Data insight driven embedded hierarchical table visualization collaborative construction method

Abstract

This invention discloses a data insight-driven embedded hierarchical table visualization human-computer collaborative construction method. It uses graph convolutional neural networks to represent the hierarchical structure of the table header, which has the ability to effectively understand the hierarchical table data structure. By finding data insights, it can discover meaningful patterns in the data. Based on deep reinforcement learning algorithms, it can effectively reduce the difficulty of manual operation. It also has the ability to interactively explore data insights, and through predefined data insights, it can discover special data patterns in the data from both single table unit data insights and multi-table unit data insights.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of data visualization, and particularly relates to a data-insight-driven embedded hierarchical table visualization human-computer collaborative construction method. BACKGROUND

[0002] Table data is a general format that plays an important role in various practical applications. For example, in the financial field, table data is used to track financial transactions, analyze trends, and help analysts make wise investment decisions. In scientific research, table data is also commonly used to store and analyze experimental results. Table data is an important tool for data organization and analysis across industries and disciplines. Tables are widely used to represent structured data because they can effectively organize and convey information in a clear and intuitive way. In particular, hierarchical table data contains multiple levels of table headers, and the hierarchical structure of these table headers improves the ability to manage data in two-dimensional space and makes it easier to analyze data across multiple dimensions, thereby facilitating more efficient data exploration. Embedded hierarchical table visualization provides a structured and organized way to present table data. Embedding visual elements in hierarchical table data can avoid distracting users and reduce the cognitive load required to explore table data. Embedded hierarchical table data visualization includes two ways: encoding a single data item within a cell as a visual element, or displaying the patterns and characteristics of a block of consecutive cells through visualization. Embedded hierarchical table data visualization helps users more efficiently understand complex data by maintaining the original table layout.

[0003] Constructing embedded hierarchical table data visualization can effectively assist users in understanding and analyzing data, but constructing embedded hierarchical table data visualization requires users to perform a series of iterative and tedious operations, including transforming hierarchical table data, selecting specific data cells, using multiple chart visualizations to visualize data, and embedding data visualizations into hierarchical table data. The above operations are not only tedious but also time-consuming. In addition, the embedded hierarchical table data visualization results constructed manually by users cannot provide comprehensive data insights into the hierarchical table data. Therefore, constructing embedded hierarchical table data visualization with rich data insights is a challenging research project, and the challenges facing this task include two aspects: first, constructing embedded hierarchical table data visualization involves multiple steps, and each step has many choices, so the design space of embedded hierarchical table data visualization is huge, and complex exploration is needed to construct effective visualization results; second, the structure of hierarchical table data is complex, and the insights into the internal data need to consider numerous variables and their relationships, so it is necessary to systematically model the patterns and insights existing in the hierarchical table data.

[0004] Based on the above technical problems existing in the prior art, the present application proposes a data-insight-driven embedded hierarchical table visualization human-computer collaborative construction method. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by providing a data insight-driven, embedded hierarchical table visualization human-computer collaborative construction method.

[0006] The present invention adopts the following technical solution:

[0007] On the one hand, a data insight-driven, embedded hierarchical table visualization human-computer collaborative construction method is provided, including:

[0008] Step 1: Define data insights for embedded hierarchical table data;

[0009] Step 2: The user uploads hierarchical table data;

[0010] Step 3: Use a graph convolutional neural network to represent the hierarchical structure of the table header and a multilayer perceptron to represent the state of the table data.

[0011] Step 4: The reinforcement learning algorithm transforms the hierarchical table data and selects the visualization area;

[0012] Step 5: Optimize the construction results of the embedded hierarchical table data visualization using the maximum reward function;

[0013] Step 6: Encourage the agent to explore the embedded hierarchical table data visualization space through the auxiliary reward module;

[0014] Step 7: Interactively generate embedded hierarchical table data visualization results and present them on the front end.

[0015] Furthermore, in step 7, users can interactively modify and update the embedded hierarchical table data visualization results.

[0016] Furthermore, step 3 includes:

[0017] The header hierarchy is modeled using a subtree-based representation method, where the hierarchical relationship is represented as the parent-child node relationship in a tree, the hierarchical state is represented as a tree structure, and a graph convolutional neural network is used to represent it to obtain the header information feature vector.

[0018] During the visualization process, the neural network understands the current state of the visualized data, uses a multilayer perceptron to represent the state of the visualized table data, and obtains a feature vector of visualization operation information.

[0019] By concatenating the feature vectors of the table header information and the feature vectors of the visualization operation information, the current state information of the table is obtained, thus enabling the neural network to correctly understand the hierarchical structure of the table data and the visualization of the table data state, and to perform operations and make decisions.

[0020] Furthermore, in step 5, the reward function includes the area ratio:

[0021] After each round of visualization generation is completed, the area ratio of the visualized data units of the intelligent agent is calculated:

[0022]

[0023] Among them, A v A represents the number of cells within the area visualized by the intelligent agent. d This indicates the number of cells in the entire table.

[0024] Furthermore, in step 5, the reward function includes a type ratio, which measures the proportion of all insight types discovered by the agent:

[0025]

[0026] Where, N d N is the number of insight types discovered. t It represents the total number of insight types.

[0027] Furthermore, in step 5, the reward function includes uniformity:

[0028] Uniformity balances the impact of various insights by calculating Shannon entropy based on the visualized area of ​​each insight, thus balancing the number of insights and producing diverse visualization results.

[0029]

[0030] Where, N t p is the number of insight types discovered. i This represents the probability that each insight covers a certain number of cells.

[0031] Furthermore, the auxiliary reward module in step 6 consists of a target network and a prediction network. The target network has randomly initialized parameters that are not updated during training, while the prediction network has randomly initialized parameters that are updated during training.

[0032] Furthermore, in step 6, the target network and the prediction network are represented as follows:

[0033]

[0034] Among them, the target network f H Used to provide a fixed embedded representation for the table header, while the prediction network It is then trained to learn the embedding representation of the header, where H is the header space. For a k-dimensional real space; the following f D and These represent the target network and the prediction network for the table content, respectively, resulting in four feature vectors embedded in the neural networks. The intrinsic reward is calculated as the mean squared error (MSE) among the four feature vectors. (Table header) Data items in the table content The intrinsic reward is defined as follows:

[0035]

[0036]

[0037] in, and It is the feature vector of the table header and table content obtained by the prediction network, while f H (x) and f D (x) is the feature vector of the header and table content obtained by the target network.

[0038] On the other hand, a data insight-driven embedded hierarchical table visualization human-computer collaborative construction system is provided, including:

[0039] The data insight definition module is used to define data insights for embedded hierarchical tabular data;

[0040] The table data upload module is used to enable users to upload hierarchical table data;

[0041] The graph convolutional neural network module is used to represent the header hierarchy;

[0042] The multilayer perceptron module is used to represent the visualization status of hierarchical tabular data;

[0043] The reinforcement learning algorithm module is used to transform hierarchical table data and select visualization areas.

[0044] A result optimization module is built to optimize the construction results of embedded hierarchical table data visualization using a maximization reward function;

[0045] An auxiliary reward module is used to encourage agents to explore the embedded hierarchical tabular data visualization space;

[0046] The visualization results presentation module is used to interactively generate embedded hierarchical table data visualization results and present the embedded hierarchical table data visualization results on the front end.

[0047] On another front, a computer device is provided, comprising: a processor and a memory, wherein the memory stores at least one program, which is loaded and executed by the processor to realize a data insight-driven embedded hierarchical table visualization human-computer collaborative construction method.

[0048] The beneficial effects of this invention are:

[0049] 1. The data insight-driven embedded hierarchical table visualization human-computer collaborative construction method described in this invention can realize the transformation and visualization embedding of hierarchical table headers for table data. It uses graph convolutional neural networks to represent the hierarchical structure state and has the ability to effectively understand the hierarchical table data structure. By finding data insights, it can discover meaningful patterns in the data. Based on deep reinforcement learning algorithms, it can effectively reduce the difficulty of manual operation.

[0050] 2. The data insight-driven embedded hierarchical table visualization human-computer collaborative construction method described in this invention has the ability to interactively explore data insights. Through predefined data insights, it can discover special data patterns in the data from both single table unit data insights and multi-table unit data insights.

[0051] 3. The data insight-driven embedded hierarchical table visualization human-computer collaborative construction method described in this invention has the ability to effectively understand the hierarchical table data structure, supports the system to parse hierarchical table data, and the graph convolutional neural network effectively represents the header structure of the hierarchical table data, and can clearly understand the header structure state.

[0052] 4. The data insight-driven embedded hierarchical table visualization human-computer collaborative construction method described in this invention reduces the difficulty for users to construct hierarchical table data visualization results. It uses reinforcement learning to replace complex and tedious manual operations, and can interactively generate visualization results with rich data insights, improve the efficiency of data acquisition, and significantly reduce the complexity of user operations. Attached Figure Description

[0053] Figure 1 This is a schematic diagram illustrating the process of transforming raw hierarchical table data into an embedded hierarchical table data visualization result in an embodiment of the present invention.

[0054] Figure 2 This is a schematic diagram illustrating the definition of data insights in an embodiment of the present invention;

[0055] Figure 3 This is a schematic diagram illustrating the process of a convolutional neural network and a multilayer perceptron understanding a table in an embodiment of the present invention;

[0056] Figure 4 This is a schematic diagram of auxiliary rewards used to assist exploration in an embodiment of the present invention;

[0057] Figure 5 This is a schematic diagram illustrating the actions taken by the reinforcement learning algorithm in an embodiment of the present invention;

[0058] Figure 6-1This is a schematic diagram of example result 1 of the generation of premium statistics related hierarchical table data in an embodiment of the present invention;

[0059] Figure 6-2 This is a schematic diagram of example result 2 of the generation of premium statistics related hierarchical table data in an embodiment of the present invention;

[0060] Figure 6-3 This is a schematic diagram of example result 3 of the generation of premium statistics related hierarchical table data in an embodiment of the present invention;

[0061] Figure 6-4 This is a schematic diagram of example result 4 of the generation of premium statistics related hierarchical table data in an embodiment of the present invention;

[0062] Figure 6-5 This is a schematic diagram of example result 5 of the generation of premium statistics related hierarchical table data in an embodiment of the present invention. Detailed Implementation

[0063] To better understand the above-mentioned objectives, features and advantages of the present invention, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present application can be combined with each other.

[0064] Example

[0065] The data insight-driven embedded hierarchical table visualization human-computer collaborative construction method includes:

[0066] Step 1: Define data insights for embedded hierarchical table data;

[0067] Step 2: The user uploads hierarchical table data;

[0068] Step 3: Use a graph convolutional neural network to represent the hierarchical structure of the table header and a multilayer perceptron to represent the state of the table data.

[0069] Step 4: The reinforcement learning algorithm transforms the hierarchical table data and selects the visualization area;

[0070] Step 5: Maximize the reward function to optimize the construction result of the embedded hierarchical table data visualization;

[0071] Step 6: Encourage agents to explore the embedded hierarchical tabular data visualization space through an auxiliary reward mechanism;

[0072] Step 7: Interactively generate embedded hierarchical table data visualization results, present the embedded hierarchical table data visualization results on the front end, and support user interactive modification and updates.

[0073] As a specific implementation method, in step 1, data insights are divided into two groups by defining continuous cells in hierarchical table data as table units: data insights for a single table unit and data insights for multiple table units.

[0074] Figure 2 The diagram illustrates the definition of data insights. The left sub-diagram shows how single-table cell data insights are defined. Single-table cell data insights include insights based on data values, insights based on data distribution, and composite data insights. Multi-table cell data insights are based on single-table cell data insights and include two types: related table cells have the same data insights or a certain table cell has significantly different data insights.

[0075] It should be noted that data insights for multiple table cells rely on insights into individual table cells and must consider the relationships between different table cells, including relationships based on name and relationships based on topology.

[0076] Furthermore, data insights for single table cells are divided into three categories: data value-based insights, data distribution-based insights, and composite data insights.

[0077] Among them, the insights based on data values ​​are related to the numerical values ​​in the table cells, including outliers, the largest percentage, the top two percentages, and outlier negative values.

[0078] Outliers are data items in a data block that differ significantly from the majority of the data. In this application, two methods are used to detect outliers:

[0079] The first method is to use quartile ranges to determine the upper and lower limits. For example, data points falling outside the defined range of Q3+3IQR or Q1-3IQR are considered outliers, where Q1 and Q3 represent the first and third quartiles, respectively, and IQR represents the difference between the two. In this embodiment, a box plot is used to represent the outlier. Another method, for example, is to fit a power law function to the data, then calculate the probability that a certain value is significantly greater than other values ​​based on the prediction error, and use a bar chart to display the outlier.

[0080] Maximum proportion refers to the degree of dominance of one random variable relative to one or more other random variables. In other words, a variable has a significant advantage over other variables (e.g., at least 50%), and this data insight can be visualized using a pie chart or radial chart.

[0081] The top two percentages refer to two variables that have a significant advantage over other variables (e.g., both have a percentage greater than or equal to 34%), which can be visualized using a pie chart or radial chart similar to the maximum percentage.

[0082] Outlier negative values ​​refer to data patterns in a dataset where the largest negative value is significantly lower than all remaining values.

[0083] The second category of insights related to numerical distribution includes evenness, trend, skewness, and kurtosis;

[0084] Evenness refers to the degree of similarity between variables. For example, it can be measured by the coefficient of variation, which is calculated by dividing the standard deviation σ by the mean μ. A lower coefficient of variation indicates that the variables are more evenly distributed. For example, evenness can be visualized using a bar chart.

[0085] To determine whether the data exhibits a consistent upward or downward pattern, in the context of time series data, the following formula can be used to detect trends:

[0086] trend=r 2 ×(1-p),

[0087] Where, r 2 It is the coefficient of determination, p represents the p-value of the probability that the observed slope value is equal to or greater than the slope of the upward or downward trend. The above formula applies to both upward and downward trends, and then the trend can be visualized using line charts or horizontal charts.

[0088] Trend change points refer to significant changes or trend shifts in values ​​within a data block. This data insight involves pinpointing specific points where the behavior or characteristics of a variable deviate significantly from its previous pattern. Trend change point insights can provide important information about key changes in trends, distributions, or relationships within the data.

[0089] The degree of asymmetry in the distribution of skewness measurements is determined by the absolute value; a larger absolute value indicates a greater deviation from symmetry. Skewness can be detected using the following formula:

[0090]

[0091] Where, μ b σ is the average value of the data. σ Let n be the standard deviation of the data, n be the number of data points, and b be the standard deviation of the data. i For the i-th data point, a density plot is used to visualize the skewness;

[0092] Kurtosis measures the degree of kurtosis within a data block. A larger absolute value indicates a more pronounced distribution. Kurtosis can be calculated using the following formula:

[0093]

[0094] Density plots are then used to visualize kurtosis;

[0095] The third category involves grouping the data and examining the relationships between different groups, including independence, correlation, and cross-correlation.

[0096] Independence is determined by dividing the dataset by row attributes, treating the rows and columns of table cells as two categorical variables, and verifying independence using a chi-square test by calculating p-values. The results of the chi-square test are then visualized using a standardized stacked bar chart.

[0097] Relationship insights divide the dataset by row attributes. Unlike independence insights, the purpose of relational insights is to verify the correlation between different rows. For example, when most rows show significant correlation (p<0.05), the multi-row data unit is considered to be composed of relational insights, and then the data insights of correlation are visualized using line charts.

[0098] Cross-correlation is calculated using the Pearson coefficient to determine the correlation between data rows / columns x and y, |ρ(x,y)|, where σ x This data insight is visualized using a scatter plot to represent the standard deviation of the data row / column x.

[0099]

[0100] Data insights for multiple table cells involve computing multiple common data insights across multiple data blocks. Its application scenario is that a user selects a specific data cell, and then other data cells are recommended based on the name or subtree. The calculation of insights for multiple table cells is based on the data insight of a single table cell.

[0101] Data insights based on multiple table cells have two aspects. They are considered to exist when all table cells contain the same type of insight, or when the insight of a certain table cell is significantly different from that of other table cells. Data insights based on multiple table cells are data insights defined based on single table cells. These data insights take into account the characteristics of hierarchical table data. Due to the hierarchical structure of the table header, there are relationships between attributes in hierarchical table data, such as same name relationships or parent-child relationships.

[0102] As a specific implementation method, in step 2, the user first uploads the tabular data to the insight-driven hierarchical tabular data visualization system. The system receives the data input and understands the hierarchical structure relationship. The system supports .xlsx data format and .csv data format.

[0103] As a specific implementation method, in step 3... Figure 3This paper demonstrates the process of understanding tables using graph convolutional neural networks (GNNs) and multilayer perceptrons (MLPs). The GNN represents the hierarchical table structure, while the MLP represents the table's changing states. The information from these two parts is concatenated as a feature of the entire hierarchical table data state and input into a reinforcement learning decision model. In tasks involving the transformation and visualization of hierarchical table data, accurately representing the table header hierarchy and visualizing the table data state affects the deep model's understanding of the data. Without this, operations and decisions cannot be performed. Table data typically consists of two parts: a header and data. The header structure frequently changes, posing a challenge to the efficiency of deep models in representing hierarchical tables. Therefore, this embodiment separates the hierarchical structure and table data into two distinct parts. Encoding can be performed using different encoders. For example, a subtree-based representation method can be used to model the header hierarchy, representing the hierarchical relationship as a parent-child node relationship in a tree. The entire hierarchical state is represented as a tree structure, and a graph convolutional neural network is used to represent it, obtaining the header information feature vector. At the same time, during the visualization process, the neural network needs to understand the current visualization data state. A multilayer perceptron is used to represent the visualization table data state, obtaining the visualization operation information feature vector. Finally, these two feature vectors are concatenated to obtain the current state information of the entire table, thus enabling the neural network to correctly understand the hierarchical structure of the table data and the visualization table data state, and perform relevant operations and decisions.

[0104] As a specific implementation method, in step 4, considering that the header structure of the hierarchical table data determines the organization method of the internal data items, this application embodiment defines an action space based on the header structure, wherein the actions include transpose, aggregation, stacking down, stacking up, row swapping, column swapping, moving the row-positioned header up, moving the row-positioned header down, moving the row-positioned header left, moving the row-positioned header right, moving the column-positioned header up, moving the column-positioned header down, moving the column-positioned header left, and moving the column-positioned header right. In addition, the actions are divided into two categories: header transformation and visualization area selection;

[0105] like Figure 1 and Figure 5 As shown, the process from the original hierarchical table data to the embedded hierarchical table data visualization result is divided into two stages of table operations. The first stage is the table header transformation stage, which involves a series of table header transformation operations to make the table reach a suitable hierarchical state. The second stage is the visualization stage, which involves a series of positioning operations to select the visualization area and then embed the corresponding visualization form in the original position.

[0106] Furthermore, the header transformation in the action space consists of five operations: transpose swaps the row headers and list headers of the entire hierarchical table; aggregation derives values ​​based on multiple rows or columns (such as summation or average) under a specific category; stack down moves the innermost list header to the innermost row header, and conversely, stack up moves the innermost row header to the innermost list header; and swap swaps two levels in the row header or list header. Using these operations, users can flexibly transform the headers and select specific data items.

[0107] It should be noted that some header transformation operations may result in empty cells in the hierarchical table because the corresponding data items for these header combinations do not exist. In this case, the header transformation process can still continue, but visualization operations cannot be performed.

[0108] The visualization operation aims to locate specific data blocks in hierarchical table data by moving row headers and list headers. The initial selection area is located in the upper left corner. The operation of moving row headers and list headers is the same, including four directions: up, down, left, and right. By selecting row headers and list headers, the table cell to be visualized can be uniquely determined. The agent uses the corresponding visualization to present the data insights discovered during the selection process. When selecting data from hierarchical table data to determine the visualization range, row headers and list headers need to be determined. The row headers and list headers of hierarchical table data are multi-level. The row headers and list headers together uniquely determine a table area. Then, the determined area can be changed by moving these headers. Table cell recommendations are based on subtree pattern and name pattern. Based on subtree pattern, a certain number of priorities are extended upwards according to the current header's level, that is, recommendations are made to the multiple generations of parent nodes of the current node in the tree structure. Based on name matching, the name of the current header is searched among all header names to find headers with the same name in the table hierarchy.

[0109] In a specific implementation, step 5 involves reinforcement learning involving agents and an environment. Each agent is a deep network used to learn a target task, while the environment is the scenario in which the agent operates. Throughout the reinforcement learning process, the agent and the environment continuously interact. The agent's goal is to obtain as many rewards as possible from the environment. In the task, the defined reward function consists of three parts, designed to optimize the diversity and effectiveness of the generated results to evaluate the quality of hierarchical tabular data visualization:

[0110] Area ratio: After each round of visualization generation, the area ratio of the visualized data units of the intelligent agent is calculated.

[0111]

[0112] Among them, A vA represents the number of cells within the area visualized by the intelligent agent. d This indicates the total number of cells in the entire table;

[0113] Type ratio: In this application embodiment, a type ratio metric is also defined to measure the proportion of all insight types discovered by the agent:

[0114]

[0115] Where, N d N is the number of insight types discovered. t It is the total number of insight types;

[0116] Uniformity, in order to balance the influence of various insights, is calculated based on the visual area of ​​each insight, using Shannon entropy, which aims to balance the number of insights and produce diverse visualization results:

[0117]

[0118] Where, N t p is the number of insight types discovered. i This represents the probability that each insight covers a certain number of cells.

[0119] As a specific implementation method, in step 6, such as Figure 4 As shown, auxiliary rewards are calculated for the hierarchical header structure and table state, respectively. These rewards are calculated by a prediction network and a target network, and their errors are used as auxiliary rewards to aid exploration. This application introduces a collaborative optimization technique based on internal and external rewards to promote the agent's exploration of new and unknown states in the environment, such as new header structures and visualization states. This technique can predict the errors in the environmental state and guide the agent's exploration. The method consists of a target network and a prediction network. The target network has randomly initialized parameters that are not updated during training, while the prediction network has randomly initialized parameters that are updated during training. As training progresses, the prediction network encounters more unexplored states, and its state representation gradually approaches the fixed representation of the observations provided by the target network. Therefore, the intrinsic reward calculated between these two representations becomes smaller. When a state is accessed less frequently, the difference between the representations of the state by the prediction network and the target network becomes greater, thereby providing a greater intrinsic reward to encourage exploration.

[0120] In the auxiliary reward module, both the graph convolutional neural network and the multilayer perceptron consist of two sub-networks with the same structure but different parameters, generating different feature representations for the same environmental state. To represent the table header, this embodiment uses two networks:

[0121]

[0122]

[0123] Among them, the target network f H Used to provide a fixed embedded representation for the table header, while the prediction network It is then trained to learn the embedding representation of the header, where H is the header space. For k-dimensional real space; similarly, f D and These represent the target network and the prediction network for the table content, respectively. After obtaining the feature vectors of the four neural network embeddings, the intrinsic reward is calculated as the mean squared error (MSE) among these feature vectors. (Table header) Data items in the table content The intrinsic reward is defined as follows:

[0124]

[0125] in, and It is the feature vector of the table header and table content obtained by the prediction network, f H (x) and f D (x) is the feature vector of the header and table content obtained by the target network.

[0126] As a specific implementation method, in step 6, during the process of generating the embedded hierarchical table data visualization result, the data will undergo two stages of processing. First, in the first stage, the table header structure of the data will be changed multiple times to obtain a better table header state. Then, in the second stage, data visualization operations are performed by moving the visualization area multiple times to the defined insight and embedding the data in that area into the corresponding visualization chart. After completing the two stages of operations, a visualization embedding result is generated. Since the program runs in parallel, multiple different visualization embedding results can be generated for users to view and select.

[0127] After the results are generated, they are displayed to users in the front-end system. Users can access the front-end system to view and edit the results. For example, users can restore the charts of certain areas to data or select data that was not selected for visualization.

[0128] To verify the technical effectiveness of the embodiments of this application, an application case is used for supplementary illustration. Specifically, data analysts from an insurance company applied the embodiments of this application. The analysts selected a hierarchical table dataset related to premium statistics, which contained information on premium income and year-on-year growth rate for each province. The attributes in the header of this hierarchical table include province, premium type, sales channel, time range, and statistical value type (total or year-on-year growth rate).Figure 6-1 to Figure 6-5 An example of a hierarchical table visualization is shown, which illustrates the insights extracted from embodiments of this application;

[0129] For single-block insights, such as Figure 6-1 As shown in box a, the total long-term insurance amount for p1 (Province 1) is an outlier within a single column, exceeding the power-law distribution of the column. This insight is visualized using a bar chart, as shown below. Figure 6-1 As shown in box b, one data item is significantly smaller than the others in the column (less than Q1-3×IQR). After examining the underlying data, the analyst discovered that the total interim insurance amount in p3 is an outlier. Figure 6-1 The middle box (C) shows a data block with aggregated insights. In the data block, data items are aggregated by calculating the average along the columns (indicated by downward arrows). This insight represents the largest proportion and is visualized using a radial graph. The analysis reveals that the service marketing channel category accounts for the largest proportion of the channels.

[0130] like Figure 6-2 As shown, the visualization results contain various insights across multiple blocks. Box d displays three recommended data blocks through a name-based mechanism. These data blocks share the same header except for the top-level header (representing the current year, previous month, and previous day). Analysts found that the two data blocks corresponding to the current day and previous month both contain outliers, while the data block corresponding to the current year does not. This suggests that the year-on-year growth rates of the four different insurance types have shown increasing variability in recent years.

[0131] exist Figure 6-3 In the middle, box e describes the skewness insights related to four recommended data blocks based on the topology mechanism. Among these four data blocks, the analyst found that the data distribution of the first data block had no obvious skewness, while the other three data blocks showed strong skewness. This indicates that the data distribution is large in scale but highly concentrated in a small range. This information tells the analyst that the total sales amount of the sales channels under the large individual insurance category is more balanced, while other channels need to be improved.

[0132] exist Figure 6-4 In the middle, box f shows the data blocks determined based on the topology mechanism. The multi-series line chart on the left illustrates the correlation between the total insurance sales channels under the major individual insurance categories, while the chart on the right does not show this correlation. This insight encourages analysts to analyze the potential relationships between these sales channels and improve strategies, such as avoiding the mutual influence between different types of insurance.

[0133] Along Figure 6-5When aggregating data items from multiple blocks in the middle g box, analysts can identify uniformity insights. The left block shows uniformity characteristics, while the right block does not. This means that the year-on-year growth rates of "large individual insurance" in p5 are similar, while the differences of "non-large individual insurance" are significant in the same month. Therefore, it is necessary to refine the policy to improve the efficiency of non-large individual insurance.

[0134] Similarly, Figure 6-5 The middle box (h) presents four data blocks, each containing an aggregated insight. The top two blocks represent one set of recommended insights, while the bottom two represent another set. All four insights fall into the top two or largest proportions and are visualized using a radial plot. Analysts found significant differences in the distribution of large and non-large individual insurance categories between p5 and p6. Specifically, within the large individual insurance category, short-term and medium-term insurance have the highest and second-highest values, respectively, but their relative sizes differ. In the non-large individual insurance category, short-term insurance has the largest proportion in p5, while medium-term insurance has the largest proportion in p6. This data insight suggests that analysts should analyze the strategic differences between p5 and p6 in detail to achieve complementarity.

[0135] The data insight-driven embedded hierarchical table visualization human-computer collaborative construction system includes:

[0136] The data insight definition module is used to define data insights for embedded hierarchical tabular data;

[0137] The table data upload module is used to enable users to upload hierarchical table data;

[0138] The graph convolutional neural network module is used to represent the header hierarchy;

[0139] The multilayer perceptron module is used to visualize the state of hierarchical tabular data.

[0140] The reinforcement learning algorithm module is used to transform hierarchical table data and select visualization areas.

[0141] A result optimization module is built to optimize the construction results of embedded hierarchical table data visualization using a maximization reward function;

[0142] An auxiliary reward module is used to encourage agents to explore the embedded hierarchical tabular data visualization space;

[0143] The visualization results presentation module is used to interactively generate embedded hierarchical table data visualization results and present the embedded hierarchical table data visualization results on the front end.

[0144] The computer device includes a processor and a memory, the memory storing at least one program, which is loaded and executed by the processor to realize a data insight-driven embedded hierarchical table visualization human-computer collaborative construction method.

[0145] This invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of this invention is defined by the appended claims.

Claims

1. A data insight-driven, embedded hierarchical table visualization human-computer collaborative construction method, characterized in that, include: Step 1: Define data insights for embedded hierarchical table data; Step 2: The user uploads hierarchical table data; Step 3: Use a graph convolutional neural network to represent the hierarchical structure of the table header and a multilayer perceptron to represent the state of the table data. The header hierarchy is modeled using a subtree-based representation method, where the hierarchical relationship is represented as the parent-child node relationship in a tree, the hierarchical state is represented as a tree structure, and a graph convolutional neural network is used to represent it to obtain the header information feature vector. A multilayer perceptron is used to represent the state of the visualized table data to obtain a feature vector of visualized operational information. By concatenating the feature vectors of the table header information and the feature vectors of the visualization operation information, the current status information of the table can be obtained. Step 4: The reinforcement learning algorithm transforms the hierarchical table data and selects the visualization area; Step 5: Optimize the construction results of the embedded hierarchical table data visualization using the maximum reward function; Step 6: Encourage the agent to explore the embedded hierarchical table data visualization space through the auxiliary reward module; Step 7: Interactively generate embedded hierarchical table data visualization results and present them on the front end.

2. The data insight-driven embedded hierarchical table visualization human-computer collaborative construction method according to claim 1, characterized in that, In step 5, the reward function includes the area ratio; After each round of visualization generation is completed, the area ratio of the visualized data units of the intelligent agent is calculated: ……（1）， in, This represents the number of cells within the area visualized by the intelligent agent. This indicates the number of cells in the entire table.

3. The data insight-driven embedded hierarchical table visualization human-computer collaborative construction method according to claim 2, characterized in that, In step 5, the reward function includes the type ratio; The type ratio is used to measure the proportion of all types of insights discovered by an agent: ……（2）， in, It is the number of insight types discovered. It represents the total number of insight types.

4. The data insight-driven embedded hierarchical table visualization human-computer collaborative construction method according to claim 3, characterized in that, In step 5, the reward function includes uniformity; Uniformity balances the impact of various insights by calculating Shannon entropy based on the visualized area of ​​each insight, thus balancing the number of insights and producing diverse visualization results. ……（3）， in, It is the number of insight types discovered. This represents the probability that each insight covers a certain number of cells.

5. The data insight-driven embedded hierarchical table visualization human-computer collaborative construction method according to claim 1, characterized in that, The auxiliary reward module in step 6 consists of a target network and a prediction network. The target network has randomly initialized parameters that are not updated during training, while the prediction network has randomly initialized parameters that are updated during training.

6. The data insight-driven embedded hierarchical table visualization human-computer collaborative construction method according to claim 1 or 5, characterized in that, In step 6, the target network and the prediction network are represented as follows: ……（4）； ……（5）， Among them, the target network Used to provide a fixed embedded representation for the table header, while the prediction network It is then trained to learn the embedding representation of the table header. For header space, for 3D real space; and Representing the target network and prediction network for the table content, four feature vectors are obtained from the neural network embeddings. The intrinsic reward is calculated as the mean squared error (MSE) among the four feature vectors. (Table header) and data items in the table content The intrinsic reward is defined as follows: ……（6）； ……（7）， in, and It is the feature vector of the table header and table content obtained by the prediction network, while and It is the feature vector of the header and table content obtained by the target network.

7. A data insight-driven embedded hierarchical table visualization human-computer collaborative construction system, characterized in that, include: The data insight definition module is used to define data insights for embedded hierarchical tabular data; The table data upload module is used to enable users to upload hierarchical table data; The graph convolutional neural network module is used to represent the header hierarchy: The header hierarchy is modeled using a subtree-based representation method, where the hierarchical relationship is represented as the parent-child node relationship in a tree, the hierarchical state is represented as a tree structure, and a graph convolutional neural network is used to represent it to obtain the header information feature vector. A multilayer perceptron is used to represent the state of the visualized table data to obtain a feature vector of visualized operational information. By concatenating the feature vectors of the table header information and the feature vectors of the visualization operation information, the current status information of the table can be obtained. The multilayer perceptron module is used to represent the visualization status of hierarchical tabular data; The reinforcement learning algorithm module is used to transform hierarchical table data and select visualization areas. A result optimization module is built to optimize the construction results of embedded hierarchical table data visualization using a maximization reward function; An auxiliary reward module is used to encourage agents to explore the embedded hierarchical tabular data visualization space; The visualization results presentation module is used to interactively generate embedded hierarchical table data visualization results and present the embedded hierarchical table data visualization results on the front end.

Images