Web resource preloading method and device based on user behaviors and medium

Abstract

The invention discloses a Web resource preloading method and device based on user behaviors and a medium. The method comprises the steps that multi-dimensional behavior data of a user in the webpage browsing process are synchronously collected; performing preprocessing and feature extraction on the multi-dimensional behavior data to obtain a multi-dimensional feature vector; fusing the multi-dimensional feature vectors to form a user intention representation vector, and judging the user intention representation vector according to a preset hybrid model to predict the operation intention of the user; determining candidate tasks according to the operation intention, determining priorities corresponding to the candidate tasks, and preloading the candidate tasks according to the priorities; and executing the pre-loaded task through the client, collecting feedback data, and performing model iteration on the hybrid model according to the feedback data to form a closed-loop optimization mechanism. According to the method, the user intention is predicted through multi-dimensional analysis, the task is preloaded, and the experience and efficiency are improved through closed-loop optimization.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a web resource preloading method based on user behavior, device and medium. BACKGROUND

[0002] In the current digital wave, the development of web applications is changing rapidly, and their complexity continues to rise. At the same time, users' requirements for page response speed are becoming increasingly stringent. Under this background, the limitations of traditional static resource loading strategies are increasingly prominent. Their simple and fixed loading mode cannot meet the actual needs of modern web applications that pursue high performance.

[0003] The preloading technology widely used at present improves the resource loading efficiency to some extent, but it mostly relies on manual configuration by developers or judgment based on simple rules. This mode not only consumes a lot of developers' efforts, but also lacks flexibility. More importantly, it fails to fully consider user personalized behavior and cannot dynamically adjust resource loading strategies according to different users' usage habits, network conditions and other factors, making it difficult to optimize resource loading. Therefore, exploring a preloading technology that has a deep understanding of user personalized behavior and can dynamically adapt has become the key to improving the performance of modern web applications. SUMMARY

[0004] To solve the above problems, the present application proposes a web resource preloading method based on user behavior, which includes: synchronously collecting multi-dimensional behavior data of users in the process of browsing web pages, the multi-dimensional behavior data including mouse trajectory, scrolling behavior, click sequence, viewport attention; pre-processing and feature extraction are performed on the multi-dimensional behavior data to obtain a multi-dimensional feature vector; the multi-dimensional feature vector is fused to form a user intention representation vector, and a pre-set hybrid model is used to judge the user intention representation vector to predict the operation intention of the user; candidate tasks are determined according to the operation intention, and the priority of the candidate tasks corresponding to the candidate tasks is determined, and the candidate tasks are preloaded according to the priority; the preloaded tasks are executed through the client, and feedback data is collected, and the hybrid model is iterated according to the feedback data to form a closed-loop optimization mechanism.

[0005] In one example, multi-dimensional behavioral data of the user during web browsing is collected synchronously, specifically including: collecting mouse trajectory data according to a pre-set sampling frequency, the mouse trajectory data including timestamps, screen coordinates, relative movement speed, and acceleration; monitoring vertical and horizontal scrolling events of the page to collect scrolling behavior data, the scrolling behavior data including timestamps, scrolling positions, scrolling directions, and scrolling speeds for each scroll, and determining a scrolling speed curve based on the scrolling behavior to determine the user's browsing mode; recording the context information of each click event to determine click sequence data, the click sequence data including the DOM path signature of the clicked element, click timestamp, mouse coordinates, whether it is a link click, target URL, and page state; and monitoring the visibility status of page elements in real time through an API to determine viewport attention data, the viewport attention data including the element's first visibility time, cumulative visibility duration, and visibility ratio change curve.

[0006] In one example, the preprocessing and feature extraction of the multi-dimensional behavioral data specifically includes: performing differential calculation on the mouse trajectory data to obtain instantaneous velocity and acceleration, and extracting hover behavior features, including hover duration and the number of times elements are entered and exited; performing speed curve analysis on the scrolling behavior data to determine fast scrolling mode and slow scrolling mode, and calculating the scrolling depth percentage; performing feature extraction on the click sequence data to obtain temporal features, which include one-hot encoding of element type, time interval with the previous click, and click position weight; and calculating attention heat values ​​on the viewport attention data to assign attention weights to each candidate target.

[0007] In one example, the multi-dimensional feature vectors are fused to form a user intent representation vector. Specifically, this includes fusing the feature vectors corresponding to the mouse trajectory data, the scrolling behavior data, the click sequence data, and the viewport attention data according to a pre-determined weighted fusion strategy, based on pre-set weights, to generate a unified user intent representation vector.

[0008] In one example, the user intent representation vector is judged according to a pre-set hybrid model to predict the user's operational intent. Specifically, this includes: the hybrid model includes a real-time rule predictor and a deep learning model inferencer; the real-time rule predictor judges the user intent representation vector to generate a preliminary prediction result and a corresponding confidence level; the deep learning model inferencer processes temporal features to generate a predicted probability of operational intent for each candidate target; and the preliminary prediction result and the predicted probability are weighted and fused to generate a final operational intent prediction result.

[0009] In one example, candidate tasks are determined based on the operational intent, and the priority of the candidate tasks is determined. Specifically, this includes: selecting candidate targets whose predicted probability exceeds a preset threshold as preloaded candidate tasks based on the final operational intent prediction result; determining the priority score corresponding to each preloaded candidate task; the priority score includes four dimensions: prediction confidence, resource size weight, network quality coefficient, and cache hit benefit; and sorting the preloaded candidate tasks according to the priority score to form a dynamic priority queue.

[0010] In one example, preloading the candidate tasks according to the priority includes: dividing the preloading operations according to the priority score of the candidate tasks and a preset threshold to implement a hierarchical preloading strategy, wherein the divided preloading operations include DNS pre-resolution and TCP pre-connection, critical resource pre-fetching, and full page pre-rendering.

[0011] In one example, the client executes preloaded tasks and collects feedback data. Specifically, the client performs DNS pre-resolution, TCP pre-connection, resource pre-fetching, and page pre-rendering operations according to the priority specified in the preload instructions to complete the preload task and obtain feedback data. The client monitors the execution of the preload task in real time, records the actual usage of preloaded resources, page loading performance indicators, and user interaction response latency. The collected feedback data is uploaded to the server, where the server cleans, deduplicates, and labels the feedback data, transforming it into a training dataset for model iteration and optimization.

[0012] On the other hand, this application also proposes a web resource preloading device based on user behavior, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the web resource preloading device based on user behavior to perform: the method described in any of the examples above.

[0013] On the other hand, this application also proposes a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured to be the method described in any of the examples above.

[0014] This application collects user behavior data from multiple dimensions, covering mouse trajectories, scrolling, clicks, and more, comprehensively and accurately reflecting user intent. Detailed preprocessing and feature extraction extract key information from diverse data sources, enhancing feature effectiveness. The fusion of multi-dimensional feature vectors forms a unified representation, strengthening the comprehensiveness of intent judgment. A hybrid model combining real-time rules and deep learning balances speed and accuracy, improving the reliability of operation intent prediction. Candidate tasks are selected and prioritized based on operation intent, allocating resources rationally to ensure important tasks are processed first. A hierarchical preloading strategy is flexible and efficient, implementing different preloading operations based on priority to optimize network resource utilization. The client executes tasks and collects feedback data, forming a closed-loop optimization mechanism. The server processes the feedback data into a training set, continuously iterating the model to adapt the preloading method to new situations, improving performance and user experience. Overall, this method forms a complete process from data collection to model optimization, accurately predicting user intent, efficiently preloading resources, and self-optimizing, providing users with a smoother and faster web browsing experience. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments of this application and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart illustrating a web resource preloading method based on user behavior in an embodiment of this application. Figure 2 This is a schematic diagram of the client subsystem architecture in an embodiment of this application; Figure 3 This is a schematic diagram of the service terminal system architecture in an embodiment of this application; Figure 4 This is a schematic diagram of a web resource preloading device based on user behavior, as described in an embodiment of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0018] Current web preloading technologies face several pressing technical challenges: First, regarding the accuracy of user intent recognition, existing technologies primarily rely on simple rule-based judgments, such as judging based on link hover duration. This approach struggles to accurately capture the user's true intent, resulting in low prediction accuracy and a large amount of unused preloaded resources, leading to bandwidth waste and performance degradation. Therefore, a technical solution capable of accurately identifying the user's next action intent is urgently needed. Second, in the fusion and processing of multi-dimensional heterogeneous data, the types of behavioral data generated when users browse web pages are diverse, encompassing mouse trajectories, scrolling events, click sequences, etc. These data have different formats, varying sampling frequencies, and complex temporal relationships. How to effectively align, extract features, and fuse these heterogeneous data to form a unified representation of user behavior is a technical challenge that urgently needs to be overcome. Third, differences in sampling frequencies from different data sources lead to inconsistent data densities, affecting the data fusion effect. A unified sampling standard needs to be established to address the problem of asynchronous data sampling frequencies. Fourth, regarding real-time inference performance on the client side, deep learning models typically have a large number of parameters and high computational complexity, making it difficult to run in real-time in resource-constrained browser environments. Existing solutions either rely entirely on server-side inference, leading to excessive latency, or use oversimplified models, sacrificing accuracy. Achieving millisecond-level client-side inference while ensuring prediction accuracy is a core technical challenge. Fifth, regarding adaptive adjustment of preloading strategies, different users have vastly different network conditions, device performance, and usage scenarios. A uniform preloading strategy cannot adapt to diverse environments. For example, while aggressive preloading is possible in high-speed WiFi environments, the same strategy will severely impact normal browsing in mobile weak network environments. How to dynamically adjust the preloading intensity and scope according to the real-time environment is a problem that must be solved. Sixth, regarding adaptation to rapidly changing user behavior, users' browsing habits and interests change over time, and website content and structure are constantly updated. Staticly trained models will gradually become outdated and ineffective. How to build an online learning mechanism that can quickly adapt to behavioral changes and continuously optimize the model without affecting online services is a key technical issue to ensure the long-term effective operation of the system. Seventh, in terms of quantitative evaluation of preloading effect, a complete indicator system needs to be established to scientifically evaluate the actual effect of intelligent preloading system. This system should include technical indicators such as prediction accuracy, resource hit rate, and latency reduction, as well as business indicators such as user satisfaction and conversion rate improvement. At the same time, a rigorous control experiment method needs to be designed to eliminate the influence of other interfering factors and accurately measure the value of the system.

[0019] like Figure 1 As shown, to solve the above problems, this application provides a method for preloading Web resources based on user behavior, the method including: S101. Synchronously collect multi-dimensional behavioral data of users during web browsing, including mouse trajectory, scrolling behavior, click sequence, and viewport attention.

[0020] Multimodal behavior fusion refers to the process of simultaneously collecting and analyzing various types of interaction data generated by users during web browsing, such as mouse trajectories, scrolling behavior, click sequences, and viewport attention. These heterogeneous data sources are then integrated into a unified user intent representation vector through techniques such as feature extraction, temporal alignment, and weight fusion. Unlike traditional prediction methods based on a single signal source, multimodal fusion can cross-validate user intent from multiple dimensions, significantly improving the accuracy and robustness of predictions. It is the core technological foundation for achieving high-precision behavior prediction.

[0021] In one embodiment, an intelligent preloading system is provided, which adopts a distributed architecture of "end-cloud collaboration," such as... Figure 2 and Figure 3 As shown, it is mainly divided into two parts: the client subsystem and the service subsystem.

[0022] In one embodiment, the data acquisition layer is the foundation of the entire system, responsible for comprehensively capturing user behavioral data generated during web browsing. This layer employs a multi-sensor fusion strategy to characterize user behavior from different dimensions.

[0023] The mouse trajectory acquisition module continuously records the user's mouse movement trajectory on the page at a high frequency, for example, a sampling interval of 100-200 milliseconds. For each sampling point, the system records a timestamp, screen coordinates (x, y), relative movement speed (vx, vy), and the DOM element identifier below the mouse pointer. By calculating the positional difference between adjacent sampling points, the instantaneous speed and acceleration of the mouse can be obtained. These parameters can reflect the strength of the user's intention. When the mouse hovers over an interactive element, such as a link or button, for more than a preset threshold, such as 800-1000 milliseconds, the system marks it as an "intentional hovering" event. Research shows that users typically have a 0.5-2 second hovering preparation period before clicking, which is the golden window for preloading. In addition, the system also records the number of times the mouse enters and leaves an element. If the user enters and leaves the same element multiple times, it suggests hesitation, and the prediction confidence should be appropriately reduced in this case.

[0024] The scrolling behavior acquisition module monitors vertical and horizontal scrolling events on the page, recording the timestamp, scroll position, scroll direction, and scroll speed for each scroll. By analyzing the scroll speed curve, it can identify the user's browsing pattern: fast scrolling usually indicates that the user is quickly browsing to find target content, while slow scrolling or frequent pauses indicate that the user is carefully reading the current content. The system also has a scroll pause detection mechanism. When scrolling stops for more than 500 milliseconds, the system captures all interactive elements in the current viewport and records the relative position of each element in the viewport. Elements in the central area of ​​the viewport receive higher attention weight. In addition, the system calculates the user's scroll depth percentage, a metric that is of significant reference value for predicting whether the user will continue scrolling down or return to the top of the page.

[0025] Click sequences are the most direct signals reflecting a user's navigation intent. The system's click sequence collection module records complete contextual information for each click event, including the DOM path signature of the clicked element, click timestamp, mouse coordinates, whether it was a link click, target URL, element text content, and the page state at the time of the click, such as scroll position and page dwell time. For link clicks, the system specifically records the relationship between the source and target pages to construct the user's page navigation map. By analyzing historical navigation sequences, common browsing path patterns can be discovered. For example, on e-commerce websites, users typically follow the path of "homepage → category page → product details page → shopping cart". Identifying the user's current stage allows for more accurate prediction of their next action.

[0026] The viewport attention module utilizes the IntersectionObserver API to monitor the visibility of page elements in real time. When an interactive element enters the viewport and reaches a certain visibility percentage, such as 50%, 75%, or 100%, the system records the element's first visibility time, cumulative visibility duration, and visibility percentage change curve. Psychological research indicates that users' attention allocation follows an "F-shaped" or "Z-shaped" reading pattern, with elements located at the top or left of the page more likely to attract attention. The system combines the element's spatial location and visibility duration to calculate an attention heatmap value for each candidate target. Generally, elements that remain in the viewport for a longer period are more likely to be clicked.

[0027] S102. Preprocess and extract features from the multi-dimensional behavioral data to obtain a multi-dimensional feature vector.

[0028] The feature processing layer plays a crucial role in transforming raw behavioral data into structured feature vectors that can be understood by machine learning models.

[0029] In terms of temporal feature extraction, for click sequences, the system extracts temporal features from the most recent N clicks, where N is 10. Each click is encoded into a multi-dimensional vector, specifically covering one-hot encoding of the element type, a binary label indicating whether it is a link, the time interval between the previous click, the percentage of scroll depth at the time of the click, and the positional weight of the click in the sequence. The time interval feature undergoes a logarithmic transformation to compress the dynamic range, ensuring that both short and long time intervals are represented appropriately. For mouse trajectories, the system extracts statistical features from the most recent M sampling points, where M is between 50 and 100, such as average speed, maximum speed, speed standard deviation, frequency of direction changes, and percentage of hovering occurrences. These features characterize the smoothness and determinism of user operations; for example, fast and smooth mouse movement often indicates a clear user objective, while a slow and zigzag trajectory may suggest the user is hesitant or exploratory.

[0030] Statistical feature calculations provide a global view of the current session. The system calculates a range of features, including but not limited to total session duration, total clicks, average page dwell time, total scroll distance, number of scrollbacks, and total hover events. These features help the model understand the user's overall behavioral patterns and activity level. Of particular note is the system's calculation of local statistical features based on the sliding window, such as click frequency in the last 5 minutes and the average speed of the last 3 scrolls. Local features can capture dynamic changes in user behavior patterns, such as a user transitioning from a fast browsing mode to a careful reading mode.

[0031] Contextual feature encoding includes the current page state and environmental information. Page context includes the current URL, page type, page depth, and the number of interactive elements within the viewport. Page type includes homepage, list page, detail page, etc., and page depth is the number of navigations from the entry page. Device context includes screen resolution, viewport size, device type, and current time period. Network context is a crucial decision-making factor. The system obtains network type (e.g., 4G, WiFi), downlink bandwidth estimate, round-trip time (RTT), and other parameters through the NetworkInformationAPI. In low-speed network environments, the system adopts a more conservative preloading strategy to avoid wasting limited bandwidth resources.

[0032] For each interactive element on the current page, especially links, the system extracts its features for prediction. Element features include their position and depth in the DOM tree, distance from the top of the viewport, whether they are currently visible, their visibility duration, the semantic vector of the text content, and historical click-through rates. In addition, the system analyzes the characteristics of the target URL, including URL path depth, whether it contains specific keywords, and whether it is an internal link. Pages known to be frequently accessed are given higher preloading priority.

[0033] S103. The multi-dimensional feature vectors are fused to form a user intent representation vector. The user intent representation vector is judged according to a pre-set hybrid model to predict the user's operation intent.

[0034] In one embodiment, the prediction engine layer serves as the core intelligent component of the system and employs a multi-layered hybrid prediction architecture.

[0035] This lightweight, real-time rule predictor makes rapid initial judgments based on heuristic rules with extremely low latency, less than 10 milliseconds. It provides numerous rule examples; for instance, if the mouse hovers over a link for more than 1 second, the probability of that link being clicked is predicted to be 0.6; if a user scrolls down at a constant speed and is about to reach the bottom of the page, the next page will be loaded; if a user quickly scrolls down a search results page, one of the first three results will be clicked. This predictor is suitable for scenarios with clear user intent and can quickly trigger preloading. When the rule predictor's confidence exceeds the threshold of 0.7, the system will directly perform preloading without waiting for the results from the deep learning model.

[0036] The deep learning model inference engine runs in the WebAssembly environment to execute more complex deep neural network models. The model architecture employs a sequence-to-sequence design. The encoder uses LSTM or GRU networks to process temporal data such as click sequences and mouse trajectories to capture the temporal dependencies of user behavior. The decoder combines the encoder output and contextual features to generate click probability predictions for each candidate target. To ensure efficient operation in the browser, the model employs several optimization techniques, including model quantization (compressing 32-bit floating-point numbers into 8-bit integers), pruning (removing unimportant connections), and knowledge distillation (training a smaller model with a larger one). After optimization, the model size is controlled between 2 and 5 MB, and the inference latency is controlled within 50 to 100 milliseconds.

[0037] The system weights and fuses the outputs of the rule predictor and the deep model. The fusion formula is: the final probability equals α multiplied by the rule probability, plus β multiplied by the model probability, plus γ multiplied by the historical statistical probability. The weighting coefficients α, β, and γ are dynamically adjusted based on historical performance. When the model prediction matches the rule prediction, the system increases the confidence level; when they contradict each other, the system decreases the confidence level or adopts a conservative strategy. Furthermore, the confidence level assessment considers the quality of the input data. If user behavior data is sparse, such as at the beginning of a new session, or if there are anomalous features, such as detected automated scripts, the system will reduce the confidence level of the prediction results.

[0038] S104. Determine candidate tasks according to the operation intention, determine the priority of the candidate tasks, and preload the candidate tasks according to the priority.

[0039] In one embodiment, the scheduling decision layer formulates the optimal preloading execution plan based on the prediction results and the system resource status.

[0040] The system calculates a comprehensive priority score for each preloaded candidate task. The specific formula is: priority score equals w1 multiplied by the predicted probability, plus w2 multiplied by the resource size weight, plus w3 multiplied by the network quality coefficient, and finally plus w4 multiplied by the cache hit benefit. Here, the resource size weight is inversely proportional to the resource size, meaning smaller resources are loaded first; the network quality coefficient reflects the current bandwidth situation, and task priority is reduced when bandwidth is low; the cache hit benefit considers the probability of the resource being accessed repeatedly. The system maintains a dynamic priority queue, sorting all pending tasks according to their scores. The queue is updated in real time whenever a new prediction result is generated. For tasks already being loaded, the system assigns them a higher hold priority to avoid frequent task cancellations and resource waste.

[0041] The system continuously monitors network conditions, including metrics such as available bandwidth, latency, and packet loss rate. By downloading probe packets or analyzing the performance data of existing requests, the system can assess the current network quality level, categorizing it into four levels: excellent, good, average, and poor. Different preloading strategies are employed at different network quality levels: In excellent network quality environments, such as high-speed WiFi, the system actively preloads, prefetching multiple candidate targets; in good network quality environments, such as 4G, the system moderately preloads, prioritizing high-probability targets; in average network quality environments, such as 3G, the system adopts a conservative preloading strategy, loading only critical resources; and in poor network quality environments, such as 2G or weak signal environments, the system disables preloading or only performs DNS pre-connection. Furthermore, the system also recognizes the user's data-saving mode settings, significantly reducing or stopping preloading activities in this mode.

[0042] The system maintains a local cache mapping table, which records metadata such as the URL, size, cache duration, and access frequency of cached resources. When deciding whether to preload, the system prioritizes resources that have not yet been cached. For cached resources that are nearing expiration, the system can perform a background refresh operation. In terms of cache capacity management, the system employs an improved LRU (Least Recently Used) strategy, combining access frequency and predicted probability for eviction decisions. High-frequency access resources that may be used again in the future are protected, while low-value resources are prioritized for eviction.

[0043] S105. Execute the preloaded task through the client and collect feedback data. Iterate the hybrid model based on the feedback data to form a closed-loop optimization mechanism.

[0044] In one embodiment, the execution layer is responsible for translating scheduling decisions into actual browser preloading operations.

[0045] The system employs a three-tiered progressive preloading strategy based on predicted confidence levels. The first tier involves DNS pre-resolution and TCP pre-connection, triggered when the confidence level exceeds 30%. This is achieved through...<linkrel="dns-prefetch"> and<linkrel="preconnect"> First, the target domain name is resolved and a TCP connection is established in advance. This operation has minimal overhead but can significantly reduce the latency of subsequent requests, especially suitable for cross-domain resources. The second level is critical resource prefetching, which is carried out when the confidence level is greater than 50%. Utilizing...<linkrel="prefetch"> The system prefetches key resources from the target page, such as JavaScript, CSS, and font files. This is done by analyzing the target page's HTML structure or querying a resource database. The prefetching operation uses low-priority network requests to avoid interfering with the normal loading of the current page. The third level is full page pre-rendering, performed when the confidence level is greater than 70% and the resource size is less than 2MB.<linkrel="prerender"> The target page is fully rendered in the background. Pre-rendering is a more aggressive optimization method, allowing the rendered page to be displayed instantly when the user clicks the link. However, due to its high resource consumption, the system strictly limits its use to only high-confidence predictions, small-sized pages, and good network conditions.

[0046] The system maintains a preloaded task queue and uses a priority scheduling algorithm to manage task execution. To prevent excessive consumption of network bandwidth and computing resources, the system limits the number of concurrent preloaded tasks, and a setting of 2 to 4 is recommended. Each task has a timeout mechanism and a cancellation mechanism. If a preloaded task is not completed for a long time, or if the user's behavior indicates a prediction error, such as scrolling to another area, the system will cancel the task to release resources. For partially completed tasks, the system will retain the downloaded resource segments to avoid starting the download from the beginning again next time.

[0047] The system also implements an adaptive rate limiting protection mechanism. By monitoring performance metrics such as page frame rate, memory usage, and CPU utilization, when performance degradation is detected, the system automatically reduces the aggressiveness of preloading or temporarily stops preloading to ensure that preloading optimization does not harm the user experience. For low-end devices or resource-constrained environments, the system dynamically adjusts its strategy based on device capability scores. For example, on devices with less than 2GB of memory, page pre-rendering is disabled, and only resource prefetching is performed.

[0048] In one embodiment, the system incorporates a complete performance evaluation mechanism to form a closed-loop optimization process.

[0049] The system continuously monitors and collects key performance indicators, including prediction accuracy (the percentage of actual clicked targets in Top-K predictions); resource hit rate (the proportion of preloaded resources actually used); reduction in page load time (derived by comparing with a baseline where preloading is disabled); bandwidth utilization efficiency (the ratio of effective traffic to total traffic); and user interaction response latency (the time from user click to a page becoming interactive). The system performs segmented analysis on these indicators according to different dimensions such as time, page type, and user group to identify performance differences in different scenarios.

[0050] New model versions or strategy adjustments need to be validated through A / B testing. The system randomly assigns users to experimental and control groups. The experimental group uses the new version, while the control group uses the old version. Statistical methods are then used to determine if the improvement is significant. Model updates employ a phased rollout strategy, initially pushing the new model to a small percentage of users, such as 1%. After observing no abnormalities in key metrics, the rollout is gradually expanded to 5%, 10%, and 50%, eventually achieving a full rollout. If the new model performs poorly, the system can quickly roll back to a stable version.

[0051] The server periodically, such as daily or weekly, retrains the model using the latest collected user behavior data. The training data undergoes privacy processing and anonymization, retaining only behavioral pattern information. After thorough offline validation, the new model is distributed to clients globally via a CDN network. For personalized scenarios, the system also supports client-side incremental learning. The client-side model is slightly adjusted based on local user feedback data to make predictions more aligned with individual habits. Incremental learning employs lightweight algorithms to avoid excessive consumption of device resources.

[0052] In one embodiment, the data processing platform of the server subsystem undertakes the important task of receiving massive amounts of user behavior data reported by clients and transforming them into high-quality training datasets.

[0053] The raw reported data contains a large amount of noise and outliers, requiring systematic cleaning. First, data integrity verification is performed, removing records with missing fields or incorrect formats. Next, anomaly detection is conducted to identify and filter abnormal behavioral patterns, such as extremely high-frequency clicking, which is likely caused by web crawlers or automated scripts; abnormal mouse trajectories, such as moving in a straight line to a pixel-precise location; and unreasonable timestamps, such as future times or time reversals. The system also implements a deduplication mechanism. Due to network instability or client retries, the same data may be reported multiple times. Deduplication is performed using combinations of session ID, timestamp, and event fingerprint to ensure that only one record is retained for each genuine event. Privacy protection is a critical aspect of data processing. The system anonymizes the reported data, including hashing user identifiers, removing the last segment of IP addresses, deleting URL parameters that may contain sensitive information, and obfuscating precise timestamps, retaining them at the hour level. While the anonymized data cannot be traced back to specific individuals, it retains behavioral pattern information.

[0054] After cleaning, the data enters the feature engineering stage, transforming it into feature vectors required for model training. This module implements feature extraction logic consistent with the client's feature processing layer, ensuring alignment of the feature spaces for training and inference. In addition to basic features, the server can also calculate global features that are difficult for the client to obtain, such as the average click-through rate of a URL among all users, the typical browsing duration of a specific page type, behavioral differences among users in different regions, and traffic patterns at different times. These global statistical features are injected into the model as supplementary information, helping to improve the model's generalization ability. Feature storage adopts a columnar storage format, such as Parquet, supporting efficient column selection and predicate pushdown, which accelerates feature retrieval. Simultaneously, feature versioning management ensures the reproducibility of model training, explicitly recording the feature version used in each training iteration.

[0055] Supervised learning requires labeled data. For user behavior prediction tasks, labeling is relatively straightforward; the user's actual click behavior serves as the positive sample label. The system automatically generates training sample pairs. The input is the sequence of behaviors and page state before the click, and the output is the clicked target element. To train the model to learn discriminative features, the system also generates negative samples, including elements that the user did not click but were visible, and elements that the user briefly hovered over but ultimately abandoned. The design of the positive and negative sample ratio affects model performance; the system employs a dynamic sampling strategy, adjusting the intensity of hard example mining during training. For some ambiguous scenarios, the system introduces a manual annotation process. For example, when a user quickly switches between two links and ultimately chooses one, determining the user's true intent requires more context. The manual annotation team reviews these boundary cases, providing high-quality supervision signals to the model.

[0056] The processed data is stored in tiers. Hot data, representing the most recent week's data, is stored in a high-speed SSD cluster, supporting frequent random access for online learning and real-time analysis. Warm data, representing the most recent month's data, is stored in a regular disk array for model training. Cold data, representing compressed historical archive data, is stored in an object storage service and retained for long-term trend analysis and model backtesting. Data is partitioned according to time and page type to improve query efficiency. The distributed storage system provides multi-replica redundancy to ensure data reliability. Data retention policies comply with privacy regulations; personal data exceeding the retention period is automatically anonymized or deleted.

[0057] In one embodiment, the model training platform is the intelligent core of the system, responsible for learning user behavior patterns from data.

[0058] The offline training system runs on a high-performance GPU cluster, periodically training new models using the full historical data. The training process begins with data loading and preprocessing, reading training data from distributed storage and then batch processing it. To expand the training samples, data augmentation techniques are employed, such as time jitter (slightly adjusting timestamps), trajectory interpolation (inserting midpoints into mouse trajectories), and feature noise injection (simulating sensor errors). These techniques help improve model robustness and avoid overfitting. In terms of model architecture, the system adopts an encoder-decoder architecture. The encoder uses a multi-layer bidirectional LSTM network, taking as input a sequence of temporal behavioral features, including click sequences, mouse trajectories, and scrolling events, and outputting a fixed-dimensional behavioral representation vector. An attention mechanism allows the model to focus on key events in the sequence. The decoder uses a fully connected neural network, taking as input the behavioral representation vector, candidate target features, and contextual features, and outputting the click probability of each candidate target. For the loss function, the system uses cross-entropy loss as a base, incorporating multiple regularization terms. L2 regularization prevents excessive weights, and Dropout randomly deactivates some neurons to prevent co-adaptation. To address the class imbalance problem, since clicks are sparse events, the system uses FocalLoss to reduce the weight of easily classified samples, allowing the model to focus more on difficult samples. During hyperparameter optimization, the system uses Bayesian optimization or grid search methods to find the optimal combination of hyperparameters, covering learning rate, batch size, number of network layers, and number of hidden units. A learning rate decay strategy is employed during training, enabling rapid learning initially and fine-tuning later. An early stopping mechanism monitors validation set performance, stopping training when performance no longer improves to avoid overfitting. Large-scale model training employs data parallelism or model parallelism, distributing training tasks across multiple GPU nodes. Gradient accumulation allows for small batch sizes, reducing memory requirements. Mixed-precision training uses a combination of FP16 and FP32 computations to accelerate training while maintaining numerical stability.

[0059] Online learning systems process newly arriving data streams, rapidly updating model parameters to adapt to changes in user behavior, making them more lightweight and agile than offline training. The system maintains a sliding window buffer, storing user behavior data from the past few hours. When the buffer reaches a threshold, incremental training is triggered. Incremental training does not start with random initialization but continues optimizing from the current model parameters, using only a small number of iterations with new data. To avoid catastrophic forgetting—the model overfitting to new data and forgetting old knowledge—the system employs Elastic Weight Consolidation (EWC) or Experience Replay techniques. EWC imposes stronger regularization constraints on important parameters, preventing significant changes. Experience Replay mixes in some historical samples during incremental training, maintaining a memory of older data. Online learning also supports reinforcement learning paradigms, where pre-loaded decisions are modeled as sequential decision problems. Rewards or penalties are given based on actual effects, such as improved page load time and user satisfaction, and policy gradient algorithms are used to optimize the decision policy.

[0060] Newly trained models cannot be deployed directly; they must undergo rigorous validation. The validation process begins with offline evaluation, assessing model performance on a reserved test set. Metrics such as accuracy, recall, F1 score, and AUC-ROC are calculated, comparing the new model with the baseline model (the current online version) to determine if there is a significant improvement. Offline evaluation also includes segmented scenario analysis, examining the model's performance across different page types, user groups, and time periods. Next, simulated replay testing is conducted using historical real traffic data. The new model makes predictions for historical scenarios, and the system evaluates the potential impact of using the new model at that time. Replay testing identifies issues with the model in extreme or boundary conditions. Finally, online A / B testing is performed. Using an A / B testing framework, a small percentage of real users (e.g., 1%) are randomly assigned to the new model group, while the remaining users continue using the old model. The system monitors key metrics for both groups in real time, including prediction accuracy, page load time, bounce rate, and conversion rate. Statistical significance testing ensures that observed differences are not random fluctuations. The new model will also run in shadow mode, i.e., shadow testing. It receives real requests but its predictions are not actually executed; they are only used for logging and performance monitoring. Shadow testing verifies the model's stability and latency performance in a production environment and helps identify potential engineering problems. Furthermore, to understand the model's decision-making logic, the system implements interpretability tools: SHAP value analysis reveals the contribution of each input feature to the prediction result; attention weight visualization shows which parts of the behavioral sequence the model focuses on; case studies select typical prediction scenarios; and manual review allows for the rationality of the model's decisions.

[0061] like Figure 4 As shown in the illustration, this application also provides a web resource preloading device based on user behavior, comprising: At least one processor; and, A memory that is communicatively connected to at least one processor; wherein, The memory stores instructions that can be executed by at least one processor to enable a user-behavior-based web resource preloading device to perform the method as described in any of the embodiments above.

[0062] This application also provides a non-volatile computer storage medium storing computer-executable instructions, which are configured as described in any of the above embodiments.

[0063] In the 1990s, improvements to a technology could be clearly distinguished as either hardware improvements (e.g., improvements to the circuit structure of diodes, transistors, switches, etc.) or software improvements (improvements to the methodology). However, with technological advancements, many methodological improvements today can be considered direct improvements to the hardware circuit structure. Designers almost always obtain the corresponding hardware circuit structure by programming the improved methodology into the hardware circuit. Therefore, it cannot be said that a methodological improvement cannot be implemented using hardware physical modules. For example, a Programmable Logic Device (PLD) (such as a Field Programmable Gate Array (FPGA)) is such an integrated circuit whose logic function is determined by the user programming the device. Designers can program a digital system themselves to "integrate" it onto a PLD, without needing chip manufacturers to design and manufacture dedicated integrated circuit chips. Moreover, nowadays, instead of manually manufacturing integrated circuit chips, this programming is mostly implemented using "logic compiler" software, which is similar to the software compiler used when writing program development code. The original code before compilation must also be written in a specific programming language, which is called a Hardware Description Language (HDL). There is not just one HDL, but many kinds, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, RHDL (Ruby Hardware Description Language), etc. Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art should also understand that by simply performing some logic programming on the method flow using the aforementioned hardware description languages ​​and programming it into an integrated circuit, the hardware circuit that implements the logic method flow can be easily obtained.

[0064] The controller can be implemented in any suitable manner. For example, it can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicon Labs C8051F320. A memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also recognize that, in addition to implementing the controller in purely computer-readable program code form, the same functionality can be achieved by logically programming the method steps to make the controller take the form of logic gates, switches, ASICs, programmable logic controllers, and embedded microcontrollers. Therefore, such a controller can be considered a hardware component, and the means included therein for implementing various functions can also be considered as structures within the hardware component. Alternatively, the means for implementing various functions can be considered as both software modules implementing the method and structures within the hardware component.

[0065] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0066] For ease of description, the above devices are described in terms of function, divided into various units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware.

[0067] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0068] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0069] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0070] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0071] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0072] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0073] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0074] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0075] Computer-readable media include both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0076] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0077] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for preloading Web resources based on user behavior, characterized in that, include: Simultaneously collect multi-dimensional behavioral data of users during web browsing, including mouse trajectory, scrolling behavior, click sequence, and viewport attention. The multi-dimensional behavioral data is preprocessed and features are extracted to obtain a multi-dimensional feature vector; The multi-dimensional feature vectors are fused to form a user intent representation vector. The user intent representation vector is then judged according to a pre-set hybrid model to predict the user's operational intent. Candidate tasks are determined based on the operational intent, and the priority of each candidate task is determined. The candidate tasks are then preloaded according to the priority. The preloaded tasks are executed by the client, and feedback data is collected. The hybrid model is then iterated based on the feedback data to form a closed-loop optimization mechanism.

2. The method according to claim 1, characterized in that, Synchronously collect multi-dimensional behavioral data of users during web browsing, specifically including: Mouse trajectory data is collected according to a pre-set sampling frequency. The mouse trajectory data includes timestamps, screen coordinates, relative movement speed, and acceleration. The system monitors vertical and horizontal scrolling events on the page to collect scrolling behavior data, including the timestamp of each scroll, scrolling position, scrolling direction, and scrolling speed. Based on the scrolling behavior, a scrolling speed curve is determined to identify the user's browsing pattern. Record the context information of each click event to determine the click sequence data, which includes the DOM path signature of the clicked element, the click timestamp, the mouse coordinates, whether it is a link click, the target URL, and the page state; The visibility status of page elements is monitored in real time via API to determine viewport attention data, which includes the first visibility time of the element, cumulative visibility duration, and visibility ratio change curve.

3. The method according to claim 2, characterized in that, The preprocessing and feature extraction of the multi-dimensional behavioral data specifically includes: The mouse trajectory data is differentially calculated to obtain instantaneous velocity and acceleration, and hovering behavior features are extracted, including hovering duration and the number of times the element enters or leaves. Speed ​​curve analysis is performed on the scrolling behavior data to determine fast scrolling mode and slow scrolling mode, and the percentage of scrolling depth is calculated; Feature extraction is performed on the click sequence data to obtain temporal features, which include one-hot encoding of element type, time interval with the previous click, and click position weight; Attention heatmap values ​​are calculated on the viewport attention data to assign attention weights to each candidate target.

4. The method according to claim 1, characterized in that, The multi-dimensional feature vectors are fused to form a user intent representation vector, specifically including: According to a predetermined weighted fusion strategy, the feature vectors corresponding to the mouse trajectory data, the scrolling behavior data, the click sequence data, and the viewport attention data are fused according to pre-set weights to generate a unified user intent representation vector.

5. The method according to claim 1, characterized in that, The user intent representation vector is judged based on a pre-set hybrid model to predict the user's operational intent, specifically including: The hybrid model includes a real-time rule predictor and a deep learning model inferencer. The real-time rule predictor judges the user intent representation vector to generate preliminary prediction results and corresponding confidence scores. The deep learning model inferencer processes temporal features to generate a predicted probability of operational intent for each candidate target; The preliminary prediction results and the predicted probabilities are weighted and fused to generate the final prediction result of the operational intent.

6. The method according to claim 5, characterized in that, Based on the stated operational intent, candidate tasks are determined, and the priorities corresponding to these candidate tasks are determined, specifically including: Based on the final operation intention prediction results, candidate targets with prediction probabilities exceeding a preset threshold are selected as preloaded candidate tasks. The priority score corresponding to each preloaded candidate task is determined. The priority score includes four dimensions: prediction confidence, resource size weight, network quality coefficient, and cache hit benefit. The preloaded candidate tasks are sorted according to the priority scores to form a dynamic priority queue.

7. The method according to claim 1, characterized in that, Preloading the candidate tasks according to the priority includes: The preloading operations are divided according to the priority scores of candidate tasks and preset thresholds to implement a hierarchical preloading strategy. The preloading operations after division include DNS pre-resolution and TCP pre-connection, critical resource pre-fetching, and full page pre-rendering.

8. The method according to claim 7, characterized in that, The preloaded tasks are executed via the client, and feedback data is collected, specifically including: The client performs DNS pre-resolution, TCP pre-connection, resource pre-fetching, and page pre-rendering operations according to the priority based on the pre-loading instruction to complete the pre-loading task and obtain feedback data; Monitor the execution of preloading tasks in real time, and record the actual usage of preloaded resources, page loading performance metrics, and user interaction response latency; The collected feedback data is uploaded to the server, where it is cleaned, deduplicated, and labeled to transform it into a training dataset for model iteration and optimization.

9. A web resource preloading device based on user behavior, characterized in that, include: At least one processor; as well as, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor to enable the user-behavior-based Web resource preloading device to perform the method as described in any one of claims 1-8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are configured to be the method as described in any one of claims 1-8.

Images