Generation method of visual data large screen based on model driving, medium and equipment

Abstract

The invention discloses a visual data large screen generation method based on model driving, a medium and equipment, and the method comprises the steps: receiving a natural language demand and a data source, analyzing visual intention information through a large language model, carrying out the cleaning and standardization of the data source, and outputting a structured data set; on the basis of the data set field and the visualization intention, matching with a pre-constructed configurable component library is carried out, and candidate visualization components are determined; taking the candidate components as layout elements, and adopting a layout engine fused with a reinforcement learning algorithm to generate an initial visual layout; rendering the initial layout and performing adjustment operation in response to user interaction feedback to obtain a final layout scheme; and finally, binding the layout scheme with the structured data set, rendering to generate an interactive visual data large screen, and outputting the visual data large screen. According to the method, the end-to-end intelligent generation from the natural language requirement to the interactive data large screen is realized, and the development efficiency and the layout quality are remarkably improved.

Description

Technical Field

[0001] This application relates to the field of data processing, specifically to a method, medium, and device for generating a model-driven visualization data dashboard. Background Technology

[0002] In the field of data visualization, traditional methods for generating data dashboards heavily rely on manual coding or the use of fixed templates, exhibiting significant limitations. First, they lack flexibility. When user business needs or the data metrics to be displayed change, existing templates or code struggle to adapt quickly, often requiring developers to redesign and recode, failing to achieve automatic mapping and generation from dynamically changing business requirements to visualization solutions. Second, the entire generation process suffers from low automation and efficiency. From understanding user intent and matching chart types to page layout, almost every step requires deep involvement and manual configuration by professionals, resulting in long development cycles and high labor costs. Furthermore, existing methods lack intelligent layout optimization capabilities. The generated initial layout typically relies on preset fixed rules or the designer's personal experience, making it difficult to achieve a comprehensive balance and optimization of automation and intelligence among multiple objectives such as layout aesthetics, information density, and user personalization preferences.

[0003] Existing technologies, such as visualization component libraries built on open-source libraries like ECharts, or some low-code design systems, while offering convenience in chart rendering, still lack the crucial step of automatically understanding user intent and intelligently matching components. Users still need to manually select and configure data and chart types. Another type of method based on image recognition primarily converts static design drafts into front-end code, but it also fails to address the core issue of automatically understanding data, matching components, and generating intelligent layouts based on unstructured natural language requirements. Summary of the Invention

[0004] In view of the above problems, this application provides a model-driven visualization data dashboard generation technology solution to solve the technical problems of poor flexibility, low automation and lack of intelligent layout optimization capabilities caused by the reliance on manual configuration and fixed templates in traditional data dashboard generation methods.

[0005] To achieve the above objectives, in a first aspect, this application provides a model-driven method for generating a visual data dashboard, the method comprising the following steps:

[0006] S1: Receive the input natural language requirement description and target data source, parse the natural language requirement through a large language model, extract the visualization intent information, clean and standardize the target data source, and output a structured dataset with unified dimension fields and indicator fields. The visualization intent information includes core business indicators, expected chart types, and style preferences.

[0007] S2: Based on the field types of the structured dataset and the visualization intent information, match them with the input data specifications and functional style characteristics of each visualization component in the pre-built configurable component library to determine at least one set of candidate visualization components;

[0008] S3: Using the determined candidate visualization components as layout elements, an initial visualization layout is generated using a layout engine incorporating reinforcement learning algorithms; wherein, the reinforcement learning algorithm is configured to use layout aesthetics, information density balance, and user historical preferences as reward signals, and to search for the optimal layout scheme by iteratively exploring different combinations of the position and size of visualization components.

[0009] S4: Render and present the initial visual layout, and in response to user interaction feedback commands, adjust the initial visual layout to obtain the final layout scheme; the adjustment operation includes: marking user-specified components as locked components with fixed positions and / or sizes, or incrementally adjusting and regenerating the layout of unlocked components;

[0010] S5: Bind the final layout scheme to the structured dataset, render and generate an interactive visual data dashboard, and output it as an independently deployable application or configuration file.

[0011] Furthermore, step S2 specifically includes:

[0012] S21: Obtain and parse the metadata description of each visualization component in the pre-built configurable component library, the metadata description including input data specifications and functional and style characteristics;

[0013] S22: The field types and number of fields in the structured dataset, the chart types and style preferences extracted from the visualization intent information, and the metadata descriptions of the parsed visualization components are matched and scored in multiple dimensions.

[0014] S23: Based on the comprehensive scoring results of multi-dimensional matching, sort the visualization components and select at least one group of visualization components with the highest ranking as the candidate visualization components.

[0015] Furthermore, step S3 specifically includes:

[0016] S31: Define the state space: Map the candidate visualization components and their corresponding initial properties to a discretized grid layout environment; wherein, the state space is defined as the position, size and canvas properties of all visualization components in the grid at the current moment;

[0017] S32: Define action space: The action space is defined as a set of executable operations, specifically including: selecting any unlocked component and moving it to an adjacent empty space in the grid, adjusting its width or height by a preset step size, or aligning it with an adjacent component;

[0018] S33: Calculate the reward function: After each layout action is executed, a comprehensive reward value is calculated based on the updated layout state; the reward function is obtained by weighted calculation of any one or more of the following: aesthetic reward, information density balance reward, and user preference reward;

[0019] S34: Policy Learning and Layout Generation: Based on the current policy network, select actions, interact with the environment and obtain rewards, iteratively update network parameters through the policy gradient method, and finally output the layout scheme with the highest cumulative reward as the initial visual layout.

[0020] Furthermore, in step S3, when the layout engine performs layout optimization, its optimization objective also includes an information transmission efficiency objective, which is constructed through the following steps:

[0021] S35: Calculate an information value weight for each candidate visualization component, wherein the information value weight is the weighted sum of the information entropy of a specific data subset extracted from the structured dataset corresponding to the candidate visualization component and the inherent visual saliency of its chart type;

[0022] S36: In the layout iteration, evaluate the degree to which the current layout scheme highlights high-information-value components, and quantify the degree of highlighting into the value of the information transmission efficiency target, so as to guide the layout engine to place key information in prominent positions.

[0023] Furthermore, the interactive feedback instructions include component locking instructions and natural language adjustment instructions;

[0024] Marking a user-specified component as a locked component with a fixed position and / or size includes: receiving a component locking instruction, marking the user-specified component as a locked component with a fixed position and / or size, and storing the position and size parameters of the locked component in a fixed parameter configuration file;

[0025] Incremental adjustment and regeneration of the layout of unlocked components includes: receiving a user's natural language adjustment instruction, parsing the semantics of the natural language adjustment instruction through a natural language processing model, converting the parsed semantics into attribute adjustment operations for the corresponding unlocked components, and regenerating the adjusted layout of the unlocked components.

[0026] Furthermore, in step S1, the natural language requirement is parsed through the large language model, and a data narrative logic chain is generated. The data narrative logic chain is a directed graph structure. The nodes of the directed graph structure are key data insights extracted from the natural language requirement and structured data. The edges of the directed graph structure define the logical relationship types between the insights. The logical relationship types include causal relationship, comparative relationship, general-to-specific relationship or temporal evolution relationship.

[0027] In step S3, the reward signal of the layout engine also includes a narrative coherence reward, which is used to incentivize the initial visual layout to form a visual narrative path in spatial arrangement that conforms to the data narrative logic chain. The narrative coherence reward is calculated in the following way:

[0028] S37: Map the data narrative logic chain to a desired visual focus sequence, which records the ideal eye movement order of the user when browsing the large screen;

[0029] S38: Based on the current layout state, calculate an actual visual flow path by simulating an eye-tracking model;

[0030] S39: Calculate the degree of consistency between the visual flow path and the visual focus sequence, and quantify this degree of consistency as a narrative coherence reward value;

[0031] The higher the degree of consistency, the greater the narrative coherence reward value.

[0032] Furthermore, the interactive feedback command is a multimodal interactive command, and the method further includes:

[0033] The multimodal interaction instructions are parsed using a multimodal fusion model and converted into specific operation parameters for the visualization components; wherein, the multimodal interaction instructions include any one or more of the following instructions:

[0034] Natural language adjustment instructions are used to describe changes in the properties of visual components;

[0035] Sketch instructions are used to specify the desired size or positional relationship of visual components through graphical outlines;

[0036] Gesture control commands are used to directly manipulate visual components via air gestures.

[0037] Furthermore, the method also includes:

[0038] Real-time monitoring of the distribution changes of the data stream bound to the final layout scheme;

[0039] When a change in data distribution is detected to exceed a preset threshold, an incremental layout adjustment process is automatically triggered. Based on the changed data characteristics, the layout of the current visualization components is dynamically optimized to focus on new data insights.

[0040] In a second aspect, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the model-driven visualization data dashboard generation method as described in the first aspect of this application.

[0041] In a third aspect, this application provides an electronic device having a computer program stored thereon, including a processor and a storage medium, wherein the computer program is stored on the storage medium, and when executed by the processor, the computer program implements the model-driven visualization data dashboard generation method as described in the first aspect of this application.

[0042] Unlike existing technologies, the above technical solution involves a model-driven method, medium, and device for generating visual data dashboards. The method includes: receiving natural language requirements and data sources; parsing the visualization intent information using a large language model; cleaning and standardizing the data sources to output a structured dataset; matching the dataset fields and visualization intent with a pre-built configurable component library to determine candidate visualization components; using the candidate components as layout elements, generating an initial visualization layout using a layout engine incorporating reinforcement learning algorithms, where the algorithm uses layout aesthetics, information density balance, and user historical preferences as reward signals, iteratively exploring component position and size combinations to search for the optimal layout scheme; rendering the initial layout and adjusting it in response to user interaction feedback to obtain the final layout scheme; finally, binding the layout scheme with the structured dataset, rendering and outputting an interactive visual data dashboard. This invention achieves end-to-end intelligent generation from natural language requirements to an interactive data dashboard, significantly improving development efficiency and layout quality.

[0043] The above description of the invention is merely an overview of the technical solution of this application. In order to enable those skilled in the art to better understand the technical solution of this application and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of this application easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of this application. Attached Figure Description

[0044] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of this application and other related content, and should not be considered as limitations on this application.

[0045] In the accompanying drawings of the instruction manual:

[0046] Figure 1 A flowchart illustrating the method for generating a model-driven visualization data dashboard according to the first exemplary embodiment of this application;

[0047] Figure 2 A flowchart illustrating the method for generating a model-driven visualization data dashboard according to the second exemplary embodiment of this application;

[0048] Figure 3 A flowchart illustrating the method for generating a model-driven visualization data dashboard according to the third exemplary embodiment of this application;

[0049] Figure 4 A flowchart illustrating the method for generating a model-driven visualization data dashboard according to the fourth exemplary embodiment of this application;

[0050] Figure 5 A flowchart illustrating the model-driven visualization data dashboard generation method described in the fifth exemplary embodiment of this application;

[0051] Figure 6 This is a schematic diagram of an electronic device according to an exemplary embodiment of this application;

[0052] The reference numerals used in the above figures are explained as follows:

[0053] 10. Electronic devices;

[0054] 101. Processor;

[0055] 102. Storage medium. Detailed Implementation

[0056] To explain in detail the possible application scenarios, technical principles, specific feasible solutions, and the objectives and effects that this application can achieve, the following detailed description is provided in conjunction with the listed specific embodiments and accompanying drawings. The embodiments described herein are only used to more clearly illustrate the technical solutions of this application, and are therefore only examples, and should not be used to limit the scope of protection of this application.

[0057] like Figure 1 As shown, in a first aspect, this application provides a model-driven method for generating a visual data dashboard, the method comprising:

[0058] S1: Receive the input natural language requirement description and target data source, parse the natural language requirement through a large language model, extract the visualization intent information, clean and standardize the target data source, and output a structured dataset with unified dimension fields and indicator fields. The visualization intent information includes core business indicators, expected chart types, and style preferences.

[0059] S2: Based on the field types of the structured dataset and the visualization intent information, match them with the input data specifications and functional style characteristics of each visualization component in the pre-built configurable component library to determine at least one set of candidate visualization components;

[0060] S3: Using the determined candidate visualization components as layout elements, an initial visualization layout is generated using a layout engine incorporating reinforcement learning algorithms; wherein, the reinforcement learning algorithm is configured to use layout aesthetics, information density balance, and user historical preferences as reward signals, and to search for the optimal layout scheme by iteratively exploring different combinations of the position and size of visualization components.

[0061] S4: Render and present the initial visual layout, and in response to user interaction feedback commands, adjust the initial visual layout to obtain the final layout scheme; the adjustment operation includes: marking user-specified components as locked components with fixed positions and / or sizes, or incrementally adjusting and regenerating the layout of unlocked components;

[0062] S5: Bind the final layout scheme to the structured dataset, render and generate an interactive visual data dashboard, and output it as an independently deployable application or configuration file.

[0063] In this embodiment, the natural language requirement description refers to the user's expression of requirements for the visual data dashboard in the form of everyday language, such as "generate a dashboard to display the monthly sales data of an e-commerce platform, focusing on the sales of the top 10 products, with a tech blue style." This description includes the user's expectations for the dashboard's business objectives, core display content, and visual style.

[0064] The target data source refers to the original data source specified by the user for generating the data dashboard. Common types include MySQL database tables, API interface returned data, Excel spreadsheets, such as sales record tables and user behavior log tables from e-commerce platforms.

[0065] Large language models refer to artificial intelligence models with natural language understanding and parsing capabilities, such as GPT-4 and deepseek. These models can extract key information from the natural language input of users and transform it into structured instructions that the system can recognize.

[0066] Visual intent information refers to the set of core demand elements parsed from natural language requirements, including core business indicators (such as product sales and user conversion rates), desired chart types (such as bar charts, line charts, and heatmaps), and style preferences (such as tech blue, minimalist style, and retro style). It is the core basis for subsequent component matching and layout generation.

[0067] A structured dataset refers to a collection of data with a unified format obtained after cleaning and standardizing the target data source. Its structure is fixed as {dimensions: Array, metrics: Array}, where “dimensions” are dimension fields (such as product name, time) and “metrics” are indicator fields (such as sales volume, sales revenue), ensuring that the data can be directly called by visualization components.

[0068] A configurable component library is a resource library with a variety of pre-built visualization components. The components support dynamic configuration of data attributes (such as data type and data format) and style attributes (such as color, border, and font). It covers common visualization chart types such as bar charts, line charts, maps, and dashboards, and each component comes with detailed metadata descriptions.

[0069] Reinforcement learning algorithms are machine learning algorithms that optimize strategies based on reward signals through agent-environment interaction. In this method, it is used to optimize the layout engine, using layout aesthetics, information density balance, and user historical preferences as rewards to explore the optimal component layout scheme.

[0070] The layout engine is the core module responsible for generating the layout of visual components. It integrates reinforcement learning algorithms and can output an initial visual layout that meets the requirements based on candidate visual components and related constraints, and supports adjustments based on user feedback.

[0071] Interactive feedback commands refer to the commands that users make to adjust the layout after viewing the initial visual layout. These include component locking commands (specifying the position / size of a component) and attribute adjustment commands (such as modifying the color and size of a component). They can be triggered by checking checkboxes or entering natural language.

[0072] Independently deployable applications or configuration files refer to the output format of the final generated visual data dashboard. Applications can run directly on servers or terminal devices, while configuration files can be imported into specific visualization platforms for loading and use, ensuring that the dashboard can be flexibly deployed and reused.

[0073] In step S1, the user's natural language description of the requirement is first input into a large language model (such as GPT-4 or deepseek). The model uses natural language processing (NLP) technology to perform semantic analysis on the requirement text and extract the visualization intent information. For example, if the user inputs "Generate an e-commerce sales dashboard to display the sales revenue of each category in Q3 2025, presenting the trend with a line chart, and a simple style," the model will parse the core business indicator as "sales revenue of each category in Q3 2025," the expected chart type as "line chart," and the style preference as "simple."

[0074] When cleaning the target data source, data cleaning algorithms (such as regular expression denoising, missing value filling, and outlier removal) are used to process the target data source. For example, invalid characters in sales data are removed, missing date fields are filled, and sales data that exceeds a reasonable range (such as negative values ​​or abnormally large amounts of data) are deleted.

[0075] The cleaned data can then be standardized. Specifically, this is done through an ETL (Extract-Transform-Load) process, which transforms the cleaned data source into a structured dataset format of {dimensions: Array, metrics: Array}. For example, from a MySQL sales table, "category name" and "sales date" can be extracted as dimension fields, and "sales amount" and "sales volume" as metric fields. If the data source is JSON data returned by an API or an Excel spreadsheet, it can also be mapped to this standard structure using appropriate conversion tools. Simultaneously, an SQL parser is used to extract field types (e.g., "sales amount" as numeric, "sales date" as date), providing a data type basis for subsequent component matching.

[0076] In step S2, the field types of the structured dataset (e.g., categorical / date type for dimension fields and numerical type for indicator fields) and visualization intent information (chart type, style preference) are matched with the input data specifications (e.g., required field types and quantities) and functional style characteristics (e.g., chart type families, configurable styles) of each component in the configurable component library to filter out at least one set of candidate visualization components that meet the requirements. For example, if the dataset contains date-type dimension fields and numerical indicator fields, and the user expects a "line chart," then components such as line charts and time-series bar charts are included as candidates.

[0077] In step S3, when constructing the layout environment, the large-screen canvas is first divided into discrete grids. The position of the visual components is represented by grid coordinates, and the size is represented by the number of grid rows and columns occupied, thus determining the canvas attributes (aspect ratio, background color). Reinforcement learning parameters include a state space, action space, and reward function. The state space describes the position (grid coordinates), size (number of grid rows / columns), and canvas attributes of all current visual components. The action space includes moving visual components (to adjacent grids), adjusting their size (by a preset step size), and aligning components (left alignment, top alignment, etc.). The reward function is a weighted average of aesthetic rewards (based on alignment, symmetry, and the golden ratio), information density balance rewards (based on regional information entropy), and user preference rewards (based on historical layout data), evaluating the quality of the layout.

[0078] In step S4, the initial visual layout is rendered and displayed to the user, who can view the component positions, sizes, styles, and data display effects. Responding to interactive feedback includes: receiving user component locking instructions (such as checking a checkbox), marking the specified component as a locked component with a fixed position / size, and storing the parameters in a fixed parameter configuration file (such as fixedParams.json), so that subsequent adjustments will not change its position / size. Alternatively, receiving user adjustment instructions (such as the natural language "increase the width of the bar chart"), adjusting the properties and regenerating the layout of unlocked components, retaining the user-confirmed settings.

[0079] In step S5, the final layout scheme is bound to the structured dataset, establishing a mapping relationship between components and data fields (e.g., binding the X-axis of the line chart to "sales date" and the Y-axis to "sales amount"). An interactive dashboard is rendered using a front-end rendering framework (e.g., ECharts, D3.js), supporting click-to-filter, data drill-down, and other operations; the final output is a deployable application (e.g., exe, web application) or a configuration file (e.g., JSON).

[0080] This application utilizes AI to automatically analyze requirements and generate layouts, reducing generation time from hours to minutes, supporting real-time data updates, and significantly lowering development costs. Simultaneously, it can dynamically adjust based on user business needs and visual preferences, addressing the lack of flexibility in traditional methods. A multi-dimensional reward mechanism based on reinforcement learning ensures aesthetically pleasing layouts, balanced information distribution, and alignment with user browsing habits, thereby improving data insight efficiency.

[0081] In some embodiments, such as Figure 2 As shown, step S2 specifically includes:

[0082] S21: Obtain and parse the metadata description of each visualization component in the pre-built configurable component library, the metadata description including input data specifications and functional and style characteristics;

[0083] S22: The field types and number of fields in the structured dataset, the chart types and style preferences extracted from the visualization intent information, and the metadata descriptions of the parsed visualization components are matched and scored in multiple dimensions.

[0084] S23: Based on the comprehensive scoring results of multi-dimensional matching, sort the visualization components and select at least one group of visualization components with the highest ranking as the candidate visualization components.

[0085] In this embodiment, metadata description refers to the structured information that defines the core attributes and constraints of the visualization component. It is stored in the component library and includes input data specifications (data structure, type, and scale required by the component) and functional and style characteristics (chart type family, configurable styles, and interactive capabilities). It is the key basis for component matching.

[0086] Input data specifications refer to the data source requirements for the normal operation of visualization components, including the required field set (the types, number, and roles of fields that the component must bind to, such as a line chart requiring one time / category field as the X-axis and at least one numerical field as the Y-axis), the optional field set (additional fields that the component supports but are not mandatory, such as using a category field to distinguish the color of points in a scatter plot), and data size constraints (the maximum number of data records that the component can efficiently process, such as a table component recommending ≤1000 records).

[0087] Features include chart type families (such as “comparison”, “trend”, “distribution”, “composition”, and “relationship”), built-in interactive capabilities (click filtering, area zooming, hover tooltips, and legend toggle); style features refer to the component’s modifiable visual attributes (color theme, axis style, legend position, and animation effects).

[0088] Multi-dimensional matching refers to evaluating the fit between visualization components and requirements from two dimensions: data adaptability (the degree of matching between field type, quantity, scale and component input data specifications) and intent conformity (the semantic similarity between user-expected chart type and component type family, and the degree of matching between style preference and component configurable style).

[0089] Comprehensive scoring refers to a method of quantitatively scoring multi-dimensional matching results. Each dimension is assigned a weight (e.g., data adaptability 0.5, intent conformity 0.5), and a weighted total score is calculated based on the degree of matching (complete match 10 points, partial match 5 points, no match 0 points), which is used for component sorting and filtering.

[0090] In step S21, the metadata files (JSON / XML format) of each visualization component are called from the configurable component library. The metadata files clearly define the input data specifications and functional and style characteristics. Then, by parsing and extracting key information from the metadata, including the required field set (type, role, quantity) and optional field set of the input data specifications, data scale constraints, and the chart type family, configurable styles, and built-in interactive capabilities of the functional and style characteristics, the data is stored as structured data for subsequent matching and calling.

[0091] In step S22, data fitness matching and scoring include:

[0092] Field type matching is as follows: Compare the field types of the structured dataset with the required field types of the component. A complete match scores 10 points, a partial match (e.g., missing optional fields) scores 5 points, and no match scores 0 points. For example, a dataset containing "Sales Date" (time type) and "Sales Amount" (numerical type) completely matches the required field types of a line chart and scores 10 points.

[0093] Field quantity matching specifically includes: checking the extent to which the dataset meets the required number of fields for the component. Meeting and exceeding the minimum requirement earns 10 points, meeting only the minimum requirement earns 8 points, and not meeting the requirement earns 0 points. For example, a line chart requires at least one Y-axis numerical field; if the dataset contains two numerical fields, it earns 10 points.

[0094] Data size matching specifically includes: determining whether the number of records in the dataset is within the component size constraint. If within the range, 10 points are awarded; if it exceeds the limit but can be sampled, 3 points are awarded; and if it far exceeds the constraint, 0 points are awarded. For example, if the table component constraint is ≤1000 records and the dataset has 800 records, 10 points are awarded.

[0095] The data adaptability score is calculated by weighting the field type (0.4), number of fields (0.3), and data size (0.3). The formula is: Data adaptability score = Field type score × 0.4 + Number of fields score × 0.3 + Data size score × 0.3.

[0096] Intent relevance matching and scoring include:

[0097] Chart type family matching specifically includes: calculating the semantic similarity between the user's expected chart type and the component type family. A perfect match (such as "Trend Display" and "Trend Category" components) scores 10 points, partial similarity scores 3-7 points, and dissimilarity scores 0 points. For example, if the user expects to "compare sales of various categories," the semantic similarity with a "Comparison Category" bar chart is 100%, scoring 10 points.

[0098] Style preference matching specifically includes: checking whether the configurable styles of components support user preferences. Full support (such as the "Tech Blue" theme) earns 10 points, manual configuration is required, earning 5 points, and no support earns 0 points.

[0099] The intent conformity score is specifically calculated by weighting the chart type family (0.6) and style preference (0.4). The calculation formula is: intent conformity score = chart type family score × 0.6 + style preference score × 0.4.

[0100] When scoring the overall score, the data adaptability (0.5) and intent conformity (0.5) are weighted and calculated. The formula is: Overall score = Data adaptability score × 0.5 + Intent conformity score × 0.5.

[0101] In step S23, when filtering candidate components, all components are sorted in descending order according to their overall scores, with higher scores indicating higher priority. By setting filtering rules (such as an overall score ≥ 8 points or selecting the top 3), components that meet the criteria are selected as candidate visualization components and output to the layout engine. For example, the top 3 overall scores for line charts (9.5 points), bar charts (9 points), and radar charts (8.5 points) are all selected as candidate components.

[0102] The above solution clarifies component constraints through metadata parsing and combines multi-dimensional quantitative scoring to ensure that candidate components are compatible with the data source and meet user intent. Since candidate components undergo rigorous adaptability screening, subsequent layouts do not require frequent component replacements, reducing iteration counts and indirectly shortening the large-screen generation cycle.

[0103] In some embodiments, such as Figure 3 As shown, step S3 specifically includes:

[0104] S31: Define the state space: Map the candidate visualization components and their corresponding initial properties to a discretized grid layout environment; wherein, the state space is defined as the position, size and canvas properties of all visualization components in the grid at the current moment;

[0105] S32: Define action space: The action space is defined as a set of executable operations, specifically including: selecting any unlocked component and moving it to an adjacent empty space in the grid, adjusting its width or height by a preset step size, or aligning it with an adjacent component;

[0106] S33: Calculate the reward function: After each layout action is executed, a comprehensive reward value is calculated based on the updated layout state; the reward function is obtained by weighted calculation of any one or more of the following: aesthetic reward, information density balance reward, and user preference reward;

[0107] S34: Policy Learning and Layout Generation: Based on the current policy network, select actions, interact with the environment and obtain rewards, iteratively update network parameters through the policy gradient method, and finally output the layout scheme with the highest cumulative reward as the initial visual layout.

[0108] In this embodiment, the discretized grid layout environment refers to dividing the large screen canvas into regular grid units according to a preset size (such as 50px × 50px). The component position is represented by grid coordinates (x column number, y row number), and the size is represented by the number of grid rows and columns occupied (such as the component occupying (2,3)-(4,5) grid, coordinates (2,3), size 3 columns wide and 3 rows high), which facilitates the accurate calculation and adjustment of component position and size.

[0109] The state space is a set of parameters describing the current state of the layout, defined as S={C1,C2,...,Cn,CanvasAttr}, where Ci is the state of the i-th candidate component (position (xi, yi), size (widthi, heighti), lock flag locked_i), and CanvasAttr is the canvas attribute (width canvasWidth, height canvasHeight, backgroundColor), which is the basis for determining the layout state.

[0110] Action space refers to the set of layout adjustment operations that an agent can perform, defined as A={Move(c,dx,dy),Resize(c,dw,dh), Align(c,targetC,alignType)};

[0111] In this function, Move(c,dx,dy) moves the unlocked component c to an adjacent grid, where dx / dy is the number of grids moved (positive or negative indicates direction), and must satisfy the conditions of not exceeding the boundaries and not overlapping. Resize(c,dw,dh) adjusts the width / height of the unlocked component c by a preset step size (e.g., 1 grid), where dw / dh is the adjustment amount (positive or negative indicates increase or decrease), and must satisfy the conditions of size ≥ minimum threshold (e.g., 2 columns × 2 rows) and not exceeding the boundaries. Align(c,targetC,alignType) aligns the unlocked component c with the target component targetC (left alignment, top alignment, etc.), and must satisfy the condition of not exceeding the boundaries after alignment.

[0112] The reward function is a function that quantifies the quality of layout actions. It consists of a weighted average of aesthetic reward Ra, information density balance reward Rd, and user preference reward Rp. The formula is R = ωa×Ra + ωd×Rd + ωp×Rp (ωa+ωd+ωp=1, default ωa=0.3, ωd=0.4, ωp=0.3).

[0113] A policy network is a neural network that outputs a policy for choosing actions. Its input is the state space S, and its output is the probability distribution of choices in the action space A. By learning, it optimizes action choices and increases the probability of obtaining high rewards.

[0114] The policy gradient method is an optimization algorithm for updating policy network parameters. It calculates the policy gradient based on the action selection probability logP(a|S) and the cumulative reward G (G=R+γ×G_next, where γ is a discount factor of 0.9), and adjusts the parameters along the gradient ascent direction to increase the selection probability of high-reward actions.

[0115] In step S31, during component state initialization, an initial state is set for each candidate component, with positions assigned sequentially from the top left corner of the canvas (to avoid overlap), and sizes set according to the component type default values ​​(e.g., columnar). Figure 4 (Columns x 3 rows, map 6 columns x 5 rows), the initial lock flag is "false" (unlocked). For example, the initial states of 3 candidate components C1 (bar chart), C2 (line chart), and C3 (map) are: C1={(0,0),(4,3),false}, C2={(4,0),(4,3),false}, C3={(0,3),(8,5),false}.

[0116] The canvas properties are set according to user needs or default configuration, such as a width of 1920px (38 columns), a height of 1080px (21 rows), and a background color of #001529.

[0117] The state space construction specifically includes: integrating component states and canvas properties to form an initial state space S0={C1,C2,C3,CanvasAttr}, which is stored in a structured format.

[0118] In step S32, the specific action rules are as follows:

[0119] Movement rules: dx / dy∈{-1,0,1} (moves at most 1 grid at a time), and must satisfy x≥0, x+width≤number of canvas columns, y≥0, y+height≤number of canvas rows after the component moves, and not overlap with other components.

[0120] Size adjustment rules: dw / dh∈{-1,0,1}, the component must be adjusted to have a width ≥ 2 columns and a height ≥ 2 rows, and must not exceed the limits.

[0121] Alignment rules: Supports left alignment, top alignment, and center alignment. The alignment must ensure that the component does not exceed the bounds and that the target visual component exists.

[0122] Action validity verification rules: Set verification conditions to include only valid actions in the executable set and avoid invalid operations (such as moving to overlapping grids).

[0123] Action space construction rules: Integrate action types and rules to form a complete action space A.

[0124] In step S33, the aesthetic reward Ra is calculated as follows:

[0125] Alignment calculation: Count the number of aligned pairs of adjacent components (horizontal / vertical). Alignment score = number of aligned pairs / total number of adjacent pairs × 10 (maximum score 10 points).

[0126] Symmetry calculation: Using the vertical center line of the canvas as the axis, count the pairs of symmetrical components. Symmetry score = total score of symmetrical pairs / (total number of components × 2 × 10 (maximum score 10 points)).

[0127] Golden Ratio Calculation: Count the number of components whose dimensions conform to the golden ratio (1:1.618 or 1.618:1). Golden Ratio Score = Number of components conforming to the golden ratio / Total number of components × 10 (maximum score 10 points).

[0128] Then, the Ra values ​​are calculated by weighting the alignment score (0.4), symmetry score (0.3), and golden ratio score (0.3). Ra = alignment score × 0.4 + symmetry score × 0.3 + golden ratio score × 0.3.

[0129] The calculation process for the information density balance reward Rd is as follows:

[0130] First, divide the canvas into regions, specifically into several sub-regions (e.g., 4 columns × 3 rows).

[0131] Then, information entropy is calculated, specifically by using the information entropy formula H=-Σ(pi×log2pi) to calculate the information entropy of each region (pi is the proportion of components / fields). The higher the entropy, the richer the information.

[0132] Then, a balance assessment is performed, specifically by calculating the standard deviation of regional information entropy σ, Rd = (1 - σ / MaxH) × 10 (MaxH is the maximum information entropy). If σ is too large, a negative reward is applied (Rd = -(σ / MaxH - 0.5) × 10), ranging from -5 to 10 points.

[0133] The calculation process for the user preference reward Rp is as follows:

[0134] First, a preference model is constructed, specifically based on users' historical layout data (such as core components often being placed in the top left corner, with a size of 4 columns × 3 rows).

[0135] Then, the matching degree is calculated. The core component matches the preference and gets 3 points, the secondary component matches and gets 1 point, and the non-matching component gets 0 points.

[0136] The formula for calculating the user preference reward Rp is as follows: Rp = total matching score / (number of core components × 3 + number of secondary components × 1) × 10 (maximum score 10 points). If there is no historical data, the default Rp = 5 points.

[0137] The formula for calculating the comprehensive reward R is as follows: R = 0.3 × Ra + 0.4 × Rd + 0.3 × Rp, calculated by weight.

[0138] In step S34, the policy network initialization process is as follows: construct a 3-layer fully connected neural network (input layer = state dimension, hidden layer = 64, output layer = action dimension), and randomly initialize the parameters.

[0139] The interactive iterative process of the policy network is as follows:

[0140] First, there is the state input, specifically the current state St is input into the policy network, and the output is the action probability distribution P(a|St).

[0141] Then, action selection is performed, specifically using an ε-greedy strategy (ε=0.1, decreasing with iteration) to select action at, with ε probability for random exploration and 1-ε probability for selecting the optimal action.

[0142] Then, the action execution steps are carried out, specifically including: executing at; if valid, updating the state to S{t+1}; if invalid, reselecting (reward Rt=0).

[0143] Then, the reward is obtained, specifically by calculating Rt according to step S33.

[0144] Then, gradient calculation and parameter update are performed. Specifically, the policy gradient ∇θlogP(at|St)×Gt is calculated based on logP(at|St) and cumulative reward Gt, and the parameter θ=θ+α×∇θlogP(at|St)×Gt (α=0.001) is updated along the gradient ascent direction.

[0145] Then, by setting a termination condition (such as 1000 iterations or 200 consecutive reward fluctuations <0.1), the layout with the highest cumulative reward is output as the initial visual layout after the condition is met.

[0146] The above solution automates the entire layout process through reinforcement learning, eliminating the need for manual intervention and addressing the issues of low efficiency and reliance on experience inherent in traditional manual layout methods. Simultaneously, by setting multi-dimensional reward functions, it ensures aesthetically pleasing layouts (alignment, symmetry), balanced information (no overload / blank space), and alignment with user habits, thereby improving user satisfaction and browsing experience.

[0147] In some embodiments, during step S3, the layout engine's optimization objective further includes an information transmission efficiency objective, such as... Figure 4 As shown, the information transmission efficiency target is constructed through the following steps:

[0148] S35: Calculate an information value weight for each candidate visualization component, wherein the information value weight is the weighted sum of the information entropy of a specific data subset extracted from the structured dataset corresponding to the candidate visualization component and the inherent visual saliency of its chart type;

[0149] S36: In the layout iteration, evaluate the degree to which the current layout scheme highlights high-information-value components, and quantify the degree of highlighting into the value of the information transmission efficiency target, so as to guide the layout engine to place key information in prominent positions.

[0150] In this embodiment, the goal of information delivery efficiency aims to improve the efficiency of users in obtaining key data insights by placing high-value information components in prominent positions on the canvas (areas where users' eyes are primarily focused, such as the upper half or the left side), and to optimize the layout in conjunction with aesthetic goals and information density balance goals.

[0151] Information value weight is an indicator that quantifies the importance of component information. It is calculated by weighting the information entropy of the corresponding data subset of the component (reflecting the richness of the data) and the inherent visual salience of the component's chart type (reflecting its ability to attract attention). The higher the weight, the more the component needs to be highlighted.

[0152] A data subset refers to a portion of data extracted from a structured dataset that is bound to a specific component. For example, the data subset corresponding to the "Top 10 Product Sales" bar chart is the filtered data of "Product Name" (dimension) and "Sales" (metric) (only the top 10 products).

[0153] Information entropy is an indicator that measures the uncertainty and richness of a subset of data. The higher the entropy value, the more obvious the data differentiation (such as the large difference in sales of the top 10 products), and the higher the decision value. The calculation formula is H=-Σ(pi×log2pi) (pi is the category proportion).

[0154] Inherent visual salience refers to the ability of a chart type to attract attention, which is determined by visual area potential (size), color contrast (difference from background), dynamic interactivity (animation / real-time updates), and visual guidance (eye-guiding effect). For example, maps and dashboards have higher salience than tables.

[0155] Prominence refers to the ease with which users notice high-value information components, and is quantified by a combination of the component's information value weight, its location weight (higher weight in prominent areas), and its size weight (higher weight for larger sizes).

[0156] Prominent positions refer to the areas on the canvas that the user's gaze focuses on first. They are divided based on eye tracking and user habits, such as primary prominent areas (upper left 1 / 3), secondary prominent areas (upper right 1 / 3 and lower left 1 / 3), etc. The higher the prominence, the greater the position weight.

[0157] In step S35, when calculating the information value weight, a data subset extraction is first performed. Specifically, based on the input data specifications (field role requirements) of the component metadata, a component-specific data subset is extracted from the structured dataset to ensure compliance with the component's data scale constraints. For example, for the "sales revenue share of each region" pie chart component, "region name (dimension)" and "sales revenue" (metric) are extracted as data subsets, and the number of records does not exceed the maximum number supported by the pie chart.

[0158] Information entropy calculation refers to calculating the information entropy H of an indicator field in a subset of data. Assuming the data subset has k categories (e.g., k regions), and pi is the proportion of the indicator for category i, the calculation formula is H = -Σ(i=1 to k)pi × log₂pi. For example, if the sales proportions of three regions are p₁=0.6, p₂=0.3, and p₃=0.1, the corresponding information entropy H is calculated as follows:

[0159] H=-(0.6×log20.6+0.3×log20.3+0.1×log20.1)≈1.02 bit.

[0160] The inherent visual salience score refers to a scoring standard established based on the characteristics of the chart type (typically, the maximum score can be set at 10 points):

[0161] Visual area potential refers to maps, dashboards (score range 8-10 points), and tables (score range 3-5 points).

[0162] Color contrast refers to heatmaps (score range 8-9) and ordinary line graphs (score range 5-7).

[0163] Dynamic interactivity refers to a dynamic dashboard (score range of 9-10 points) and a static bar chart (score range of 5-6 points).

[0164] Visual guidance refers to funnel charts (score range 7-8 points) and tables (score range 3-4 points).

[0165] For example, a map score of 10, a dynamic dashboard score of 9, and a broken line... Figure 6 point.

[0166] The information value weight calculation process is as follows: Set the information entropy weight ωh = 0.5 and the visual salience weight ωs = 0.5. The formula is W = ωh × (H / Hmax) + ωs × (S / Smax), where Hmax = log₂k (theoretical maximum entropy), S is the component score, and Smax = 10. For example, for a map component, H = 1.02, Hmax = log₂3 ≈ 1.58, S = 10, W = 0.5 × (1.02 / 1.58) + 0.5 × (10 / 10) ≈ 0.82.

[0167] In step S36, when evaluating the information transmission efficiency target and guiding the layout, the prominent area is first divided and weighted. Specifically, the canvas is divided into prominent areas of different levels, and a position weight P_pos is assigned (the higher the prominence, the greater the weight).

[0168] First-level prominent region: the upper left 1 / 3, P_pos1=1.0.

[0169] Secondary prominent region: upper right 1 / 3 and lower left 1 / 3, P_pos2=0.7.

[0170] Level 3 prominent region: the middle 1 / 3 of the upper half and the right 1 / 3 of the lower half, P_pos3=0.4.

[0171] Non-obvious region: the middle 1 / 3 of the lower half, P_pos4=0.1.

[0172] For example, a 38-column × 21-row canvas with first-level region coordinates x∈[0,12] and y∈[0,10].

[0173] The component prominence is then calculated using the following steps: Combining the component's information value weight W, location weight P_pos, and size weight P_size (P_size = component width × height / total number of columns × total number of rows in the canvas), the calculation formula is S_c = W × (0.6 × P_pos + 0.4 × P_size). For example, a map with W = 0.82 located in a first-level area (P_pos = 1.0), with a size of 6 × 5 and a canvas size of 38 × 21, has P_size = 30 / 798 ≈ 0.0376, and S_c = 0.82 × (0.6 × 1.0 + 0.4 × 0.0376) ≈ 0.504.

[0174] Then, the target value of information transmission efficiency is calculated. The calculation process is as follows: the target value E is the weighted sum of the component prominence (the weight is the component information value weight W). The calculation formula is E=Σ(W_i×S_ci) / ΣW_i (n is the number of components). The higher E is, the higher the information transmission efficiency.

[0175] Then, layout guidance is implemented, specifically by incorporating E into the reward function, adjusting it to R=ωa×Ra+ωd×Rd+ωp×Rp+ωe×E (ωe=0.2, other weights are adjusted to a total of 1). When E is low (e.g., E<0.3), the agent prioritizes actions that increase E (e.g., moving high-W components to prominent areas, increasing their size), guiding the layout to focus on key information.

[0176] The above solution allows for the priority placement of high-value information components in prominent areas, allocating resources (prominent locations, large sizes) based on information value, thereby enhancing the decision-making support capabilities of large screens. In complex large screens with multiple components and dimensions (such as enterprise operation screens), it effectively avoids information overload and chaos, helping users quickly establish a "core-secondary-auxiliary" information cognition framework, suitable for scenarios such as executive decision-making and real-time monitoring.

[0177] In some embodiments, the interactive feedback instructions include component locking instructions and natural language adjustment instructions;

[0178] Marking a user-specified component as a locked component with a fixed position and / or size includes: receiving a component locking instruction, marking the user-specified component as a locked component with a fixed position and / or size, and storing the position and size parameters of the locked component in a fixed parameter configuration file;

[0179] Incremental adjustment and regeneration of the layout of unlocked components includes: receiving a user's natural language adjustment instruction, parsing the semantics of the natural language adjustment instruction through a natural language processing model, converting the parsed semantics into attribute adjustment operations for the corresponding unlocked components, and regenerating the adjusted layout of the unlocked components.

[0180] In this embodiment, incremental layout regeneration refers to performing local layout optimization only on the non-locked components that have undergone attribute adjustment operations. If other non-locked components do not overlap with or visually conflict with the adjusted components, their original parameters remain unchanged. If there is a conflict (such as the adjusted component occupying the grid area of ​​other components), only the non-locked components in the conflict area are slightly adjusted (such as moving 1-2 columns to the right) to avoid significant changes in the overall layout and ensure that the user's existing preference settings (such as the position and style of non-conflicting components) are preserved.

[0181] The above solution, through component locking commands, ensures that core components (such as KPI dashboards) are not accidentally adjusted, meeting the stability requirements for displaying key information in business scenarios. Simultaneously, natural language adjustment commands lower the operational threshold, allowing for personalized layouts without requiring specialized technical knowledge, thus enhancing the user experience. The incremental adjustment and regeneration mechanism only modifies the target and conflicting components, reducing repetitive operations.

[0182] In some embodiments, in step S1, the natural language requirement is parsed through the large language model, and a data narrative logic chain is generated. The data narrative logic chain is a directed graph structure. The nodes of the directed graph structure are key data insights extracted from the natural language requirement and structured data. The edges of the directed graph structure define the logical relationship types between the insights. The logical relationship types include causal relationship, comparative relationship, general-to-specific relationship, or temporal evolution relationship.

[0183] In step S3, the reward signal from the layout engine also includes a narrative coherence reward, used to incentivize the initial visual layout to form a visual narrative path in spatial arrangement that conforms to the data narrative logic chain, such as... Figure 5 As shown, the narrative coherence reward is calculated in the following way:

[0184] S37: Map the data narrative logic chain to a desired visual focus sequence, which records the ideal eye movement order of the user when browsing the large screen;

[0185] S38: Based on the current layout state, calculate an actual visual flow path by simulating an eye-tracking model;

[0186] S39: Calculate the degree of consistency between the visual flow path and the visual focus sequence, and quantify this degree of consistency as a narrative coherence reward value;

[0187] The higher the degree of consistency, the greater the narrative coherence reward value.

[0188] In this embodiment, the data narrative logic chain refers to a directed graph structure composed of key data insights extracted from user natural language needs and structured data, and the logical relationships between these insights. Nodes are key data insights (such as "sales revenue increased by 15% year-on-year in Q3 2025", "East China region accounted for the highest sales share", and "electronic products are the main driver of growth"), while edges represent logical relationship types (causal relationship, comparative relationship, general-to-specific relationship, and temporal evolution relationship), used to guide the layout to form a visual narrative path that conforms to human cognitive habits.

[0189] Key data insights refer to information points that are of core value to business decisions, extracted from user needs (such as "focusing on the reasons for sales growth") and structured data (such as sales data broken down by region and category). They must simultaneously meet the two conditions of "strongly correlated with user needs" and "data-supported" and are the core components of the data narrative logic chain.

[0190] A causal relationship refers to a situation where the preceding insight is the cause and the following insight is the result. For example, "increased sales volume in the electronics category" is the cause of "increased sales revenue in East China," indicating a causal relationship between these two insights. A comparative relationship refers to a data comparison between two or more insights (year-on-year / month-on-month / cross-dimensional comparison). A total-to-specific relationship refers to a situation where the preceding insight represents the overall data, and the following insight represents the data after breaking down the overall data. A temporal evolution relationship refers to insights arranged in chronological order, reflecting the trend of data changes over time.

[0191] A visual focus sequence refers to the ideal eye movement order when users browse a large screen, generated based on the data narrative logic chain mapping. Each element in the sequence corresponds to a visualization component (i.e., the component that carries the insight data) of a key data insight point. The order is consistent with the logical relationship flow of the insight points in the logic chain. For example, if the logic chain is "total sales revenue → East China sales revenue → East China electronic product sales revenue" (total-to-score relationship), then the visual focus sequence is "total sales revenue dashboard → East China bar chart → East China electronic product line chart".

[0192] Eye-tracking models are computational models that simulate the eye movement patterns of humans when browsing large visual screens. Based on visual attention mechanisms (such as the color contrast, size, position, and dynamic effects of components attracting attention), the model takes the current layout state (the position, size, style, and data content of all components) as input and outputs the actual eye movement path of the user (i.e., the browsing order of components). The path is represented by component IDs arranged in the order of browsing.

[0193] Narrative coherence reward refers to the reward signal used to evaluate whether the initial visualization layout conforms to the data narrative logic chain in terms of spatial arrangement. It is quantified by calculating the degree of consistency between the actual visual flow path and the visual focus sequence. The higher the consistency, the greater the reward value. It is an important supplementary dimension to the reinforcement learning reward function.

[0194] In practical applications, the system reads the directed graph structure of the data narrative logic chain, extracts the "relatedComponentId" corresponding to each node according to the direction of the directed edges (from the source node to the target node), and arranges them in order to form a visual focus sequence. For example, based on the above directed graph, the visual focus sequence is: ["dashboard_total_sales", "bar_chart_east_sales", "line_chart_east_electronics"], indicating that the ideal browsing order for users is "total sales dashboard → East China region bar chart → East China electronic products line chart".

[0195] Then, the eye-tracking model is invoked, taking the current layout state as input (position coordinates, size, color contrast, and dynamic effect markers of all components). Based on visual attention weights (components with high color contrast, large size, prominent location, and dynamic effects have higher attention weights), the model simulates the user's eye movement process during browsing and outputs the actual visual flow path (component ID sequence). For example, in the current layout, if the total sales dashboard is located in the upper left corner (prominent area) with high color contrast, the East China region bar chart is located on the right side of the dashboard, and the East China electronics product line chart is located below the bar chart, then the actual visual flow path might be: ["dashboard_total_sales", "bar_chart_east_sales", "line_chart_east_electronics"], which is completely consistent with the visual focus sequence.

[0196] Then, sequence similarity algorithms (such as edit distance algorithm and longest common subsequence algorithm) are used to calculate the consistency between the actual visual flow path and the visual focus sequence. The specific steps are as follows: Let the visual focus sequence be S (length L), and the actual visual flow path be P (length M, M≥L, because the path may include browsing of non-focus components). Extract the subsequence P' that overlaps with elements of S in P, calculate the longest common subsequence length LCS of P' and S, and the consistency degree = LCS / L (value range 0-1). The formula for calculating the narrative coherence reward value R_n is as follows:

[0197] R_n = Consistency level × 10 (the maximum score is 10 points, consistent with the scoring scale of other reward sub-items).

[0198] For example, if S = ["A", "B", "C"] and P = ["A", "D", "B", "E", "C"], then P' = ["A", "B", "C"], LCS = 3, consistency level = 3 / 3 = 1, R_n = 10 points; if P = ["B", "A", "C"], then P' = ["B", "A", "C"], LCS = 2 (such as "A", "C" or "B", "C"), consistency level = 2 / 3 ≈ 0.67, R_n ≈ 6.7 points.

[0199] Reward Function Integration and Application: The narrative coherence reward R_n is integrated into the reward function of step S3. The adjusted reward function formula is: R = ωa × Ra + ωd × Rd + ωp × Rp + ωn × R_n, where ωn is the weight of the narrative coherence reward (default value is 0.2, and other weights are adjusted proportionally to ensure ωa + ωd + ωp + ωn = 1, such as ωa = 0.25, ωd = 0.35, ωp = 0.2). During the reinforcement learning iteration process, the system will prioritize layout actions that can improve R_n (such as adjusting the position of the focus component to make the browsing order closer to the visual focus sequence), and finally generate an initial visualization layout that conforms to the data narrative logic.

[0200] The above solution connects fragmented visualization components into a "data story" that aligns with business logic through a data narrative logic chain and visual focus sequence. This avoids the problems of isolated components and difficulty in understanding data relationships found in traditional layouts, helping users quickly establish a data cognitive framework. The narrative coherence and reward-guided layout create a visual path that conforms to human eye movement habits, reducing eye jumps and ineffective searches during browsing and improving the efficiency of acquiring core data insights. This is particularly suitable for large screens in complex business scenarios (such as multi-dimensional sales analysis and cross-departmental operational monitoring). A logically clear visual narrative path helps users quickly grasp the causal, comparative, and total-to-specific relationships in the data, enabling a more accurate understanding of the business issues behind the data (such as the core driving regions and categories for sales growth), providing stronger evidence for business decisions.

[0201] In some embodiments, the interactive feedback instruction is a multimodal interactive instruction, and the method further includes:

[0202] The multimodal interaction instructions are parsed using a multimodal fusion model and converted into specific operation parameters for the visualization components; wherein, the multimodal interaction instructions include any one or more of the following instructions:

[0203] Natural language adjustment instructions are used to describe changes in the properties of visual components;

[0204] Sketch instructions are used to specify the desired size or positional relationship of visual components through graphical outlines;

[0205] Gesture control commands are used to directly manipulate visual components via air gestures.

[0206] By employing the above methods, the limitations of traditional single-text interaction can be overcome. Users can choose the optimal interaction method based on the usage scenario (such as using gestures / sketches when it is inconvenient to input text in a meeting), which is especially suitable for non-technical personnel (such as business executives and operations staff), reducing the operational threshold and improving interaction efficiency. Multimodal commands can complement each other's strengths and weaknesses (such as text describing component types, sketches precisely defining dimensions, and gestures adjusting positions in real time), avoiding the ambiguity of single-modal commands (such as the text "a little bigger" being precisely quantified through the sketch's "2:1 aspect ratio"), and improving the accuracy of command parsing.

[0207] In some embodiments, the method further includes:

[0208] Real-time monitoring of the distribution changes of the data stream bound to the final layout scheme;

[0209] When a change in data distribution is detected to exceed a preset threshold, an incremental layout adjustment process is automatically triggered. Based on the changed data characteristics, the layout of the current visualization components is dynamically optimized to focus on new data insights.

[0210] In this embodiment, after the final layout scheme is generated, the system establishes a real-time binding relationship between "components and data fields". It obtains the data stream (incremental update data or full update data of structured dataset) bound to the component in real time through data interfaces (such as JDBC connection to the database or API interface polling) and updates the data content displayed by the component synchronously.

[0211] Then, the system calculates the data distribution characteristics. It automatically calculates the key distribution characteristics of the current data stream periodically (e.g., every 5 minutes) or when a data update is detected. These characteristics are compared with the baseline characteristics calculated previously (characteristics from the layout generation or previous adjustments). For numerical indicator fields (e.g., sales revenue, sales volume), the system calculates the mean, variance, maximum value, minimum value, and year-on-year / month-on-month growth rate. For category dimension fields (e.g., region, category), the system calculates the percentage, ranking, and number of newly added / deleted categories for each category. For time-series data fields (e.g., time-sales revenue), the system calculates the trend slope (growth / decline rate), periodic fluctuation amplitude, and number of inflection points. For example, if the baseline characteristic is "East China region's sales revenue share is 35% (ranked 1st)," and the current characteristic is "South China region's sales revenue share is 38% (ranked 1st), East China region's share is 28% (ranked 2nd)," then the category percentages and rankings will change.

[0212] When determining the magnitude of change, the difference (magnitude of change) between the current feature and the baseline feature is compared with a preset threshold: If the difference is less than or equal to the preset threshold: the data distribution change is considered minor, no layout adjustment is needed, and only the data content displayed by the component is updated. If the difference is greater than the preset threshold: the data distribution is considered to have changed significantly, triggering an incremental layout adjustment process, and the current feature is updated to the new baseline feature. For example, if the preset threshold for category proportion is ±5%, and the proportion of the East China region decreases from 35% to 28%, the difference of 7% > 5%, triggering the adjustment process.

[0213] When implementing incremental layout adjustments, the first step is to assess the impact of data changes and identify affected components. Specifically, this involves analyzing the fields corresponding to changes in data distribution (e.g., changes in the regional dimension's proportion corresponding to the "Region" field), querying the "component-data field" binding relationship, and filtering out all visualization components associated with that field (i.e., affected components). For example, changes in the Region field affect components such as the "Regional Sales Bar Chart," "Regional Heat Map," and "Regional Ranking Table."

[0214] Then, the impact is further assessed. If the component only displays data content (such as a table that only updates percentage values) and there is no need to adjust attributes, it is marked as a "data update only component". If the component needs to adjust attributes due to changes in data characteristics (such as the bar chart corresponding to the first-ranked region needs to be highlighted, or the component size is insufficient due to the addition of a new category), it is marked as a "component that needs attribute adjustment".

[0215] Then, for components requiring attribute adjustments, an adjustment strategy is formulated based on the new data characteristics:

[0216] If the data insight focus changes (e.g., South China becomes the top-selling region): adjust the component that carries the focus (e.g., the highlighted bar chart for South China) to a more prominent position (e.g., move it from the right to the left prominent area), increase its size (e.g., increase the width by 1-2 columns), and enhance its style (e.g., increase color contrast, add dynamic flashing effects) to highlight the new focus.

[0217] If the amount of data changes (e.g., adding 5 new categories, causing the table component to not display completely), adjust the component size (e.g., increase the table height, or change from a single-column layout to a double-column layout) to ensure that the data is displayed completely.

[0218] If the time-series trend changes (e.g., sales change from growth to decline), adjust the component type (e.g., change from a line chart to an area chart with warning colors) or style (e.g., change the color of the growth segment to green and the decline segment to red) to strengthen the visual indication of the trend change.

[0219] Then, for the affected components after the attribute adjustments, check whether their new parameters (position, size) overlap or visually obstruct other components (especially locked components). If there is no conflict, apply the adjusted attributes directly to complete the local layout update. If there is a conflict, only make minor adjustments to the unlocked components in the conflict area (such as moving them 1-2 grid columns to the right or down, or slightly reducing their size) to avoid the conflict, while ensuring that the parameters of unaffected components are not modified.

[0220] After completing attribute adjustments and conflict resolution, the system renders the updated layout in real time and pushes a notification to the user that "the layout has been optimized due to data changes" (such as a pop-up window in the upper right corner of the interface). The user can view the adjusted layout and manually fine-tune it if there are any objections through multimodal interaction commands.

[0221] This solution monitors data distribution changes in real time and automatically adjusts the layout, ensuring that the large screen displays not only the latest data but also the most valuable data insights. This avoids the information misleading effects of traditional large screens where "data is updated but the layout lags behind" (e.g., core indicators have changed but are still displayed in non-prominent positions). It is particularly suitable for real-time monitoring scenarios (such as real-time sales during e-commerce promotions and real-time operational monitoring in government). Simultaneously, this solution can automatically focus on new data insights, helping users quickly capture the business dynamics behind data changes (such as market share shifts and sudden growth points), improving the large screen's responsiveness to business changes and providing more timely support for real-time decision-making.

[0222] In a second aspect, the present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the model-driven visualization data dashboard generation method as described in the first aspect of the present invention.

[0223] The computer-readable storage medium may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory.

[0224] The non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), a magnetic random access memory (FRAM), a flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD ROM); the magnetic surface memory may be a disk storage device or a magnetic tape storage device.

[0225] The volatile memory may be random access memory (RAM), which serves as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), synchronous static random access memory (SSRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synclink dynamic random access memory (SLDRAM), and direct memory bus random access memory (DRRAM). The computer-readable storage media described in the embodiments of the present invention are intended to include these and any other suitable types of memory.

[0226] like Figure 6 As shown, in a third aspect, the present invention provides an electronic device 10, including a processor 101 and a storage medium 102, wherein a computer program is stored on the storage medium, and when the computer program is executed by the processor, it implements the model-driven visualization data screen generation method as described in the first aspect of the present invention.

[0227] In some embodiments, the processor may be implemented by software, hardware, firmware, or a combination thereof, and may be a circuit, one or more of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, or a microprocessor, thereby enabling the processor to execute some or all of the steps or any combination of the steps in the model-driven visualization data dashboard generation method described in the various embodiments of this application.

[0228] Finally, it should be noted that although the above embodiments have been described in the text and drawings of this application, this should not limit the scope of patent protection of this application. Any technical solutions that are based on the essential concept of this application and utilize the content described in the text and drawings of this application, resulting in equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this application.

Claims

1. A method for generating a model-driven visual data large screen, characterized in that, The method comprises: S1: receiving an input natural language requirement description and a target data source, analyzing the natural language requirement by a large language model, extracting visualization intent information, and cleaning and standardizing the target data source, to output a structured data set with unified dimension fields and index fields, the visualization intent information including core business indexes, expected chart types, and style preferences; S2: based on the field types of the structured data set and the visualization intent information, matching the input data specifications and functional style characteristics of each visualization component in the pre-built configurable component library, to determine at least one group of candidate visualization components; S3: taking the determined candidate visualization components as layout elements, using a layout engine integrated with a reinforcement learning algorithm to generate an initial visualization layout; wherein the reinforcement learning algorithm is configured to use layout aesthetics, information density balance, and user historical preferences as reward signals, and searches for an optimal layout scheme by iteratively exploring different position and size combinations of the visualization components; S4: rendering and presenting the initial visualization layout, and adjusting the initial visualization layout in response to user interaction feedback instructions to obtain a final layout scheme; the adjustment operation includes marking user-specified components as locked components with fixed positions and / or sizes, or incrementally adjusting and regenerating the layout of non-locked components; S5: binding the final layout scheme with the structured data set, rendering an interactive visualization data wall, and outputting an independently deployable application or configuration file.

2. The method of claim 1, wherein, Step S2 specifically comprises: S21: obtaining and analyzing the metadata description of each visualization component in the pre-built configurable component library, the metadata description including input data specifications and functional and style characteristics; S22: multi-dimensionally matching the field types and field quantities of the structured data set with the chart types and style preferences extracted from the visualization intent information, and the metadata description of the parsed visualization components, and comprehensively scoring; S23: based on the comprehensive scoring results of multi-dimensional matching, sorting the visualization components, and selecting at least one group of visualization components ranked at the top as the candidate visualization components.

3. The method of claim 1, wherein, Step S3 specifically comprises: S31: defining a state space: mapping the candidate visualization components and their corresponding initial attributes into a discretized grid layout environment; wherein the state space is defined as the positions, sizes, and overall layout canvas attributes of all visualization components in the grid at the current time; S32: defining an action space: the action space is defined as a set of executable operations, specifically including: selecting any non-locked component and moving it to an adjacent empty position in the grid, adjusting its width or height by a preset step, or aligning it with adjacent components; S33: calculating a reward function: after each layout action is executed, a comprehensive reward value is calculated according to the updated layout state; the reward function is calculated by weighting any one or more of aesthetic rewards, information density balance rewards, and user preference rewards; S34: Strategy learning and layout generation: based on the current policy network, select actions, interact with the environment and obtain rewards, and update network parameters iteratively through policy gradient method, and finally output the layout scheme with the highest cumulative reward as the initial visualization layout.

4. The method of claim 3, wherein, In step S3, when the layout engine performs layout optimization, the optimization target further includes an information transmission efficiency target, The information transmission efficiency target is constructed by the following steps: S35: Calculate an information value weight for each candidate visualization component, which is the weighted sum of the information entropy of the specific data subset corresponding to the candidate visualization component extracted from the structured data set and the inherent visual saliency of its chart type; S36: In the layout iteration, evaluate the prominence of high information value components in the current layout scheme, and quantify the prominence as the value of the information transmission efficiency target to guide the layout engine to place key information in prominent positions.

5. The method of claim 1, wherein, The interactive feedback instruction includes a component locking instruction and a natural language adjustment instruction; Marking the user-specified component as a locked component with fixed position and / or size includes receiving the component locking instruction, marking the user-specified component as a locked component with fixed position and / or size, and storing the position and size parameters of the locked component in a fixed parameter configuration file; Incremental adjustment and regeneration of the layout of non-locked components includes receiving the user's natural language adjustment instruction, parsing the semantics of the natural language adjustment instruction through a natural language processing model, converting the parsed semantics into attribute adjustment operations for the corresponding non-locked component, and regenerating the layout of the adjusted non-locked component.

6. The method of claim 1, wherein the model-driven visualization data wall generation is based on a model-driven architecture (MDA) approach. In step S1, the natural language requirement is parsed by the large language model, and a data narrative logic chain is also generated, The data narrative logic chain is a directed graph structure, the nodes of the directed graph structure are key data insight points extracted from the natural language requirement and structured data, and the edges of the directed graph structure define the logical relationship types between the insight points, including causal relationship, contrast relationship, total and partial relationship, or time sequence evolution relationship; In step S3, the reward signal of the layout engine also includes a narrative coherence reward for encouraging the initial visualization layout to form a visual storytelling path that conforms to the data narrative logic chain in spatial arrangement, and the narrative coherence reward is calculated by the following way: S37: Map the data narrative logic chain to an expected visual focus sequence, which records the ideal visual line movement sequence of the user when browsing the large screen; S38: Based on the current layout state, calculate an actual visual flow path through a visual line tracking model; S39: Calculate the consistency degree between the visual flow path and the visual focus sequence, and quantify this consistency degree as the narrative coherence reward value; The higher the consistency degree, the greater the narrative coherence reward value.

7. The method of claim 1, wherein, The interactive feedback instruction is a multi-modal interactive instruction, and the method further includes: The multi-modal interaction instruction is analyzed by a multi-modal fusion model, and the multi-modal interaction instruction is converted into specific operation parameters of the visualization component; wherein the multi-modal interaction instruction comprises any one or more of the following instructions: a natural language adjustment instruction for describing attribute changes of the visualization component; a sketch instruction for specifying expected dimensions or position relationships of the visualization component through a graphical outline; a gesture control instruction for directly manipulating the visualization component through an air gesture.

8. The method for generating model-driven visualized big screen data according to claim 1, characterized in that, The method further comprises: monitoring distribution changes of a data stream bound to the final layout scheme in real time; when detecting that the data distribution changes exceed a preset threshold, automatically triggering an incremental layout adjustment process, and dynamically optimizing the current visualization component layout based on changed data characteristics to focus on a new data insight focus.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The program, when executed by a processor, implements the model-driven generation method of the visualization data large screen according to any one of claims 1 to 8.

10. An electronic device on which a computer program is stored, characterized by The program, when executed by a processor, implements the model-driven generation method of the visualization data large screen according to any one of claims 1 to 8.

Images