Edge intelligence experience implementation method and apparatus, and medium and device

Abstract

Provided in the present application are an edge intelligence experience implementation method and apparatus, and a medium and a device, which are used for performing user experience perception, decision-making and handling on a front end for a front-end code suitable for various types of clients. The method comprises: starting a client, sending an edge intelligence experience configuration request to a server, acquiring an edge intelligence experience configuration and storing same locally; and running an edge intelligence kit in a front-end code, comprising: starting a perception engine in the edge intelligence kit, so as to acquire target data in real time; triggering a decision-making engine in the edge intelligence kit, so as to perform experience decision-making for the target data on the basis of the edge intelligence experience configuration; and running a contact engine in the edge intelligence kit, so as to trigger a corresponding handling scheme on the basis of the experience decision-making result. The embodiments of the present application can realize front-end intellectualization.

Description

End smart experience implementation method, device, medium and equipment TECHNICAL FIELD

[0001] The present application relates to the technical field of application development, in particular to an end smart experience implementation method, device, medium and equipment. BACKGROUND

[0002] Users use terminals (such as smart phones) to work, communicate, shop, etc. through various clients (such as APPs, web pages, websites, mini programs, etc.), which is an important part of modern life. The instant experience of users when using the client may involve interface design, page loading, function response, ease of use, smoothness, user feedback and a series of functions. How to improve application development and improvement based on improving user experience is a technical problem that needs to be considered by those skilled in the art.

[0003] SUMMARY

[0004] Therefore, the present application provides an end smart experience implementation method, device, medium and electronic equipment, which mainly aims to improve user experience by implementing end smart.

[0005] According to one aspect of the present application, an end smart experience implementation method is provided for sensing, decision-making and handling of user experience in the front end for front-end code applicable to multiple types of clients. The method is applied to one of any type of client in multiple types of clients, and includes:

[0006] Starting the client, sending an end smart experience configuration request to the server, obtaining the end smart experience configuration and storing it locally;

[0007] Running the end smart kit in the front-end code, including: starting the perception engine in the end smart kit, obtaining target data in real time; triggering the decision engine in the end smart kit, making experience decisions for the target data according to the end smart experience configuration; and running the touch engine in the end smart kit, triggering the corresponding handling scheme according to the experience decision result.

[0008] According to another aspect of the present application, an end smart experience implementation device is provided for sensing, decision-making and handling of user experience in the front end for front-end code applicable to multiple types of clients. The device is applied to one of any type of client in multiple types of clients, and includes:

[0009] The configuration request unit is configured to start the client, send an end smart experience configuration request to the server, obtain the end smart experience configuration and store it locally;

[0010] The end intelligent execution unit is configured to run an end intelligent suite in the front-end code, including: starting a perception engine in the end intelligent suite to obtain target data in real time; triggering a decision engine in the end intelligent suite to make experience decisions for the target data according to the end intelligent experience configuration; and running a touch engine in the end intelligent suite to trigger a corresponding treatment scheme according to the experience decision result.

[0011] According to an aspect of the present application, a storage medium is provided, and the storage medium stores a computer program. The computer program is configured to execute the end intelligent experience implementation method when running.

[0012] According to an aspect of the present application, an electronic device is provided, including a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to execute the end intelligent experience implementation method. BRIEF DESCRIPTION OF DRAWINGS

[0013] FIG. 1 shows a flowchart of an end intelligent experience implementation method according to an embodiment of the present application;

[0014] FIG. 2 shows an end intelligent experience treatment scheme according to an embodiment of the present application;

[0015] FIG. 3 shows a flowchart of an end intelligent experience monitoring method according to an embodiment of the present application;

[0016] FIG. 4 shows a complex event configuration and matching logic diagram in end intelligent experience monitoring according to an embodiment of the present application;

[0017] FIG. 5 shows a timing matching flowchart in end intelligent experience monitoring according to an embodiment of the present application;

[0018] FIG. 6 shows a flowchart of an end intelligent experience layering method according to an embodiment of the present application;

[0019] FIG. 7 shows an end intelligent experience layering implementation logic diagram according to an embodiment of the present application;

[0020] FIG. 8 shows a flowchart of an end intelligent experience enhancement method according to an embodiment of the present application;

[0021] FIG. 9 shows a flowchart of another end intelligent experience implementation method according to an embodiment of the present application;

[0022] FIG. 10 shows a structure diagram of an end intelligent experience implementation device according to an embodiment of the present application. DETAILED DESCRIPTION

[0023] In order for those skilled in the art to better understand the scheme of the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application.

[0024] Each type of client has a user experience problem, so how to improve the user experience is a subject that needs continuous research in application development.

[0025] Traditional HTML 5 application development is efficient, but the user performance experience is not good enough, so attempts have been made to develop web applications with similar user experience to native applications using web technology. The emergence of MiniApp has quickly won the favor of users. MiniApp is a small, installation-free, fast-loading program that usually runs in a host application or operating system (such as MiniApp, fast application), or a JS native application that supports cross-end deployment. Due to the limitations of the architecture, it is difficult for MiniApp itself to achieve a smooth experience comparable to native applications, especially for large MiniApps. The dynamic experience degradation during user operation is a difficult problem. Since the user experience problem of MiniApp is more prominent, the embodiments of the present application take MiniApp as an example for illustration, but it can be understood that it is applicable to other types of clients.

[0026] In existing schemes, the front-end collects data, the server calculates data and feeds back to the front-end to improve user experience. This method can be referred to as a server-side intelligence or cloud-side intelligence (cloud-side AI) scheme. Although it has the characteristic of strong data processing capability, it has the limitations of high delay and high maintenance cost.

[0027] Therefore, the present inventors innovatively propose to improve user experience through end-side intelligence (also known as "terminal-side intelligence" or "front-end intelligence", hereinafter referred to as "end-side intelligence"). End-side intelligence can be understood as having the ability to think and make decisions on the end side. In short, end-side intelligence is real-time perception, calculation, decision-making and intervention on the end side. By giving the end machine learning capabilities, it brings user experience improvement. The end-side intelligence experience in the embodiments of the present application deeply understands and understands user behavior through data analysis and machine learning, globally detects and coordinates various user device links and complex running scenarios, and through the implementation of feedback loops and timely adjustment and intervention, it is a comprehensive solution to optimize the user experience of client products. It has the characteristics of immediacy, intelligence and disposability.

[0028] Through analysis and practice, the end intelligence has outstanding technical advantages: 1, real-time: the real-time on the end is difficult to match on the cloud, for many interactive and streaming scenarios, the delay on the cloud is almost unable to meet the real-time experience demand; 2, offline capability: although 4G and 5G have been popular, in actual life, such as on a car, an airplane, or in a suburb, network connection is often unstable or even unavailable, at this time, the end intelligence can well avoid the problem of poor network; 3, data privacy: due to the requirement of data privacy, many user data cannot be uploaded to the cloud, so can only be used on the end; 4, cost: although cloud computing provides massive computing power, the cost of AI computing and network bandwidth is very high for a large number of users. At present, the popularity of end-side devices, such as a large number of mobile phones, smart cars, set-top boxes and other devices with computing power, can be used to reduce the cost of cloud deployment.

[0029] The present inventors continue to study and find that there are several challenges in realizing end intelligence. The first is the package size challenge. The end-side technology involves the installation package size and storage resources of the APP, which puts forward higher requirements for the ability of end computing engine slimming, end model size, end model compression, dynamic publishing, etc. The second is the performance challenge. Since the computing resources of mobile devices are used, the balance between user experience and computing efficiency needs to be considered, and continuous optimization needs to be carried out for mobile end device CPU / GPU, memory, power consumption, instruction set, thread scheduling, etc. The third is the adaptation challenge. In the face of device and system version fragmentation, end intelligent technology needs to be adapted a lot to ensure the stability and usability of runtime. Therefore, the embodiment of the present application proposes an end intelligent experience implementation scheme, which aims to overcome the defects of cloud AI and optimize the implementation effect of the scheme from the above challenges.

[0030] In addition, in existing network applications, the same network product often corresponds to different types of clients (different channel releases), for example, a takeout product has multiple types of client forms such as APP, small program, website, and web page. For the front-end code of the same product, if different end intelligent experience mechanisms are adopted for various different types of clients, it will have a great impact on code development and release efficiency. Therefore, the embodiment of the present application also further improves in this aspect in order to realize a set of end intelligent experience implementation scheme applicable across ends. In order to achieve a scheme applicable to multiple types of clients, the embodiment of the present application relies on a programming language (such as JavaScript, JS) applicable across ends for front-end code development. Since JS development is compatible with APP, small program, H5, and various types of clients, the cross-end purpose can be achieved. In the specific implementation process, it also needs to consider processing the source code to remove the differences between various types of clients and other problems.

[0031] Referring to FIG. 1, it is a flowchart of an implementation method of an end user experience in an embodiment of the present application. The method is used for the perception, decision and handling of user experience in the front end for the front-end code applicable to multiple types of clients. The method is applied to any one of the multiple types of clients, and includes the following steps S101-S102.

[0032] In an embodiment of the present application, the front-end code (for example, JavaScript (JS) front-end code) applicable to multiple types of clients (also known as "cross-end") is obtained by statically compiling the source code to remove the differences between various types of clients. Static compilation refers to the compiler extracting part of the corresponding static library that needs to be called by the executable file when compiling the executable file, and linking it to the executable file, so that the executable file does not depend on the dynamic link library when running. In an embodiment of the present application, static compilation of the source code mainly processes the part of the JS source code that is strongly constrained and cannot be dynamically modified, which can include three aspects: 1. Module reference, for example, replacing the module reference in the JS source code and modifying the suffix name; 2. Template attribute mapping or syntax compatibility processing, for example, in JS, template attribute mapping usually refers to mapping the properties of an object to a template string to generate a complete string. This mapping can be used to create dynamic HTML strings or other types of text; in JS, the purpose of syntax compatibility processing is to ensure that the source code can run in different environments, and the syntax differences that may exist in different environments need to be considered. Syntax compatibility processing can be performed in multiple ways, for example, using strict mode, using JSON objects for serialization and deserialization, using optional chaining operators, using specific syntax, using template imports or exports, etc.; 3. Configuration mapping, such as through page configuration. Thus, through the above several aspects of operation, the differences between various types of clients are smoothed out, providing a basis for implementing cross-end end user experience.

[0033] S101: Start the client, send an end user experience configuration request to the server, obtain the end user experience configuration and store it locally;

[0034] S102: Run the end user experience suite in the front-end code, including: starting the perception engine in the end user experience suite to obtain target data in real time; triggering the decision engine in the end user experience suite to make experience decisions for the target data based on the end user experience configuration; and running the touch engine in the end user experience suite to trigger the corresponding handling scheme according to the experience decision result.

[0035] In an embodiment of the present application, in order to reduce the processing pressure of the end, the end user experience configuration is preformed on the server. After starting the client, the end user experience configuration is loaded and stored locally, providing preparation for subsequent decision-making and other operations on the local.

[0036] The pre-configuration of the end intelligent experience on the service side (or cloud side) can be understood as including two-dimensional configuration. 1. Event rule configuration: the service side needs to predefine various user behavior events and related rules, including the triggering conditions of events, the execution logic of rules, etc. 2. Timing strategy configuration: timing on the scene refers to the expression of performing a certain specific operation under the condition of meeting a series of conditions. In implementation, timing is a complex event (composite event) rule plus trigger configuration; the strategy is the condition that needs to be met or the composite event rule model deployment: for the strategy that needs the support of machine learning model, the service side is responsible for the deployment and delivery of the model. These models are trained based on a large amount of historical data and are used to predict user behavior or evaluate user experience quality.

[0037] In the development of client code, the end intelligent suite is implanted. The end intelligent suite can be understood as a set of functional modules for solving the end intelligent experience, which deeply understands and understands user behavior, combines data analysis, machine learning and other technologies, comprehensively detects and coordinates the device environment and scene running scene of the user, and optimizes the client experience through real-time feedback loop and timely adjustment intervention. The end intelligent suite has a built-in perception engine, decision engine and touch point engine. According to the end intelligent experience configuration loaded locally, the corresponding engine is automatically triggered at the corresponding processing node to realize real-time perception, real-time decision and real-time disposal of data.

[0038] In an implementation manner, the end intelligent suite can be developed based on the end intelligent experience architecture. The end intelligent experience framework can include the following and part. 1. End intelligent experience solution layer, a set of solution strategies and tools for specific problems, which rely on the monitoring and decision-making capabilities provided by the underlying capabilities to specifically solve various problems encountered by users in the use process. According to the experience dimension classification, the end intelligent experience solution layer can include three functional modules of experience monitoring, experience layering and experience enhancement.

[0039] 2. End intelligent experience plug-in system, for example, a plug-in market and a plug-in development platform can be provided, allowing third parties or internal developers to develop plug-ins for specific functions. This system can be regarded as an extension platform, relying on the services and APIs provided by the underlying architecture and solutions to develop and distribute plug-ins.

[0040] 3. End intelligent experience kernel, the basis of the entire system operation, which can be composed of perception, modeling, decision-making and disposal, responsible for collecting and processing information, and making intelligent responses. This kernel directly supports the intelligent implementation of experience monitoring, experience enhancement and experience layering.

[0041] 4. Cross-end intelligent layer, which is the high-level function layer of the system, which integrates unified event flow, model reasoning, feature engineering, decision engine and touch point engine capabilities.

[0042] The logical relationship between the above-mentioned parts and levels reflects the service levels from the base to the high level, each level provides support and service for the upper level, and they work together to ensure the overall intelligentization and efficient operation of the system.

[0043] As can be seen, the end intelligent kit developed through the end intelligent experience architecture can be understood as being constructed with a unified experience base. Unlike the scattered point solution (such as developing a single specific solution for a specific experience problem), the end intelligent experience implementation solution models a unified experience base. This way of using a unified experience base has the benefits of cost reduction and efficiency improvement. In terms of scenarios, only the base needs to be called and rendered to meet individual scenario needs, which has the advantages of scalability and individualization.

[0044] The following will detail the three main actions (data collection, decision-making, and disposal) in S102.

[0045] Data collection

[0046] The end intelligent kit includes a perception engine for collecting data. Perceiving real-time data of users on the end is the basis of end intelligent experience. The target data in S102 is a collection of different types of data needed to be perceived in different experience solutions (such as experience monitoring, experience layering, and experience enhancement, which will be detailed later). Generally speaking, the target data may include user basic data, user behavior data, scenario data, performance data, etc.

[0047] User basic data can be obtained through some APIs, such as using getSysteminfo to obtain the device model, using getNetworkType to obtain the current network type, and using onNetworkStatusChange to listen to the network status change event, etc.

[0048] User behavior data can be tracked by UT, and placed in different event streams according to the event number of the behavior for tracking and reporting. User behavior data is relatively complex and requires certain calculation logic. Taking the "page stay duration event" as an example, the following points are involved: (1) start timing when the user enters the page, and trigger the page stay event when the user browses in the page for more than xx seconds; (2) when the user stays in the page for less than xx seconds, i.e. exits the page, the stay duration will be paused, and the timing will continue when the user returns to the page; (3) at most only one page stay event is triggered within one page visit (pageAppear to pageDisAppear).

[0049] For scenario data, page visit (2001), click (2101), and exposure (2201) behaviors may be involved, and other behaviors are placed in custom behaviors, and different scenario data is sorted out according to certain scenario rules.

[0050] For performance data, for example, interface time consumption can be collected and reported by the underlying network SDK, and rendering-related performance data can rely on the client framework and container (for example, the applet framework and container), combined with the rendering scenario of the scene / page, to complete data collection and reporting at each stage. Taking the applet as an example, it is specifically divided into: (1) applet performance plug-in: collecting application / page startup and standard life cycle time consumption, and automatically reporting, wherein (1.1) extended capability: through the applet performanceApi, some container startup performance data is provided, mainly the startup starting point time and route time consumption; (1.2) for some channels lacking container collection API, unable to obtain the startup time point, the scene T2 time consumption can be customized with the application onLaunch as the starting point; (2) applet performance component mini-performance collects scene T2 time consumption end point, wherein (2.1) settlement method: the component is based on the API of IntersectionObserver, connected by the scene to the last module of the first full-screen rendering, the component records the event (& time) when it is seen (and rendered) in the view as the settlement point of the simulated T2, and there is no additional performance loss. In addition, the applet framework provides the onError method of the App object, which can be used to listen to global script errors, and the onPromiseError monitors all uncaught Promise exceptions. As for the monitoring of page white screen, a custom way is adopted, that is, a timer is started after the interface request is completed and the data is returned, and if there is no element successfully rendered in the subsequent seconds, this situation is regarded as page white screen and corresponding statistics are performed.

[0051] II. Decision

[0052] The end intelligent suite includes a decision engine for experience decision, that is, making experience decisions according to downloaded end intelligent experience configurations and target data collected in real time. The end intelligent experience configurations can include event configuration information or model configuration information, so as to perform event matching according to the event configuration information or prediction according to the model. The event configuration information can be complex event configuration information (corresponding to the "experience monitoring" scheme, described below), and the model can be an experience hierarchical model (corresponding to the "experience hierarchical" scheme, described below), or a full-link motion line prediction model (corresponding to the "experience enhancement" scheme, described below). In the embodiments of the present application, event configuration and model training and deployment are offline, and reasoning is performed on the front end, which can achieve faster response speed, reduce network request delay, protect user data privacy, and still provide services in a network connection-free environment.

[0053] III. Disposal

[0054] The end-to-end intelligent suite includes a touchpoint engine, which triggers experience actions based on experience decision results. In the end-to-end intelligent experience solution provided by the embodiments of this application, after detecting problems through collection and experience decision-making, the problem can be resolved or reported directly on the end-to-end, providing comprehensive and flexible action capabilities. For example, in some cases, user interaction may require UI forms such as pop-ups and bottom-bars. Real-time messages can be pushed within the application interface to inform users of important information or operational feedback, such as real-time message prompts and user guidance. The end-to-end intelligent experience framework provides comprehensive and highly customizable UI components, which are embedded in the host as dynamic plug-ins or subpackages, allowing for rapid delivery of information to users. The end-to-end intelligent experience framework can intercept network request data when sending or receiving it, customize request headers, or process return data, enabling flexible network intervention. The end-to-end intelligent experience framework can dynamically proxy and rewrite basic library APIs in different scenarios to extend or modify existing behavior. For example, it can proxy changes to hashToUrl to return images of varying quality depending on the device's operating state.

[0055] Therefore, through the above three steps, real-time perception and processing of user experience can be achieved on the terminal. The following example illustrates these steps using a specific example of "experience layering".

[0056] Step 1: Preparation and model training. Data collection: Collect behavioral data on user page access, including page load time, user device type, network environment, and other information. Feature extraction: Extract key features from the collected data, such as device performance indicators (CPU, memory usage), network speed, etc. Model training: Use these features and page load time as training data to build a machine learning model to predict the operating status of different models. Model publishing: Convert the trained model to ONNX format and publish the model file to the CDN for the decision engine to load and use for prediction.

[0057] Step 2: Real-time feature collection. Data collection: Monitor and collect data such as page load time, device CPU usage, and memory usage in real time while users use the app. Data normalization: Adjust the data value range to a specific range, usually between 0 and 1 or -1 to 1.

[0058] Step 3: Use the model to make decisions. The collected feature data is used to determine the model's operating status based on the model's prediction results. If the prediction result exceeds the set threshold, it is judged to be in a low operating state (poor operating state or average operating state).

[0059] Step 4: Scene handling. Interface loading optimization: for the models determined to be in a poor running state, automatically optimize the network loading strategy, such as delaying the loading of non-critical interfaces, etc.; content adaptation: dynamically adjust the page content according to the device running state, such as reducing the image quality or simplifying the page elements; data feedback: continue to collect the handling results and user experience feedback as new data input, and continuously optimize the model and strategy.

[0060] As previously described, according to the user experience angle, three dimensions of experience monitoring, experience stratification and experience enhancement can be included. Referring to FIG. 2, an end intelligent experience solution schematic diagram provided by an embodiment of the present application is shown. Experience monitoring is to actively discover the experience problems of users in the process of using the client (for example, whether a specific complex event has occurred) through real-time collection, analysis and early warning of user experience data, and to perform real-time feedback and on-site handling, thereby improving user experience. Experience stratification is to evaluate the running state of user equipment from the actual use angle of users, which is different from the previous rigid classification of equipment according to factory hardware information. The intelligent experience stratification takes the multi-dimensional time-consuming data of actual application running as the basis, uses an experience stratification model (obtained through end intelligent experience configuration), performs comprehensive real-time weighted scoring on user equipment data, and performs corresponding handling according to the scoring. Experience enhancement: based on real-time interaction behavior information of users, using a full-link motion line prediction model (obtained through end intelligent experience configuration), real-time user future trajectory prediction is realized, and intervention such as preloading of sub-packages and pages that users may access is performed to improve user experience surprise.

[0061] Therefore, in one implementation manner, the above-mentioned handling scheme includes: an experience monitoring handling scheme, an experience stratification handling scheme, and / or an experience enhancement handling scheme.

[0062] The experience monitoring handling scheme includes: when a complex event occurs, according to an experience decision, performing real-time message pushing and / or page jumping operation in the interface in a UI interaction manner.

[0063] The experience layering treatment scheme includes: when the device running state is determined as a low running state according to the experience decision, performing an interface loading optimization and / or a page content dynamic adjustment scheme. The prediction result of the experience decision can be divided into three levels: a running state is good, a running state is general, and a running state is poor. The running state is good belongs to a high running state, and the running state is general and the running state is poor belong to a low running state. The running state is general can be understood as a medium level of the running state, and has a space for experience improvement, and the running state is poor needs experience improvement urgently. A person skilled in the art can set and distinguish the running state of each level according to specific performance parameters. When the low running state is determined, the corresponding treatment scheme can be performed, for example, the interface loading optimization scheme is used for improvement, including optimizing a network loading strategy and the like; for example, the page content dynamic adjustment scheme can include: reducing picture quality, simplifying page elements, and / or simplifying page interaction.

[0064] The experience enhancement treatment scheme includes: obtaining a user future behavior track according to the experience decision, determining a subpackage resource or a page resource that the user is likely to access, and performing a preloading operation on the subpackage resource or the page resource.

[0065] In an implementation manner, the execution logical relationship between different treatment schemes can be determined by setting priorities of different treatment schemes. Therefore, the method further includes: setting different priorities corresponding to different treatment schemes, and determining the execution logical relationship between different treatment schemes according to the priorities. For example, the experience layering treatment scheme corresponds to a first priority; and whether to perform the experience monitoring treatment scheme and / or the experience enhancement treatment scheme is determined according to a processing result of the experience layering treatment scheme. The processing manner has the following advantages: first, the experience layering treatment scheme is used for treatment, for example, the picture quality is reduced for a low-performance device, and after the experience layering treatment, if the device running state is still in a poor state (which can be predicted again by the layering experience model), it is considered that the experience monitoring and the experience enhancement are not suitable at this time, so as to save resource overhead and reduce data processing pressure.

[0066] In addition, in an implementation manner, after the treatment scheme is executed, the following can be further included: determining the executed treatment scheme and collecting user feedback data, and sending the treatment scheme and the user feedback data to a server as new inputs to optimize the end intelligent experience configuration of the server. In this way, the treatment result and the user experience feedback are continuously collected as new data inputs, and the model and the strategy are continuously optimized.

[0067] In addition, in order to optimize the implementation effect, a white box + black box mode is used for decision determination.

[0068] White-box model: The online runtime weight is already specific, and there is no need to load the inference model for inference. The white-box model is mainly used in the cold start stage when the user opens the client (such as a mini-program) to preliminarily screen low-performance devices to avoid additional burdens on the device caused by model running. This strategy ensures that subsequent processes can be further diagnosed and optimized through higher-level models. Black-box model: This is a deep network architecture constructed through methods such as multi-layer nonlinear transformation and abstract global feature learning, which simulates the hierarchical processing of information in the human brain nervous system to solve classification, regression, and other problems. For example, in the embodiment of the present application, the full-link dynamic line prediction model can mainly rely on a black-box model, such as a deep learning network (e.g., LSTM), to predict the user's behavior path. This model attempts to predict the user's next possible operation or preference by analyzing the user's historical behavior data. Since deep learning models usually involve complex network structures and a large number of parameters, their internal decision logic is opaque to users, so they are called black-box models. This type of model plays an important role in predicting user future behavior trajectories and implementing intelligent preloading, aiming to improve the smoothness and personalization level of user experience. The experience layering model uses more white-box models, such as GBDT algorithms, to manage user experience in layers according to the device's running status. By analyzing the device's performance parameters (such as memory usage, processor performance, etc.), the model can assess the device's performance level in real time and adjust the application's performance accordingly, such as reducing image quality or simplifying interactive effects, to adapt to devices of different performance levels. The decision-making process of the white-box model is transparent, allowing developers to clearly understand the model's decision logic and basis, facilitating optimization and maintenance.

[0069] Therefore, in an implementation manner, the experience decision of experience monitoring, experience layering, or experience enhancement is made by the white-box model or the black-box model, and a preliminary decision is made according to the white-box model, and it is determined whether to further make a fine decision by the black-box model according to the preliminary decision result of the white-box model.

[0070] In an implementation manner, the experience layering model is set to a white-box model mode; in the cold start stage of the client, the experience layering model in the white-box model mode is run, and it is determined whether to continue executing the end intelligent suite or closing the end intelligent suite according to the prediction result of the experience layering model; correspondingly, the full-link dynamic line prediction model is set to a black-box model mode; in the hot start stage of the client, the full-link dynamic line prediction model in the black-box model mode is run.

[0071] As can be seen, the end intelligent experience implementation method provided in the embodiments of the present application emphasizes a cross-end, instant, and intelligent experience optimization method, which not only focuses on improving a certain aspect of performance, but also starts from the overall perspective of user experience, dynamically adjusts the experience by real-time sensing of the user state, and provides better services. Its core advantages can include:

[0072] 1. Cross-end intelligent experience: Support cross-end use, such as applet, different technology stacks such as web end, etc., wherein the front-end code is converted to multi-end support code through static compilation, which can provide consistent optimization effect on different platforms and devices.

[0073] 2. Instant reaction ability: It can realize real-time perception of user behavior and device state, such as network condition, memory usage, etc., and make instant adjustments according to these information. This instant reaction ability can continuously optimize the experience during user use, rather than only one-time optimization at page loading;

[0074] 3. Intelligent decision and automatic optimization: The end intelligent experience framework has built-in black box model and white box model, which can make intelligent analysis according to the collected data, so as to make targeted optimization decisions. This data-driven optimization method can more accurately identify and solve performance bottlenecks.

[0075] 4. Global strategy support: A series of global strategies are provided, such as image quality degradation, package preloading, weak network switching, etc., which can be flexibly used in different scenarios to achieve end-to-end performance optimization.

[0076] In summary, compared with traditional performance optimization, the traditional method often focuses on technical optimization for specific problems, such as code splitting, lazy loading, resource compression, etc. While the embodiment of the application is based on the end intelligent experience framework, it focuses more on real-time data analysis and intelligent decision-making, from the perspective of user experience, to provide a comprehensive, dynamic and adaptive optimization solution. This method not only solves some limitations encountered in traditional performance optimization, but also provides more personalized and high-quality services according to actual use scenarios and user needs.

[0077] Next, the end intelligent experience implementation method provided by the embodiment of the application is exemplarily described from the aspects of experience monitoring, experience layering and experience enhancement.

[0078] First, referring to FIGS. 3-5, the experience monitoring is described.

[0079] Experience monitoring is to actively discover user experience problems (such as whether a specific complex event has occurred) in the use of the client by real-time collection, analysis and early warning of user experience data, and to provide real-time feedback and on-site disposal, thereby improving user experience.

[0080] Referring to FIG. 3, an end intelligent experience monitoring method flowchart provided by the embodiment of the application is shown.

[0081] S301: Start the client, send a client intelligent experience configuration request to the server, obtain complex event configuration information and initialize it. The complex event configuration information includes event configuration information and timing configuration information of at least one target complex event;

[0082] S302: Start the perception engine, capture the embedded data based on the unified event stream processing tool in the front-end code, and determine whether the target complex event corresponding to the event configuration information is triggered based on the embedded data;

[0083] S303: In response to the triggering of the target complex event, the decision engine is run to match the timing of the tracking data according to the timing configuration information to determine whether the timing of the target complex event is met;

[0084] S304: After the timing is successfully matched, the contact engine is run to trigger the experience monitoring and handling plan corresponding to the target complex event.

[0085] Complex events can be understood as combined events. For example, in the practical application of big data analytics, you might need to detect the occurrence of "consecutive login failures," which is actually a combination of "login failures" and "login failures." Another example is detecting user order and payment behavior, which is also a combined event: a "payment event" occurs some time after an "order event," also with a time constraint. Combined events like these are called complex events.

[0086] In an embodiment of the present application, complex event configuration information can be configured in advance or in real time through an end-intelligent configuration platform. The end-intelligent configuration platform can be a visualization platform that can be set on the server. Through the end-intelligent configuration platform, users (such as R&D personnel) can configure complex event information. For example, the end-intelligent configuration platform provides configuration of timing decisions, including event configuration and timing configuration. Event configuration refers to the matching rules of user events on the end, such as using "spm==abc&&click" to correspond to a click event at a fixed position. Timing configuration is a sequence of events, which is used to perform complex matching on the occurrence sequence of user event streams on the end, such as "click to enter the store" then "add to cart" then "return to homepage". In particular, in order to perform personalized configuration for different types of clients, the end-intelligent configuration platform can be used to configure corresponding complex event sets for different types of clients, wherein the complex event set includes multiple complex events, and each complex event includes event configuration information and timing configuration information.

[0087] See Figure 4 for a schematic diagram of the complex event configuration and matching logic in end-to-end intelligent experience monitoring. In this example, complex event configuration is first performed on the configuration management platform, including event configuration and timing configuration. Event configuration (which can be understood as feature configuration) refers to the accumulation and statistics of user behavior, such as the number of store visits within three days; timing configuration refers to specific conditions that trigger decisions and responses, such as visiting a store twice and adding an item to a purchase. After configuration is completed on the configuration management platform, the configuration information is written to a database (DB), which is then read from the DB by a configuration gateway service. After the client is started, the client automatically loads the configuration information from the configuration gateway service. After obtaining the configuration information, the client's end-to-end intelligent suite makes relevant decisions based on this configuration information. Specifically, the decision engine performs event mapping / filtering, pattern matching, and behavior accumulation to complete the timing decision. After the decision engine determines that the timing is successfully matched, it runs the contact engine to trigger the corresponding action plan. For example, real-time message push in the interface through UI interaction or automatic page redirection for the current operation failure event (for example, redirecting to the front-end checkout counter for multiple payment failures).

[0088] When the client is started, it automatically loads the complex event configuration information. For example, in step S301, the client sends a request to the server to obtain and initialize the complex event configuration information. The complex event configuration information includes event configuration information and timing configuration information for at least one target complex event. The complex event configuration request may include latitude and longitude information, client information, and a user ID. The client information may refer to the client type. Upon receiving the complex event configuration request, the server determines the client type based on the client information, matches the complex event configuration information corresponding to the client type, and issues the corresponding complex event configuration information.

[0089] In the embodiments of the present application, the unified event stream processing tool refers to a standardized event stream processing tool based on super location tracking. It can classify user behaviors into standardized events based on tracking information from various types of clients, including tracking information, event ID (eventId), parameter information (properties), and other information fields. Functions in other pages or plug-ins can obtain the current user's operations on the client by monitoring the unified events.

[0090] Among them, super location tracking includes the following settings.

[0091] Event Marker

[0092] For example, abcd is used to mark an event on a page, where a represents the user-side level, b represents a page on the side, c represents a block on a page, and d represents a specific click position on a block.

[0093] Event Type

[0094] The above-mentioned abcd positions can only mark a specific location, and distinguishing different event types and recording different general parameter information based on specific event types can cover all statistical needs, and also help reduce the difficulty and resource consumption of data processing. For example, the following three types can be used: Page type: In order to count user behavior at the page granularity, when entering and exiting the page, it is necessary to record parameters such as user ID, device ID, timestamp, entry or exit, etc., so that the UV, PV, user stay time and other data of each page can be counted; Click type: Long press, sliding, etc. are all counted as click types; Exposure type: refers to the exposure at a specific module level, such as product exposure and video exposure. By analyzing click-type events together, the click rate of a product, post, or video can be calculated.

[0095] Jump relationship

[0096] For redirect relationships, simply record the relationship between the previous hop and the previous two hops in the URL. For example, if someone clicks the first item on the shopping cart page, a single click event can be reported. However, the previous hop might be from a specific item in the product details, in which case the previous hop can be recorded. Alternatively, the previous two hops might be from an item in the "Guess You Like" feed on the homepage, in which case both hops can be recorded. This connects the user's movement paths.

[0097] Traffic Channel

[0098] For example, if a channel on the homepage contains multiple pages, traffic channels can be introduced. For example, the multiple pages in a flash sale channel can be marked as a traffic channel. Then, by simply removing duplicate visitors from these three pages, the total number of unique visitors for the channel can be obtained. Other statistics such as clicks and total page views can be processed similarly using statistical logic.

[0099] Among them, after the unified event stream processing tool captures the buried point data in the above step S302, it can also include: performing client type analysis and format analysis on the buried point data, and standardizing the buried point data so that the data of different types of clients have a unified format and structure. Among them, the standardized processing of the buried point data can specifically include: parsing the buried point information, event ID, and parameter information of the buried point data; encapsulating the buried point information, event ID and parameter information in a unified format, and converting them into event information in a standard format.

[0100] It can be seen that the unified event stream processing tool in the embodiment of the present application can be understood as a client-side standardized event stream processing tool, which uses the super-position buried points in the cross-end (APP, applet or H5) source code to capture buried point data, and encapsulates the buried point data and then converts it into standardized events. By introducing a unified event stream processing tool, the target complex event triggered by the monitoring processing tool is monitored. When the target complex event occurs, the filtering and matching of the complex event timing decision engine is triggered. This unified event stream processing method is an important part of realizing cross-end processing of complex events. By capturing buried point data based on super-position buried points, event information of various types and locations can be captured on the client side, and by standardizing the buried point data, it is possible to subsequently use a unified complex event timing decision engine to perform complex event detection and disposal based on the same set of complex event matching rule bases and the same matching logic.

[0101] In an embodiment of the present application, a decision engine is used to make timing decisions for complex events (which may be referred to as a "timing decision engine"), which may be based on a self-developed complex event matching rule library and declared as a complex event processing component of a static function, wherein the self-developed complex event matching rule library may be developed in JavaScript to be compatible with multiple types of clients. The advantage of the self-developed complex event matching rule library is that it can be personalized configured or modified according to scenario requirements, and specific personalized rule matching is performed based on this complex event matching rule library. Among them, in order to realize a cross-end adaptive timing decision engine, the complex event processing component can be declared as a static function through the aforementioned JS static compilation processing method to remove the differences between multiple ends and obtain a cross-end timing decision engine.

[0102] In one implementation, the specific processing process of the complex event timing decision engine is as follows: according to pre-configured complex event matching rules, event processing and timing matching are performed on the target complex event, wherein the complex event matching rules include event rules, pattern rules, and timing rules. For example, for the triggered target complex event, event filtering is performed according to the event rules, and event conversion is performed on the filtered events to obtain event IDs in a standard format; pattern filtering is performed on the converted events according to the pattern rules, and events corresponding to the timing rules are retained, and pattern matching is performed on the retained events, wherein rule matching is performed based on multiple rule rules and item rules corresponding to the pattern rules to determine whether a round of pattern matching is successful, and whether the pattern matching is successful is determined based on the progress of multiple rounds of pattern matching; after the pattern matching is successful, fatigue detection is performed according to the timing rules, wherein the latest fatigue detection is performed based on the fatigue cache information, and if the fatigue detection passes, the timing matching is determined to be successful.

[0103] In the complex event processing, the pattern matching can be understood as two aspects: the characteristics of each simple event and the combination relationship between simple events. In addition, the functions of the pattern can be extended, such as time limit of the matching detection, whether each simple event can appear repeatedly, whether to skip the following matching after encountering a match for the pattern of the repeatable event, and the like. The pattern matching detects the continuous events or the discontinuous but sequential events according to the proximity relationship and the conditions in the events. The pattern can also have a time limit, and if the matching condition is not met within the set time range, the pattern matching will time out.

[0104] Referring to FIG. 5, an example of the process of the opportunity decision engine is shown.

[0105] S501, event triggering: the target complex event triggered by the unified event stream processing tool is listened to.

[0106] S502, event filtering: filtering is performed according to the event rule, for example, if only the click event is configured, all the exposure events and the scroll events will be filtered, thereby reducing the matching pressure.

[0107] S503, event conversion: the filtered event is converted into a unified event that can be matched, in which the event of the front end is mapped to the ID of the event configured in the background.

[0108] S504, pattern filtering: there are multiple complex rules in an opportunity, and the events configured by each rule are different. In order to reduce the matching pressure, further filtering is performed, and only the events used by the rule are left.

[0109] S505, pattern matching: for example, the events are matched in turn by using the depth-first traversal rule, and multiple rule rules are included. The pattern matching further includes rule matching and item matching, in which the rule rule refers to a rule in the complex rule, and includes multiple items; the item matching: the item is one-to-one corresponding to the event configured in the background, and the underlying event matching is performed, and the matching result is returned according to the level after completion.

[0110] S506, viewing the matching progress: the matching progress is increased by one each time a pattern is matched, and the next link is entered after all the patterns are matched.

[0111] S507, matching result: determining whether the matching is successful or failed.

[0112] S508 fatigue detection: after matching, the fatigue degree is detected, such as the rule is set as: triggering three times of ordering within 3 days; if the ordering is less than or equal to 3 times within 3 days, the fatigue degree is true, and the subsequent callback behavior is allowed; if the ordering is more than 3 times within 3 days, the fatigue degree is false, and the subsequent callback behavior is not allowed any more. Among them, the historical fatigue degree can be cached locally, and the next time the client is loaded.

[0113] Step 509: decision making is completed.

[0114] In an implementation manner, after the time matching is successful, the callback function (for example, callback function) pre-registered by the complex event time decision engine can be automatically called, and the contact engine is started through the callback function to execute the disposal scheme corresponding to the target complex event. Among them, as described before, the end intelligent configuration platform can realize complex event configuration, and corresponding to this, the time matching success information of the target complex event can be received through the end intelligent configuration platform, and the corresponding disposal scheme is determined according to the scene attribute of the target complex event and / or the end attribute of the target complex event corresponding type client. Among them, the scene attribute can include application scene type information, and the client end attribute is used to distinguish the client type, therefore, through the above-mentioned manner, the individualized disposal scheme for the target complex event can be realized, wherein the disposal scheme can be, for example, the prompt window or the floating layer.

[0115] Taking "automatic switching to front-end cash register after three consecutive payment failures of the user" as an example, the experience monitoring is exemplarily illustrated.

[0116] Step 1: in the end intelligent configuration platform, the user payment failure event (if it has been configured, it can be reused) is configured, and it is assumed that the event is event F;

[0117] Step 2: in the end intelligent configuration platform, the number of times of user payment failure event is configured as three times, and the number of times of event D matching is 3;

[0118] Step 3: after the client is started, the client requests configuration information and initializes;

[0119] Step 4: after the user fails to pay once (the unified event flow processing tool determines through burying), the time decision engine is triggered to perform time matching, and the number of times is recorded as 1;

[0120] Step 5: after the user fails to pay twice, the time decision engine is triggered to perform time matching, and the number of times is recorded as 2;

[0121] Step 6: after the user fails to pay three times, the time decision engine is triggered to perform time matching, and the number of times is recorded as 3;

[0122] Step 7: After the matching test is passed, the registered callback function (such as callback function) is called;

[0123] Step 8: Run the touchpoint engine and automatically switch to the front-end cashier. The front-end cashier allows you to select a specific payment platform on the merchant page to complete payment, avoiding redirects to intermediate pages and helping to optimize the online banking payment experience. The front-end cashier supports integration on the Web, WAP, and App platforms.

[0124] In the aforementioned example of a user failing to log in three times in a row, the timing decision engine is a matching engine based on matching rules. It performs only data parsing and rule matching. It can be thought of as a core component declared as a static function but caching a portion of its execution. The association between matching rules and complex events can be pre-configured on the server side, while the intermediate steps in the execution process (triggering complex events and matching timing) are not linked and can be considered a black-box implementation. Taking the example of "monitoring for unresponsiveness after repeated clicks on the same location on a page," the monitoring strategy is first configured. For example, the monitoring strategy is configured based on the scenario, specifying matching rules and triggering timing: Event matching logic: using the a and b positions to distinguish pages, the c position to distinguish scenarios, and the d position to distinguish pits. These timing configurations cover all clickable areas on the page; the triggering strategy: three or more consecutive clicks within two seconds; then, user anomalies are reported in real time through tracking points, and actionable anomalies are intervened in real time. For example, targeted actions can be taken at different points, such as page downgrade or redirection to an H5 page details page. Therefore, compared with traditional monitoring that is difficult to perceive occasional anomalies, the experience monitoring in the embodiment of the present application can discover problems in the first time based on user sessions, and intervene and handle them before customer complaints are generated, thereby improving user experience.

[0125] In addition, taking "intelligent error correction of bill of lading mobile phone number" as an example, in the scenario where the user places an order, a decision is made based on the similarity score distribution. When it is determined that the user is likely to have filled in the wrong mobile phone number (for example, the score is between 0.7-1), the mobile phone number is corrected to remind the user to modify the mobile phone number with one click (this can be achieved on the front end). For example, correction is performed through bubble prompts, and the mobile phone number that the user will fill in can be predicted and the one-click filling function can be provided.

[0126] In addition, there are many other scenario instances of experience monitoring, such as automatic switching of mobile networks in weak WIF scenarios, etc., which are not described one by one in the embodiments of this application.

[0127] In one implementation, the above method also includes the following steps: determining whether the target complex event is a real-time behavior event or a cumulative behavior event; if it is a real-time behavior event, obtaining user behavior data in real time, and performing complex event triggering judgment and timing matching based on the buried data corresponding to the user behavior data in real time; if it is a cumulative behavior timing, obtaining user behavior data in multiple times, accumulating the user behavior data, and performing complex event triggering judgment and timing matching based on the buried data corresponding to the currently accumulated user behavior data. It can be seen that through the above method, using user behavior data as the basis for matching can achieve more intelligent decision-making. Unlike processing real-time user behavior events for users, processing cumulative behavior events (for example, ordering 5 times in a week) can achieve domain linkage: accumulated user data can be used in the process of timing decision-making, and a complete access solution can be provided.

[0128] It can be seen that through the above-mentioned experience monitoring technical solution, the embodiment of the present application targets JavaScript front-end code applicable to multiple types of clients, detects complex events and handles them at the front end, wherein after the client is started, a complex event configuration request is sent to the server, the complex event configuration information is obtained and initialized, and then based on the unified event stream processing tool in the front-end code, the buried data is captured, and it is determined whether the target complex event corresponding to the event configuration information is triggered according to the buried data; then, in response to the triggering of the target complex event, the decision engine is run to match the timing for the buried data according to the timing configuration information to determine whether the timing of the target complex event is met; finally, after the timing is successfully matched, the callback function is triggered to execute the handling solution corresponding to the target complex event. In the embodiment of the present application, compared with the method of mainly relying on the server to detect and handle complex events, complex event processing is performed on the client side, thereby performing real-time rule matching for the data obtained in real time, thereby reducing the processing time, and having the advantages of low latency and low maintenance cost.

[0129] Furthermore, unlike the existing complex event processing solutions that are only applicable to one type of client, the embodiments of the present application also make improvements in this regard, in order to achieve a set of complex event processing solutions that are applicable across multiple clients. In order to achieve a complex event processing solution that is applicable to multiple types of clients, the embodiments of the present application rely on JavaScript (JS) for front-end code development. Since JS development is compatible with various types of clients such as APP, mini-programs, and H5, it can achieve the purpose of cross-end. Through a set of event processing mechanisms, it can be universally applicable to various types of clients, greatly reducing the development pressure.

[0130] In addition, in order to conveniently personalize complex event configuration for different types of clients, the matching rule of the complex event can be configured through a terminal intelligent configuration platform, so that the strategy is more flexible and easy to use. With the help of the terminal intelligent configuration platform, the simple implementation mode of "input configuration" and "output time matching result" is achieved.

[0131] Next, referring to FIGS. 6-7, the experience layering is explained.

[0132] Referring to FIG. 6, it is a flow chart of the terminal intelligent experience layering method provided by the embodiment of the application.

[0133] S601: Start the client, send a terminal intelligent experience configuration request to the server, and obtain an experience layering model;

[0134] S602: Start the perception engine, obtain device hardware static data, device real-time state data, and scene real-time running data;

[0135] S603: Trigger the decision engine, input the obtained device hardware static data, device real-time state data, and scene real-time running data into the experience layering model, and predict the device running state;

[0136] S604: Run the contact engine, trigger the corresponding experience layering disposal scheme according to the prediction result of the device running state.

[0137] In actual application, it is tried to distinguish the running state of the user device, and give optimization on the strategy to the device with low running state, so as to achieve normal running state and user experience. Taking a small program as an example, due to the small program container and framework, and the complicated scene logic, it is difficult to achieve consistent experience of each model in performance and experience. Based on the long tail theory, the focus of solution is placed on the low-performance device, and a small amount of resources is used to assist in solving the experience problem. In the existing mode, the judgment for the low-performance device is based on the following key factors: hardware specifications, price, performance, and function, etc. These judgment standards are relatively static, and cannot make more detailed distinction according to the real-time running state of the user device. For example, when a high-performance mobile phone opens a few large game apps and then opens a target small program, the page is very laggy, and the traditional discrimination mode still considers that it is a high-performance device at this moment. The present embodiment can make real-time judgment in combination with the performance of page rendering. In the experience layering scheme of the present embodiment, the running state of the user device is evaluated from the actual use of the user, which is different from the past rigidly according to the hardware information of the factory. The intelligent layering uses the multi-dimensional time consumption data of the actual running of the application as the basis, uses the terminal intelligent experience layering model to comprehensively and real-timely score the user device data, and makes corresponding disposal according to the score.

[0138] For the experience layering scheme, data collection and processing need to be performed on the server side first, and the experience layering model needs to be trained and then distributed to the front end. During the client-side running process, based on the obtained target data, the experience layering model is used to perform layering prediction in real time on the front end, and scene degradation and other treatments are performed based on the user equipment running state. The following introduces the scheme from two angles of data collection and model structure determination.

[0139] I. Data collection and processing

[0140] Data collection and processing is a key link for model training. For data processing, the following process division can be performed: feature dimension division, data collection and reporting, data reading, data cleaning, data division, and data preprocessing. The standard input of the model is obtained through these steps.

[0141] (1) Data dimension division

[0142] Unlike existing low-performance device determination schemes, the data in the present scheme includes not only static data such as mobile phone models and pixel ratios, but also real-time dynamic running data of the mobile phone, such as the current available memory, the number of alarms, the loading time of each page, and the change of the memory in each stage. The running state of the mobile phone is more real-time, and the scheme can capture the moment when the running state is not good in time, and assist in degradation to prevent the running state from deteriorating, and even to improve.

[0143] Taking the experience layering model as an example, the collected data can be divided into static data and dynamic data. The static data includes the hardware conditions of the user's mobile phone, such as the mobile phone model, configuration, and pixel ratio. The dynamic data is divided into device-related and scenario-related data.

[0144] In the embodiments of the present application, the collected data can be divided into device hardware static data, device real-time state data, and scenario real-time running data. The device hardware static data includes, for example, mobile phone models, memory capacities, and other parameters. The device real-time state data refers to dynamically changing device data, such as power or network state parameters. The scenario real-time running data refers to data related to scenario running, including but not limited to home page scenarios and store scenarios, such as home page loading time parameters. The embodiments of the present application do not limit the specific parameters of each type of data, such as parameters a, parameters b, and parameters c.

[0145] (2) Data collection and reporting

[0146] The data collection method is different according to the feature dimension. Static data is read only once for caching, reducing the time consumption, and dynamic data is extracted at the collection node to ensure accuracy.

[0147] Taking a small program as an example, the data reporting method supports different dimensions:

[0148] Commonly buried point data provides dimensional report, suitable for T+1 offline analysis scene;

[0149] Highway can return data at minute level, suitable for online real-time analysis scene;

[0150] Answer has high flexibility, can customize data form, suitable for more rich scene;

[0151] Mtop entrainment, high flexibility, high real-time, but limited by interface qps and trigger opportunity, suitable for end cloud collaborative scene;

[0152] (3) Data preprocessing

[0153] The quality of data directly determines the prediction effect and generalization ability of the model, and determines the ceiling of the model. Therefore, data processing involves several dimensions: accuracy, completeness, consistency, effectiveness, credibility, and interpretability. The data of the experience stratification model still contains a lot of noise, although it has undergone basic data cleaning.

[0154] Therefore, preprocessing is needed before the model, which can obtain standard, clean and smooth data through filling missing values, smoothing noise data, smoothing or deleting outliers, and solving inconsistent data. In addition, continuous data can be discretized by binning, and continuous data can be divided into corresponding barrels according to certain rules. The binning rules include unsupervised binning (equal frequency binning, equal distance binning), supervised binning (chi-square binning, decision tree binning), etc. Equal frequency and equal distance are commonly used. Combined with the characteristics of experience stratification data, equal frequency binning can be adopted.

[0155] II. Model structure and training

[0156] In the embodiments of the present application, a convolutional neural network (CNN) can be used to construct an experience stratification model. The training process of CNN includes forward propagation (the network calculates the output according to the input data and the current weight), loss calculation (based on the output and the real label), back propagation (calculate the gradient of each layer according to the loss), and weight update (use gradient descent algorithm or other optimization algorithm to adjust the network weight).

[0157] Through further research and exploration, it is found that the first problem to be solved by the experience layering model is classification, and the classic algorithms such as GBDT, CNN, DNN, etc. are candidate models. Convolutional neural network (CNN) and deep neural network (DNN) are tried in turn. They have their own advantages and disadvantages. CNN is mainly designed to process spatially continuous data (such as images), and can effectively process spatial data such as images through the introduction of convolution operation and pooling operation, but may not be as efficient as DNN on other types of data. DNN is simpler and more general in structure, but may face huge parameter space and overfitting problems, but can be optimized to avoid such problems. Based on the collection of mobile static data and dynamic data such as memory, the overall performance on DNN is better than that on CNN. Combining the advantages of the two algorithms, in the optimization scheme, CNN can be used for feature extraction, and then the extracted features are input into DNN for final model training.

[0158] In an implementation manner, the experience layering model is trained by a "white box + black box model". As previously described, the white box model has specific weights during online operation, and does not need to load the inference model for inference. The white box model is mainly used for preliminary screening of low-performance devices in the cold start stage when the user opens the client (such as the applet), so as to avoid additional burden on the device caused by model running. This strategy ensures that further running state diagnosis and experience optimization can be performed through higher-level models in the subsequent process. The black box model is a deep network architecture constructed by methods such as multi-layer nonlinear transformation and abstract global feature learning, which simulates the hierarchical processing of information in the human brain nervous system and is used to solve classification, regression and other problems. Specifically, the experience layering model can be obtained by "training GBDT structure by white box model + training DNN structure by black box model". The advantage of this way is that, first, the white box model (such as GBDT algorithm) is used to manage the user experience according to the running status of the device. By analyzing the performance parameters of the device (such as memory usage, processor performance, etc.), the model can evaluate the performance level of the device in real time, and adjust the performance of the application accordingly, such as reducing the picture quality or simplifying the interaction effect, to adapt to devices of different performance levels. The decision-making process of the white box model is transparent, allowing developers to understand the decision-making logic and basis of the model, facilitating optimization and maintenance. Then, the black box model (such as DNN) usually involves complex network structure and a large number of parameters, and its internal decision-making logic is not transparent to users, which has strong prediction ability and aims to improve the smoothness and personalization level of user experience. Therefore, by training different network structures by "white box + black box" to obtain the experience layering model, the characteristics of each model can be utilized, the white box model can be used for preliminary screening of low-performance devices, and the black box model can be used for accurate prediction, so as to realize real-time hierarchical processing scheme.

[0159] Therefore, in an implementation mode, a first experience score model is obtained by training a first network structure (for example, GBDT or CNN), and a second experience score layer model is obtained by training a second network structure (for example, DNN). In the decision process, white-box pre-screening is performed by the first experience score layer model, and black-box fine screening is performed by the second experience score layer model, so as to predict the device running state. For example, through white-box pre-screening, it is preliminarily determined that 15% of the devices are in a poor running state and can be degraded for disposal. Then, through black-box fine screening, the remaining 85% of the devices are finely determined, and 30% of them can be degraded for disposal. This experience score layer mode combining white-box and black-box can comprehensively perform device layer disposal. Experiments prove that 50% of the devices can be significantly throttled, the average memory alarm frequency is reduced by 10%, and the client running fluency and user experience are greatly improved.

[0160] In the end intelligent experience layering scheme, another technical key is to publish and deploy the experience layering model trained on the service end, so that the client can obtain the experience layering model during running and use the model for real-time prediction. Compared with the way of obtaining data from the front end on the service end or cloud end and then performing model prediction, performing prediction on the front end can reduce the step of data reporting, thereby accelerating the prediction speed and reducing the delay. Moreover, this model prediction mode on the front end is not limited to the network state. Even when the network is jammed or fails, model prediction can be completed relying on the front-end data processing capability. Therefore, running the experience layering model on the client is a technical point.

[0161] Referring to FIG. 7, an end intelligent experience layering implementation logic diagram provided by an embodiment of the present application is shown. The whole process includes four main links of model preparation, environment preparation, parameter preparation, and inference scoring.

[0162] The model preparation link is to obtain an experience layering model by offline training. As described above, data collection and processing are performed on the service end, and the model structure is determined, so as to train the model structure by using the processed data and obtain the experience layering model.

[0163] The environment preparation link refers to environment preparation of running the model on the client side, mainly including starting the engine. Taking the applet as an example, first, according to the applet starting, the model is downloaded according to the terminal intelligent configuration information (for example, through the configuration capability of the terminal intelligent configuration platform, the model and related control parameters are configured on the platform, and then the model is saved in a DB, and the model is read from the DB through a configuration gateway service; after the client starts, the client automatically loads the model from the configuration gateway service, which can be referred to FIG. 4 and the description), and then Tensor preparation is performed; then, through functions such as access to a graphics library (for example, webGL), a programming scheme (for example, WASM, a technical scheme using a non-Java programming language to write code and capable of running on a browser), image processing (webGPU), etc., to realize dynamic switching of JS engine and NA engine.

[0164] In the parameter preparation link, mainly the model inference parameter preparation, which can include steps such as computation graph analysis, operator loading, weight loading, and loading readiness.

[0165] In the inference running link, mainly inputting the inference parameters into the model to obtain the prediction result, which can include steps such as Tensor input, input preprocessing, operator operation, and result obtaining.

[0166] Therefore, through the above four links, the experience layering model can be used for real-time inference and scoring, thereby determining the real-time running state of the user equipment (such as a mobile phone).

[0167] It should be noted that taking the applet as an example, since the model file format that can be run in the applet scenario is limited to.xnn, other formats need to be optimized for model structure and compiled for heterogeneous models, and the essence is the conversion of operator operation. Limited by the running environment and framework of the applet, the complexity of the model should be simple, and complex model structures are one that the model file is large and takes a long time to load, and two that the operation is complex and consumes machine resources. Therefore, when designing the model, the principle of simple calculation is followed to avoid using complex operators and complex processes.

[0168] Therefore, after obtaining the experience layering model, the above method can further include: an environment preparation link: based on the downloaded experience layering model, Tensor preparation is performed, the logic corresponding to the model data preprocessing stage is mapped to JS, and an XNN parser is loaded to realize dynamic JS engine and NA engine; a parameter preparation link, based on the model format supported by the client, model parameter preparation is performed, including model structure optimization and heterogeneous model compilation.

[0169] Model structure optimization refers to the process of improving efficiency and reducing resource consumption by refining the architecture of a model. This typically involves the following aspects: Pruning: removing unimportant weights or neurons to reduce model size. Quantization: reducing the precision of model parameters to decrease model size and speed up inference, such as converting from 32-bit floating-point numbers to 8-bit integers. Simplification: simplifying the model by removing overly complex parts to improve running efficiency. Knowledge distillation: transferring the knowledge of a large model to a smaller one. AutoML for model compression: using automated algorithms to find optimal model compression methods.

[0170] Heterogeneous model compilation: closely related to model structure optimization, refers to the process of compiling deep learning models for different hardware platforms, such as GPUs, CPUs, TPUs, or FPGAs. As hardware devices become more diverse, a model that runs efficiently on one hardware may not perform well on another. Heterogeneous compilation adapts the model to the characteristics of different hardware by converting it, thereby achieving model portability and performance optimization.

[0171] In summary, model structure optimization focuses on reducing complexity and improving efficiency within the model, while heterogeneous model compilation emphasizes model portability and performance optimization across different hardware platforms. Both aspects are crucial for deploying deep learning models in real-world production environments.

[0172] In an implementation manner, the experience hierarchical model is pre-trained on a server and published to a content distribution network. The decision engine accesses the content distribution network according to configuration information of the experience hierarchical model, and loads the experience hierarchical model to the local. For example, through the configuration capability of the end intelligent configuration platform, the model and related control parameters are configured on the platform, and then the model is saved in a DB, and the model is read from the DB through a configuration gateway service; after the client is started, the client automatically loads the model from the configuration gateway service. Please refer to FIG. 4 and the description.

[0173] For example, the experience layering model prediction result is divided into three levels of good running state, general running state and poor running state. For the latter two levels, disposal schemes such as path simplification, rich interaction degradation and function degradation can be taken. Path simplification can be to close non-critical functions or reduce the priority of non-critical functions. For example, if the non-critical function has no effect on the user's order placement or order fulfillment, the function can be directly closed, or if the non-critical function has a small effect on the user's order placement or order fulfillment, the priority of the function can be reduced. Rich interaction degradation can include reducing the animation frame rate, reducing the picture resolution, simplifying the interaction effect, etc. Reducing the animation frame rate: the higher the animation frame rate, the higher the requirement for machine performance, so the animation frame rate can be adjusted in real time according to the performance of the user's machine to achieve smooth running effect; reducing the picture resolution: adjusting the picture resolution in real time according to the performance of the user's machine to reduce the consumption of picture rendering on memory; simplifying the interaction effect: some interaction effects have high requirements for machine performance, and these interaction effects (such as lottie animation effect) can be simplified or closed in real time according to the performance of the user's machine. Function degradation refers to degrading the function in the client that does not block the user's order placement or order fulfillment.

[0174] The experience layering scheme provided by the embodiments of the present application is a cross-end, instant, dynamic, intelligent and full-link low-performance device discrimination strategy. It not only focuses on the static data of the device, but also starts from various dimensional data in the running process of the device, and provides more accurate data basis for the judgment of low-performance devices through real-time sensing calculation. Compared with the traditional method of evaluating low-performance devices, the traditional method is more fixed and less flexible, while the embodiments of the present application focus on real-time data analysis and algorithm decision-making, starting from the real user experience, and providing a more intelligent, comprehensive and accurate evaluation scheme. The scheme breaks through the limitations of the traditional scheme and has the characteristics of individualization, intelligence and instantaneity.

[0175] Finally, the experience enhancement scheme is introduced in combination with FIG. 8.

[0176] Referring to FIG. 8, a flow chart of an end intelligent experience enhancement method provided by the embodiments of the present application is shown.

[0177] S801: Start the client, send an end intelligent experience configuration request to the server, and obtain a full-link dynamic line prediction model;

[0178] S802: Start the sensing engine, and obtain user behavior data, user basic data, scene data and performance data;

[0179] S803: Trigger the decision engine, input the obtained user behavior data, user basic data, scene data and performance data into the full-link dynamic line prediction model, and predict the future behavior trajectory of the user;

[0180] S804: running the contact engine, triggering the corresponding experience enhancement treatment scheme according to the prediction result of the future behavior trajectory of the user.

[0181] Experience enhancement refers to predicting the future trajectory of the user in real time based on real-time interaction behavior information of the user, using a full-link motion line prediction model, realizing real-time user future trajectory prediction, preloading and other interventions on the sub-packages and pages that the user may access, and improving user experience surprise.

[0182] For the experience enhancement scheme, data collection and processing need to be performed on the server side, and the full-link motion line prediction model needs to be trained and completed, and then the full-link motion line prediction model is distributed to the front end. During the running process of the client, based on the target data obtained, the future behavior prediction is performed in real time through the full-link motion line prediction model on the front end, and the sub-package preloading and other treatments are performed based on the future behavior prediction result.

[0183] Regarding the two aspects of data collection and model structure determination, similar to the principles and processing processes of the experience layering scheme, reference can be made to the related descriptions of the experience layering scheme. The difference lies in the division of data feature dimensions and the selection of model structures. Since experience enhancement mainly considers predicting future user behavior, the data feature dimensions mainly include user behavior data, user basic data, scene data, and performance data. The various data acquisition methods can be referred to the previous description. For the network structure of the full-link motion line prediction model, through research and practice, a network structure with strong computing power can be selected, such as a deep learning network, for example, LSTM, etc. This model attempts to predict the user's next possible operation or preference by analyzing the user's historical behavior data, in order to predict the user's behavior path. Since deep learning models usually involve complex network structures and a large number of parameters, their internal decision logic is not transparent to users, so a black box model can be used. Therefore, in one implementation, the full-link motion line prediction model corresponding to the experience enhancement decision is set to a black box model mode, and the full-link motion line prediction model in the black box model mode is run during the hot start stage of the client.

[0184] Regarding the publishing and deployment of the full-link motion line prediction model trained on the server side, and the real-time prediction of the client after downloading and locally, reference can be made to the related descriptions of the experience layering scheme (Figure 7), which is similar in principle and implementation process. Therefore, it is not repeated here.

[0185] In addition, in an implementation, the full-link motion line prediction model is pre-trained on the server side and published to the content distribution network, wherein the decision engine accesses the content distribution network according to configuration information of the experience layering model, and loads the experience layering model to the local. For example, the model and related control parameters are configured on the platform through the configuration capability of the end intelligent configuration platform, and then the model is saved in a DB, and the model is read from the DB through a configuration gateway service; after the client is started, the client automatically loads the model from the configuration gateway service. Please refer to FIG. 4 and the description.

[0186] For example, through the full-link motion line prediction model, the future behavior trajectory of the user can be predicted, and corresponding experience enhancement treatment can be performed on the prediction result. For example, the experience enhancement treatment scheme includes: determining the package resources or page resources that the user is likely to access according to the future behavior trajectory of the user obtained through experience decision, and performing preloading operation on the package resources or page resources.

[0187] In a specific scenario, intelligent preloading can be performed on channel packages. In the client interface, the next jump page of the homepage is usually a package, and hard full preloading not only wastes traffic, but also occupies memory, and even causes freezing. By using the end-to-end intelligent experience enhancement scheme, the next jump address is predicted, and the package preloading is completed according to the nearest principle, so that the contradiction between memory and experience can be well balanced.

[0188] Corresponding to the above-mentioned end intelligent experience implementation method applied in the client, the embodiment of the present application also provides an end intelligent experience implementation method applied in the server. Please refer to FIG. 9, which is a flow chart of the end intelligent experience implementation method applied in the server.

[0189] The end intelligent experience implementation method is used for configuring and contacting the end intelligent experience front-end code, wherein the front-end code is suitable for multiple types of clients and is used for perceiving, deciding and disposing user experience in the front end, and the method is applied in the server, and includes the following steps.

[0190] S901: receiving the end intelligent configuration request sent by the client, matching the end intelligent configuration and delivering it to the client, so that the client performs end intelligent experience decision and disposal;

[0191] The experience decision and disposal performed by the client include: running the end intelligent suite in the front-end code, including: starting the perception engine in the end intelligent suite to obtain target data in real time; triggering the decision engine in the end intelligent suite to make experience decision according to the end intelligent experience configuration and the target data; and running the contact engine in the end intelligent suite to trigger the corresponding disposal scheme according to the experience decision result.

[0192] S902: in response to the triggering of the callback function, receiving the disposal scheme execution result fed back by the client.

[0193] As to the method for implementing the end intelligent experience executed on the server side, reference can be made to the foregoing principles and descriptions, which are not repeated here.

[0194] Referring to FIG. 10, a structure schematic diagram of an end intelligent experience implementation device provided by an embodiment of the present application is shown.

[0195] The device is used for the perception, decision and disposal of user experience on the front end for the front end code applicable to multiple types of clients, and is applied to one of any type of clients in the multiple types of clients, and comprises:

[0196] The configuration request unit 1001 is configured to start the client, send an end intelligent experience configuration request to the server, obtain the end intelligent experience configuration and store it to the local;

[0197] The end intelligent execution unit 1002 is configured to run the end intelligent suite in the front end code, comprising: starting the perception engine in the end intelligent suite to obtain the target data in real time; triggering the decision engine in the end intelligent suite to make experience decisions for the target data according to the end intelligent experience configuration; and running the contact engine in the end intelligent suite to trigger the corresponding disposal scheme according to the experience decision result.

[0198] The specific examples of each unit in the embodiment can refer to the examples described in the above-described embodiments and optional implementation manners, which are not repeated here.

[0199] Embodiments of the present application also provide a storage medium having a computer program stored therein, wherein the computer program is configured to execute the steps in any of the above method embodiments when running.

[0200] Optionally, in the present embodiment, the above storage medium can be configured to store the computer program for executing the above method embodiments.

[0201] Optionally, in the present embodiment, the above storage medium can include but is not limited to: a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk and various computer program storage media.

[0202] Embodiments of the present application also provide an electronic device comprising a memory having a computer program stored therein and a processor configured to run the computer program to execute the steps in any of the above method embodiments.

[0203] Optionally, the electronic device can further include a transmission device connected to the processor, and an input and output device connected to the processor.

[0204] Optionally, in the embodiment, the processor can be configured to execute the steps in any of the method embodiments described above by using a computer program.

[0205] Optionally, the specific examples in the embodiment can refer to the examples described in the above embodiments and optional implementation manners, and the embodiment will not be described here.

[0206] The serial numbers of the embodiments of the application are only for description, and do not represent the advantages and disadvantages of the embodiments.

[0207] In the above embodiments of the application, the description of each embodiment has its own focus, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0208] In the several embodiments provided by the present application, it should be understood that the disclosed technology can be implemented in other ways. Of course, the device embodiment described above is only schematic. For example, the division of the units is only a logical function division, and there can be another division manner in actual implementation. For example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed coupling or direct coupling or communication connection between the units can be indirect coupling or communication connection through some interface, or electrical or other forms.

[0209] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place, or can be distributed on a plurality of network units. According to actual needs, part or all of the units can be selected to achieve the purpose of the embodiment.

[0210] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0211] The integrated unit, if implemented in the form of a software function unit and sold or used as an independent product, can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application essentially or say the part that contributes to the prior art or the whole or part of the technical solutions can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the method described in the various embodiments of the present application. The aforementioned storage medium includes various media that can store program codes, such as a U disk, a ROM, a RAM, a mobile hard disk, a magnetic disk or an optical disk.

[0212] The above is only the preferred embodiment of the present application, and it should be pointed out that for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, which should be considered as the protection scope of the present application.

Claims

1. A method for realizing terminal intelligent experience, characterized in that: The method is used to perceive, decide, and handle user experience on the front end for front-end code applicable to multiple types of clients. The method is applied to one client of any type among the multiple types of clients, including: Start the client, send a request for the client intelligent experience configuration to the server, obtain the client intelligent experience configuration and store it locally; Running the end-intelligent suite in the front-end code includes: starting the perception engine in the end-intelligent suite to obtain target data in real time; triggering the decision engine in the end-intelligent suite to make experience decisions for the target data based on the end-intelligent experience configuration; and running the touchpoint engine in the end-intelligent suite to trigger corresponding disposal plans based on the experience decision results.

2. The method according to claim 1, characterized in that The processing scheme includes: an experience monitoring processing scheme, an experience layering processing scheme, and / or an experience enhancement processing scheme; The experience monitoring and handling solution includes: when a complex event is determined to have occurred based on the experience decision, real-time message push and / or page jump operation is performed in the interface in a UI interactive manner; The experience layering solution includes: when the device operating state is determined to be a low operating state based on the experience decision, performing interface loading optimization and / or page content dynamic adjustment solutions; The experience enhancement solution includes: obtaining a user's future behavior trajectory based on the experience decision, determining subpackage resources or page resources that the user may access, and preloading the subpackage resources or page resources.

3. The method according to claim 2, characterized in that Also includes: Set different priorities for different disposal plans, and determine the execution logic relationship between different disposal plans based on the priorities; among them, setting the experience layered disposal plan corresponds to the first priority; Based on the processing results of the experience stratification solution, determine whether to execute the experience monitoring solution and / or the experience enhancement solution.

4. The method according to claim 2, characterized in that Also includes: Use white-box models or black-box models to make experience decisions for experience monitoring, experience layering, or experience enhancement, as well as make preliminary decisions based on the white-box model and determine whether to use the black-box model to make further refined decisions based on the preliminary decision results of the white-box model.

5. The method according to claim 4, characterized in that Also includes: Setting the experience layering model corresponding to the experience layering decision to a white box model mode, running the experience layering model in the white box model mode during the cold start phase of the client, and determining whether to continue executing the end intelligent suite or shut down the end intelligent suite based on the prediction result of the experience layering model; The full-link traffic flow prediction model corresponding to the experience enhancement decision is set to a black box model mode, and during the client hot start phase, the full-link traffic flow prediction model of the black box model mode is run.

6. The method according to any one of claims 1 to 5, characterized in that The terminal intelligent experience configuration includes complex event configuration information, and the complex event configuration information includes event configuration information and timing configuration information of at least one target complex event; The perception engine in the startup end intelligent suite obtains target data in real time, including: starting the perception engine, capturing embedded data based on the unified event stream processing tool in the front-end code, and determining whether the target complex event corresponding to the event configuration information is triggered based on the embedded data; The decision engine in the trigger-end intelligent suite makes an experience decision for the target data based on the end intelligent experience configuration, including: in response to the triggering of the target complex event, running the decision engine to match the timing of the embedded data based on the timing configuration information to determine whether the timing of the target complex event is met; The contact engine in the operating end intelligent suite triggers a corresponding handling plan based on the experience decision result, including: after the timing is successfully matched, running the contact engine to trigger the experience monitoring handling plan corresponding to the target complex event.

7. The method according to claim 6, characterized in that After the unified event stream processing tool captures the tracking data, it also includes: Perform client type and format analysis on tracking data, and standardize the tracking data so that data from different types of clients have a unified format and structure. Among them, the standardized processing of the buried point data includes: parsing the buried point information, event ID, and parameter information of the buried point data; packaging the buried point information, event ID, and parameter information in a unified format, and converting them into event information in a standard format.

8. The method according to claim 6, characterized in that The decision engine is a complex event processing component based on a self-developed complex event matching rule library and declared as a static function, wherein the self-developed complex event matching rule library is developed using JavaScript to be compatible with multiple types of clients.

9. The method according to any one of claims 1 to 5, characterized in that The terminal intelligent experience configuration includes an experience layering model; The sensing engine in the startup terminal intelligent suite is started to obtain target data in real time, including: starting the sensing engine to obtain static data of device hardware, real-time status data of the device, and real-time operation data of the scene; The decision engine in the triggering end intelligent suite makes an experience decision for the target data based on the end intelligent experience configuration, including: triggering the decision engine to input the acquired static data of device hardware, real-time status data of the device, and real-time operation data of the scenario into the experience hierarchical model to predict the operation status of the device; The touchpoint engine in the operation-end intelligent suite triggers a corresponding handling plan based on the experience decision result, including: running the touchpoint engine to trigger a corresponding experience layered handling plan based on the prediction result of the device operation status.

10. The method according to claim 9, characterized in that The experience layering model is obtained by extracting features from data samples based on a convolutional neural network and inputting the extracted features into a deep neural network model for training.

11. The method according to claim 9, characterized in that The experience layering model is obtained by training a first network structure to obtain a first experience layering model and training a second network structure to obtain a second experience layering model. In the decision-making process of the decision engine, the first experience layering model is used as a white box model for white box pre-screening and the second experience layering model is used as a black box model for black box fine screening, so as to predict the operating status of the device.

12. The method according to any one of claims 1 to 5, characterized in that The terminal intelligent experience configuration includes a full-link dynamic line prediction model; The perception engine in the startup terminal intelligent suite is used to obtain target data in real time, including: starting the perception engine to obtain user behavior data, user basic data, scenario data, and performance data; The decision engine in the trigger-end intelligent suite makes experience decisions based on the target data according to the end intelligent experience configuration, including: triggering the decision engine to input the acquired user behavior data, user basic data, scenario data, and performance data into the full-link traffic flow prediction model to predict the user's future behavior trajectory; The touchpoint engine in the operating end intelligent suite triggers a corresponding disposal plan based on the experience decision result, including: running the touchpoint engine to trigger a corresponding experience enhancement disposal plan based on the prediction result of the user's future behavior trajectory.

13. A method for realizing terminal intelligent experience, characterized in that: Used to configure and touchpoint the front-end code for the intelligent experience of the client, wherein the front-end code is applicable to multiple types of clients and performs perception, decision-making, and processing of the user experience on the front end. The method is applied to the server and includes: Receive an end-intelligent configuration request sent by a client, match the end-intelligent configuration, and send it to the client, so that the client executes the following steps: run the end-intelligent suite in the front-end code, wherein the perception engine in the end-intelligent suite is started to obtain target data in real time; trigger the decision engine in the end-intelligent suite to make an experience decision for the target data based on the end-intelligent experience configuration; and run the touchpoint engine in the end-intelligent suite to trigger a corresponding disposal plan based on the experience decision result; and In response to the triggering of the callback function, the client receives the execution result of the disposal plan.

14. A device for realizing intelligent experience on a terminal, characterized in that: The device is used to perceive, decide, and process user experience at the front end for front-end codes applicable to multiple types of clients. The device is applied to one client of any type among the multiple types of clients, including: The configuration request unit is used to start the client, send a terminal intelligent experience configuration request to the server, obtain the terminal intelligent experience configuration and store it locally; The end-intelligent execution unit is used to run the end-intelligent suite in the front-end code, including: starting the perception engine in the end-intelligent suite to obtain target data in real time; triggering the decision engine in the end-intelligent suite to make experience decisions for the target data based on the end-intelligent experience configuration; and running the touchpoint engine in the end-intelligent suite to trigger the corresponding disposal plan according to the experience decision results.

15. A storage medium, characterized in that: The storage medium stores a computer program, wherein the computer program is configured to execute the method according to any one of claims 1 to 13 when executed.

16. An electronic device comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the method according to any one of claims 1 to 13.

Images