized personalized recommendation results.

[0033] Specifically, distributed caching clusters are a key technology for improving the performance of personalized recommendation methods. Multi-replica storage mechanisms ensure high availability and fault tolerance of data. For example, an e-commerce platform stores personalized recommendation results for users across multiple cache nodes distributed throughout the country. Each recommendation result has at least three replicas, distributed in different geographical locations to cope with potential node failures. Load balancing algorithms select the optimal cache node by analyzing the source of user requests. For example, user requests from Beijing may be routed to cache nodes in North China, while requests from users in Guangzhou are directed to nodes in South China. This location-based routing strategy can significantly reduce network latency and improve user experience. Furthermore, this application uses a sliding window mechanism to monitor the access pressure of cache nodes in real time. If the node access pressure exceeds a preset access threshold, dynamic load balancing adjustments are triggered. The cache hit rate is analyzed based on user access logs. If the hit rate is lower than a preset hit rate value, the caching strategy is updated and the replica distribution of the optimized personalized recommendation results is adjusted. The user access logs record detailed data on the frequency of user requests to cache nodes, access time, resource paths, and cache hit status. Load balancing parameters are optimized using gradient descent to dynamically adjust user request routing strategies. Collaborative filtering is employed to calculate user preference similarity, storing recommendations for similar users in the nearest cache node. A sliding window mechanism is used to monitor cache node access pressure in real time. Assuming an e-commerce platform sets a 5-minute sliding window, updating every minute, if the access volume to a node suddenly increases from an average of 1000 times per minute to 3000 times per minute within this window, exceeding the preset threshold of 2000 times per minute, dynamic load balancing adjustments will be triggered. This may involve rerouting some user requests to other less busy nodes or adding new cache nodes to distribute the load. User access logs record detailed data on user request frequency, access time, resource path, and cache hit status (e.g., hit / miss) for cache nodes, used for real-time analysis of node pressure, triggering load balancing strategies, and optimizing cache distribution to improve recommendation system performance. Cache hit rate is a crucial metric for evaluating the effectiveness of caching strategies. By combining data from user access logs, if the cache hit rate of a certain type of recommendation result falls below a preset threshold of 80%, the caching strategy will be updated. This might include increasing the caching time for popular content or adjusting the distribution of recommendation results across different nodes. For example, frequently accessed recommendation results from the last 24 hours might be copied to more cache nodes to improve the overall hit rate. Gradient descent is used to optimize load balancing parameters. The method continuously adjusts the routing weights of user requests based on historical data. For instance, if a surge in access volume in East China is detected during a certain period, the method will automatically increase the weight of cache nodes in that region, distributing more requests to these nodes. This dynamic adjustment better handles traffic fluctuations and ensures the stability of the method.Collaborative filtering algorithms also play a crucial role in cache optimization. The method calculates the similarity of preferences between users and stores recommendations for similar users in nearby cache nodes. For example, users A and B, who like technology products, are identified as similar users, and the method stores their recommendations in the same or adjacent cache nodes. This strategy not only improves cache hit rate but also reduces data duplication between nodes, enhancing the overall efficiency of the method. Through the comprehensive application of these technologies, personalized recommendation methods can significantly improve response speed and stability while maintaining recommendation quality. This not only improves user experience but also brings higher user retention rates and economic benefits to the platform. Furthermore, this data-driven approach provides reliable technical support for continuous optimization of the method.

[0034] Reference Figure 2 This application also provides a computer device, which may be a server, and its internal structure may be as follows: Figure 2 As shown. The computer device includes a processor, memory, network interface, and database connected via a system bus. The processor in this computer is designed to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database of the computer device is used to store data such as user behavior characteristics and recommendation results. The network interface of the computer device is used to communicate with external terminals via a network connection. When the computer program is executed by the processor, it implements the personalized recommendation method based on real-time user behavior of any of the above embodiments.

[0035] Those skilled in the art will understand that Figure 2 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer equipment on which the present application is applied.

[0036] This invention also provides a computer-readable storage medium storing a computer program thereon. When executed by a processor, the computer program implements the personalized recommendation method based on real-time user behavior described in any of the above embodiments. It is understood that the computer-readable storage medium in this embodiment can be a volatile readable storage medium or a non-volatile readable storage medium.

[0037] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the methods described above. Any references to memory, storage, databases, or other media used in this application and in the embodiments can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual-rate SDRAM (SSRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0038] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A personalized recommendation method based on real-time user behavior, characterized in that, The method includes: Real-time user behavior data is collected and streamed. A message queue is used to buffer and smooth traffic fluctuations to obtain user behavior feature data. This user behavior feature data is then input into a distributed computing architecture, where it is processed through sharded storage and parallel computing tasks to obtain optimized user behavior feature data. Flink-SQL is used to perform real-time statistics and aggregation on the optimized user behavior feature data to obtain statistical values. These statistical values ​​are calculated as: Statistical Value = Number of Purchases / Number of Views × 100%. Real-time aggregation calculations are then performed on these statistical values ​​to obtain a scheduling value, which is calculated as: Scheduling Value = Click-Through Rate × ... The statistical value is multiplied by the time period weight. Distributed computing is performed on multiple computing nodes based on the scheduling value to obtain the calculated value. For this calculated value, a preset popularity analysis algorithm is used to analyze popular items, resulting in an item value. The item value is calculated as: Item Value = Calculated Value × {α × Clicks + β × Favorites + γ × Purchases + δ × Shares - θ}, where α, β, γ, and δ are weighting coefficients, and θ is a time decay factor. A preset user interest analysis model is used to analyze user interests based on the item value, resulting in an interest value. The interest value is calculated as: Interest Value = Total Behavior Score × Item Value / Maximum Item Value of Similar Items, where Total Behavior Score = ... The interest value is calculated by multiplying the number of behaviors by their weights and the time decay factor, and then inputting the interest value into a preset user preference analysis model for processing. The preference value is calculated as: Preference Value = Σ Behavior Type Weight × Interest Value × Category Coefficient. If the preference value is greater than a preset preference threshold, user preference profile data is generated based on this value. Using this user preference profile data, a personalized recommendation engine generates targeted recommendation results in real time. To address the risks of failure and data loss during the operation of the personalized recommendation method, a fault-tolerant recovery mechanism and data backup disaster recovery are adopted. Load balancing is achieved through dynamic resource scheduling to determine the operating status of the personalized recommendation method and generate reliable results. The report assesses the user growth rate and analyzes user growth data. If the user growth rate exceeds a preset threshold, it is considered a significant increase in user scale. Distributed storage and computing nodes are dynamically added, and data transmission paths are optimized using the principle of proximity computing. Resource allocation is adjusted to address the user growth of the personalized recommendation method. User feedback on recommendation results is continuously collected, and the recommendation results are optimized based on this feedback to obtain optimized personalized recommendation results. The optimized personalized recommendation results are cached in a distributed cache cluster using a multi-replica storage mechanism, and user requests are routed to the nearest cache node through load balancing.

2. The method according to claim 1, characterized in that, The process involves collecting real-time user behavior data, streaming the data, and smoothing out traffic fluctuations using a message queue to obtain user behavior feature data, including: Real-time user behavior data is obtained from the data acquisition layer to obtain the raw behavior data stream; Use Flink to build a real-time data stream processing pipeline; The raw behavioral data stream is buffered using Kafka to smooth out instantaneous traffic fluctuations, and the buffered data stream is then input into the real-time data stream processing pipeline. In Flink, window functions are used to perform real-time aggregation processing on the buffered data stream to obtain aggregated user behavior data. The aggregated user behavior data is obtained and classified using the K-nearest neighbor classification algorithm to obtain classification result data that represents user behavior patterns. The classification result data is obtained, and the random forest algorithm is used to filter the features of the classification result data through the feature importance screening method to obtain user behavior feature data that reflects user behavior characteristics. The feature importance screening method refers to calculating the feature importance score through the random forest algorithm and screening out features that have a significant impact on the target variable.

3. The method according to claim 1, characterized in that, The step of inputting the user behavior feature data into a distributed computing architecture and processing the user behavior feature data through sharded storage and parallel computing tasks to obtain optimized user behavior feature data includes: User behavior feature data is input into a distributed computing architecture. According to a preset sharding strategy, the user behavior feature data is sharded and stored using a distributed storage architecture. The user behavior feature data is evenly distributed to at least one storage node to obtain the sharded data. For the sharded data, start at least one parallel computing task and use the MapReduce computing framework to process the sharded data to obtain the MapReduce processed data. The data processed by MapReduce is obtained, and the KMeans algorithm is used to perform cluster analysis on the data processed by MapReduce to obtain the grouping results of user behavior. Principal component analysis (PCA) was used to reduce the dimensionality of the grouped user behavior data, resulting in optimized user behavior feature data.

4. The method according to claim 1, characterized in that, To address the risks of failures and data loss during the operation of the personalized recommendation method, a fault-tolerant recovery mechanism and data backup disaster recovery are adopted. Load balancing is achieved through dynamic resource scheduling. The operating status of the personalized recommendation method is assessed, and an availability assessment report is generated, including: Obtain the runtime log data of the personalized recommendation method, wherein the runtime log data refers to the collection of events and data automatically recorded during the runtime of the personalized recommendation method; By employing a pre-established fault-tolerant recovery mechanism, the operation log data is analyzed to detect the operation status of the personalized recommendation method and obtain the operation status detection results. According to the preset data backup strategy, the optimized user behavior feature data in the personalized recommendation method is backed up in full on a regular basis. Combined with the preset disaster recovery strategy, deploy full backup data in a remote data center; Based on the operational status detection results, if at least one fault point is detected, a preset data backup and disaster recovery mechanism is triggered, the user request is switched to the backup data, the lost data caused by the at least one fault point is restored, and the restored data is obtained. Based on the recovered data, a pre-established dynamic resource scheduling algorithm is used to analyze the load of the personalized recommendation method during operation and obtain the current load value. For the current load value, the computing resources of the personalized recommendation method are dynamically allocated to obtain the computing resource allocation result; A pre-established load balancing algorithm is used to dynamically adjust the various resources of the personalized recommendation method based on the computing resource allocation results, thereby obtaining a resource balance result; Based on the resource balancing results, the usability index of the personalized recommendation method is calculated to obtain the method usability value; Determine whether the availability value of the method is greater than a preset availability threshold. If the availability value of the method is greater than the preset availability threshold, then the personalized recommendation method is determined to be in a high availability state, and an availability assessment report is generated.

5. The method according to claim 1, characterized in that, The process of acquiring user growth data and analyzing the user growth rate, determining a significant user growth if the growth rate exceeds a preset threshold, involves dynamically increasing distributed storage and computing nodes, optimizing data transmission paths using the principle of data proximity for computation, reducing network overhead, and adjusting resource allocation to address user growth in personalized recommendation methods. This includes: Retrieve user growth data from the log database and use linear regression to analyze the user growth rate. If the growth rate is less than or equal to the preset growth rate threshold, the user scale is considered stable and no action is taken. If the growth rate is greater than the preset growth rate threshold, the user scale is considered to have grown significantly. If it is determined that the user base has grown significantly, the storage capacity will be dynamically increased based on the user base growth rate. Deploy computing nodes in densely populated user areas, allocate computing tasks based on the user's geographical location, optimize data transmission paths using the shortest path algorithm, reduce data transmission distance, and lower network transmission overhead. Load data is obtained from a pre-set monitoring system, and idle computing resources are allocated using a dynamic resource scheduling algorithm to improve computing efficiency. Based on the load data, a round-robin algorithm is used to adjust the resource allocation results to ensure the stability and efficiency of the personalized recommendation method.

6. The method according to claim 1, characterized in that, The process of continuously collecting user feedback data on the recommendation results and optimizing the recommendation results based on the feedback data to obtain optimized personalized recommendation results includes: Click-through rate, dwell time, and conversion rate data of recommendation results are extracted from user behavior logs. The data is cleaned and features are extracted using a streaming computing architecture. The user behavior logs record detailed user operations and interaction data. A sliding window mechanism is used to aggregate the click-through rate, dwell time, and conversion rate data of the recommendation results to obtain the temporal characteristics of user behavior; If the feature value of the time-series feature exceeds the preset feature value threshold, the engine update mechanism is triggered, and the time-series feature that exceeds the feature value threshold is input into the personalized recommendation engine. The parameters of the personalized recommendation engine are optimized by using the gradient descent algorithm, and the user preference profile data in the personalized recommendation engine is updated to obtain the updated user preference profile data. New recommendation results are generated by using updated user preference profile data through a personalized recommendation engine; The collaborative filtering algorithm is used to calculate the similarity between the updated user preference profile data and other user preference profile data. Target user preference profile data with a similarity greater than a preset similarity threshold is selected, and the target recommendation results corresponding to the target user preference profile data are obtained. The new recommendation results are combined with the target recommendation results to update the recommendation results, resulting in optimized personalized recommendation results.

7. The method according to claim 1, characterized in that, The step of caching the optimized personalized recommendation results to a distributed cache cluster, employing a multi-replica storage mechanism, and using load balancing to route user requests to the nearest cache node to improve the access speed of the optimized personalized recommendation results includes: The distributed cache cluster obtains the optimized personalized recommendation results and uses a multi-replica storage mechanism to store the recommendation results on multiple cache nodes; The source of user requests is analyzed through load balancing algorithms, and the optimal cache node is determined based on geographical location and network latency information. By routing user requests to the optimal cache node and retrieving optimized personalized recommendation results from that node, the access speed of the optimized personalized recommendation results is improved.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images