An analysis task perception method fusing a deep learning model and rules

Abstract

The application discloses an analysis task perception method fusing a deep learning model and rules, first designs and realizes a workshop system, holds a workshop to obtain data, screens and collates operation sequence data, annotates analysis task labels for views according to notes written by users, and obtains effective data; determines input and output formats of the model according to user operation classification, the input is a reorganized user operation sequence, and the output is scores predicted for each task; uses the model to output prediction scores of each analysis task according to a current operation sequence of the user, and determines prediction results of each analysis task according to rules between column preferences of the user, current column attributes and the analysis task. The application makes up for the deficiency of task perception in existing task-driven automatic visualization, can predict the current analysis task of the user, makes task perception more convenient and intelligent, and makes the perception of the analysis task more humanized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of automatic information visualization, in particular to an analysis task perception method fusing a deep learning model and rules. BACKGROUND

[0002] With the development of Internet technology and the advent of the big data era, how to quickly analyze large-scale data sets and obtain insights and knowledge from them is crucial. In recent years, the development of data visualization technology has provided help for the analysis of data sets. With the aid of data visualization tools to create views, not only can data be analyzed intuitively and conveniently, but also the data analysis capabilities of non-data analysis professionals can be enhanced.

[0003] In order to further simplify the operation of visualization tools, automatic visualization technology has emerged. Most automatic visualization systems first require users to upload data sets in table format, and the system automatically generates meaningful views. Users can also specify data columns, and the system generates corresponding views and recommends related views. The generation and recommendation methods are various, such as obtaining a search space through enumeration and pruning using predefined rules, training machine learning models to obtain better column combinations, etc.

[0004] However, for most users, they do not have professional visualization knowledge and do not understand the data to be analyzed. The above process still requires users to repeatedly try to compare the pros and cons of views generated by different column combinations, which is tedious and time-consuming. Since users' analysis of data is usually task-oriented, a method based on user-given analysis tasks for view recommendation has recently appeared, but this method still requires users to specify analysis tasks and cannot automatically perceive user tasks. SUMMARY

[0005] To solve the above problems, the purpose of the present application is to perceive the analysis task of the user in the process of the user exploring data using an automatic visualization system, and to realize analysis task prediction based on a deep learning model and rules. The present application also supports user adjustment of rules and can recommend in real time according to adjusted rules to help users quickly and individually explore data. The specific technical solutions are as follows:

[0006] An analysis task perception method fusing a deep learning model and rules, comprising the following steps:

[0007] S1: Data acquisition and processing

[0008] A workshop system is designed and implemented, a workshop is held to obtain data, operation sequence data is screened and arranged, valid information is retained, operation sequences and their corresponding view data are obtained; analysis task labels are annotated for views according to notes written by users, and task label data is obtained;

[0009] S2: Data analysis

[0010] By the operation sequence data, view data and task label data obtained in step S1, the views are converted into corresponding visual encodings, and the significance relationship between each visual encoding and analysis task is obtained using statistical test method, the visual encodings with significant relationship are screened out, and the user operations are screened and classified accordingly, to prepare for model construction and training;

[0011] S3: Model construction and training

[0012] According to the classification of user operations in S2, the user operation sequence in the workshop and the analysis task data corresponding to the view are processed to determine the input and output formats of the model, the input is the reorganized user operation sequence, and the output is the predicted score of each task; a deep learning model is designed and implemented; the processed data is split into training set and test set for training and testing the model, and the final model and its score are obtained;

[0013] S4: Task prediction result fusion

[0014] According to the current operation sequence of the user, the model in S3 is used to determine the predicted score of the analysis task, and according to the column preference of the user and the rules between the current column attribute and the analysis task, the predicted result of the analysis task is determined; the obtained predicted results are fused to obtain the final analysis task prediction result.

[0015] Further, in step S1, the data acquisition and processing are specifically as follows:

[0016] S11: Design and implement the workshop system, which includes loading data, adding views, modifying visual encodings of views, writing notes for views, collecting views, operation burying, and view recommendation;

[0017] S12: Hold a workshop to make the user familiar with the workshop system, so that the user can freely explore the data set through the system, add the created view which can complete the user's analysis task to the collection, and write notes to record ideas;

[0018] S13: Obtain "operation sequence-view sequence" data according to the operation record in the burying data;

[0019] S14: Obtain "view-analysis task" data according to the note manually annotated by the user for the analysis task label corresponding to the view;

[0020] S15: Integrate the data in S13 and S14 to obtain "operation sequence-task set sequence" data.

[0021] Furthermore, in step S2, the data analysis specifically includes:

[0022] S21: Convert the views in the "view-analysis task" data obtained in S14 into corresponding visual codes to obtain the "view-visual code set-analysis task" data;

[0023] S22: Use statistical testing methods to obtain the significant association between each visual code and the task;

[0024] S23: Filter out visual codes with significant correlations, categorize user operations based on their impact on the visual codes, and retain four categories of operations: modifying chart types, aggregation, coding, and others.

[0025] Furthermore, the specific scheme for obtaining the saliency correlation between each visual encoding and the task is as follows:

[0026] S221: Construct statistical test hypotheses for any visual code; construct the following hypotheses:

[0027] H0: Visual encoding and analysis tasks are independent of each other;

[0028] H1: Visual encoding and analysis tasks are not independent of each other;

[0029] S222: Construct an RxC contingency table, where n represents the sample size and f... i,j This represents the value of the cell corresponding to the i-th row and j-th column, represented by F. Y,i To represent the sum of the values ​​in the i-th row of cells, use F. X,j This represents the sum of the values ​​in the j-th column, where r is the row number and c is the column number;

[0030] S223: Calculate the degrees of freedom d and the theoretical frequency e of each cell. ij :

[0031] d = (r-1)*(c-1)

[0032]

[0033] The chi-square test is met if all of the following conditions are met:

[0034] 1) The degree of freedom d is greater than or equal to 1;

[0035] 2) Theoretical frequency e ij The number of cells with a value less than 5 accounts for less than or equal to 20% of the total number of cells;

[0036] If the chi-square test conditions are met, proceed to step S224; otherwise, terminate the calculation.

[0037] S224: Calculate the statistic: calculate X by the following formula 2 :

[0038]

[0039] wherein, X 2 represents the statistic;

[0040] Further, the p-value is obtained according to the probability density function of the chi-square distribution:

[0041]

[0042] wherein, d is the degree of freedom, x represents X 2 , Γ is the gamma function, and p represents the significance;

[0043] If p<0.005, it is considered that the H0 hypothesis is not established, and the H1 hypothesis is accepted, that is, the visual coding and analysis task has a significant correlation.

[0044] Further, in step S3, the model construction and training are specifically:

[0045] S31: For each "operation sequence-task set sequence" data, according to the user operation category defined in S2, the data is filtered, classified, and the operation sequence is split and reorganized according to the given format, and the task set data is encoded, to obtain the input and output data of the model;

[0046] S32: Design and implement an LSTM model, the input is the reorganized operation sequence, and the output is the score of each analysis task predicted by the model;

[0047] S33: Split the data into a training set and a validation set, and use five-fold cross-validation to obtain the final model and score.

[0048] Further, the specific process of obtaining the input and output data of the model is as follows:

[0049] S311: A sliding window method is used, considering the user's continuous behavior of creating charts, according to the average length l of the operation sequence of the user on each data set and the average number m of charts created in the workshop, the window is set to be slightly larger than l / m, denoted as w;

[0050] S312: Extract the w operations closest to the current operation as part of the model input, denoted as F raw ;

[0051] S313: According to the proportion of the operation sequence data obtained in the workshop according to the modification of the chart type, aggregation, encoding and other operations, for each type of operation, a number of operations of the type before the current operation and closest to the current operation are extracted in the operation sequence, and are sequentially composed into a sequence, denoted as F gi ; and the number of extracted operations is positively correlated with the proportion of the type of operation;

[0052] S314: The sequence F gi obtained for each type of operation is sequentially spliced to obtain another part of the model input, denoted as F grouped ;

[0053] S315: F raw and F grouped are spliced to obtain the input data of the model;

[0054] S316: The set of analysis tasks of the view corresponding to the position of the current sliding window is encoded and combined using one-hot encoding to obtain the output data of the model.

[0055] Further, the model training and verification are as follows:

[0056] S331: The data is randomly shuffled and split into a training set and a test set in proportion;

[0057] S332: The training set is evenly divided into f parts, and f times of training and verification are performed to obtain f model parameters and their corresponding scores, wherein the kth training uses the kth, (k+1)th, (k+2)th, (k+3)th, and (k+4)th parts of data as the training set, and the (k+4)th part of data as the verification set;

[0058] S333: Select the model parameter with the highest score in the f times of training as the final model parameter, and test it on the test set to obtain the final score of the model.

[0059] Further, in step S4, the task prediction result fusion is as follows:

[0060] S41: Obtain the prediction score of the analysis task using the model in S3 according to the user operation sequence;

[0061] S42: Use statistical methods to determine whether there is a column preference in the user operation process. If there is a column preference, obtain the prediction result of the analysis task according to the attribute of the column selected by the current user and the corresponding rule of the analysis task, and fuse the model prediction score and the prediction result according to the rule as the final prediction result, otherwise, use the model prediction score as the final prediction result.

[0062] Further, whether the column selection in the user operation process has a preference is judged using a statistical method, and the method is as follows:

[0063] S421: constructing a statistical test hypothesis: the following hypothesis is constructed:

[0064] H3: the user has no preference for the column selection;

[0065] H4: the user has a preference for the column selection;

[0066] S422: counting the number of times of selection of each column in the user operation process, respectively n i , and the total number of times of selection, denoted as N; if there is no column preference, it is considered that the selection probability of each column is equal, and the selection probability is denoted as P;

[0067] S423: if the following conditions are met at the same time, the sample size test is passed:

[0068] 1) N >= 40; that is, the total number of times of column selection of the user is greater than 40 times;

[0069] 2) min (n i ) >= 5; that is, the number of times of selection of each column is not less than 5 times;

[0070] 3) N*P >= 5; that is, the theoretical frequency of each column is not less than 5;

[0071] If the sample size test is passed, step S424 is entered, otherwise it is considered that the user has no preference for the column selection;

[0072] S424: correlation test: the statistical quantity X 2 is calculated by the following formula:

[0073]

[0074] According to the probability density function of the chi-square distribution, the p value representing significance is further obtained, if p < 0.005, it is considered that the H3 hypothesis is not established, and the H4 hypothesis is accepted, that is, the user has a preference for the column selection.

[0075] The beneficial effects of the present application are:

[0076] 1) The present application makes up for the deficiency of task perception in the existing task-driven automatic visualization. The existing task-driven visualization method can recommend related views according to the analysis task specified by the user, however, the user in reality is difficult to clearly propose the analysis task without knowing the data set. In the face of this problem, the processed user operation sequence is input into a deep learning model to predict the analysis task of the user, which can predict the current analysis task of the user, making the task perception more convenient and intelligent.

[0077] 2) The present application makes up for the deficiency of manual intervention in the existing automatic visualization system. On the basis of predicting the analysis task through user operation sequence, the user can adjust the default column attribute and the association rule between tasks, and the system output will integrate the model output score and the prediction result according to the current rule, so that the perception of the analysis task is more humanized. BRIEF DESCRIPTION OF DRAWINGS

[0078] Figure 1 is the whole flow framework diagram of the analysis task perception method of the present application fusing deep learning model and rule.

[0079] Figure 2 is the schematic diagram of the workshop system in the present application.

[0080] Figure 3 is the deep learning model structure schematic diagram in the present application.

[0081] Figure 4 is the implementation effect schematic diagram of the present application. DETAILED DESCRIPTION

[0082] The present application realizes the multi-angle analysis of the time sequence behavior of the network shopping platform user through the effective information visualization method, the multi-view linkage strategy and the flexible interaction means, helps the advertisement analyst to find the periodic behavior rule, the behavior mode under different time scales and the specific time sequence relationship between the advertisement click and other shopping behaviors. The technical scheme includes data acquisition and processing, visualization mapping, visualization layout and implementation, and interaction design. The specific steps are as follows:

[0083] Step one: data acquisition and processing

[0084] The workshop system is designed and implemented, the workshop is held to obtain data, the operation sequence data is screened and arranged, the note written by the user is used to label the analysis task label of the view, and the effective data is obtained.

[0085] 1. Design and implement the workshop system: the workshop system includes data loading, adding view, modifying the visual coding of view, writing note for view, collecting view, operation burying point and view recommendation and other functions (as shown in the drawing). Figure 2

[0086] 2. Hold the workshop: after the user is familiar with the workshop system, the user is required to freely explore the cars and birdstrict data set using the system, the view created by the user to complete the analysis task is added to the collection, and the note is written to record the idea as much as possible.

[0087] ​3. Data processing: Process the user operation sequence data and view note data obtained by the workshop.

[0088] (1) User operation sequence processing: In the process of exploring a certain data set, the user will generate an operation sequence, obtain the current view at the time of the last "favorite view" operation after each operation, and finally obtain "operation sequence-view sequence" data.

[0089] (2) View note data processing: For all user favorite views, combine the view content and the note corresponding to the view, manually label the possible analysis tasks it contains, including data conversion, comparison, association, trend, anomaly, distribution, and value confirmation, to obtain "view-analysis task set" data. For views with multiple analysis tasks labeled, split them into multiple data, i.e., each data contains only one analysis task, and finally obtain "view-analysis task" data.

[0090] (3) Integration data: Integrate the "operation sequence-view sequence" data in (1) and the "view-analysis task set" data in (2) to obtain "operation sequence-task set sequence" data.

[0091] Step 2: Data analysis

[0092] After data acquisition and processing, further data analysis is performed to prepare for model construction and training.

[0093] 1. Obtain the visual encoding of the view: Convert the view in the "operation sequence-view" data obtained in S14 into the corresponding visual encoding, and further obtain "view-visual encoding set-analysis task" data, where the visual encoding involved is shown in Table 1;

[0094] Table 1 Visual encoding explanation

[0095] Visual encoding meaning English abbreviation Chart type mark Whether the x-axis is encoded x_enabled Name of the column that the x-axis encodes x_column Whether the x-axis uses an aggregate function x_aggregate Whether the column that the x-axis encodes is nominal x_nominal Whether the column that the x-axis encodes is quantitative x_quantitative Whether filtering is used on the x-axis x_filter Whether the y-axis is encoded y_enabled Name of the column that the y-axis encodes y_column Whether the y-axis uses an aggregate function y_aggregate Whether the column that the y-axis encodes is nominal y_nominal Whether the column that the y-axis encodes is quantitative y_quantitative Whether filtering is used on the y-axis y_filter Whether color is encoded color_enabled Name of the column that color encodes color_column Whether color encoding uses an aggregate function color_aggregate Whether the column that color encodes is nominal color_nominal Whether the column that color encodes is quantitative color_quantitative Whether filtering is used on color encoding color_filter

[0096] 2. Chi-square independence test: Obtain the significant association between each visual encoding and task in the "view-visual encoding set-analysis task" data.

[0097] (1) For any visual encoding, construct a statistical test hypothesis, construct the following hypothesis:

[0098] H0: Visual encoding and analysis task are independent of each other.

[0099] H1: Visual encoding and analysis task are not independent of each other.

[0100] (2) Construct an RxC contingency table, where n represents the sample size, f i,jFij represents the value of the cell corresponding to the ith row and jth column Y,i Fi represents the sum of the values of the cells in the ith row X,j Fj represents the sum of the values of the cells in the jth column, r is the number of rows, and c is the number of columns.

[0101] (3) Calculate the degrees of freedom and the theoretical frequency. The degrees of freedom are calculated by the formula (r-1)*(c-1), and the theoretical frequency of each cell is calculated by the formula If the following conditions are met simultaneously, the chi-square test conditions are met:

[0102] ① The degrees of freedom are greater than or equal to 1;

[0103] ② The proportion of the number of cells with a theoretical frequency less than 5 to the total number of cells is less than or equal to 20%. If the chi-square test conditions are met, the calculation continues, otherwise the calculation is terminated.

[0104] (4) Calculate the statistic. Calculate X 2 :

[0105]

[0106] where X 2 represents the statistic;

[0107] and the p-value is further obtained according to the probability density function of the chi-square distribution:

[0108]

[0109] where d is the degrees of freedom, x is X 2 , Γ is the gamma function, and p represents the significance. If p<0.005, it is considered that the H0 hypothesis is not established, and the H1 hypothesis is accepted, i.e. the visual encoding and analysis task has a significant correlation.

[0110] 3. User operation classification: filter out visual encodings with significant correlation, filter and classify user operations according to the influence of user operations on visual encodings, obtain four types of operations: chart type, aggregation, encoding and other operations.

[0111] Step three: model construction and training

[0112] According to the classification of user operations, the "operation sequence-task set sequence" data obtained in step one is processed to determine the input and output formats of the model, where the input is the reorganized user operation sequence and the output is the predicted score of each task. Design and implement a deep learning model. Use the processed data to train and test the model to obtain the final model and its score;

[0113] 1、"Operation Sequence-Task Set Sequence" Data Processing: For each piece of data, according to the user operation category defined in step two, the data is filtered, classified, and the user operation sequence is split and reorganized according to the given format, and the task set is encoded, obtaining the input and output data of a pair of models.

[0114] (1) Using the sliding window method, considering the user's continuous behavior of creating charts, according to the average length l of the operation sequence of the user on each data set obtained in the workshop and the average number m of creating charts, the window is set to be slightly larger than l / m, i.e. 15;

[0115] (2) Extract the 15 operations closest to the current operation as part of the model input, denoted as F raw ;

[0116] (3) Calculate the proportion of chart type, aggregation, encoding and other four types of operations in the operation sequence data obtained in the workshop, set as f i , for each type of operation, extract 15*f i operations of that type before the current operation and closest to the current operation in the operation sequence and form a sequence in order, denoted as F gi ;

[0117] (4) According to the order of "chart type-aggregation-encoding-other", splice the sequence F gi obtained by each operation to obtain another part of the model input, denoted as F grouped ;

[0118] (5) Splice F raw and F groued to obtain the input data of the model;

[0119] (6) Use one-hot encoding to encode and combine the analysis task set corresponding to the view where the current sliding window is located, to obtain the output data of the model.

[0120] 2、Implementation of the model: design and implement an LSTM model (as shown in Figure 3 ), the input is the current operation sequence of the user, and the output is the score of each analysis task predicted by the model.

[0121] 3、Model training and verification: split the processed "operation sequence-task set sequence" data into training set and validation set, and use five-fold cross-validation to obtain the final model and score.

[0122] (1) Randomly shuffle the data and split it into training set and test set according to the ratio of 8:2;

[0123] (2) Split the training set into 5 parts, and perform 5 training and validation to obtain 5 model parameters and their corresponding scores. In the kth training, the kth, (k+1) %5, (k+2) %5, (k+3) %5 parts of data are used as the training set, and the (k+4) %5 part of data is used as the validation set.

[0124] (3) Select the model parameter with the highest score in the 5 training as the final model parameter, and test it on the test set to obtain the final score of the model.

[0125] Step four: task prediction result fusion

[0126] According to the user's current operation sequence, use the model in S3 to output the prediction score of each analysis task, and determine the prediction result of each analysis task according to the user's column preference and the rule between the current column attribute and the analysis task. Fuse the prediction results of the above two aspects to obtain the final analysis task prediction result.

[0127] 1. Obtain model prediction result: according to the user's current operation sequence, use the model in S3 to output the prediction score of each analysis task;

[0128] 2. Obtain rule prediction result: use statistical method to judge whether the column selection in the user operation process has preference, if there is column preference, then according to the attribute of the column selected by the current user and the corresponding rule of the analysis task, obtain the prediction result of the analysis task.

[0129] 3. If there is column preference, fuse the model prediction score and the prediction result according to the rule as the final prediction result, otherwise, use the model prediction score as the final prediction result.

[0130] (1) Construct statistical test hypothesis. Construct the following hypothesis:

[0131] H3: The user has no preference for column selection.

[0132] H4: The user has preference for column selection.

[0133] (2) Count the number of times each column is selected in the user operation process, respectively n i , and the total number of selections, denoted as N. If there is no column preference, it is considered that the selection probability of each column is equal, and the selection probability is denoted as P.

[0134] (3) If the following conditions are met at the same time, the sample size test is passed:

[0135] 4) N≥40; that is, the total number of column selections by the user is greater than 40 times;

[0136] 5) min(n i) ≥ 5; i.e. the selection frequency of each column is not less than 5 times;

[0137] 6) N*P ≥ 5; i.e. the theoretical frequency of each column is not less than 5;

[0138] If the sample size test is passed, step S424 is entered, otherwise it is considered that the user has no preference for the selection of the column.

[0139] (4) Correlation test. Calculate X 2 :

[0140]

[0141] And according to the probability density function of the chi-square distribution, the p value representing significance is further obtained, if p < 0.005, it is considered that the H3 hypothesis is not established, and the H4 hypothesis is accepted, i.e. the user has a preference for the selection of the column.

Claims

1. An analysis task-aware method fusing a deep learning model and rules, characterized in that, Comprising the following steps: S1: data acquisition and processing Design and implement a workshop system, hold a workshop to obtain data, screen and organize operation sequence data, retain valid information, obtain operation sequences and their corresponding view data; label analysis tasks according to the notes written by users, and obtain task label data; S2: data analysis Through the operation sequence data, view data and task label data obtained in step S1, the view is converted into the corresponding visual coding, and the significance relationship between each visual coding and the analysis task is obtained using statistical test method, the visual coding with significant relationship is screened out, and the user operation is screened and classified accordingly, to prepare for model construction and training; S3: model construction and training According to the classification of user operation in S2, the user operation sequence and the corresponding view analysis task data in the workshop are processed to determine the input and output format of the model, the input is the reorganized user operation sequence, and the output is the score of each task prediction; design and implement a deep learning model; split the processed data into training set and test set for model training and testing, and obtain the final model and its score; S4: task prediction result fusion According to the current operation sequence of the user, the prediction score of the analysis task is determined using the model in S3, and the prediction result of the analysis task is determined according to the column preference of the user and the rules between the current column attribute and the analysis task; fuse the obtained prediction results to obtain the final analysis task prediction result.

2. The analytical task perception method integrating deep learning models and rules according to claim 1, characterized in that, In step S1, the data acquisition and processing is specifically: S11: design and implement a workshop system, which includes loading data, adding views, modifying visual coding of views, writing notes for views, collecting views, operation burying points and view recommendation; S12: hold a workshop to make users familiar with the workshop system, so that they can freely explore the data set through the system, add the created view which can complete the user's analysis task to the collection, and write notes to record ideas; S13: obtain "operation sequence-view sequence" data according to the operation record in the burying point data; S14: obtain "view-analysis task" data according to the analysis task label manually labeled by the user for the view according to the note written by the user for the view; S15: integrate the data in S13 and S14 to obtain "operation sequence-task set sequence" data. 3.The method of claim 2, wherein, In step S2, the data analysis is specifically: S21: convert the view in the "view-analysis task" data obtained in S14 into the corresponding visual coding to obtain "view-visual coding set-analysis task" data; S22: use statistical test method to obtain the significant association relationship between each visual coding and the task; S23: screen out the visual coding with significant association relationship, classify the user operation according to the influence of the user operation on the visual coding, and retain the operations of modifying chart type, aggregation and coding.

4. The method of claim 3, wherein the analysis task is one of a plurality of analysis tasks, and the plurality of analysis tasks are associated with a plurality of rules, and the plurality of rules are associated with a plurality of deep learning models. The scheme for obtaining the significant association relationship between each visual coding and the task is specifically: S221: Statistical test hypothesis for any visual encoding construction; construct the following hypothesis: H0: Visual encoding and analysis task are independent of each other; H1: Visual encoding and analysis task are not independent of each other; S222: Construct an RxC contingency table, where n represents the sample size, f i,j represents the value of the cell corresponding to the ith row and jth column, and F Y,i represents the sum of the values of the cells in the ith row, and F X,j represents the sum of the values of the cells in the jth column, r is the number of rows, and c is the number of columns. S223: Calculate the degrees of freedom d and the theoretical frequency e of each cell ij : d = (r-1)*(c-1) If the following conditions are met at the same time, it meets the chi-square test condition: 1) The degree of freedom d is greater than or equal to 1; 2) theoretical frequency e ij The proportion of the number of cells less than 5 in the total number of cells is less than or equal to 20%. If it meets the chi-square test condition, go to step S224, otherwise terminate the calculation; S224: Calculate the statistic: Calculate X by the following equation 2 : wherein X 2 represents a statistical quantity; Further obtain the p value according to the probability density function of the chi-square distribution: where d is the degree of freedom, x represents X 2 , Γ is the gamma function, and p represents the significance. If p < 0.005, it is considered that the H0 hypothesis is not established, and the H1 hypothesis is accepted, that is, the visual encoding and analysis task have significant correlation. 5.The method of claim 2, wherein, In step S3, the model construction and training are specifically: S31: For each "operation sequence-task set sequence" data, according to the user operation category defined in S2, the data is filtered, classified, and the operation sequence is split and reorganized according to the given format, and the task set data is encoded, to obtain the input and output data of the model; S32: Design and implement an LSTM model, input is the reorganized operation sequence, output is the score of each analysis task predicted by the model; S33: Split the data into training set and validation set, and use five-fold cross-validation method to obtain the final model and score.

6. The method of claim 5, wherein the analysis task is one of a plurality of analysis tasks, and the plurality of analysis tasks are associated with a plurality of rules, and the plurality of rules are associated with a plurality of deep learning models. The specific process of obtaining the input and output data of the model is as follows: S311: Use the sliding window method, considering the user's continuous behavior of creating charts, according to the average length l of the user's operation sequence on each data set and the average number m of creating charts obtained in the workshop, set the window to be slightly larger than l / m, denoted as w; S312: extract the w operations closest to the current operation as part of the model input, denoted as F raw ; S313: According to the proportion of the operation sequence data obtained in the workshop according to the modification of the chart type, the aggregation, and the encoding operation, for each type of operation, a number of operations of the type before the current operation and closest to the current operation are extracted in the operation sequence, and are sequentially composed into a sequence, denoted as F gi , and the number of extracted operations is positively correlated with the proportion of the type of operation. S314: Obtain the sequence F of each type of operation gi Carry out sequential splicing to obtain another part of the model input, denoted as F grouped ; S315: F raw and F grouped perform splicing to obtain input data of the model; S316: Use one-hot encoding to encode and combine the analysis task set of the view corresponding to the current sliding window position, to obtain the output data of the model.

7. The analytical task perception method integrating deep learning models and rules according to claim 5, characterized in that, The model training and verification are specifically as follows: S331: Randomly shuffle the data and split it into training set and test set according to the proportion; S332: Divide the training set into f parts, and perform f times of training and verification to obtain f model parameters and their corresponding scores, wherein the kth training uses the kth, (k+1)th, (k+2)th, (k+3)th, and (k+4)th parts of the data as the training set, and the (k+4)th part of the data as the validation set; S333: Select the model parameters with the highest score in the f times of training as the final model parameters, and test them on the test set to obtain the final score of the model. 8.The method of claim 1, wherein, In step S4, the task prediction result fusion is specifically as follows: S41: Obtain the prediction score of the analysis task using the model in S3 according to the user operation sequence; S42: Use statistical methods to determine whether there is a preference for column selection in the user operation process. If there is a column preference, obtain the prediction result of the analysis task according to the attribute of the column selected by the current user and the corresponding rule of the analysis task, and fuse the model prediction score and the prediction result according to the rule as the final prediction result, otherwise, use the model prediction score as the final prediction result.

9. The analytical task perception method integrating deep learning models and rules according to claim 8, characterized in that, The statistical method is used to determine whether there is a preference for column selection in the user operation process, which is specifically as follows: S421: Construct a statistical test hypothesis: construct the following hypothesis: H3: The user has no preference for column selection; H4: The user has preference for column selection; S422: count the number of times each column is selected during the user's operation, respectively n i and the total number of selections, denoted as N; if there is no column preference, it is considered that the selection probability of each column is equal, and the selection probability is denoted as P; S423: If the following conditions are met, the sample size test is passed: 1) N > 40; that is, the total number of column selections by the user is not less than 40 times; 2) min(n i ) ≥ 5; that is, the selection frequency of each column is not less than 5 times; 3) N*P > 5; that is, the theoretical frequency of each column is not less than 5; If the sample size test is passed, go to step S424, otherwise it is considered that the user has no preference for column selection; S424: Relevance test: Calculate the statistic X by the following formula 2 : According to the probability density function of the chi-square distribution, the p value representing significance is further obtained. If p < 0.005, it is considered that the H3 hypothesis is not established, and the H4 hypothesis is accepted, that is, the user has preference for column selection.

Images