JP7049211B2 - Data analyzer and data analysis method - Google Patents

Description

The present invention relates to a data analyzer and a data analysis method.

Conventionally, so-called data mining, which attempts to obtain useful knowledge that has been unknown until now from a large amount of information, has been known. In a general data mining method, each step of a preprocessing step, a feature extraction step, a model learning step, and a postprocessing step is executed in order. In the preprocessing step, data necessary for analysis is collected, and in order to enhance the effect of analysis, unnecessary noise is removed from the data and missing items are filled. This work is called ETL: Extract / Transport / Load. In the feature extraction step, the features existing in the original data are processed to generate new features that are more useful for analysis. In the model learning step, the data prepared in the preprocessing step and the feature extraction step are input to the machine learning algorithm, and the analysis result is obtained. In the post-processing step, the analysis result obtained in the model learning step is output to an external device, and a report is created based on the analysis result.

By the way, the design of the optimum feature amount requires trial and error by a skilled analyst, which causes a long time for data analysis. On the other hand, for example, as disclosed in Non-Patent Document 1, a technique for automatically generating a new feature amount from data is known. In the techniques disclosed in these documents, a series of predefined operators (+,-, ×, ÷, etc.) are comprehensively applied to the original features to create a large number of new features. It can be generated automatically.

"A Randomized Exhaustive Propositionalization Approach for Molecule Classification" May 26, 2010

However, in the technique of the above literature, the number of new features obtained as a result of the combination of a plurality of operators is enormous. Since the huge number of features includes a large amount of features that are not effective for analysis, there is a problem that the subsequent model learning step takes time.

In addition, even if a new feature quantity that helps improve the accuracy of analysis is generated, the relationship between the generated feature quantity and the objective variable and the relationship with other feature quantities cannot be intuitively understood, so that it is actually In order to obtain knowledge from the analysis results, additional analysis work was required for individual features. That is, even if the techniques of the above documents are used, the problem that data analysis takes time is still not solved because manual work by an analyst is required in order to obtain useful findings.

The present invention has been made in view of this point, and an object of the present invention is to obtain a feature amount from a large amount of data, enable display of a feature amount of high importance, and help to obtain new knowledge. Is to be able to do.

In order to achieve the above object, the first invention comprises a data input unit for receiving input of analysis target data including a plurality of feature quantities and objective variables in a data analysis apparatus for analyzing analysis target data, and the plurality of data input units. A prediction model generation unit that generates a prediction model for predicting the objective variable from the feature amount, an importance calculation unit that calculates the importance of the prediction by the prediction model for each of the plurality of feature amounts, and the importance. Based on the importance calculated by the degree calculation unit, the display unit that displays the feature amount with higher importance and the feature selected from the feature amounts displayed on the display unit according to the input of the user. The basic diagram generator that divides the analysis target data into a plurality of clusters based on the amount value and generates a basic diagram showing the relationship between each cluster and the representative value of the objective variable of each cluster, and the basic diagram. A cluster selection unit that accepts the selection of any one of the indicated clusters, a feature amount that contributes to the prediction of the objective variable in the cluster selected by the cluster selection unit, and all clusters or the clusters. In a non-selected cluster not selected by the selection unit, the display unit includes a comparison diagram generation unit that generates a comparison diagram showing a comparison with a feature amount that contributes to the prediction of the objective variable, and the display unit generates the basic diagram. It is characterized in that the basic diagram generated by the unit and the comparison diagram generated by the comparison diagram generation unit can be displayed.

According to this configuration, when the analysis target data is input, a prediction model for predicting the objective variable from a plurality of features included in the analysis target data is generated. The feature amount may be one included in the analysis target data or a newly generated one. Then, the importance to the prediction by the prediction model is calculated for each of the plurality of feature quantities, and among the calculated plurality of feature quantities, the feature quantity having the higher importance is displayed on the display unit. At this time, a feature amount having a lower importance may be displayed. In any case, the user can grasp the feature amount of high importance separately from the feature amount of low importance by looking at the display unit. When the user selects a desired feature amount from the feature amounts displayed on the display unit, the analysis target data is divided into a plurality of clusters based on the value of the selected feature amount.

In addition, a basic diagram showing the relationship between each cluster and the representative value of the objective variable of each cluster is generated, and when the selection of any one cluster is accepted from the plurality of clusters shown in this basic diagram, it is selected. A comparison diagram showing a comparison between the feature quantities that contribute to the prediction of the objective variable in the clusters and the feature quantities that contribute to the prediction of the objective variables in all clusters or non-selected class stars is generated. Then, the basic diagram and the comparison diagram are displayed on the display unit.

By looking at the basic diagram, the user can know the relationship with the representative value of the objective variable for each of the clusters formed by dividing the data to be analyzed, so it is possible to obtain new knowledge based on this. can.

In addition, when there is a cluster that the user wants to pay attention to, when the cluster is selected, not only the feature amount that contributes to the prediction of the objective variable in the cluster but also the feature amount that contributes to the prediction of the objective variable in all the clusters. Since it is possible to make a comparison and a comparison with a feature amount that contributes to the prediction of the objective variable in a non-selected class star, new findings can be obtained based on this.

The basic diagram and the comparison diagram may be displayed on the display unit at the same time, or may be displayed at different timings. The "figure" also includes graphs, tables, and characters.

The second invention is characterized in that the display unit is configured to display a plurality of feature quantities having higher importance side by side in descending order of importance.

According to this configuration, a plurality of features having higher importance are displayed on the display unit in a ranking format, so that the user can make a relative comparison of the importance of the plurality of features. The direction in which the importance is arranged in descending order may be, for example, a vertical direction or a horizontal direction when viewed from the user.

A third aspect of the invention is characterized in that the display unit is configured to simultaneously display a feature amount having the highest importance and a plurality of feature amounts less than the highest importance.

According to this configuration, it is possible to easily grasp which feature amount has the highest importance and the difference between the highest importance and the next highest importance.

A fourth aspect of the invention is characterized in that the comparison diagram generation unit is configured to generate the comparison diagram showing a feature amount having a significantly high contribution only in the cluster selected by the cluster selection unit. do.

According to this configuration, features that are highly relevant to the selected cluster are automatically shown, so that new findings can be obtained based on a plurality of features.

In the fifth aspect of the invention, the comparison diagram generation unit is configured to generate the comparison diagram showing the feature quantities not selected by the user among the feature quantities having higher importance. It is a feature.

According to this configuration, although it was not selected by the user, it is considered that the features having higher importance have a large influence on the objective variable, so the features having higher importance are also used as a comparison diagram. By showing, new findings can be obtained based on a plurality of features.

In the sixth invention, the comparison diagram generation unit is configured to generate the comparison diagram showing the feature amount that contributes to the prediction of the objective variable in the defective cluster in which the value of the analysis target data does not exist. It is characterized by being.

According to this configuration, when the analysis target data group for which no value exists is used as a missing cluster, this missing cluster may also contribute to the prediction of the objective variable, and the comparison diagram shows the prediction of the objective variable in the missing cluster. New findings can be obtained by showing the amount of features that contribute.

The seventh invention is a data analysis method for analyzing analysis target data, in which a data input step for accepting input of analysis target data including a plurality of feature quantities and objective variables and a prediction of the objective variable from the plurality of feature quantities. The importance calculated by the prediction model generation step for generating the prediction model, the importance calculation step for calculating the importance of the prediction by the prediction model for each of the plurality of features, and the importance calculation step. Based on the degree, the value of the feature amount selected according to the input of the user from the feature amount display step for displaying the feature amount having higher importance and the feature amount displayed in the feature amount display step. Based on this, the analysis target data is divided into a plurality of clusters, and a basic diagram generation step for generating a basic diagram showing the relationship between each cluster and the representative value of the objective variable of each cluster, and the plurality shown in the basic diagram. A cluster selection step that accepts the selection of any one of the clusters, a feature amount that contributes to the prediction of the objective variable in the cluster selected in the cluster selection step, and selection by all clusters or the cluster selection step. In the unselected clusters that have not been selected, the comparison diagram generation step that generates a comparison diagram showing the comparison with the feature amount that contributes to the prediction of the objective variable, the basic diagram generated in the basic diagram generation step, and the comparison. It is a data analysis method characterized by including a comparison diagram display step capable of displaying a comparison diagram generated in the diagram generation step.

According to the present invention, a feature amount of high importance contained in a large amount of data can be displayed on the display unit. Then, the relationship between the representative value of the objective variable and the representative value of the objective variable can be known for each of the clusters formed by dividing the data to be analyzed, and the features that contribute to the prediction of the objective variable in the selected cluster and all clusters or non-selected clusters. Since it is possible to compare with the features that contribute to the prediction of the objective variable in the cluster of, it is possible to help to obtain new findings.

It is a figure which shows the schematic structure of the data analysis apparatus which concerns on embodiment of this invention. It is a block diagram of a data analyzer. It is a flowchart which shows the procedure of data analysis. It is a flowchart which shows the operation of a data analyzer. It is a figure which shows the user interface for data input. It is a figure which shows the analysis target data stored in each of two files in a table format. It is a figure which shows the data flow display user interface. It is a figure which shows the data flow display user interface which preview-displayed the analysis target data. It is a figure which shows the data processing user interface. It is a figure which shows the data flow display user interface which displayed the processed data as an icon. It is a figure which shows the feature quantity automatic generation user interface. It is a figure explaining the calculation method of the ratio which has a correspondence relationship between two analysis target data. It is a figure explaining the case which the sum of the price columns is calculated and added to "master". It is a figure explaining the case which determines the row to be aggregated based on the combination with the row representing a date. It is a figure which shows the feature amount display user interface. It is a figure which shows the feature quantity display user interface when "the total number of visits of the last 30 days" is selected. It is a figure which shows the feature amount display user interface when "Compare average value of values" is selected from a detail menu. It is a figure which shows the feature amount display user interface which the comparison table is displayed. It is a figure which shows the feature quantity display user interface which displayed the prediction accuracy comparison graph. It is a figure which shows the feature quantity display user interface when "the number of days elapsed from the last visit date" is selected. It is a figure which shows the feature amount display user interface which the influence degree comparison graph is displayed. It is a figure which shows the feature amount display user interface when "comparison of the contribution degree of a feature amount" is selected from a detail menu. It is a figure which shows the feature amount display user interface which the contribution display graph is displayed. It is a figure explaining the calculation method of the degree of contribution. It is a figure which shows the feature amount display user interface which is displayed when the feature amount "gender" is selected. It is a figure which shows the feature amount display user interface which the explanation graph of the difference is displayed. It is a figure which shows the feature amount display user interface which displayed the graph about the change of the contribution degree of a feature amount.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. It should be noted that the following description of the preferred embodiment is essentially merely an example and is not intended to limit the present invention, its application or its use.

FIG. 1 is a diagram showing a schematic configuration of a data analysis device 1 according to an embodiment of the present invention, and FIG. 2 is a block diagram of the data analysis device 1. The data analysis device 1 is a device that analyzes the data to be analyzed, and by using the data analysis device 1, the data analysis method according to the present invention can be executed. The data analysis device 1 and the data analysis method can be used, for example, when performing so-called data mining, which attempts to obtain useful findings previously unknown from a large amount of information.
(Data analysis procedure)
A general data analysis procedure will be described with reference to the flowchart shown in FIG. After the start, the preprocessing step is executed in step SA1. In the preprocessing step, data necessary for analysis, that is, data to be analyzed is collected, and in order to enhance the effect of analysis, unnecessary noise is removed from the data and missing items are filled. This work is called ETL: Extract / Transport / Load. In step SA2, a feature extraction step is executed. In the feature extraction step, the features existing in the original data are processed to generate new features that are more useful for analysis. In step SA3, a model learning step is executed. In the model learning step, the data prepared in the preprocessing step and the feature extraction step are input to the machine learning algorithm, and the analysis result is obtained. In step SA4, a post-processing step is executed. In the post-processing step, the analysis result obtained in the model learning step is output to an external device, and a report is created based on the analysis result.

(Overall configuration of data analyzer 1)
As shown in FIGS. 1 and 2, the data analysis apparatus 1 includes an apparatus main body 2, a monitor 3, a keyboard 4, and a mouse 5, and the monitor 3, the keyboard 4, and the mouse 5 include the apparatus main body 2. It is connected to the. For example, the data analysis device 1 can be set by installing a program that executes the control contents described later on a general-purpose personal computer, and the data analysis device 1 is configured by the dedicated hardware in which the program is installed. You can also do it. The device main body 2 and the monitor 3 may be integrated, or a part of the device main body 2 may be built in the monitor 3.

The data analysis device 1 has a built-in communication module (not shown) and is configured to enable communication with the outside. This makes it possible to download data from an external server via the Internet line.

Further, the keyboard 4 and the mouse 5 are an operation means for operating the data analysis device 1, an input means for inputting various information, a selection means for performing a selection operation, and the like. In addition to the keyboard 4 and the mouse 5, or in place of the keyboard 4 and the mouse 5, a touch panel type input device, a voice input device, a pen type input device, and the like can also be used.

(Configuration of monitor 3)
The monitor 3 shown in FIG. 1 is composed of, for example, an organic EL display, a liquid crystal display, or the like, and can be referred to as a display unit by itself, or the monitor 3 and the display control unit 3a shown in FIG. 2 are combined. It can also be called a display unit. The display control unit 3a may be built in the monitor 3 or may be built in the device main body 10. The display control unit 3a is composed of a display DSP or the like for displaying an image on the monitor 3. The display control unit 3a may include a video memory such as a VRAM that temporarily stores image data when displaying an image. The display control unit 3a transmits a control signal for displaying a predetermined image to the monitor 3 based on a display command (display command) sent from the CPU 11a of the main control unit 11 described later. For example, various user interfaces and icons, and control signals for displaying the user's operation contents using the keyboard 4 and the mouse 5 on the monitor 3 are also transmitted. Further, a pointer or the like that can be operated by the mouse 5 can also be displayed on the monitor 3.

By using the monitor 3 as a touch operation panel type monitor, the monitor 3 can be provided with various information input functions.

(Overall configuration of device body 2)
The apparatus main body 2 shown in FIG. 1 includes a control unit 10 and a storage unit 30. The storage unit 30 is composed of a hard disk drive, a solid state drive (SSD), or the like. The storage unit 30 is connected to the control unit 10 and is controlled by the control unit 10 so that various data can be stored and the stored data can be read out.

(Control unit 10)
Although not specifically shown, the control unit 10 can be configured by an MPU, a system LSI, a DSP, dedicated hardware, or the like. The control unit 10 is equipped with various functions as described later, and these may be realized by a logic circuit or may be realized by executing software.

As shown in FIG. 2, the control unit 10 includes a main control unit 11, a data input unit 12, a prediction model generation unit 13, an importance calculation unit 14, a cluster selection unit 15, and a display diagram generation unit 16. , The explanation text automatic generation unit 18 is provided. Although each part of the control unit 10 is described separately as described above, the same part may be configured to execute a plurality of types of processing, or may be further subdivided and linked to each other for one processing. May be configured to run.

Each of the above hardware is connected so as to be capable of two-way communication or one-way communication via an electrical communication path (wiring) such as a bus.

The main control unit 11 performs numerical calculation and information processing based on various programs, and also controls each hardware unit. The main control unit 11 stores a CPU 11a that functions as a central arithmetic processing unit, a work memory 11b such as a RAM that functions as a work area when the main control unit 11 executes various programs, a startup program, an initialization program, and the like. It is provided with a program memory 11c such as a ROM, a flash ROM, or an EEPROM.

The data input unit 12 is a part that accepts input of data to be analyzed including a plurality of feature quantities and objective variables. The data input user interface 50 shown in FIG. 5 is displayed on the monitor 3, and the data input user interface is displayed. Accepts user operations made on 50. User operations include keyboard 4 operations and mouse 5 operations (including button clicks, drag and drop, wheel rotation, and the like).

Here, the analysis target data is data including a plurality of feature quantities and objective variables, but the user is aware that the plurality of feature quantities originally exist in the analysis target data (existing feature quantities). Any data may be used as long as it does not include data implicitly contained in the data to be analyzed (potential feature amount), newly generated data (new feature amount), or any one or more of them. A part of the data to be analyzed may be missing. In that case, the work of filling in the missing items may be performed as described later.

The data input user interface 50 shown in FIG. 5 is provided with a file selection button 50a, a database selection button 50b, a URL designation button 50c, a database display area 50d, and a read start button 50e. The arrangement of each button can be set freely.

For example, if a file storing analysis target data is stored in an external storage device or storage unit 30 and is on the desktop or in an open folder, the user drags the file to the database display area 50d. & Drop operation. As a result, the file name in which the analysis target data is stored is displayed in the database display area 50d. After that, when the read start button 50e is pressed, the file displayed in the database display area 50d is read into a predetermined area of the storage unit 30 and saved.

When the data to be analyzed is on the database, the user presses the database selection button 50b. When the database selection button 50b is pressed, a setting screen (not shown) for connecting to the database is displayed, and the user is urged to enter the table name and, if necessary, the password. After that, when the read start button 50e is pressed, the analysis target data is read and saved in the predetermined area of the storage unit 30 in a predetermined file format, and the file name storing the analysis target data is displayed in the database display area 50d. File.

When the data to be analyzed is on the Internet or a server, the user presses the URL designation button 50c. When the URL designation button 50c is pressed, a URL input screen (not shown) is displayed, prompting the user to input the URL. After that, when the read start button 50e is pressed, the analysis target data is downloaded, read into a predetermined area of the storage unit 30 in a predetermined file format, and saved, and the file name in which the analysis target data is stored is the database display area 50d. Is displayed in.

There may be only one file or a plurality of files storing the data to be analyzed. FIG. 5 shows a case where the analysis target data stored in each of the two files "transaction" and "master" is read. The file may be read by a method other than the above three methods. The above file format is a csv format, but other formats may be used.

FIG. 6 shows the analysis target data stored in the “master” and the “transaction”, respectively. In this example, "master" is customer information, and the age and gender are recorded in association with the customer ID, and whether or not the customer is separated from the customer in association with the customer ID is "separation". Recorded in a column. "Release" means that you visited the store once, but then stopped visiting. A customer recorded as "1" in the "separation" column is a estranged customer, and a customer recorded as "0" is a non-separated customer.

In the "transaction", each purchase history is recorded, and in the "customer ID" column, the correspondence relationship between "master" and each row is associated. In the "transaction", the purchase date, category and price are recorded in association with the customer ID. The data described above are examples prepared for the purpose of explaining the present invention, and do not limit the scope of application of the present invention. INDUSTRIAL APPLICABILITY In addition to the purchase history data, the present invention can analyze various analysis target data.

When the reading of the data to be analyzed is completed, the data flow display user interface 51 shown in FIG. 7 is displayed on the monitor 3, and the user's operation performed on the data flow display user interface 51 is received. The data flow display user interface 51 is provided with a data flow display area 51a, a lower display area 51b, a feature amount automatic generation button 51c, a data processing button 51d, and a data output button 51e.

The read analysis target data is displayed as an icon in the data flow display area 51a. In the lower display area 51b, a relationship diagram (generally called an ER diagram) between the read analysis target data is displayed. The data input unit 12 shown in FIG. 2 detects whether or not a common column exists in the plurality of read data to be analyzed. Whether or not there is a common column can be determined based on, for example, the name of the column. In this example, as shown in FIG. 6, since a column named "customer ID" exists in both "master" and "transaction", the data input unit 12 detects this. Then, the data input unit 12 determines that the "customer ID" of the "master" and the "customer ID" of the "transaction" are related, and is displayed in the lower display area 51b as shown in FIG. In the relationship diagram, the "customer ID" of "master" and the "customer ID" of "transaction" are connected by a line and displayed.

The above processing is a processing for accepting input of analysis target data, and is performed by the data input unit 12. The data input unit 12 may include an input device such as a keyboard 4 and a mouse 5. Further, the above processing corresponds to the data input step of step SB1 in the flowchart shown in FIG.

In step SB2 following step SB1, the analysis target data input in the data input step is processed. First, from the analysis target data displayed in the data flow display area 51a of the data flow display user interface 51 shown in FIG. 7, the data to be processed is selected. The data selection operation may be, for example, an operation of clicking an icon corresponding to the data to be processed. When the icon is clicked, the relationship diagram of the lower display area 51b is deleted, and the selected analysis target data is displayed in the lower display area 51b as shown in FIG. At this time, only a part of the selected analysis target data can be preview-displayed in the lower display area 51b.

When the "transaction" is previewed and viewed, the words "daily delivery" and "daily delivery A" are mixed in the "category" column. Since the values of "daily delivery product" and "daily delivery product A" are different from each other, they are treated as different categories at the time of model learning described later, but if they are to be treated as the same in data analysis, these are treated. Higher accuracy can be obtained by unifying the notational fluctuations of. This is called data processing.

When performing data processing, the data processing button 51d of the data flow display user interface 51 shown in FIG. 8 is pressed. When the data processing button 51d is pressed, the data processing user interface 52 shown in FIG. 9 is displayed on the monitor 3, and the user's operation performed on the data processing user interface 52 is received. The data processing user interface 52 includes a processing target data display area 52a, a missing value removal / fill-in-the-blank button 52b, a data replacement button 52c, a column deletion button 52d, a column addition button 52e, a formula input button 52f, and a column. An information display button 52g and an end button 52h are provided. The selected analysis target data is displayed in a table format in the processing target data display area 52a. You can also scroll if the displayed table is large.

The missing value removal / fill-in-the-blank button 52b is a button operated when removing the missing value of the analysis target data and filling in the hole in the analysis target data. By operating the missing value removal / fill-in-the-blank button 52b, any missing value can be removed or a missing item can be added. The data replacement button 52c is a button operated when replacing an existing value with another value. For example, by selecting the cell of "Daily delivery A" in the "Category" column, pressing the data replacement button 52c, and entering "Daily delivery" as the replacement character (value), "Daily delivery" in the same column The value "A" is automatically replaced with the value "daily delivery". The column deletion button 52d is a button operated when deleting an arbitrary column. After selecting an arbitrary column, the selected column is deleted by pressing the column delete button 52d. The column addition button 52e is a button operated when adding an arbitrary column. After selecting any column in the table, pressing the column addition button 52e adds a column to the front or back of the selected column. The formula input button 52f is a button for inputting a formula in a cell. The value is calculated by the formula input by the operation of the formula input button 52f. The column information display button 52g is a button for adding predetermined information to an arbitrary column.

The data to be analyzed can be processed by operating the missing value removal / fill-in-the-blank button 52b, the data replacement button 52c, the column deletion button 52d, the column addition button 52e, the formula input button 52f, and the column information display button 52g. Therefore, the data analyzer 1 can perform ETL processing, and by performing ETL processing, the data to be analyzed can be transformed into a form more suitable for analysis. As a result, the accuracy of the analysis result can be further improved. When the series of operations is completed, the end button 52h of the data processing user interface 52 shown in FIG. 9 is pressed.

When the end button 52h is pressed, the data flow display user interface 51 is displayed again, as shown in FIG. A new icon "transaction-1" is displayed in the data flow display area 51a of the data flow display user interface 51. This "transaction-1" is the analysis target data after processing the "transaction", and the "transaction-1" and the "transaction" are connected by a relation line, and the relationship is graphically displayed. .. Therefore, the user can intuitively understand that "transaction-1" is the processed data of "transaction" and that "transaction-1" is the processed data. The above series of ETL processing is the data processing processing in step SB2 of the flowchart shown in FIG.

When the series of ETL processing is completed, the process proceeds to the step of inputting the setting related to the feature amount generation in step SB3 of the flowchart shown in FIG. When the feature amount automatic generation button 51c of the data flow display user interface 51 shown in FIG. 10 is pressed, the feature amount automatic generation user interface 53 shown in FIG. 11 is displayed on the monitor 3 and is performed on the feature amount automatic generation user interface 53. Accepts user operations.

The feature amount automatic generation user interface 53 is provided with a data flow display area 53a, a lower display area 53b, a relationship designation area 53c, an objective variable designation area 53d, a detailed setting area 53e, and an execution button 53f. ing. In the data flow display area 53a, for example, the data flow is displayed in the same manner as the data flow display area 51a of the data flow display user interface 51 shown in FIG. Similar to the lower display area 51b of the data flow display user interface 51 shown in FIG. 7, the lower display area 53b displays a relationship diagram between the data to be analyzed.

The relationship designation area 53c shown in FIG. 11 is an area for designating a correspondence relationship between a plurality of analysis target data. Here, the correspondence between "transaction-1" and "master" is specified. As an example, as described above, since the "customer ID" of "transaction-1" and the "customer ID" of "master" are in a corresponding relationship, the "customer ID" and "master" of "transaction-1" Although "customer ID" is specified as a correspondence relationship, it is not limited to this, and any column of "transaction-1" and any column of "master" can be specified as having a correspondence relationship. , The names do not have to match.

The relationship may be specified by selecting it from the list of features, by dragging it on the relationship diagram to draw a line, or by directly entering the name. It may be specified by. If there is only one data, the correspondence between the data to be analyzed is not specified.

At the bottom of the relationship designation area 53c, there is a numerical value displayed as a percentage. This number indicates what percentage of the total rows correspond between the two analyzed data according to the specified column. For example, as shown in FIG. 12, one "1", "2", "3", and "4" exist in the "ID" column of the first analysis target data, and the second analysis target. Assuming that there are three "1" s, one "3", and one "5" in the "ID" column of the data, "1" of the first analysis target data is the second analysis target data. Corresponding to "1", "3" of the first analysis target data corresponds to "3" of the second analysis target data, and other IDs do not correspond. In this case, 50% (2/4 rows) of the first analysis target data and 80% (4/5 rows) of the second analysis target data correspond to the "ID" column, which is 6 as a whole. / 9 lines, that is, about 67% corresponds. Such a calculation is performed, and the calculation result is displayed at the lower part of the relationship designation area 53c.

The objective variable designation area 53d is an area for designating the objective variable to be analyzed. By specifying the objective variable, the table to be aggregated can be determined. Here, in order to analyze whether or not the customer is separated, the "separation" column of "master" is specified as the objective variable. Therefore, the features of "transaction-1" can be aggregated and added as new features of "master". The objective variable may be specified by selecting it from the list of features, by dragging it from the list, or by directly entering the name. May be good. Further, when the objective variable is specified, the data analysis device 1 internally sets a flag indicating that the objective variable has been specified. By setting this flag, the features described later are not generated from the objective variable itself.

The above series of processes is the process of inputting the settings related to the feature amount generation in step SB3 of the flowchart shown in FIG.

In the following step SB4, a process of generating a feature amount is executed. This process is started by pressing the execution button 53f of the feature amount automatic generation user interface 53.

That is, as a method of generating a feature amount when a plurality of analysis target data are specified in steps SB1 and SB3 and a plurality of analysis target data relationships are specified in step SB3, for example, AJ Knobbe, "Propositionalisation and Aggregates", The techniques described in 2001 can be used. The example shown in FIG. 13 is a case where the total is calculated for the "price" column of "transaction" and added as the feature amount of "master". The purchase history in which the "customer ID" of the "transaction" is "1" is extracted, and the values of the "price" are summed to obtain the total purchase price. In "master", a column of "total purchase amount" is generated, and a total value is automatically input in a column of "total purchase amount" in which "customer ID" is "1". This is possible because the row correspondence between "master" and "transaction" is linked by the "customer ID" column.

The row to be aggregated may be determined based on, for example, a combination with a row representing a date. For example, as shown in FIG. 14, the purchase history in which the "customer ID" of "transaction" is "1" is extracted, and the values of "price" within the last 10 days starting from February 1, 2018 are totaled. .. In "master", a column of "purchase amount within the last 10 days" is generated, and in the column of "purchase amount within the last 10 days" with "customer ID" of "1", "price within the last 10 days". The total value of "" is automatically entered. Similar processing can be performed for all customer IDs.

When the data to be analyzed is time-series data, the aggregation period can be switched to, for example, 10 days, 20 days, and 30 days in this way. By switching the aggregation period, various types of features can be generated even with the same aggregation function. The unit of the period may be seconds, minutes, hours, months, or years. For example, it is possible to acquire the total aggregation period of the data to be analyzed and automatically divide the acquired total aggregation period into a plurality of aggregation periods. When the total aggregation period of the data to be analyzed is 30 days, it can be divided into 0 to 10 days, 11 to 20 days, and 21 to 30 days. The division of the aggregation period can be configured to be performed when the derivation diagram generation unit 16b, which will be described later, generates the derivation diagram.

After the features are generated, the importance of the objective variable for prediction is evaluated for each feature. The prediction of the objective variable is executed by the prediction model generation unit 13 shown in FIG. This is the predictive model generation step. The prediction model generation unit 13 generates a prediction model for predicting the objective variable from a plurality of feature quantities. Predictive models can be created, for example, by multiple regression analysis or decision trees. There are various generation methods, but all of them can use conventionally known methods.

After generating the prediction model, the importance of the prediction by the prediction model is calculated for each of the plurality of features. This is a process executed by the importance calculation unit 14 shown in FIG. 2, and is an importance calculation step. The importance of the feature amount is known by the name of Feature Importance, and various calculation methods have been proposed, and any calculation method may be used. For example, in the linear multiple regression model represented by the following equation (1), the coefficients a1, a2, ... For each feature amount x1, x2, ... may be obtained and the absolute value thereof may be used as the value of importance.

y = a1x1 + a2x2 + ... + anxn ... (1)
Alternatively, instead of using the coefficient as it is, a normalization coefficient corrected so that the sizes of the feature quantities x1, x2, ... Are uniform may be used for the importance.

After performing such an analysis of importance, the control unit 10 extracts the one having the higher importance. At this time, for the feature quantities having different aggregation periods, even if a plurality of feature quantities are in the higher order of importance, only the most important one is left and the others are deleted. For example, if the "purchase amount of the last 30 days", "purchase amount of the last 20 days", and "purchase amount of the last 10 days" are in the top of the feature amount at the same time, if they are presented to the user as they are, It is difficult for the user to understand because the features having almost the same meaning are displayed in duplicate. Further, if a feature amount having the same meaning occupies the upper part, other effective feature amounts will be expelled from the upper part, and it will be difficult to obtain effective knowledge.

The number in the "higher rank" can be, for example, 2 or more, 5 or more, 10 or more, and 20 or more. The number to be included in the "upper" may be automatically set by the control unit 10 or may be set by the user to any number.

The above series of processes is the feature amount generation process of step SB4 in the flowchart shown in FIG. When the generation of the feature amount is completed, the process proceeds to step SB5. Step SB5 is a step of outputting information on the feature amount generated in step SB4.

When the generation of the feature amount is completed and the upper feature amount is extracted, the feature amount display user interface 60 shown in FIG. 15 is displayed on the monitor 3, and the user's operation performed on the feature amount display user interface 60 is performed. Accept.

The feature amount display user interface 60 is provided with a feature amount display area 60a and a detailed information display area 60b. The feature amount display area 60a and the detailed information display area 60b can be arranged so as to be arranged on the left and right sides of the monitor 3. The upper feature amount is displayed in the feature amount display area 60a. That is, the monitor 5 displays the feature amount having a higher importance based on the importance calculated by the importance calculation unit 14. This step is a feature quantity display step.

The monitor 5 is configured to display a plurality of features having higher importance side by side in descending order of importance, that is, to display a ranking. In this example, the feature quantity having the highest importance and the feature quantity having the highest importance are displayed at the same time, and the feature quantity having the highest importance is the highest. Multiple features that are located in and less than the highest importance are arranged in descending order according to the magnitude of importance. The feature quantities may be arranged in ascending order or may be arranged so as to be arranged in the left-right direction. The feature amount to be displayed in the feature amount display area 60a may be a predetermined number, and the number is not particularly limited.

The feature amount display area 60a is provided with a name display unit 60c for displaying the name of the feature amount and an importance display unit 60d for displaying the importance. The name of the feature amount existing in "transaction" and "master" is displayed on the name display unit 60c. The name of the feature amount existing in "transaction" and "master" can be displayed as it is on the name display unit 60c, but it may be changed by the user.

On the importance display unit 60d, the importance is displayed so as to be arranged next to each feature amount in a bar graph format. When displaying in a bar graph format, the longer it is, the more important it can be. In addition to the various graph formats, the weight degree may be displayed numerically on the importance display unit 60d, or the numerical value and the graph may be displayed at the same time.

The feature amount display area 60a is provided with a feature amount addition button 60f. When the feature amount addition button 60f is pressed, a feature amount other than that displayed in the feature amount display area 60a is displayed, and the user selects a feature amount to be displayed in the feature amount display area 60a from the displayed feature amounts. You can do it. The selected feature amount is displayed in the feature amount display area 60a. It is also possible to delete or hide the feature amount displayed in the feature amount display area 60a, which is considered unnecessary.

(Detailed operation of control unit 10)
The feature amount displayed in the feature amount display area 60a of the feature amount display user interface 60 shown in FIG. 15 can be selected by the user. When the user performs an input operation as to which feature amount is to be selected by using the mouse 5 or the like, as shown in FIG. 16, the first basic figure 70 is displayed in the detailed information display area 60b of the feature amount display user interface 60. Is displayed. The first basic diagram 70 is generated by the basic diagram generation unit 16a of the display diagram generation unit 16 shown in FIG. This process is the basic diagram generation step. That is, the basic diagram generation unit 16a is a portion that generates a basic diagram showing the relationship between the first feature amount selected according to the user's input from the feature amounts displayed on the monitor 3 and the objective variable. Is. The first feature amount is the feature amount selected by the user, and in the example shown in FIG. 16, it is the “total number of days of visit to the store in the last 30 days”. The feature amount selected by the user is highlighted so that it can be easily distinguished from other feature amounts. Further, in FIG. 70, the analysis target data is divided into a plurality of clusters based on the value of the feature amount selected according to the input of the user, and each cluster and the representative of the objective variable of each cluster are represented. It is a figure which shows the relationship with a value.

Hereinafter, the first basic diagram 70 will be described in detail. The first basic figure 70 can be divided into three regions, an upper region 71, an intermediate region 72, and a lower region 73. In the upper region 71, a line graph is displayed in which the vertical axis is the ratio of separation and the horizontal axis is the total number of days visited in the last 30 days. The percentage of customers who have left the store is the percentage of customers who have visited the store once but have not visited the store since then. On the horizontal axis, 0 to 30 days are divided into 6 periods (sections) and displayed. The line graph shows the relationship between the selected features and the objective variable (separation).

In the intermediate region 72, a bar graph is displayed in which the vertical axis is the number of data and the horizontal axis is the total number of days visited in the last 30 days. The horizontal axis of this bar graph is shared with the horizontal axis of the line graph displayed in the upper region 71. In this bar graph, the number of data existing in each section is represented, and it is a frequency distribution table. That is, the control unit 10 divides the analysis target data into a plurality of periods based on the value of the first feature amount to form a cluster, and generates a data number display graph showing the number of analysis target data belonging to each cluster. It is configured to do. The monitor 3 is configured to be able to display the generated data number display graph.

From the line graph of the upper area 71 and the bar graph of the intermediate area 72, "The greater the total number of store visits in the last 30 days, the lower the ratio of data with separation = 1", "The total number of store visits is 21 to 25 days." In the case of, the ratio of separation = 1 is the minimum, but the number of applicable data is small. "

Further, the control unit 10 calculates an average value of the ratio of separation = 1. The average value can be the average value of all data, which is 21% in this example. Furthermore, in the section where it is determined that there is a significant difference in the mean value of the objective variable compared to the overall mean value (21%), the series is highlighted and the explanatory text is further shown at the bottom of the graph, that is, the lower region 73. Is added to.

If there is no significant difference, it will not be highlighted. "There is no significant difference" means that the representative value of the section does not have a sufficient difference from the representative value of the section to be compared, taking into consideration the variation of data and the number of data. Refers to. Further, the highlighting is a display that can be distinguished from other characters, for example, by thickening the character, changing the color of the character from the color of another character, or coloring the background of the character.

The explanatory text is a text explaining the relationship between the selected first feature amount and the objective variable, and is automatically generated by the explanatory text automatic generation unit 18 shown in FIG. The description automatic generation unit 18 is, for example, a first feature amount selected based on the name of the analysis target data received by the data input unit 12, the name of the feature amount, the relative difference of the feature amount, the number of data, and the like. Can generate statements that explain the degree of influence that has on the objective variable. This makes it easier for the user to understand the relationship between the feature amount and the objective variable, and saves the user the trouble of creating a comment.

Here, the determination of the significant difference will be described. Indicators such as p-values and t-tests commonly used in statistics can be used to determine significant differences. For example, the t-test is calculated by the following equation 1. x and s are the average and standard deviation of the target values in the series of interest, μ is the overall average of the target values, and n is the number of data in the series. It can be determined that the larger this value is, the more significant the difference is.

By using these indicators, it is possible not only to simply look at the degree of deviation from the overall average, but also to take into account the amount of data and the variation of individual data, so it means that a large deviation has occurred by chance. It is possible to distinguish whether a certain difference is occurring.

Highlighting using this statistical test allows the user to focus on truly meaningful locations from the data and quickly obtain useful insights from the automatically generated features.

When an arbitrary section is selected from the graph displayed in the upper area 71, the detailed menu 74 is displayed as shown in FIG. The detail menu 74 is a menu for comparing the data included in the selected section with the overall tendency, and more detailed information can be displayed according to the menu selection. The detailed menu 74 includes three menus, "Compare average values", "Compare contributions of features", and "Calculate explanation of difference".

As shown in FIG. 17, for example, when the menu "Compare average values" is selected for the section of "total number of visits to the store in the last 30 days to 5 days", the selection is made as shown in FIG. A comparison table 75 comparing the average value of each feature amount is displayed for each of the section data and all the data.

Specifically, when the user performs an operation (click operation, etc.) to select one of the clusters from the plurality of clusters shown in the first basic FIG. 70 shown in FIG. 16, this operation is performed by the cluster shown in FIG. It is accepted by the selection unit 15. This process is the cluster selection step. After the selection is accepted by the cluster selection unit 15, the comparison diagram generation unit 16c shown in FIG. 2 is selected by the feature quantity that contributes to the prediction of the objective variable and all clusters or the cluster selection unit 15 in the selected cluster. In the unselected clusters, a comparison table 75 (shown in FIG. 18) showing a comparison with the features that contribute to the prediction of the objective variable is generated. This process is the comparison diagram generation step.

Although the details will be described later, the comparison diagram generation unit 16c can also be configured to generate a comparison diagram showing a feature amount having a significantly high contribution only in the cluster selected by the cluster selection unit 15. Further, the comparison diagram generation unit 16c can be configured to generate a comparison diagram showing the feature quantities not selected by the user among the feature quantities having higher importance. Further, the comparison diagram generation unit 16c can be configured to generate a comparison diagram showing the feature amount that contributes to the prediction of the objective variable in the defective cluster in which the value of the analysis target data does not exist.

The comparison table 75 is a comparison diagram, and may be a comparison diagram displaying a graph in addition to the tabular format. The monitor 3 is configured to be able to display the first basic diagram 70 and the comparison table 75 simultaneously or separately, and the comparison diagram display step of displaying the first basic diagram 70 and the comparison table 75 simultaneously or separately. Can be executed.

In the comparison table 75, only the ones having a significantly large deviation in the mean value are displayed from all the feature quantities including the ones having a low importance. The p-value and t-test can be used to determine the significance here as before.

The feature amount to be displayed in the comparison table 75 can be configured to be selectable by the user. At the bottom of the comparison table 75, a feature amount addition button 75a is provided. By pressing the feature amount addition button 75a, any feature amount can be added as a comparison axis. The feature amount displayed in the comparison table 75 can be configured to be deleteable by the user.

In this example, the mean value is displayed in the comparison table 75, but statistics other than the mean value such as variance, median value, and minimum value may be displayed.

As a result, when an interesting data group (customer group with a high separation rate in this example) is found using a certain feature amount as a starting point, it is possible to delve into in detail what the unique properties of the data group are. , New findings can be obtained.

Furthermore, since the feature quantities that are the core of comparison and analysis are automatically generated, for example, "customers who have a short total number of visits in the last 30 days have a particularly small purchase price of vegetables and daily deliveries." It is easy to obtain deep knowledge that is difficult to notice unless you are an analyst.

Further, the graph displayed when the feature amount is selected may include one or more graphs other than the graph in addition to the relationship with the objective variable. For example, when "total number of days visited in the last 30 days" is specified, as shown in FIG. 19, the prediction accuracy comparison regarding "comparison of accuracy with the case of totaling in a period other than 30 days" as the second graph. Graph 76 can be displayed. The prediction accuracy comparison graph 76 is located below the first basic drawing 70 shown in FIG. 18, and therefore the vertical scroll bar 60e is operated from the state where the first basic drawing 70 is displayed in the detailed information display area 60b. By scrolling down, the prediction accuracy comparison graph 76 can be displayed. The first basic figure 70 and the prediction accuracy comparison graph 76 may be displayed at the same time. As described above, the derivative diagram display step of displaying the first basic diagram 70 and the prediction accuracy comparison graph 76 simultaneously or separately can be executed.

The prediction accuracy comparison graph 76 shown in FIG. 19 can also be referred to as a third derivative diagram. The third derivative diagram shows the accuracy of the prediction model generated based on the first feature amount (in this example, "total number of days visited in the last 30 days"), and belongs to the same category as the first feature amount, and It is a figure which shows the comparison with the accuracy of the prediction model generated based on the 1st feature quantity and another feature quantity which has a different aggregation period. "The aggregation period is different" is shown on the horizontal axis of the prediction accuracy comparison graph 76, and specifically, "10 days", "20 days", ... This third derivative diagram is generated by the derivative diagram generation unit 16b shown in FIG. This process is the derivation diagram generation step. The total number of days visited in the last 10 days, the total number of days visited in the last 20 days, and the total number of days visited in the last 30 days belong to the same category of "total number of days visited".

When the derivation diagram generation unit 16b generates the third derivation diagram, the accuracy of the prediction model predicted based on another feature amount that belongs to the same category as the first feature amount and has a different aggregation period is determined. It can be compared with the accuracy of the prediction model generated based on the feature amount of 1. This makes it possible for the user to grasp the relative difference in accuracy of each prediction model and then easily determine which prediction model should be used for analysis to improve the accuracy.

Further, the derivative diagram generation unit 16b has the accuracy of each prediction model generated based on a plurality of different feature quantities at the time of generation of the third derivative diagram, and the prediction generated based on the first feature quantity. Generate a diagram that is comparable to the accuracy of the model. Further, the derivative diagram generation unit 16b has, as another feature amount, the accuracy of the prediction model generated based on the feature amount having a predetermined degree of importance at the time of generating the third derivative figure, and the first feature amount. Generates a diagram that makes it possible to compare with the accuracy of the predictive model generated based on. Specifically, it is a graph, but it may also be a figure that can be compared numerically.

For example, for features that are aggregated by time or number of days, as described above, it is configured to generate multiple patterns of features with different aggregation periods and automatically select the most suitable one. .. In FIG. 19, by presenting the accuracy difference from the unselected aggregation period as a relative value, it is possible to show "what happened when the total number of store visits in another aggregation period was used as a feature quantity".

As the accuracy, the F value, AUC, etc. generally used in data analysis are used. The accuracy may be an absolute value, or may be an index that combines viewpoints other than accuracy, such as processing time and data amount. As a result, it is possible to give a more convincing feeling to the user who thinks "Why was the total number of visits to the store 30 days output instead of 20 days or 40 days?". In addition, it is possible to provide a judgment criterion by presenting the degree of influence of the amount of data in response to the request that "if the accuracy does not change so much, the amount of data to be aggregated should be reduced as much as possible".

FIG. 20 shows a case where “the number of days elapsed from the last visit date” is selected from the feature quantities displayed in the feature quantity display area 60a of the feature quantity display user interface 60. The second basic figure 77 is displayed in the detailed information display area 60b of the feature amount display user interface 60. The second basic diagram 77 is generated by the basic diagram generation unit 16a of the display diagram generation unit 16 shown in FIG. 2, and this process is the basic diagram generation step. That is, the basic diagram generation unit 16a generates a basic diagram showing the relationship between the first feature amount selected according to the user's input from the feature amounts displayed on the monitor 3 and the objective variable. The first feature amount is, in the example shown in FIG. 20, "the number of days elapsed since the last visit date".

The second basic figure 77 can be divided into three regions, an upper region 77a, an intermediate region 77b, and a lower region 77c. In the upper region 77a, a line graph is displayed in which the vertical axis is the ratio of separation and the horizontal axis is the number of days elapsed from the last visit date. On the horizontal axis, 0 to 30 days are divided into 6 periods (sections) and displayed.

In the intermediate region 77b, a bar graph is displayed in which the vertical axis is the number of data and the horizontal axis is the number of days elapsed since the last visit date. The horizontal axis of this bar graph is shared with the horizontal axis of the line graph displayed in the upper region 77a. Further, in the lower region 77c, the explanatory text automatically generated by the explanatory text automatic generation unit 18 is displayed.

In the example shown in FIG. 20, the relationship between the selected feature amount and the objective variable is also displayed, but the emphasized series and the content of the explanatory text described in the lower region 77c are the properties of the feature amount. Correspondingly, it is different from the example shown in FIG.

Further, as shown in FIG. 21, the second graph for explaining the feature amount in detail is an influence degree comparison graph 78 showing a complex effect with another feature amount. Such an influence degree comparison graph 78 can also be generated, and this graph 78 can be called a simultaneous display graph in which the first feature amount and the second feature amount are simultaneously displayed. The generated simultaneous display graph can be displayed on the monitor 3. The influence degree comparison graph 78 may be displayed at the same time as the second basic diagram 77, or may be displayed separately. The step of displaying the influence degree comparison graph 78 and the second basic diagram 77 is a derivative diagram display step.

The influence degree comparison graph 78 can also be called a first derivative diagram. In the first derivative diagram, the data to be analyzed is divided into a plurality of clusters based on the value of the first feature amount (in this example, "the number of days elapsed from the last visit date"), and the representative value of the objective variable of each cluster. A second feature amount having a representative value of the objective variable determined to have a significant difference from the representative value of the objective variable of the first feature amount in at least one cluster (this example). Is a diagram showing the relationship between the second feature amount and the objective variable by extracting “total number of days visited in the last 30 days”). This first derivative diagram is generated by the derivative diagram generation unit 16b shown in FIG. This process is the derivation diagram generation step.

That is, when the derivative diagram generation unit 16b shown in FIG. 2 generates the influence degree comparison graph 78 (shown in FIG. 21) which is the first derivative diagram, for example, the first feature amount divided into a plurality of periods. The mean value, median value, etc. are calculated as the representative value of the objective variable for each period, and the objective variable having a significant difference from the representative value of the objective variable of the first feature amount in an arbitrary period. A second feature amount having a representative value of can be extracted, and the relationship between the second feature amount and the objective variable can be shown. As a result, it is possible to obtain findings that could not be obtained only with the first feature amount. The condition for distributing the analysis target data can be called an interval, and the interval may be not only the period but also, for example, gender, place, and the like.

Further, the derivative diagram generation unit 16b divides the analysis target data into a plurality of clusters based on the value of the first feature amount, and generates a data number display graph 78a showing the number of analysis target data belonging to each cluster. It is configured in. When the analysis target data is divided into a plurality of clusters, the derivative diagram generation unit 16b acquires the total aggregation period of the analysis target data, and automatically divides the acquired total aggregation period into a plurality of aggregation periods. This can save the user the trouble of dividing work.

The data number display graph 78a is a bar graph displayed below the first derivative diagram (impact degree comparison graph 78). The monitor 3 is configured to be able to display the data number display graph 78a generated by the derivative diagram generation unit 16b. According to the data number display graph 78a, the number of data to be analyzed can be grasped.

Further, it can be configured to notify the user that there are clusters in which the number of data to be analyzed is the first predetermined number or less. For example, if there is a cluster with less than a few percent of the total number of data, the user can be notified that the number of data constituting the cluster is less than a few percent of the total number of data. It is a material for judging the reliability of the quantity.

When the derivation diagram generation unit 16b generates the first derivation diagram, the feature amount whose number of data to be analyzed is equal to or less than the second predetermined number is significant with respect to the representative value of the objective variable of the first feature amount. It is configured to determine that there is no difference. The second predetermined number and the first predetermined number may be the same, or one may be less than the other. For example, if there is a feature amount with only a few percent or less of the total number of data, it can be considered that the reliability of the objective variable is poor, and in such a case, it is determined that there is no significant difference. It can be prevented from being used for analysis.

The influence comparison graph 78 shows that even if the data is in the same section when viewed only with the selected features, further division of the data by different features causes a large difference in the relationship with the objective variable. ing. In the line graph displayed in the upper region 77a shown in FIG. 20, it was possible to read the fact that "the smaller the number of days elapsed from the last visit date, the lower the ratio of separation = 1". Further, from the influence degree comparison graph 78 shown, it is possible to obtain a new finding that "even if the number of days elapsed from the last visit date is small, the customer who has a total visit time of less than 10 days has a high turnover rate".

In this influence degree comparison graph 78, whether or not the difference in the section due to the difference in another feature amount is significant is tested, and only the section determined to be significant is highlighted. Here, the feature amount of "total number of days of visit to the store in the last 30 days" is combined with the selected feature amount, but when the graph of the combination is displayed for all the other feature amounts, the number displayed becomes enormous.

Therefore, after calculating the significance value for each combination, only those having a particularly high significance are selected and displayed. In the example shown in FIG. 16, this influence degree comparison graph 78 was not described, because no significant difference occurred in the series regardless of the combination evaluated. That is, if a significant difference occurs in the series, the influence degree comparison graph 78 is automatically generated and displayed.

At this time, the feature amount to be combined may be a feature amount having high importance or a feature amount having low importance, and can be determined regardless of the degree of importance.

Also, regarding which section to divide the feature amount to be combined (5 days in the case of FIG. 21), the significance is verified by each of the various division methods, and the value having the highest significance is selected. be able to. For example, it is possible to verify the significance of each of the case of dividing into 5 days and the case of dividing into 10 days, and the control unit 10 may automatically perform the method of dividing such an interval. , The user may manually do it.

In the feature amount display user interface 60 shown in FIG. 22, a detailed menu 74 is displayed by selecting an arbitrary section from the graph displayed in the upper area 71 thereof, and further, the detailed menu 74 is displayed. , "Compare the contributions of features" is selected. Then, the contribution display graph 79 shown in FIG. 23 can be displayed on the feature amount display user interface 60. In this contribution display graph 79, the contribution of each feature amount is compared between the selected data group and the whole, which corresponds to a comparison diagram. Therefore, the contribution display graph 79 is generated by the comparison diagram generation unit 16c. This step is a comparison diagram generation step. Further, the contribution display graph 79 and the second basic diagram 77 can be displayed on the monitor 3 at the same time or separately. The process of displaying the contribution display graph 79 and the second basic diagram 77 at the same time or separately is the comparison diagram display step.

Here, the calculation method of the contribution of the feature amount will be described in detail. In the analysis method called linear multiple regression analysis, the prediction formula (2) expressed in the following format is used.

y = a1x1 + a2x2 + a3x3 + ... + anxn + b ... (2)
y represents a predicted value, x1, x2, ... Represents a value (explanatory variable) of each feature amount, and a1, a2, ... Represents a coefficient for each feature amount. Further, b is a constant term. In the multiple regression analysis, the coefficient a and the constant term b are learned so that the value of y approaches the objective variable for each data.

From the above equation (2), the predicted value y consists of the terms (a1x1, a2x2, ...) For each feature amount and the constant term. It can be interpreted that the larger the absolute value of the term is, the greater the influence on the prediction.

Therefore, the product aixi of the coefficient and the value of the feature amount is defined as the contribution degree with respect to the feature amount i. By averaging this contribution to the predictions for a plurality of data, it is possible to calculate the average contribution to the predictions of the data group for any data group.

For example, as shown in FIG. 24, when y = 3x1-2x2 + 1, when averaging the four data, the feature x1 has a slightly positive contribution to the predicted value, and the feature x2 has a slightly positive contribution to the predicted value. It will have a large negative contribution.

In this example, the procedure for calculating the contribution by the linear multiple regression analysis in the above equation was explained, but even if it is a non-linear algorithm, the same calculation can be performed by approximating each prediction with a linear model. Yes (eg Scott M. Lundberg, “Consistent Individualized Feature Attribution for Tree Ensembles” (2018)).

The contribution display graph 79 shown in FIG. 23 calculates the contribution degree defined above for the selected data, calculates all the data, and compares the two. According to the contribution display graph 79, for example, there is a tendency that "the minimum purchase amount does not affect the separation so much as a whole, but it has a large effect on the data group of interest" and "however, the influence of the total purchase amount is still larger". This will be a new finding.

The features displayed in this example are narrowed down to two, the features with higher importance and the features with significantly higher contribution only in the selected data group (minimum purchase price). In addition to this, only those with a significantly high difference in contribution may be displayed. The feature amount to be displayed may be deleted by the user.

When "gender" is selected from the feature quantities displayed in the feature quantity display area 60a of the feature quantity display user interface 60 shown in FIG. 15, the third basic diagram 80 shown in FIG. 25 is the feature quantity display user interface 60. Is displayed in. The third basic diagram 80 is generated by the basic diagram generation unit 16a of the display diagram generation unit 16 shown in FIG. 2, and this process is the basic diagram generation step. In the third basic figure 80, unlike the example shown in FIG. 16 and the example shown in FIG. 20, since the feature amount is not a numerical value, the expression of the third basic figure 80 is changed to a bar graph. Further, the "missing" of the item represents the case where the value does not exist in the original analysis target data. In the example shown in FIG. 25, since no significant difference was found in the value of the objective variable due to the difference in gender, none of the series was highlighted, but if a significant difference was found, it could be highlighted.

Here, when the series "missing" shown in FIG. 25 is selected to display the detailed menu 74 as shown in FIG. 17, and then "calculate the explanation of the difference" is selected from the detailed menu 74, it is shown in FIG. The difference explanation graph 81 can be displayed on the feature amount display user interface 60. Explanation graph 81 of this difference shows which feature amount mainly explains the difference of the value of the objective variable when the whole is compared with the selected data group, and corresponds to the comparison diagram. Therefore, the difference explanatory graph 81 shown in FIG. 26 is generated by the comparison diagram generation unit 16c. This step is a comparison diagram generation step. Further, the difference explanatory graph 81 and the third basic diagram 80 can be displayed on the monitor 3 at the same time or separately. Explanation of Differences The process of displaying the graph 81 and the third basic diagram 80 at the same time or separately is the comparison diagram display step. The above-mentioned contribution value is used in the calculation for explanation.

Here, the calculation method of the explanation of the difference will be described. First, for all the data and each of the selected data groups, the above-mentioned contribution is calculated for each feature amount.

y_all = B1all + B2all + B3all +… + Bnall + b ・ ・ ・ (3)
y_sel = B1sel + B2sel + B3sel +… + Bnsel + b ・ ・ ・ (4)
y_all and y_sel represent the average predicted value for all data / selected data. Biall and Bisel represent the contribution of all data / selected data for the i-th feature.

When (4)-(3) is calculated,
y_sel = y_all + (B1sel --B1all) + (B2sel --B2all) +… + (Bnsel --Bnall)
... (5)
Will be. The second and subsequent terms on the right side represent the difference in contribution between the selected data and all data (this is referred to as Di). In addition, y_sel and y_all are considered to contain prediction errors with respect to the average values y_sel_true and y_all_true of the actual objective variables.

y_sel_true = y_sel + δ_sel ・ ・ ・ (6)
y_all_true = y_all + δ_all ・ ・ ・ (7)
Substituting (6) and (7) into (5)
y_sel_true = y_all_true + D1 + D2 +… Dn + (δ_all --δ_sel) ・ ・ ・ (8)
Will be. From equation (8), the mean value (y_sel_true) of the objective variable of the selected data is the sum of the three elements of the mean value of the objective variable of all the data, the difference in the contribution of each feature, and the prediction error. I can explain.

In the above graph, among the differences D1, D2, ... Of the degree of contribution, the ones with particularly large values are displayed individually, and the remaining elements and the prediction error are collectively displayed in the "Other / Prediction error" item. There is.

In this way, by displaying the graph 81 that explains the difference in the objective variable according to the difference in the degree of contribution, the question "Why is the separation rate of the user group of gender = defect as high as 54%?" "The high divergence rate of users with gender = deficiency is partially affected by the'total purchase price'", "However, about 15% of the difference is explained by the difference in the main features. You can get the insight that you can't do it. "

However, as the gender-deficient sequence was not originally highlighted, there is not enough information to say that gender = deficiency has a large divergence, and the above insights may be incorrect. The wording that calls attention is also displayed together with the graph 81.

Further, as another graph for explaining gender, as shown in FIG. 27, a graph 82 relating to a change in the contribution of the feature amount can be displayed. This displays the same type of graph for gender as that displayed by clicking on the graph in the example shown in FIG.

The graph 82 shown in FIG. 27 can also be referred to as a second derivative diagram. The second derivative diagram is a diagram showing the feature quantities that contribute to the prediction of the objective variable in a predetermined cluster among a plurality of clusters. The cluster is the same as the first derivative diagram, and in this example, the gender is "male", "female", and "defective", but it can also be a period. The second derivative diagram is generated by the derivative diagram generation unit 16b shown in FIG. This process is the derivation diagram generation step. After the derivation diagram generation step, a derivation diagram display step that displays the basic diagram and the second derivative diagram simultaneously or separately can be executed.

Further, in this example, the feature amount contributing to the prediction of the objective variable in a predetermined cluster and the first feature amount in which all the clusters are combined can be displayed at the same time.

When the derivation diagram generation unit 16b generates the second derivation diagram, when the cluster is divided into male and female as described above, the degree of contribution to the prediction of the objective variable may change significantly depending on the gender. By showing the features that contribute to the prediction of such objective variables, new findings can be obtained.

Further, based on the graph shown in FIG. 27, "Gender alone does not affect the mean value of the objective variable, but the degree of contribution of other features varies greatly depending on the gender, which is important as a result. It is a feature quantity. "

(Action and effect of the embodiment)
As described above, according to the data analysis device 1 and the data analysis method according to this embodiment, the feature amount of high importance contained in a large amount of data can be displayed on the monitor 3. Then, a basic diagram showing the relationship between the feature amount and the objective variable, a derivative diagram showing the relationship between a plurality of feature quantities and the objective variable, and a derivative diagram showing the feature amount contributing to the prediction of the objective variable in a predetermined cluster. , A derivative diagram or the like showing an accuracy comparison of prediction models generated based on features with different aggregation periods can be displayed on the monitor 3.

In addition, it is possible to know the relationship between the representative value of the objective variable for each of the clusters formed by dividing the data to be analyzed, and the features that contribute to the prediction of the objective variable in the selected cluster, and all clusters or non-selected clusters. It is possible to compare with the features that contribute to the prediction of the objective variable in the cluster of. Therefore, it can help to obtain new knowledge.

(Other embodiments)
The above embodiments are merely exemplary in all respects and should not be construed in a limited way. Further, all modifications and modifications belonging to the equivalent scope of the claims are within the scope of the present invention.

In the above embodiment, a line graph or a bar graph is used to show the relationship between the objective variable and the feature amount, but other expression methods such as a scatter diagram may be used. In the case of a scatter plot, instead of clicking on a series of graphs, you can select a portion of the data by dragging and selecting an area in the graph to see more detailed graphs.

The first derivative diagram (influence comparison graph 78 shown in FIG. 21), the second derivative diagram (graph 82 regarding the change in the contribution of the feature amount shown in FIG. 27), and the third derivative diagram (prediction shown in FIG. 19). Any one of the accuracy comparison graphs 76) and the basic drawings 70, 77, 80 may be displayed on the monitor 3 at the same time, or may be displayed at different timings. The derivation diagram generation unit 16b may generate all of the first derivation diagram, the second derivation diagram, and the third derivation diagram, and the order of generation is not particularly limited. Further, all of the first derivative diagram, the second derivative diagram, and the third derivative diagram may be displayed on the monitor 3.

Further, although only three basic drawings 70, 77, and 80 are shown in the present embodiment, the basic figure generation unit 16a is generated according to the feature amount selected by the user. When switching between the basic diagrams 70, 77, 80 and the first derivative diagram, the second derivative diagram, and the third derivative diagram, an operation means such as a screen switching button can be used.

As described above, the data analysis device and the data analysis method according to the present invention can be used when attempting to obtain useful knowledge previously unknown from a large amount of information.

1 Data analyzer 3 Monitor (display)
3a Display control unit 11 Main control unit 12 Data input unit 13 Predictive model generation unit 14 Importance calculation unit 15 Cluster selection unit 16a Basic diagram generation unit 16b Derived diagram generation unit 16c Comparison diagram generation unit 17 Name automatic generation unit 18 Description Automatic Generation unit 30 Storage unit 70, 77, 80 Basic diagram 76 Predictive accuracy comparison graph (third derivative diagram)
75 Comparison table (comparison diagram)
78 Impact comparison graph (first derivative)
79 Contribution display graph (comparison diagram)
81 Difference explanation graph (comparison diagram)
83 Graph regarding changes in the contribution of features (second derivative diagram)

Claims (7)

In a data analyzer that analyzes data to be analyzed,
A data input unit that accepts input of analysis target data including multiple features and objective variables,
A prediction model generation unit that generates a prediction model for predicting the objective variable from the plurality of features.
An importance calculation unit that calculates the importance of the prediction by the prediction model for each of the plurality of features,
A display unit that displays features having higher importance based on the importance calculated by the importance calculation unit, and a display unit.
The data to be analyzed is divided into a plurality of clusters based on the value of the feature amount selected according to the input of the user from the feature amounts displayed on the display unit, and each cluster and the objective variable of each cluster are divided. A basic diagram generator that generates a basic diagram showing the relationship with the representative values of
A cluster selection unit that accepts the selection of any one of the multiple clusters shown in the basic diagram, and a cluster selection unit.
Features that contribute to the prediction of the objective variable in the cluster selected by the cluster selection unit and features that contribute to the prediction of the objective variable in all clusters or non-selected clusters that are not selected by the cluster selection unit. It is equipped with a comparison diagram generation unit that generates a comparison diagram showing comparison with a quantity.
The display unit is a data analysis device characterized in that the basic diagram generated by the basic diagram generation unit and the comparison diagram generated by the comparison diagram generation unit can be displayed. In the data analyzer according to claim 1,
The display unit is a data analysis device characterized in that a plurality of feature quantities having higher importance are arranged and displayed in descending order of importance. In the data analyzer according to claim 2,
The display unit is a data analysis device characterized in that the feature amount having the highest importance and a plurality of feature amounts less than the highest importance are simultaneously displayed. The data analyzer according to any one of claims 1 to 3.
The data analyzer is characterized in that the comparison diagram generation unit is configured to generate the comparison diagram showing a feature amount having a significantly high contribution only in the cluster selected by the cluster selection unit. The data analyzer according to any one of claims 1 to 4.
The data analysis device is characterized in that the comparison diagram generation unit is configured to generate the comparison diagram showing the feature quantities not selected by the user among the feature quantities having higher importance. .. The data analyzer according to any one of claims 1 to 5.
The comparison diagram generation unit is characterized in that it is configured to generate the comparison diagram showing the feature amount that contributes to the prediction of the objective variable in the defective cluster in which the value of the analysis target data does not exist. Data analyzer. In the data analysis method that analyzes the data to be analyzed,
A data input step that accepts input of analysis target data including multiple features and objective variables,
A prediction model generation step for generating a prediction model for predicting the objective variable from the plurality of features, and a prediction model generation step.
An importance calculation step for calculating the importance of the prediction by the prediction model for each of the plurality of features, and
A feature amount display step that displays a feature amount having a higher importance based on the importance calculated by the importance calculation step, and a feature amount display step.
The analysis target data is divided into a plurality of clusters based on the value of the feature amount selected according to the input of the user from the feature amounts displayed in the feature amount display step, and each cluster and each cluster A basic diagram generation step that generates a basic diagram showing the relationship with the representative value of the objective variable,
A cluster selection step that accepts the selection of any one of the multiple clusters shown in the basic diagram,
Features that contribute to the prediction of the objective variable in the cluster selected in the cluster selection step and features that contribute to the prediction of the objective variable in all clusters or non-selected clusters not selected by the cluster selection step. A comparison diagram generation step that generates a comparison diagram showing a comparison with a quantity,
A data analysis method comprising: a comparison diagram display step capable of displaying a basic diagram generated in the basic diagram generation step and a comparison diagram generated in the comparison diagram generation step.

Images