APP interface visual communication adaptive optimization method and system based on user satisfaction

Abstract

The invention discloses an APP interface visual communication adaptive optimization method and system based on user satisfaction, and the method comprises the steps: constructing a visual element-satisfaction association prediction model, taking historical satisfaction score, operation feedback and scene feature data as a training basis, and carrying out the training through a gradient boosting tree algorithm; collecting real-time scene features and initial interaction data of a target user, and generating candidate visual element parameters and prediction scores; constructing an instant preference vector, and screening a to-be-verified visual scheme in combination with cosine similarity; a local gray level display scheme is adopted, and multi-modal feedback is collected; constructing an evaluation matrix, and calculating a real-time satisfaction comprehensive score by using an analytic hierarchy process; determining a scheme and incrementally updating the model if the standard is reached, and regenerating parameters if the standard is not reached; the system comprises six components such as a model iteration engine and a real-time data collector, and dynamic adaptation is achieved cooperatively. According to the method, user preferences and scenes can be accurately matched, the visual experience and the operation efficiency are improved, continuous optimization can be achieved through model iteration, and the method is suitable for various APPs needing personalized interfaces.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of software interface design, and particularly relates to an APP interface visual communication adaptive optimization method and system based on user satisfaction. BACKGROUND

[0002] With the rapid development of mobile internet technology, APP has become the core carrier for users to obtain information and complete service interaction, and interface visual communication effect directly determines user experience perception and use willingness. The current APP user group presents significant individualized demand differences, for example, the demand of elderly users for font clarity and color contrast is higher than that of young users, the requirement of users in outdoor scenes for screen brightness and visual element recognition is obviously different from that in indoor scenes, and the preferences of users with different operation habits for interface layout and interaction effects are also different. However, the traditional APP interface design mostly adopts a “unified template” mode, which cannot dynamically adjust the visual communication scheme according to the individual characteristics of users and the use scene, leading to problems such as “visual fatigue”, “operation mis-touch” and “low information acquisition efficiency” of some users, and ultimately causing the decrease of user retention rate and the decrease of core function conversion rate. Under this background, there is an urgent need for an APP interface visual communication optimization technology that takes user satisfaction as the core orientation and can adaptively match user demand and scene characteristics, accurately captures user preferences and dynamically adjusts visual element parameters, solves the contradiction between “one-size-fits-all” interface design and individualized user demand, and improves the overall user experience and market competitiveness of APP.

[0003] The traditional APP interface visual communication optimization scheme mainly relies on the experience of designers, collects feedback through regular user research and focus group interviews, and then iterates the interface version based on the feedback results. Specifically, the traditional scheme usually follows the process of “demand analysis—design and development—gray test—full online”, for example, for font size optimization, the designer will determine 2-3 sets of alternative schemes based on industry general standards (such as iOS human-computer interaction guidelines and Android design specifications), collect satisfaction scores through small-scale user testing, and finally select the highest scoring scheme for full deployment. The advantages of such a scheme are that the design logic is clear, the landing cost is relatively low, and it can meet the basic needs of mass users to a certain extent, and is suitable for APP products with simple functions and high homogeneity of user groups. However, the disadvantages are also very prominent: on the one hand, the optimization cycle is long, and it usually takes weeks or even months from research to online, which cannot respond to real-time demand changes of users; on the other hand, the scheme lacks individualized adaptation ability, for example, the same font size scheme cannot meet the “enlargement demand” of elderly users and the “simplicity demand” of young users, leading to low satisfaction of some sub-user groups; in addition, the traditional scheme relies on manual judgment, and it is difficult to quantitatively analyze the correlation between “visual element parameters” and “user satisfaction”, and the optimization effect has great uncertainty.

[0004] With the development of artificial intelligence and big data technology, the existing technology gradually evolves towards the direction of "data-driven" interface optimization. Some technologies collect user click path, dwell time and other interaction data to build user preference models and achieve preliminary individual adjustment of interface elements. For example, some e-commerce APPs adjust the layout position of the homepage recommendation module based on the user's historical click on the product category; some information APPs optimize the font style and line spacing of the article detail page based on the user's reading time. Compared with traditional experience-driven solutions, such technologies have certain improvements in individualization and response speed, can capture user short-term preferences through real-time interaction data, and reduce subjective bias caused by human intervention. However, the existing technology still has obvious limitations: first, the data dimension is single, focusing on user operation behavior data, ignoring the influence of scene feature data (such as environmental lighting, device holding posture) and physiological feedback data (such as eye gaze trajectory, touch pressure) on user satisfaction, leading to the disconnection between optimization scheme and actual use scene; second, it lacks a closed-loop iteration mechanism, most technologies can only achieve "one-time optimization", cannot continuously adjust model parameters based on user's subsequent satisfaction feedback, and the optimization effect is difficult to maintain for a long time; third, the model precision is insufficient, the existing technology mostly uses simple statistical analysis methods (such as frequency statistics, association rule mining) to build preference models, which cannot accurately quantify the non-linear relationship between visual element parameters and user satisfaction, resulting in low accuracy of individualized recommendation.

[0005] Chinese patent (CN117850786A) discloses a "method and system for generating adaptive user interfaces based on user preferences". This technology collects historical user interaction data through a differential privacy mechanism, extracts interface preference features using a deep learning model, constructs a modular programmatic interface generation model, and uses a policy gradient algorithm to train an intelligent agent model to optimize interface parameters. Although it has some innovation in user privacy protection and modular interface generation, its technical solution has serious shortcomings and cannot meet the actual needs of adaptive optimization of APP interface visual communication. First, the core objective is "generating personalized interfaces", but it does not use "user satisfaction rating" as the core optimization indicator. It only indirectly infers preferences through user interaction behavior, lacks direct quantification and feedback of user subjective satisfaction, which may cause the optimization direction to deviate from the actual needs of users. Secondly, while the technology uses reinforcement learning to update the intelligent agent model, it lacks a closed-loop mechanism of "multimodal feedback collection—comprehensive satisfaction evaluation—incremental model iteration." This prevents dynamic adjustment of visual element parameters based on multi-dimensional feedback such as user eye gaze, voice evaluation, and touch pressure, resulting in limited optimization accuracy. Furthermore, it fails to consider the impact of scene characteristics on the interface's visual communication. For example, it doesn't adjust visual elements (such as brightness and animation loading methods) for factors like ambient lighting and network latency, leading to poor user experience consistency across different scenarios. The technology's interface generation relies on the combination of modular components and doesn't model the relationship between specific parameters of visual elements (colors, fonts, icons, etc.) and user satisfaction. This prevents fine-grained adaptive adjustment of visual element parameters and fails to meet users' personalized needs for interface visual details. Summary of the Invention

[0006] To address the aforementioned technical issues, this application discloses an adaptive optimization method and system for APP interface visual communication based on user satisfaction, wherein the optimization method specifically includes:

[0007] S1. Construct an APP interface visual element-satisfaction correlation prediction model. The training data of the prediction model includes historical user satisfaction rating data for each visual element, user operation feedback data, and scene feature data. The training data is trained using the gradient boosting tree algorithm to output the predicted satisfaction rating corresponding to the candidate visual element parameters.

[0008] S2. Collect real-time scene feature data and initial visual interaction data of the target user currently using the APP, and input the real-time scene feature data into the visual element-satisfaction association prediction model to generate candidate visual element parameters and corresponding predicted satisfaction scores.

[0009] S3. Construct the instant preference vector of the target user based on the initial visual interaction data, calculate the matching degree between the candidate visual element parameters and the instant preference vector using the cosine similarity algorithm, and select the candidate visual element parameters with a matching degree ≥ a preset threshold and the highest predicted satisfaction score as the visual solution to be verified.

[0010] S4. Display the visual solution to be verified in partial grayscale on the APP interface, and at the same time start the multimodal feedback acquisition module to collect objective and subjective behavioral feedback data of the target user during the use of the visual solution to be verified.

[0011] S5. Based on the feedback data collected by the multimodal feedback acquisition module, a real-time satisfaction evaluation matrix is ​​constructed. The weight values ​​of each feedback type are calculated by the analytic hierarchy process. Combined with the feedback data scores of each visual element category, the real-time comprehensive satisfaction score of the visual solution to be verified is calculated.

[0012] S6. If the real-time satisfaction score is greater than or equal to the preset satisfaction threshold, the visual scheme to be verified is taken as the final visual scheme for the current target user, and the visual element-satisfaction association prediction model is updated by incremental training. If the real-time satisfaction score is less than the preset satisfaction threshold, return to S2, regenerate the candidate visual element parameters and repeat the above steps until a final visual scheme that meets the preset satisfaction threshold is obtained.

[0013] Preferably, the user operation feedback data in S1 includes eye gaze core area data and operation error data during operation; the operation error data in the user operation feedback data is calculated using an operation deviation quantification formula, assuming the standard coordinates of the target interactive element are (x0, y0), and the user's actual operation coordinates are (x0, y0). i ,y i If the single operation error E is... i The calculation formula is: Average operational error of historical users on a certain visual element It is the arithmetic mean of the errors from multiple operations performed by the current user, i.e. Where n represents the total number of times the current user interacts with the current visual element; the core area data of eye fixation is quantified using the fixation coverage index. Let the area of ​​the core area of ​​the visual element be S0, and the projected area of ​​the user's gaze trajectory within the core area be S... i Then the fixation coverage C i The calculation formula is: Average gaze coverage of historical users on current visual elements It is the arithmetic mean of the coverage over multiple fixations, i.e. Where m represents the total number of times the current user gazes at the current visual element.

[0014] Preferably, during the training of the gradient boosting tree algorithm in S1, Z-score standardization is applied to continuous feature data. Let the original data of a certain continuous feature be X = {x1, x2, ..., x...}. n The standardized data X′={x′1,x′2,...,x′} n The calculation formula is: Where μ is the mean of the original data, i.e. σ is the standard deviation of the original data, i.e. During the model iteration process, the loss function for each regression tree is the squared loss function. Let F be the predicted value of the model in the t-th iteration. t-1 (x i The actual satisfaction rating is y. i Then the residual r ti =y i -F t-1 (x i ), the t-th regression tree h t (x) with residual r ti The training objective is to minimize the squared loss. Final model predictions Where T is the total number of regression trees and γ is the learning rate.

[0015] Preferably, the initial visual interaction data in S2 includes the target user's first click location data on the current APP interface, page dwell time data, and swipe speed data; the swipe speed data is calculated by the ratio of swipe displacement to time difference. Let the coordinates of two consecutive touch points in the user's swipe operation be (x1, y1) and (x2, y2), and the corresponding timestamp be t. a and t b Then the formula for calculating the sliding speed v is: The swipe speed in the initial visual interaction data is the average of multiple swipe speeds. Where k is the total number of sliding operations.

[0016] Preferably, the instant preference vector in S3 includes a click preference dimension, a dwell preference dimension, and an operation rhythm preference dimension; the click preference dimension is represented by the click probability vector of each functional area. Let the APP interface be divided into p functional areas, and the number of clicks by the target user on the j-th area be c. j Total number of clicks: Then the click probability of the j-th region Click preference vector V c =(P1,P2,...,P) p The dwell preference dimension is represented by a vector of dwell time percentages in each region. Let t be the dwell time of a user in the j-th region. jThe total stay duration is Then the percentage of time spent in the j-th region Dwell preference vector V t =(R1,R2,...,R) p ); Operational rhythm preference dimension is based on average sliding speed and average click interval Constructing a two-dimensional vector in t i Let V be the timestamp of the i-th click; let V be the instantaneous preference vector. c V t V r A normalized vector V concatenated according to preset dimension weights norm The normalization formula is: Where ||V|| is the L2 norm of vector V, i.e. q is the total dimension of vector V.

[0017] Preferably, in S3, the matching degree between the candidate visual element parameter feature vector and the instantaneous preference vector is calculated using cosine similarity. Let the instantaneous preference vector be V = (v1, v2, ..., v...). q The feature vector of a candidate visual element parameter is U = (u1, u2, ..., u...). q If the matching degree is ), then the matching degree is . Where U·V is the vector dot product, i.e. U is the L2 norm of the eigenvector U, i.e. V is the L2 norm of the immediate preference vector V; during the screening process, the matching degree Sim(U,V) of all candidate parameters is calculated, and candidate parameters with Sim(U,V)≥θ are retained, where θ is a preset matching threshold. The predicted satisfaction score S is selected from the retained candidate parameters. pred The highest parameter is taken as the visual solution to be verified, that is, the solution to be verified satisfies S. pred =max{S pred1 ,S pred2 ,...,S predn}, where n is the number of candidate parameters that meet the matching criteria.

[0018] Preferably, the multimodal feedback acquisition module in S4 includes an eye-tracking unit, a voice feedback unit, and a touch pressure unit; when the eye-tracking unit calculates the screen coordinates of the gaze hotspot, it uses the pupil center-corneal reflection method, assuming the terminal device screen resolution is (W, H), and the pupil center coordinates in the eye image acquired by the front-facing camera are (p...). x ,p y The coordinates of the corneal reflection point are (c x ,c y If the gaze point is located at a horizontal coordinate S on the screen, then the gaze point's horizontal coordinate S is the coordinate of the gaze point on the screen.x The calculation formula is Vertical coordinates Where a, b, d, and e are the parameters after camera and screen calibration, respectively; the touch pressure value collected by the touch pressure unit is used to quantify the operation stability through pressure deviation rate. Let the multiple touch pressure values ​​of a user on a certain interactive element be {P1, P2, ..., P...} k Average touch pressure Pressure deviation rate A lower deviation rate indicates higher operational stability.

[0019] Preferably, in the S5, the row dimension of the real-time satisfaction evaluation matrix is ​​the feedback type, including gaze feedback, voice feedback, and touch feedback, and the column dimension is the visual element category; regarding the gaze feedback scoring, let R be the percentage of gaze duration in the core area of ​​the visual element. g Then observe feedback rating Where α and β are weighting coefficients, E g E represents the average distance between the fixation point and the core region. max The length is the diagonal of the screen; for voice feedback scoring, it is determined by the emotion tendency value S. emo Keyword matching degree M k The calculation formula is: S v =γ×S emo +(1-γ)×M k γ is the weighting coefficient; regarding touch feedback scoring, let the touch pressure deviation rate be δ, and the deviation rate of touch duration from the standard duration be η, then the touch feedback score S t =1-(λ) s ×δ+μ s ×η), λ s μ s These are the weight coefficients; when calculating the feedback weights using the analytic hierarchy process (AHP), let the judgment matrix be A = (a ij ), where a ij The importance of the i-th type of feedback relative to the j-th type of feedback is represented by the largest eigenvalue λ of the judgment matrix. max The corresponding eigenvectors w = (w1, w2, w3) are then normalized to obtain the feedback weights W. g =w1,W v =w2,W t =w3, calculate the real-time comprehensive satisfaction score in These are the average scores for all visual element categories under each feedback type.

[0020] Preferably, in step S6, when incrementally training and updating the prediction model, weighted fusion of training data is used, and the sample size of the new training data is n. new The original training data sample size is n old Then the sample weights of the new training set Sample weights of the original training set Ensure a balance between the contributions of new and old data to model training; when returning to S2 to regenerate candidate parameters, revise the association rules based on the previous feedback data, assuming the visual element parameter vector of the previous candidate parameters is U. last The corresponding real-time satisfaction rating S last <τ, where τ is a preset threshold, then the range of the corrected parameter generation is U. new ∈[U last -ΔU,U last +ΔU], where ΔU is the correction step size, calculated using the formula: k u Based on the step size coefficient, S last The lower the value, the larger ΔU becomes, thus expanding the range of parameter adjustments and increasing the probability of finding an effective solution.

[0021] An adaptive optimization system for APP interface visual communication based on user satisfaction scores includes a model iteration engine, a real-time data collector, a visual solution generation and filtering device, a multimodal feedback collector, a satisfaction evaluator, and an interface solution deployer.

[0022] The model iteration engine receives historical user satisfaction scores, historical operation feedback, and historical scene feature data, and trains a visual element-satisfaction association prediction model through the gradient boosting tree algorithm. At the same time, it receives new training data, including real-time scene features, instant preference vectors, schemes to be verified, and comprehensive scores, for incremental training, updates the prediction model, and pushes the trained or updated model to the visual scheme generation filter through a one-way model interface.

[0023] The real-time data acquisition device captures real-time scene feature data such as ambient light intensity, device holding posture, network latency, and initial visual interaction data such as the target user's first click position, page dwell time, and scrolling speed. The real-time scene feature data and the initial visual interaction data are pushed to the visual solution generation filter through a bidirectional data interface.

[0024] The visual scheme generation filter receives the prediction model from the model iteration engine and the real-time scene feature data from the real-time data collector. It calls the model to generate multiple sets of candidate visual element parameters and corresponding prediction satisfaction scores. At the same time, it receives the initial visual interaction data from the real-time data collector, constructs an instant preference vector, and filters out the visual schemes to be verified. The schemes to be verified are pushed to the multimodal feedback collector and the interface scheme deployer through a one-way data interface.

[0025] The multimodal feedback collector is based on the visual solution to be verified. After integrating multimodal feedback data through an eye-tracking unit, a voice processing unit, and a touch sensing unit, it pushes the data to the satisfaction evaluator through a one-way data interface.

[0026] The satisfaction evaluator receives feedback data from the multimodal feedback collector, constructs a real-time satisfaction evaluation matrix, calculates the weight of each feedback type using the analytic hierarchy process, and calculates a real-time comprehensive satisfaction score by combining the feedback scores of each visual element. The comprehensive score is then pushed to the interface solution deployer via a one-way data interface.

[0027] The interface solution deployer receives the solutions to be verified from the visual solution generation filter and the comprehensive score from the satisfaction evaluator. It then sends a full rendering command to the APP interface and packages the real-time scene feature data, instant preference vector, solutions to be verified, and comprehensive score into new training data, which is pushed to the model iteration engine through a one-way data interface. Simultaneously, it calculates the parameter correction step size, generates parameter adjustment range commands, and feeds them back to the visual solution generation filter through a one-way data interface to trigger the regeneration of candidate parameters, thus completing closed-loop optimization in conjunction with the model iteration engine.

[0028] Compared with the prior art, the technical solution of this application has the following technical effects:

[0029] This invention utilizes a gradient boosting tree algorithm, trained on historical user satisfaction ratings, operational feedback, and scene feature data. It standardizes continuous features using Z-scores and iteratively optimizes the regression tree model with a squared loss function. The constructed visual element-satisfaction correlation prediction model accurately outputs the predicted satisfaction rating corresponding to candidate visual element parameters when real-time scene feature data of the target user is input. This provides a scientific quantitative basis for subsequent visual solution generation, effectively establishing a mapping relationship between visual element parameters and user satisfaction, and improving the accuracy of the initial visual solution generation.

[0030] This invention collects initial visual interaction data such as the target user's first click location, page dwell time, and scrolling speed to construct an instant preference vector containing a click probability vector, a dwell time percentage vector, and an operation rhythm vector. After normalization, the matching degree between candidate visual element parameters and the preference vector is calculated using a cosine similarity algorithm. The visual solution to be verified that meets the matching degree standard and has the highest prediction score is selected. This invention can quickly transform real-time user interaction behavior into specific visual adaptation requirements, ensuring that the generated visual solution is highly consistent with the user's current usage habits.

[0031] This invention utilizes a multimodal feedback acquisition module composed of an eye-tracking unit, a voice feedback unit, and a touch pressure unit to acquire various types of feedback data, including user gaze hotspots, gaze duration, subjective voice evaluations, touch pressure values, and touch duration. It constructs a real-time satisfaction evaluation matrix with rows representing gaze, voice, and touch feedback, and columns representing various visual elements. Then, it calculates the weights of each feedback type using the analytic hierarchy process (AHP), and combines the scores of visual elements under different feedback types to calculate a comprehensive score. This achieves a comprehensive and objective evaluation of the visual solution's satisfaction, avoiding evaluation bias caused by single feedback data.

[0032] This invention packages real-time scene features, instant preference vectors, solutions, and scores into new training data, and uses a weighted fusion method to incrementally train the prediction model, balancing the contribution of new and old data. If the target is not met, the step size is adjusted based on the difference between the previous score and the threshold, and the parameter generation range is adjusted for re-optimization. This closed-loop mechanism allows the prediction model to continuously absorb new data and iterate and upgrade, while ensuring that the visual solution is always optimized in the direction of meeting the user satisfaction threshold, achieving dynamic and continuous improvement of visual communication effects.

[0033] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings.

[0034] The above and other objects, advantages and features of this application will become more apparent to those skilled in the art from the following detailed description of specific embodiments in conjunction with the accompanying drawings. Attached Figure Description

[0035] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. In all drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.

[0036] Based on the description of the figures and their corresponding technical content in the document, the titles of the figures are as follows:

[0037] Figure 1 A flowchart illustrating an adaptive optimization method for the visual communication of an app interface based on user satisfaction.

[0038] Figure 2 System architecture diagram for adaptive optimization of APP interface visual communication based on user satisfaction;

[0039] Figure 3 This is a schematic diagram illustrating the working principle and data flow of a multimodal feedback data acquisition device.

[0040] Figure 4 Comparison of the visual communication effects of the interface under different ambient lighting intensities;

[0041] Figure 5 A chart comparing user operation frequency in different app usage scenarios;

[0042] Figure 6 A comparison chart showing user visual comfort ratings for different font sizes;

[0043] Figure 7 A comparison chart showing the changes in user satisfaction scores between this application and the traditional solution during the testing period;

[0044] Figure 8 This is a comparison of the daytime and nighttime visual communication effects of the APP interface under this application scheme. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. In the following description, specific details such as specific configurations and components are provided merely to help fully understand the embodiments of this application. Therefore, those skilled in the art should understand that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. In addition, for clarity and brevity, descriptions of known functions and structures are omitted in the embodiments.

[0046] It should be understood that the phrase "an embodiment" or "this embodiment" throughout the specification means that a specific feature, structure, or characteristic related to the embodiment is included in at least one embodiment of this application. Therefore, "an embodiment" or "this embodiment" appearing throughout the specification does not necessarily refer to the same embodiment. Furthermore, these specific features, structures, or characteristics can be combined in any suitable manner in one or more embodiments.

[0047] Furthermore, reference numerals and / or letters may be repeated in different examples within this application. Such repetition is for the purpose of simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or settings discussed.

[0048] In this article, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can mean: A exists alone, B exists alone, and A and B exist simultaneously. The term " / and" in this article describes another type of relationship between related objects, indicating that two relationships can exist. For example, A / and B can mean: A exists alone, and A and B exist alone. In addition, the character " / " in this article generally indicates that the related objects before and after it are in an "or" relationship.

[0049] In this article, the term "at least one" is merely a description of the relationship between related objects, indicating that there can be three relationships. For example, "at least one of A and B" can mean: A exists alone, A and B exist simultaneously, or B exists alone.

[0050] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion.

[0051] Example 1

[0052] This embodiment mainly describes an adaptive optimization method for the visual communication of an app interface based on user satisfaction, such as... Figure 1 As shown, it specifically includes:

[0053] S1. Construct an APP interface visual element-satisfaction correlation prediction model. The training data of the prediction model includes historical user satisfaction rating data for each visual element, user operation feedback data, and scene feature data. The training data is trained using the gradient boosting tree algorithm to output the predicted satisfaction rating corresponding to the candidate visual element parameters.

[0054] S2. Collect real-time scene feature data and initial visual interaction data of the target user currently using the APP, and input the real-time scene feature data into the visual element-satisfaction association prediction model to generate candidate visual element parameters and corresponding predicted satisfaction scores.

[0055] S3. Construct the instant preference vector of the target user based on the initial visual interaction data, calculate the matching degree between the candidate visual element parameters and the instant preference vector using the cosine similarity algorithm, and select the candidate visual element parameters with a matching degree ≥ a preset threshold and the highest predicted satisfaction score as the visual solution to be verified.

[0056] S4. Display the visual solution to be verified in partial grayscale on the APP interface, and at the same time start the multimodal feedback acquisition module to collect objective and subjective behavioral feedback data of the target user during the use of the visual solution to be verified.

[0057] S5. Based on the feedback data collected by the multimodal feedback acquisition module, a real-time satisfaction evaluation matrix is ​​constructed. The weight values ​​of each feedback type are calculated by the analytic hierarchy process. Combined with the feedback data scores of each visual element category, the real-time comprehensive satisfaction score of the visual solution to be verified is calculated.

[0058] S6. If the real-time satisfaction score is greater than or equal to the preset satisfaction threshold, the visual scheme to be verified is taken as the final visual scheme for the current target user, and the visual element-satisfaction association prediction model is updated by incremental training. If the real-time satisfaction score is less than the preset satisfaction threshold, return to S2, regenerate the candidate visual element parameters and repeat the above steps until a final visual scheme that meets the preset satisfaction threshold is obtained.

[0059] Furthermore, the user operation feedback data in S1 includes eye gaze core area data and operation error data during operation; the operation error data in the user operation feedback data is calculated using an operation deviation quantification formula, assuming the standard coordinates of the target interactive element are (x0, y0), and the user's actual operation coordinates are (x0, y0). i ,y iIf the single operation error E is... i The calculation formula is: Average operational error of historical users on a certain visual element It is the arithmetic mean of the errors from multiple operations performed by the current user, i.e. Where n represents the total number of times the current user interacts with the current visual element; the core area data of eye fixation is quantified using the fixation coverage index. Let the area of ​​the core area of ​​the visual element be S0, and the projected area of ​​the user's gaze trajectory within the core area be S... i Then the fixation coverage C i The calculation formula is: Average gaze coverage of historical users on current visual elements It is the arithmetic mean of the coverage over multiple fixations, i.e. Where m represents the total number of times the current user gazes at the current visual element.

[0060] Furthermore, during the training of the gradient boosting tree algorithm in S1, continuous feature data is standardized using Z-score. Let the original data of a certain continuous feature be X = {x1, x2, ..., x...}. n The standardized data X′={x′1,x′2,...,x′} n The calculation formula is: Where μ is the mean of the original data, i.e. σ is the standard deviation of the original data, i.e. During the model iteration process, the loss function for each regression tree is the squared loss function. Let F be the predicted value of the model in the t-th iteration. t-1 (x i The actual satisfaction rating is y. i Then the residual r ti =y i -F t-1 (x i ), the t-th regression tree h t (x) with residual r ti The training objective is to minimize the squared loss. Final model predictions Where T is the total number of regression trees and γ is the learning rate.

[0061] Furthermore, the initial visual interaction data in S2 includes the target user's first click location on the current APP interface, page dwell time, and swipe speed data. The swipe speed data is calculated by the ratio of swipe displacement to time difference. Let the coordinates of two consecutive touch points during the user's swipe operation be (x1, y1) and (x2, y2), and the corresponding timestamp be t. a and t b Then the formula for calculating the sliding speed v is: The swipe speed in the initial visual interaction data is the average of multiple swipe speeds. Where k is the total number of sliding operations.

[0062] Furthermore, the instant preference vector in S3 includes click preference dimension, dwell preference dimension, and operation rhythm preference dimension; the click preference dimension is represented by the click probability vector of each functional area. Let the APP interface be divided into p functional areas, and the number of clicks by the target user on the j-th area be c. j Total number of clicks: Then the click probability of the j-th region Click preference vector V c =(P1,P2,...,P) p The dwell preference dimension is represented by a vector of dwell time percentages in each region. Let t be the dwell time of a user in the j-th region. j The total stay duration is Then the percentage of time spent in the j-th region Dwell preference vector V t =(R1,R2,...,R) p ); Operational rhythm preference dimension is based on average sliding speed and average click interval Constructing a two-dimensional vector in t i Let V be the timestamp of the i-th click; let V be the instantaneous preference vector. c V t V r A normalized vector V concatenated according to preset dimension weights norm The normalization formula is: Where ||V|| is the L2 norm of vector V, i.e. q is the total dimension of vector V.

[0063] Furthermore, the matching degree between the candidate visual element parameter feature vector and the immediate preference vector in S3 is calculated using cosine similarity. Let the immediate preference vector be V = (v1, v2, ..., v q The feature vector of a candidate visual element parameter is U = (u1, u2, ..., u...). q If the matching degree is ), then the matching degree is . Where U·V is the vector dot product, i.e. U is the L2 norm of the eigenvector U, i.e. V is the L2 norm of the immediate preference vector V; during the screening process, the matching degree Sim(U,V) of all candidate parameters is calculated, and candidate parameters with Sim(U,V)≥θ are retained, where θ is a preset matching threshold. The predicted satisfaction score S is selected from the retained candidate parameters.pred The highest parameter is taken as the visual solution to be verified, that is, the solution to be verified satisfies S. pred =max{S pred1 ,S pred2 ,...,S predn}, where n is the number of candidate parameters that meet the matching criteria.

[0064] Furthermore, the multimodal feedback acquisition module in S4 includes an eye-tracking unit, a voice feedback unit, and a touch pressure unit; when calculating the screen coordinates of the gaze hotspot, the eye-tracking unit uses the pupil center-corneal reflection method. Assuming the terminal device screen resolution is (W, H), and the pupil center coordinates in the eye image captured by the front-facing camera are (p...). x ,p y The coordinates of the corneal reflection point are (c x ,c y If the gaze point is located at a horizontal coordinate S on the screen, then the gaze point's horizontal coordinate S is the coordinate of the gaze point on the screen. x The calculation formula is Vertical coordinates Where a, b, d, and e are the parameters after camera and screen calibration, respectively. a is the horizontal deviation compensation value between the pupil center and the corneal reflection point in the eye image; b is the mapping ratio coefficient between horizontal pixels in the eye image and horizontal pixels on the screen; d is the vertical deviation compensation value between the pupil center and the corneal reflection point in the eye image; and e is the mapping ratio coefficient between vertical pixels in the eye image and vertical pixels on the screen. The touch pressure value collected by the touch pressure unit quantifies the operation stability through the pressure deviation rate. Let the multiple touch pressure values ​​of a user on a certain interactive element be {P1, P2, ..., P...}. k Average touch pressure Pressure deviation rate A lower deviation rate indicates higher operational stability.

[0065] Furthermore, in S5, the row dimension of the real-time satisfaction evaluation matrix represents the feedback type, including gaze feedback, voice feedback, and touch feedback, while the column dimension represents the visual element category. Regarding the gaze feedback scoring, let R be the percentage of gaze duration in the core area of ​​the visual element. g Then observe feedback rating Where α and β are weighting coefficients, E g E represents the average distance between the fixation point and the core region. max The length is the diagonal of the screen; for voice feedback scoring, it is determined by the emotion tendency value S. emo Keyword matching degree M k The calculation formula is: S v =γ×S emo +(1-γ)×M kγ is the weighting coefficient; regarding touch feedback scoring, let the touch pressure deviation rate be δ, and the deviation rate of touch duration from the standard duration be η, then the touch feedback score S t =1-(λ) s ×δ+μ s ×η), λ s μ s These are the weight coefficients; when calculating the feedback weights using the analytic hierarchy process (AHP), let the judgment matrix be A = (a ij ), where a ij The importance of the i-th type of feedback relative to the j-th type of feedback is represented by the largest eigenvalue λ of the judgment matrix. max The corresponding eigenvectors w = (w1, w2, w3) are then normalized to obtain the feedback weights W. g =w1,W v =w2,W t =w3, calculate the real-time comprehensive satisfaction score in These are the average scores for all visual element categories under each feedback type.

[0066] Furthermore, in S6, when incrementally training and updating the prediction model, weighted fusion of training data is used, with the sample size of the new training data set being n. new The original training data sample size is n old Then the sample weights of the new training set Sample weights of the original training set Ensure a balance between the contributions of new and old data to model training; when returning to S2 to regenerate candidate parameters, revise the association rules based on the previous feedback data, assuming the visual element parameter vector of the previous candidate parameters is U. last The corresponding real-time satisfaction rating S last <τ, where τ is a preset threshold, then the range of the corrected parameter generation is U. new ∈[U last -ΔU,U last +ΔU], where ΔU is the correction step size, calculated using the formula: k u Based on the step size coefficient, S last The lower the value, the larger ΔU becomes, thus expanding the range of parameter adjustments and increasing the probability of finding an effective solution.

[0067] This implementation details a visual element-satisfaction correlation prediction model built using a gradient boosting tree algorithm, with user satisfaction as the core orientation. It generates candidate solutions by combining real-time scene and interaction data, and then uses cosine similarity screening, multimodal feedback evaluation, and hierarchical analysis scoring to achieve accurate adaptation of visual solutions. This model can dynamically match user preferences and usage scenarios, improve the visual comfort of the interface and operational efficiency, and continuously optimize the model through incremental training to improve user satisfaction and the conversion effect of the APP's core functions.

[0068] Example 2

[0069] This embodiment describes in detail an adaptive optimization system for the visual communication of an app interface based on user satisfaction, such as... Figure 2 As shown, it includes a model iteration engine, a real-time data collector, a visual solution generation filter, a multimodal feedback collector, a satisfaction evaluator, and an interface solution deployer.

[0070] The model iteration engine is responsible for two core tasks. First, it receives historical user satisfaction scores, historical operation feedback, and historical scene feature data. It uses the gradient boosting tree algorithm to standardize continuous feature data with Z-score and iteratively trains the regression tree with the squared loss function as the training objective to generate a visual element-satisfaction association prediction model. Second, it receives new training data (including real-time scene features, instant preference vectors, schemes to be verified, and comprehensive scores), performs incremental training by weighted fusion of training data, updates the prediction model, and pushes the trained or updated model to the visual scheme generation filter through a one-way model interface.

[0071] The real-time data acquisition unit is responsible for capturing two types of data in real time. The first is real-time scene feature data such as the ambient light intensity of the target user, the device holding posture, and network latency (obtained by calculating the difference between the time it takes for the terminal device to send data packets, the time it takes for the server to receive and return response packets, and the time it takes for the terminal to receive response packets). The second is initial visual interaction data such as the target user's first click position, page dwell time, and swipe speed (obtained by calculating the ratio of the displacement of continuous touch points to the time difference and taking the average). The real-time scene feature data and the initial visual interaction data are pushed to the visual solution generation filter through a bidirectional data interface.

[0072] The visual solution generation filter is responsible for three core operations: First, it receives the prediction model from the model iteration engine and the real-time scene feature data from the real-time data collector, and calls the model to generate at least multiple sets of candidate visual element parameters and corresponding predicted satisfaction scores; Second, it receives the initial visual interaction data from the real-time data collector, calculates the click probability vector, dwell time ratio vector, and operation rhythm vector of each functional area, and generates an instant preference vector after normalization; Third, it calculates the cosine similarity between the candidate parameter feature vector and the instant preference vector, filters out the visual solutions to be verified with a matching degree ≥ a preset threshold and the highest predicted score, and pushes the solutions to be verified to the multimodal feedback collector and the interface solution deployer through a one-way data interface.

[0073] like Figure 3 As shown, the multimodal feedback collector is responsible for integrating the multimodal feedback data based on the visual solution to be verified through an eye-tracking unit (which calculates the screen coordinates of the gaze point using the pupil center-corneal reflection method and collects gaze hotspot and gaze duration data), a speech processing unit (which obtains subjective speech evaluation data through sentiment analysis and keyword matching), and a touch sensing unit (which calculates the touch pressure deviation rate and collects touch pressure value and touch duration data). The data is then pushed to the satisfaction evaluator through a one-way data interface.

[0074] The satisfaction evaluator is responsible for receiving feedback data from the multimodal feedback collector, constructing a real-time satisfaction evaluation matrix with row dimensions of gaze / voice / touch feedback and column dimensions of color / font / icon / layout / motion effect elements, calculating the maximum eigenvalue and corresponding eigenvector of the judgment matrix through the analytic hierarchy process to determine the weight of each feedback type, and calculating the real-time comprehensive satisfaction score by combining the feedback scores of each visual element, and pushing the comprehensive score to the interface solution deployer through a one-way data interface.

[0075] The interface solution deployer is responsible for two key actions. First, it receives the solutions to be verified from the visual solution generation filter and the comprehensive score from the satisfaction evaluator. If the score is greater than or equal to a preset threshold, it sends a full rendering command to the APP interface and packages the real-time scene feature data, instant preference vector, solutions to be verified, and comprehensive score into new training data, which is then pushed to the model iteration engine through a one-way data interface. Second, if the score is less than the preset threshold, it calculates the parameter correction step size based on the difference between the real-time satisfaction score of the previous candidate parameter and the preset threshold, generates a parameter adjustment range command, and feeds it back to the visual solution generation filter through a one-way data interface to trigger the regeneration of candidate parameters. Finally, it works with the model iteration engine to complete closed-loop optimization.

[0076] This embodiment details how, through multi-component collaboration, a model iteration engine accurately trains and predicts the model, a real-time data collector captures scene and interaction data, a visual solution generation filter efficiently selects suitable solutions, a multimodal feedback collector comprehensively collects feedback, a satisfaction evaluator scientifically calculates scores, and an interface solution deployer dynamically implements or iteratively optimizes the solution. This enables dynamic adaptation of the interface visuals to user needs and scenarios, significantly improving user satisfaction and operational experience. Furthermore, it allows for continuous model optimization, making the APP's visual communication more intelligent and efficient.

[0077] Based on Embodiment 1 or 2, this embodiment describes in detail the adaptive optimization of APP interface visual communication for day-night switching scenarios, specifically as follows:

[0078] Using a comprehensive information app with over 5 million daily active users as the application platform, and targeting four typical lighting environment scenarios for users during the day (strong outdoor light, normal indoor light, weak light, and dark light at night, with the daytime scenario defined as 9:00-18:00 and the nighttime scenario defined as 18:00-9:00 the next day), the interface visual communication is adapted to the specific scenarios through real-time scene data collection, dynamic generation of visual solutions, multimodal feedback evaluation, and model iteration.

[0079] The app collected scene feature data (including ambient light intensity and device holding posture) from 1 million users over a period of 6 months, visual element satisfaction rating data (1-5 points for 4 categories of elements: color, font, layout, and animation), and operation feedback data (eye gaze core area data and operation error data) as the training dataset for the model iteration engine. The trained visual element-satisfaction association prediction model achieved a 92% accuracy rate in predicting satisfaction ratings for "high contrast color" and "bold font" in daytime scenes, and a 91% accuracy rate in predicting satisfaction ratings for "low saturation color" and "low brightness animation" in nighttime scenes.

[0080] The APP client integrates a real-time data collector (including a light sensor and a touch sensor) and a multimodal feedback collector (including a front-facing camera eye-tracking module and a voice input module). The server deploys a model iteration engine, a visual solution generation filter, a satisfaction evaluator, and an interface solution deployer. All components interact with each other through standardized API interfaces, and the data transmission latency is controlled within 50ms to ensure the real-time adjustment of the visual solution.

[0081] 2,000 users were selected as the test group (using the optimization scheme of this application), and 2,000 users were selected as the control group (using the traditional fixed day-night interface scheme (non-adaptive interface visual configuration scheme)). The two groups of users were kept consistent in age (18-60 years old), device type (iOS / Android each accounting for 50%), and usage time (2-3 hours per day). The test period was 30 days, and the interface interaction data and satisfaction data of the two groups of users in different lighting environments were recorded every day.

[0082] For color and brightness adaptation driven by ambient light intensity, a real-time data acquisition device uses the device's built-in light sensor to capture ambient light intensity values ​​(unit: lux) in real time. Specifically, the ambient light intensity for daytime outdoor strong light scenes is ≥5000 lux, for daytime indoor normal light scenes it is 100-5000 lux, for nighttime low light scenes it is 10-100 lux, and for nighttime dark light scenes it is ≤10 lux. The visual solution generation filter uses the different light intensity data to call the prediction model to generate corresponding color and brightness parameters. Specific adaptation schemes and effect data are as follows:

[0083] Color parameter adaptation: In daytime scenes, a high-contrast, high-saturation color scheme is generated (primary color contrast ≥ 7:1, saturation ≥ 60%) to ensure text-background differentiation; in nighttime scenes, a low-contrast, low-saturation color scheme is generated (primary color contrast ≤ 3:1, saturation ≤ 30%) to avoid strong light stimulation. In daytime outdoor bright light scenes, the test group achieved a text recognition accuracy of 98.5%, while the control group (fixed contrast 5:1) achieved 89.2%. In low-light nighttime scenes, the test group's user visual fatigue feedback rate was only 3.2%, while the control group (fixed saturation 50%) achieved a visual fatigue feedback rate of 21.7%.

[0084] like Figure 4 As shown in the figure, it can be clearly observed that as the light intensity increases from 10 lux (dark night light) to 10000 lux (strong daylight), the text recognition accuracy of the test group remains stable at over 96%, with a fluctuation range of only 2.3%. The text recognition accuracy of the control group increases from 82.1% to 90.3% with the increase of light intensity, with a fluctuation range of 8.2%. Especially when the light intensity is <100 lux or >5000 lux, the accuracy of the control group is significantly lower than that of the test group, which confirms the color adaptation stability of the proposed solution under different light intensities.

[0085] Brightness parameter adaptation: In daytime scenarios, the base brightness of the interface dynamically adjusts with the light intensity. For every 1000 lux increase in light intensity, the brightness increases by 5% (maximum brightness does not exceed 80% of the device's maximum brightness). In nighttime scenarios, the base brightness of the interface is fixed at 20%-30% of the device's maximum brightness, and can be further reduced based on user touch feedback (down to a minimum of 10%). In daytime outdoor bright light scenarios, the test group manually adjusted the brightness 2.1 times / day, while the control group (fixed brightness 60%) manually adjusted it 8.7 times / day. In nighttime low light scenarios, the test group experienced glare feedback 0.3 times / user / day, while the control group (fixed brightness 40%) experienced glare feedback 4.5 times / user / day.

[0086] like Figure 5 As shown in the figure, the length of the bars represents the number of adjustments, and the horizontal axis is divided into four scenarios: outdoor daytime, indoor daytime, low-light nighttime, and dark-light nighttime. It can be clearly seen that the number of brightness adjustments in the test group is significantly lower than that in the control group in all scenarios. The difference is largest in the outdoor daytime scenario (2.1 adjustments in the test group vs. 8.7 adjustments in the control group), followed by the difference in the dark-light nighttime scenario (0.3 adjustments in the test group vs. 4.5 adjustments in the control group). This indicates that the brightness adaptive solution of this application can accurately match the lighting requirements of different scenarios, reducing the manual operation cost for users.

[0087] The system employs operation feedback-driven font and layout adaptation. A multimodal feedback collector captures real-time user eye gaze data (gaze hotspots, gaze duration) and touch data (touch pressure value, touch duration) in daytime and nighttime scenarios. A satisfaction evaluator calculates satisfaction scores for font and layout parameters based on this data. The interface solution deployer adjusts font size, line spacing, and core button layout according to the scores. Specific adaptation solutions and effect data are as follows:

[0088] Font parameter adaptation: In daytime scenarios, the default font size is set to 16-18pt, and the line spacing is 1.2 times the font size; in nighttime scenarios, the default font size is set to 18-20pt, and the line spacing is 1.4 times the font size. The font weight is also adjusted from regular (400) to medium (500) to improve text readability at night. In daytime scenarios, the average time users gazed at the text area was 1.8 seconds per line in the test group, compared to 2.5 seconds per line in the control group (fixed 16pt, 1.1 times line spacing). In nighttime scenarios, the number of times users re-gazed due to blurred fonts was 0.5 times per page in the test group, compared to 3.2 times per page in the control group (fixed 16pt, regular weight).

[0089] like Figure 6As shown in the figure, the reading time corresponding to different font sizes (16pt, 18pt, 20pt) is displayed in line form, divided into two curves: daytime (solid line) and nighttime (dashed line). It can be observed that in daytime scenarios, the reading time for the test group with 16-18pt font size is basically the same (1.8-1.9 seconds / line), and lower than the 2.3-2.5 seconds / line for the control group with the same font size. In nighttime scenarios, the reading time for the test group with 18-20pt font size is 2.0-2.1 seconds / line, significantly lower than the 2.8 seconds / line for the control group with 16pt font size. This indicates that the font adaptation scheme of this application can optimize reading efficiency in different scenarios, especially in nighttime scenarios.

[0090] Layout parameter adaptation: In daytime scenarios, core interactive buttons (such as "favorite", "share", and "comment") adopt a compact layout with a button spacing of 8-10 dp, concentrated in the bottom 1 / 3 area of ​​the screen; in nighttime scenarios, the spacing between core buttons is increased to 12-15 dp, the button size is increased by 10%, and a faint glow effect (brightness ≤5 cd / m²) is added to the button edges. 2 This improves button recognition and touch accuracy. In daytime scenarios, the average touch error of the core button in the test group was 1.2mm, while that in the control group (fixed 10dp spacing, standard size) was 2.5mm. In nighttime scenarios, the false touch rate of the core button in the test group was 1.8%, while that in the control group (fixed 10dp spacing, no lighting effect) was 9.3%.

[0091] Combination Figure 5 and Figure 6 Further investigation revealed that in nighttime scenarios, due to the coordinated adaptation of fonts and layout, the average time for users to complete a single page operation (from opening the page to completing the core operation) was 8.5 seconds in the test group, compared to 14.2 seconds in the control group, representing a 40.1% improvement in operation efficiency. In daytime scenarios, the average time for users to complete a single page operation was 7.2 seconds in the test group, compared to 8.8 seconds in the control group, representing an 18.2% improvement in operation efficiency. This indicates that the layout adaptation of the proposed solution has a more significant effect on improving operation efficiency in nighttime scenarios, and also confirms the necessity of differentiated adaptation of visual solutions between daytime and nighttime scenarios.

[0092] Scene switching response and model iteration optimization effects: The interface solution deployer monitors scene switching events in real time (e.g., a user moves from outdoor daytime to indoor nighttime, with light intensity dropping sharply from 6000 lux to 50 lux), triggering a rapid switching of the visual solution. The switching response time is controlled within 300ms. Simultaneously, the user feedback data after the switch (satisfaction rating, operation data) is packaged into new training data and pushed to the model iteration engine for incremental training, improving the model's adaptation accuracy to scene switching. Specific effect data is as follows:

[0093] Regarding scene switching response speed, in the test group, the average switching time of the interface visual solution during day-night scene transitions was 280ms, with 95% of switching events having a response time ≤300ms, meaning users could not perceive the switching delay. In contrast, the control group, using a fixed solution, required users to manually trigger night mode, with an average trigger time (from the user perceiving the change in light to completing the mode switch) of 8.3 seconds. Furthermore, 30% of users reported "forgetting to manually switch," resulting in a poor nighttime user experience. During the testing period, the test group's user satisfaction rating for the timeliness of scene switching was 4.7 out of 5, while the control group's was 2.3.

[0094] The model iteration optimization effect was as follows: as the testing cycle progressed, the model iteration engine received a total of 1.2 million new training data points. Through weighted fusion (new data weight 0.6, old data weight 0.4), the prediction model was updated, resulting in a reduction in the prediction error for "color contrast-satisfaction" from 8.5% to 3.2% in daytime scenes, and a reduction in the prediction error for "font size-satisfaction" from 9.1% to 2.8% in nighttime scenes. Data from the 30th day of testing showed that the overall satisfaction score of the test group users for the interface visual communication was 4.6 points, an increase of 21.1% from the initial (day 1) score of 3.8 points; the overall satisfaction score of the control group remained stable at 3.1-3.3 points, without significant improvement.

[0095] like Figure 7 As shown, it is clear that the test group's score showed a continuous upward trend over the days, rising from 3.8 points on day 1 to 4.6 points on day 30, with the rate of increase gradually slowing down (the model tended to stabilize in the later stages); the control group's score fluctuated around 3.2 points without a significant upward trend, indicating that the model iteration mechanism of this application can continuously optimize the visual solution and ensure long-term improvement in user experience under day-night switching scenarios, while traditional fixed solutions cannot adapt to the dynamic changes in user needs.

[0096] like Figure 8 As shown, the APP interface switches between day and night scenes. In the daytime scene... Figure 8 (a) The interface uses a bright, light color scheme with a white background. Information cards and other elements are brightly colored, and the text has high contrast with the background, allowing users to read the content clearly in well-lit environments, reducing eye strain and improving information retrieval efficiency; in nighttime scenes… Figure 8 (b) The interface is switched to a dark theme, the background is changed to dark gray, the screen brightness is reduced, and the color matching of text and cards is adjusted. This avoids the stimulation of the eyes by strong light and ensures the recognizability of the content, so that users can have a good visual experience under different lighting conditions.

[0097] This embodiment verifies the technology of this application by applying a visual communication adaptive optimization scheme of "day-night scene switching" in a comprehensive information app, from four dimensions: color brightness, font layout, scene response, and model iteration. In terms of color brightness adaptation, the test group achieved a daytime outdoor text recognition accuracy of 98.5% (compared to 89.2% in the control group), and a nighttime low-light visual fatigue feedback rate of only 3.2% (compared to 21.7% in the control group). In terms of font layout adaptation, the test group completed a single-page operation in the nighttime scene in 8.5 seconds (compared to 14.2% in the control group), with a false touch rate of 1.8% (compared to 9.3% in the control group). In terms of scene response, the test group's switching response time was 280ms (compared to 8.3 seconds in the control group, which required manual switching). In terms of model iteration, the test group's overall satisfaction score increased from 3.8 to 4.6 over 30 days (compared to no improvement in the control group).

[0098] The above data fully demonstrates that the proposed solution can not only achieve good visual communication effects in both daytime and nighttime switching scenarios, but also continuously optimize scene adaptation accuracy through real-time data collection and model iteration. It solves the problems of "difficult recognition, complicated operation, and poor experience" of traditional fixed interface solutions in different lighting environments, significantly improves the user's interface experience and operation efficiency in all scenarios, and has high practical application value.

[0099] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. For those skilled in the art, the present invention can have various modifications and variations. Any changes, modifications, substitutions, integrations, and parameter changes made to these embodiments within the spirit and principles of the present invention, without departing from the principles and spirit of the present invention, through conventional substitutions or to achieve the same function, fall within the scope of protection of the present invention.

Claims

1. An APP interface visual communication adaptive optimization method based on user satisfaction, characterized in that, The method comprises the following steps: S1, constructing an APP interface visual element-satisfaction correlation prediction model, the training data of the prediction model comprising historical user satisfaction score data of each visual element, user operation feedback data and scene feature data, the training data being trained by a gradient boosting tree algorithm, and a predicted satisfaction score corresponding to a candidate visual element parameter being output; S2, collecting real-time scene feature data and initial visual interaction data of a target user currently using an APP, inputting the real-time scene feature data into the visual element-satisfaction correlation prediction model, and generating a candidate visual element parameter and a corresponding predicted satisfaction score; S3, constructing an instant preference vector of the target user based on the initial visual interaction data, calculating the matching degree of the candidate visual element parameter and the instant preference vector by a cosine similarity algorithm, and screening a candidate visual element parameter with a matching degree greater than or equal to a preset threshold and a highest predicted satisfaction score as a to-be-verified visual scheme; S4, performing local grayscale display of the to-be-verified visual scheme on the APP interface, and starting a multi-modal feedback collection module to collect objective and subjective behavior feedback data of the target user during use of the to-be-verified visual scheme; S5, constructing a real-time satisfaction evaluation matrix based on the feedback data collected by the multi-modal feedback collection module, calculating the weight values of each feedback type by an analytic hierarchy process, and calculating a real-time satisfaction comprehensive score of the to-be-verified visual scheme in combination with feedback data scores of each visual element category; S6, if the real-time satisfaction comprehensive score is greater than or equal to a preset satisfaction threshold, the to-be-verified visual scheme is taken as a final visual scheme of the current target user, and the visual element-satisfaction correlation prediction model is updated in an incremental training manner; if the real-time satisfaction comprehensive score is less than the preset satisfaction threshold, returning to S2 to regenerate a candidate visual element parameter and repeating the above steps until a final visual scheme meeting the preset satisfaction threshold is obtained. 2.The user satisfaction based APP interface visual communication adaptive optimization method of claim 1, wherein, The user operation feedback data in S1 comprises eye fixation core area data and operation error data during operation; The operation error data in the user operation feedback data is calculated using the operation deviation quantification formula. Let the standard coordinates of the target interactive element be (x0, y0), and the actual coordinates of the user's operation be (x0, y0). i ,y i If the single operation error E is... i The calculation formula is: Average operational error of historical users on a certain visual element It is the arithmetic mean of the errors from multiple operations performed by the current user, i.e. Where n represents the total number of times the current user interacts with the current visual element; the core area data of eye fixation is quantified using the fixation coverage index. Let the area of ​​the core area of ​​the visual element be S0, and the projected area of ​​the user's gaze trajectory within the core area be S... i Then the fixation coverage C i The calculation formula is: Average gaze coverage of historical users on current visual elements It is the arithmetic mean of the coverage over multiple fixations, i.e. Where m represents the total number of times the current user gazes at the current visual element. 3.The user satisfaction based APP interface visual communication adaptive optimization method of claim 1, wherein, During the training of the gradient boosting tree algorithm in S1, continuous feature data is standardized using Z-score. Let the original data of a certain continuous feature be X = {x1, x2, ..., x...}. n The standardized data X′={x′1,x′2,...,x′} n The calculation formula is: Where μ is the mean of the original data, i.e. σ is the standard deviation of the original data, i.e. During the model iteration process, the loss function for each regression tree is the squared loss function. Let F be the predicted value of the model in the t-th iteration. t-1 (x i The actual satisfaction rating is y. i Then the residual r ti =y i -F t-1 (x i ), the t-th regression tree h t (x) with residual r ti The training objective is to minimize the squared loss. Final model predictions Where T is the total number of regression trees and γ is the learning rate. 4.The user satisfaction based APP interface visual communication adaptive optimization method of claim 1, wherein, The initial visual interaction data in S2 includes first click position data of the target user on the current APP interface, page staying time length data and sliding speed data; the sliding speed data is calculated by the ratio of sliding displacement and time difference, assuming that the coordinates of two continuous touch points in the user sliding operation are (x1, y1) and (x2, y2), and the corresponding time stamps are t a and t b The calculation formula of the sliding speed v is: The sliding speed in the initial visual interaction data is the average value of multiple sliding speeds Where k is the total number of sliding operations. 5.The user satisfaction based APP interface visual communication adaptive optimization method of claim 1, wherein, The instant preference vector in S3 includes click preference dimension, dwell preference dimension, and operation rhythm preference dimension; the click preference dimension is represented by the click probability vector of each functional area. Let the APP interface be divided into p functional areas, and the number of clicks by the target user on the j-th area be c. j Total number of clicks: Then the click probability of the j-th region Click preference vector V c =(P1,P2,...,P) p The dwell preference dimension is represented by a vector of dwell time percentages in each region. Let t be the dwell time of a user in the j-th region. j The total stay duration is Then the percentage of time spent in the j-th region Dwell preference vector V t =(R1,R2,...,R) p ); Operational rhythm preference dimension is based on average sliding speed and average click interval Constructing a two-dimensional vector in t i Let V be the timestamp of the i-th click; let V be the instantaneous preference vector. c V t V r A normalized vector V concatenated according to preset dimension weights norm The normalization formula is: Where ||V|| is the L2 norm of vector V, i.e. q is the total dimension of vector V. 6.The user satisfaction based APP interface visual communication adaptive optimization method of claim 1, wherein, The matching degree of the candidate visual element parameter feature vector in S3 and the instant preference vector is calculated by cosine similarity. Assuming that the instant preference vector is V = (v1, v2,..., vn), the feature vector of a certain candidate visual element parameter is U = (u1, u2,..., un), and the matching degree is Sim(U, V) = (U · V) / (||U|| · ||V||). q q wherein U · V is the vector dot product, i.e. ||U|| is the L2 norm of the feature vector U, i.e. ||V|| is the L2 norm of the instant preference vector V; in the screening process, the matching degrees Sim(U, V) of all candidate parameters are calculated, the candidate parameters with Sim(U, V) ≥ θ are retained, θ is a preset matching threshold, and the candidate parameter with the highest predicted satisfaction score S pred is selected from the retained candidate parameters as the to-be-verified visual scheme, i.e., the to-be-verified scheme satisfies S pred = max{S pred1 , S pred2 ,..., S predn}, wherein n is the number of candidate parameters with the matching degree meeting the standard.​​ 7. The method of claim 1, wherein, The multi-modal feedback collection module in the S4 includes an eye tracking unit, a voice feedback unit, and a touch pressure unit; when the eye tracking unit calculates the gaze hotspot screen coordinates, the pupil center-corneal reflection method is adopted, the terminal device screen resolution is set as (W, H), the pupil center coordinates in the eye image collected by the front camera are (p x ,p y ), the corneal reflection point coordinates are (c x ,c y ), and the horizontal coordinate S x of the gaze point on the screen is calculated according to the following formula The vertical coordinate Wherein a, b, d, and e are parameters after the camera and the screen are calibrated; the touch pressure value collected by the touch pressure unit is quantified by the pressure deviation rate to stabilize the operation, the multiple touch pressure values of a user on a certain interactive element are set as {P1, P2,..., P k}, the average touch pressure is calculated according to the following formula The lower the pressure deviation rate is, the higher the operation stability is. 8.The user satisfaction based APP interface visual communication adaptive optimization method of claim 1, wherein, The row dimension of the real-time satisfaction evaluation matrix in S5 is feedback type, including gaze feedback, voice feedback and touch feedback, and the column dimension is visual element category; as for the gaze feedback score, let the gaze time ratio of the core area of the visual element be R g , then the gaze feedback score is R , where a and β are weight coefficients, E g is the average distance between the gaze point and the core area, E max is the length of the screen diagonal; as for the voice feedback score, the sentiment tendency value S emo and the keyword matching degree M k are calculated, and the formula is: S v =γ×S emo +(1-γ)×M k , where γ is a weight coefficient; as for the touch feedback score, let the touch pressure deviation rate be δ, and the deviation rate of the touch time length and the standard time length be η, then the touch feedback score S t =1-(λ s ×δ+μ s ×η), where λ s and μ s are weight coefficients; when the analytic hierarchy process is used to calculate the feedback weight, let the judgment matrix be A=(a ij ), where a ij represents the importance of the i-th feedback relative to the j-th feedback, the maximum eigenvalue λ max of the judgment matrix and the corresponding eigenvector w=(w1,w2,w3) are calculated, the eigenvector is normalized to obtain the feedback weights W v =w1,W t =w2 and W new =w3, and the real-time satisfaction comprehensive score is calculated , where are the average scores of all visual element categories under each feedback type. 9.The user satisfaction based APP interface visual communication adaptive optimization method of claim 1, wherein, In step S6, when incrementally training and updating the prediction model, weighted fusion of training data is used, with the sample size of the new training data set being n. new The original training data sample size is n old Then the sample weights of the new training set Sample weights of the original training set Ensure a balance between the contributions of new and old data to model training; When the candidate parameters are regenerated at S2, the association rules are corrected based on the last feedback data, and let the visual element parameter vector of the last candidate parameters be U last The corresponding real-time satisfaction score S last <τ, where τ is a preset threshold, and the parameter generation range after correction is U new ∈[U last -ΔU, U last +ΔU], where ΔU is a correction step, and the calculation formula is k u is a basic step coefficient, and S last is lower, ΔU is larger, so as to expand the parameter adjustment range and improve the probability of finding an effective solution.

10. An adaptive optimization system for APP interface visual communication based on user satisfaction, applicable to any one of claims 1-9, characterized in that, The system comprises a model iteration engine, a real-time data collector, a visual scheme generation and screening device, a multi-modal feedback collector, a satisfaction evaluator and an interface scheme deployer; The model iteration engine receives historical user satisfaction scores, historical operation feedback and historical scene feature data, trains a visual element-satisfaction correlation prediction model by a gradient boosting tree algorithm, receives new training data including real-time scene features, instant preference vectors, to-be-verified schemes and comprehensive scores for incremental training, updates the prediction model, and pushes the trained or updated model to the visual scheme generation and screening device through a one-way model interface; The real-time data collector captures real-time scene feature data of environmental light intensity, device holding posture and network delay of a target user, and initial visual interaction data of a first click position, page dwell time and sliding speed of the target user, and pushes the real-time scene feature data to the visual scheme generation and screening device and the initial visual interaction data to the visual scheme generation and screening device through a two-way data interface. The visual scheme generation filter receives the prediction model of the model iteration engine, real-time scene feature data of the real-time data collector, calls the model to generate a plurality of groups of candidate visual element parameters and corresponding prediction satisfaction scores; simultaneously receives initial visual interaction data of the real-time data collector, constructs an instant preference vector, and screens out a to-be-verified visual scheme, and pushes the to-be-verified scheme to the multi-modal feedback collector and the interface scheme deployer through a unidirectional data interface; The multi-modal feedback collector integrates multi-modal feedback data based on the to-be-verified visual scheme through an eye tracking unit, a voice processing unit and a touch sensing unit, and pushes the multi-modal feedback data to the satisfaction evaluator through a unidirectional data interface; The satisfaction evaluator receives feedback data of the multi-modal feedback collector, constructs a real-time satisfaction evaluation matrix, calculates the weight of each feedback type through an analytic hierarchy process, calculates a real-time satisfaction comprehensive score in combination with feedback scores of each visual element, and pushes the comprehensive score to the interface scheme deployer through a unidirectional data interface; The interface scheme deployer receives the to-be-verified scheme of the visual scheme generation filter and the comprehensive score of the satisfaction evaluator, sends a full-amount rendering instruction to an APP interface, packs real-time scene feature data, an instant preference vector, the to-be-verified scheme and the comprehensive score into new training data, and pushes the new training data to the model iteration engine through a unidirectional data interface; simultaneously calculates a parameter correction step, generates a parameter adjustment range instruction, and feeds back the parameter adjustment range instruction to the visual scheme generation filter through a unidirectional data interface, so as to trigger the regeneration of candidate parameters, and complete closed-loop optimization in cooperation with the model iteration engine.

Images