Missing data complementary device and missing data complementary method

Abstract

To provide a missing data complementary device and a missing data complementary method which improve the overall prediction accuracy of learning data including missing data, while reducing the missing data to be complemented more than before.SOLUTION: A missing data complementary device includes: a complementation priority flagging section for giving a complementation priority flag to at least one entry included in a predetermined ratio defined by a complementation amount adjustment parameter in each datum of a specific string data item; and a priority order determination section for counting the number of the complementation priority flags given to each entry in each entry included in a data table, so as to determine integrated complementation priority rank order in descending order of the number of the complementation priority flags.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present invention relates to a missing data completion device and a missing data completion method, and is suitable for application to a missing data completion device related to a technique for completing missing data, for example.

Background Art

[0002] In recent years, with the development of technologies using artificial intelligence, research on machine learning using learning data has been actively conducted. Such learning data may include missing data in a part thereof, and various studies have been made on the completion of missing data so as to cope with such cases. Patent Document 1 discloses a technique for completing missing data as preprocessing of machine learning. In the technique disclosed in Patent Document 1, for the purpose of improving the completion accuracy of missing data, specifically, first, a correlation matrix calculation unit calculates a correlation matrix between attributes using all learning records, and a regression completion unit, for a missing attribute, when there is an attribute whose absolute value of the correlation value is greater than a correlation threshold, performs regression completion using the attribute whose absolute value of the correlation value is greater than the correlation threshold.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, in the technique disclosed in Patent Document 1, as described above, only regression completion is performed for missing data, and no consideration has been given to how much missing data should be completed to improve the overall prediction accuracy of learning data including missing data.

[0005] This invention was made in consideration of the above points, and aims to propose a missing data imputation device and a missing data imputation method that can improve the overall prediction accuracy of training data containing missing data while minimizing the amount of missing data that needs to be imputed. [Means for solving the problem]

[0006] To solve the above problems, the present invention provides a data table in which a plurality of column data items are defined in the column direction and a plurality of entries in the row direction, each consisting of data from the plurality of column data items, and missing data exists in a specific column data item in some of the entries; a parameter definition unit that defines a interpolation amount adjustment parameter used to adjust the proportion of data in the specific column data item that needs interpolation; a missing data prediction unit that predicts the missing data in the specific column data item based on the data of other column data items other than the specific column data item in the plurality of entries constituting the data table, and the data of the specific column data item; an extraction unit that extracts at least one of the entries that are included in a predetermined proportion defined by the interpolation amount adjustment parameter from the data of the specific column data item; a interpolation priority flag assignment unit that assigns an interpolation priority flag to at least one of the entries included in the predetermined proportion; and a priority ranking determination unit that counts the number of entries to which the interpolation priority flag has been assigned for each entry in the data table, and determines the integrated interpolation priority ranking in descending order of the number of interpolation priority flags.

[0007] Furthermore, in the present invention, a parameter definition step defines a parameter definition unit for a parameter amount adjustment parameter used to adjust the proportion of data that needs to be imputed for a specific column data item in a data table in which a plurality of column data items are defined in the column direction and a plurality of entries in the row direction composed of the data of each of the plurality of column data items, and missing data exists for a specific column data item in some of the entries. The missing data prediction unit predicts the specific column data based on the data of other column data items other than the specific column data item in the plurality of entries that constitute the data table, and the data of the specific column data item. The system includes: a missing data prediction step that predicts the missing data for an item; an extraction step in which an extraction unit extracts at least one entry from each data item of a specific column data item that is included in a predetermined proportion defined by the completion amount adjustment parameter; a completion priority flag assignment step in which a completion priority flag assignment unit assigns a completion priority flag to at least one entry included in the predetermined proportion; and a priority ranking determination step in which a priority ranking determination unit counts the number of entries to which the completion priority flag has been assigned for each entry included in the data table and determines the integrated completion priority ranking in descending order of the number of completion priority flags. [Effects of the Invention]

[0008] According to the present invention, it is possible to improve the overall prediction accuracy of training data containing missing data while minimizing the amount of missing data that needs to be imputed. [Brief explanation of the drawing]

[0009] [Figure 1] This is a system configuration diagram showing an example of the configuration of a missing data completion device according to the first embodiment. [Figure 2] This flowchart shows an example of the procedure for processing missing data. [Figure 3] Figure 2 is a flowchart illustrating an example of the procedure for handling missing data. [Figure 4]This figure shows an example of the contents of a data table. [Figure 5] This figure shows an example of the contents of a data table. [Figure 6] This figure shows an example of the contents of a data table. [Figure 7] This figure shows an example of the contents of a data table. [Figure 8] This figure shows an example of the contents of a data table. [Figure 9] This figure shows an example of the contents of a data table. [Figure 10] This figure shows an example of the contents of a data table. [Figure 11] This figure shows an example of the characteristics of overall prediction accuracy relative to the effort required for data interpolation measurement. [Modes for carrying out the invention]

[0010] One embodiment of the present invention will be described in detail below with reference to the drawings.

[0011] First, let me explain the purpose of this embodiment. In recent years, high-precision risk prediction models using a wide variety of data have been developed, but when combining such data and inputting it into a risk prediction model, the existence of missing data can be a problem. Information collection to fill in such missing data is sometimes done individually, but the amount of work involved in such information collection is large. On the other hand, in order to achieve the target prediction accuracy in a risk prediction model, it is not always necessary to collect information to fill in all the missing data. For example, if entries represented by 20% of the total IDs (hereinafter simply referred to as "entries") are related to 80% of the risks, then it is sufficient to obtain the target predictiveness by supplementing only the missing data of those 20% of entries. Therefore, in this embodiment, a supplementation priority flag is assigned to entries that are estimated to be related to 80% of the risks, and an "integrated supplementation priority flag" is assigned according to the number of entries that have been assigned the supplementation priority flag, i.e., the number of supplementation priority flags, with a rank corresponding to that number. With these flags, this embodiment enables the supplementation of missing data that contributes to high prediction accuracy with less work. The details will be explained below.

[0012] Figure 1 is a system configuration diagram showing an example of the configuration of a missing data completion device 100 according to the first embodiment. The missing data completion device 100 is, for example, a computer and includes a data table 10, a parameter definition unit 20, a missing data prediction unit 30, an extraction unit 40, a completion priority flag assignment unit 50, and a priority ranking determination unit 60, and preferably includes a display unit 70. The display unit 70 may be integrated with the missing data completion device 100 or it may be a separate unit.

[0013] Data table 10 has multiple column data items defined in the column direction, and multiple entries in the row direction, each consisting of data from the multiple column data items. In data table 10, some of the entries have missing data in specific column data items (flood risk, exchange rate risk).

[0014] The parameter definition unit 20 defines a complement amount adjustment parameter X, and this complement amount adjustment parameter X is a parameter used to adjust the ratio for which complement is required for the data of specific column data items (for example, "flooding risk", "exchange rate risk" described later). In the present embodiment, a predetermined ratio, for example, the top 30% is defined as this complement amount adjustment parameter X, and preferably, the top 70% is further defined.

[0015] The missing data prediction unit 30 predicts missing data of specific column data items based on the data of other column data items other than the specific column data items and the data of the specific column data items in the plurality of entries constituting the data table 10.

[0016] The missing data prediction unit 30, for example, uses the data of other column data items other than the specific column data items in the plurality of entries constituting the data table 10 as explanatory variables and the data of the specific column data items as target variables to learn a model. The missing data prediction unit 30 predicts the missing data of the specific column data items by machine learning using this model.

[0017] The extraction unit 40 extracts at least one entry included in a predetermined ratio defined by the complement amount adjustment parameter X from each data of the specific column data items.

[0018] The complement priority flag assignment unit 50 assigns a complement priority flag to at least one entry included in a predetermined ratio defined by the complement amount adjustment parameter X as described above in the data table 10.

[0019] The priority order determination unit 60 counts the number of entries to which the complement priority flag is assigned for each entry included in the data table 10, and determines the integrated complement priority order in descending order of the number of assigned complement priority flags.

[0020] The display unit 70 displays, for each entry containing missing data, the data for each specific column data item and a list of integrated completion priority rankings. The content of this display will be described later.

[0021] The data completion priority flag assignment unit 50 assigns a data completion priority flag for each of the predetermined percentages (e.g., 30%, 70%) defined as data completion adjustment parameters. The display unit 70 displays the data completion measurement effort based on the list data described above, and the prediction accuracy of the data for specific column data items (e.g., "flood risk" and "exchange rate risk," described later) by the missing data prediction unit 30.

[0022] The missing data imputation device 100 has the configuration described above, and next, a missing data imputation method as an example of the operation of the missing data imputation device 100 will be described. The missing data imputation method has multiple column data items defined in the column direction and multiple entries composed of the data of multiple column data items in the row direction, and a parameter definition step in which an imputation amount adjustment parameter X is defined in the parameter definition unit to adjust the proportion of data of a specific column data item in a data table 10 in which missing data exists in a specific column data item in some of the entries, and a missing data prediction unit 30 based on the data of other column data items other than the specific column data item in the multiple entries that make up the data table 10, and the data of the specific column data item (for example, excluding missing data) The system includes: a missing data prediction step that predicts missing data for a specific column data item; an extraction step in which an extraction unit 40 extracts at least one entry from each of the data for the specific column data item that is included in a predetermined proportion defined by the completion amount adjustment parameter X; a completion priority flag assignment step in which a completion priority flag assignment unit 50 assigns a completion priority flag to at least one entry included in the predetermined proportion; and a priority ranking determination step in which a priority ranking determination unit 60 counts the number of entries to which a completion priority flag has been assigned for each entry included in the data table 10 and determines the integrated completion priority ranking in descending order of the number of completion priority flags.

[0023] Figure 2 is a flowchart showing an example of the procedure for processing missing data, and Figure 3 is a flowchart showing an example of the procedure for processing missing data shown in Figure 2 (step S20 in Figure 2). The following explanation will refer to the data table 10 shown in Figures 4 to 9.

[0024] In step S10 shown in Figure 2, the data table 10 and the interpolation amount adjustment parameter X are input to the missing data interpolation device 100. The parameter definition unit 20 sets the input interpolation amount adjustment parameter X to, for example, the top 30%.

[0025] On the other hand, as shown in Figure 4, for example, the data table 10 has column data items such as business ID (IDentifier), country of location, long-term interest rate, distance from river, flood risk, priority of supplementary flood perspective, exchange rate risk, priority of supplementary exchange perspective, and Integration A priority order for completion is defined. Data table 10 consists of data for each column data item and manages entries that are identified by the aforementioned business ID (IDentifier). In data table 10 shown in Figure 4, for example, there is missing data indicated by "-" in the case of flood risk and exchange rate risk, which are examples of specific column data items.

[0026] In step S20, missing data processing is performed. This missing data processing is performed by the missing data prediction unit 30. In step S21 in Figure 3, the missing data prediction unit 30 predicts the data for the specific column data item (missing value indicated by "-" in the figure).

[0027] Specifically, for example, regarding flood risk, the missing data prediction unit 30 creates a model that predicts specific column data items where missing data exists, such as flood risk data, from combinations of entries that do not contain missing data. The combinations of entries that do not contain missing data refer, for example, to the entries with business IDs "AAAA001", "AABB002", "DDFF004", "AAFF005", and "CCGG006" shown by the thick border line 10A in Figure 5. In this embodiment, flood risk = S is assumed to be influenced by the country of location and the distance from the river. In this embodiment, these two (country of location and distance from the river) are used as explanatory variables, and flood risk is used as the dependent variable.

[0028] On the other hand, for example, with respect to exchange rate risk, the missing data prediction unit 30 creates a model that predicts specific column data items where missing data exists, such as exchange rate risk data, from combinations of entries that do not contain missing data. The combinations of entries that do not contain missing data refer, for example, to the entries with business IDs "AAAA002", "CCDD003", "DDFF004", and "CCGG006" shown by the thick border line 10B in Figure 6. In this embodiment, exchange rate risk = K is assumed to be influenced by the long-term interest rate of the country where the system is located. In this embodiment, this (long-term interest rate of the country where the system is located) is used as an explanatory variable, and exchange rate risk is used as the dependent variable.

[0029] The missing data prediction unit 30, for example, learns a model using data from other column data items in multiple entries constituting the data table 10, excluding a specific column data item, as explanatory variables, and data from a specific column data item (e.g., flood risk, exchange rate risk) as the target variable. Using this model, it predicts missing data for the specific column data item through machine learning (corresponding to the "(predicted)" portion shown in Figure 7).

[0030] In step S22 of Figure 3, the extraction unit 40 extracts at least one entry from each data point of a specific column data item (e.g., flood risk, exchange rate risk) that falls within a predetermined percentage (e.g., top 30%) defined by the interpolation adjustment parameter X.

[0031] In step S23 of Figure 3, the interpolation priority flag assignment unit 50 assigns a interpolation priority flag to at least one entry that falls within a predetermined percentage (e.g., the top 30%) defined by the interpolation amount adjustment parameter X (see Figure 8). In the illustrated example, entries to which the interpolation priority flag has been assigned are indicated with "◎".

[0032] Next, in step S30 of Figure 2, the priority ranking determination unit 60 counts the number of entries in the data table 10 that have been assigned a supplemental priority flag, and determines the integrated supplemental priority ranking in descending order of the number of supplemental priority flags (see Figure 9).

[0033] In step S40, the display unit 70 displays a list of entries with missing data × (data for each column data item, integrated completion priority ranking).

[0034] In step S50, the display unit 70 displays pairs of interpolation effort and overall prediction accuracy for multiple interpolation amount adjustment parameters X.

[0035] Figure 10 shows an example of missing data imputation processing performed with the imputation amount adjustment parameter X = 70%. 10 In the examples shown, explanations of the same content as in the examples shown in Figures 8 and 9 above will be omitted.

[0036] In the example shown in Figure 10, the interpolation adjustment parameter X = 70%, and the extraction unit 40 extracts specific column data. item Extract at least one entry from each data point (e.g., flood risk, exchange rate risk) that falls within the top 70%.

[0037] The completion priority flag assignment unit 50 is as described above. Figure In addition to the extracted data 10D in step 9, a supplementation priority flag is also assigned to the flood risk data 10E for the entry with business ID "CCDD003" (corresponding to the "◎" in the flood risk supplementation priority shown in the diagram).

[0038] The priority ranking determination unit 60 counts the number of entries with a supplemental priority flag assigned to each entry in the data table 10, and determines the integrated supplemental priority ranking in descending order of the number of supplemental priority flags. As a result, the integrated supplemental priority ranking is as follows: 1st place is the entry with business ID "SSDD007" with 2 "◎" flags, 2nd place is the entry with business ID "CCDD003" with 1 "◎" flag, and 3rd place is the entry with business ID "AAFF005" with 1 "◎" flag. Therefore, the example shown in Figure 10 differs from the example shown in Figure 9 in that it also includes the entry with business ID "CCDD003".

[0039] Figure 11 shows an example of the characteristics of overall prediction accuracy with respect to data interpolation measurement effort. In the example shown, the target prediction accuracy (corresponding to "Target Accuracy" in the figure) is 0.75, and the figures show the cases when the interpolation adjustment parameter X = 70% and when the interpolation adjustment parameter X = 30%. In the example shown, "1ID" refers to interpolation using an entry corresponding to one business establishment ID (hereinafter also referred to as "1ID interpolation"), "2ID" refers to interpolation using entries corresponding to two business establishment IDs (hereinafter also referred to as "2ID interpolation"), and "3ID" refers to interpolation using entries corresponding to three business establishment IDs (hereinafter also referred to as "3ID interpolation").

[0040] Here, we are trying to determine the number of entries corresponding to the business IDs to be supplemented, while taking into account the target accuracy mentioned above. In this embodiment, the missing data prediction unit 30 considers how to set the value of the supplementation amount adjustment parameter X (here, either 30% or 70%) in order to obtain a prediction accuracy of "0.75" as an example of the target accuracy. That is, the missing data prediction unit 30 selects one predetermined percentage from a plurality of predetermined percentages (e.g., 30%, 70%) defined as the supplementation amount adjustment parameter X so that the overall prediction accuracy of the training data including missing data satisfies the target accuracy.

[0041] In the illustrated example, the data interpolation measurement effort increases in the order of "1 ID," "2 ID," and "3 ID," and the overall prediction accuracy improves accordingly. The prediction accuracy for "2 ID" is, for example, 0.76, the prediction accuracy for "1 ID" is less than 0.75 (for example, 0.67), and the prediction accuracy for "3 ID" is greater than 0.75 (for example, 0.82).

[0042] In this embodiment, when the interpolation amount adjustment parameter X = 30%, if we aim to achieve the prediction accuracy of the target accuracy "0.75" mentioned above, interpolation with 1 ID will be insufficient, so interpolation with 2 IDs is used.

[0043] On the other hand, in this embodiment, when the interpolation amount adjustment parameter X = 70%, in order to achieve the prediction accuracy of the target accuracy "0.75" mentioned above, interpolation with 1 ID is insufficient, so interpolation with 3 IDs is used.

[0044] As described above, by setting the interpolation amount adjustment parameter X to achieve a prediction accuracy corresponding to the aforementioned target accuracy of "0.75", the missing data prediction unit 30 can improve the overall prediction accuracy of the training data including missing data using the model described above.

[0045] As described above, the missing data completion device 100 according to this embodiment comprises: a data table 10 in which a plurality of column data items are defined in the column direction and a plurality of entries in the row direction consisting of the data of each of the plurality of column data items, wherein missing data exists for a specific column data item (e.g., flood risk, exchange rate risk) in some of the entries; a parameter definition unit 20 that defines a completion amount adjustment parameter X used to adjust the proportion of data of a specific column data item that needs to be completed; a missing data prediction unit 30 that predicts missing data for a specific column data item based on the data of other column data items other than the specific column data item in the plurality of entries constituting the data table 10, and the data of the specific column data item; an extraction unit 40 that extracts at least one entry that is included in a predetermined proportion (e.g., 30%) of the data of the specific column data item as defined by the completion amount adjustment parameter X; a completion priority flag assignment unit 50 that assigns a completion priority flag to at least one entry included in the predetermined proportion; and a priority ranking determination unit 60 that counts the number of entries to which a completion priority flag has been assigned for each entry included in the data table 10 and determines the integrated completion priority ranking in descending order of the number of completion priority flags.

[0046] In this way, the extraction unit 40 extracts at least one entry that is included in a predetermined proportion defined by the interpolation amount adjustment parameter X, assigns an interpolation priority flag to it, and determines the integrated interpolation priority order in descending order of the number of interpolation priority flags. Therefore, by referring to the integrated interpolation priority order, it is possible to accurately identify missing data that has a deep impact on prediction accuracy, thereby improving the overall prediction accuracy of the training data including missing data while keeping the amount of missing data to be interpolated to a minimum.

[0047] The missing data imputation device 100 according to this embodiment includes a display unit 70 that displays data for each specific column data item and a list of integrated imputation priority rankings for each entry containing missing data. In this way, by referring to the data for each specific column data item and the list of integrated imputation priority rankings for each entry containing missing data, it is possible to reduce the amount of missing data that needs to be imputed while improving the overall prediction accuracy of the training data containing missing data.

[0048] In this embodiment, the data completion priority flag assignment unit 50 assigns a data completion priority flag at each of several predetermined percentages (e.g., 30%, 70%) defined as the data completion amount adjustment parameter X, and the display unit 70 displays the data completion measurement effort based on the list data and the prediction accuracy of the data for a specific column data item by the missing data prediction unit 30. In this way, by referring to the data completion measurement effort based on the list data and the prediction accuracy of the data for a specific column data item by the missing data prediction unit 30, it is possible to reduce the amount of missing data to be completed while improving the overall prediction accuracy of the training data including missing data.

[0049] In this embodiment, the missing data prediction unit 30 selects one predetermined proportion from a plurality of predetermined proportions defined as the imputation amount adjustment parameter X so that the overall prediction accuracy of the training data including missing data meets the target accuracy. In this way, the overall prediction accuracy of the training data including missing data can be increased to above the target accuracy while keeping the amount of missing data to be imputed to a minimum.

[0050] In this embodiment, the missing data prediction unit 30 learns a model using data from columns other than a specific column data item in multiple entries constituting the data table 10 as explanatory variables, and data from a specific column data item (for example, the flood risk and exchange rate risk mentioned above) as the target variable, and uses the model to predict missing data for the specific column data item by machine learning. In this way, the missing data prediction unit 30 can predict missing data using the model while keeping the amount of missing data to be filled in to a minimum and improving the overall prediction accuracy of the training data including missing data.

[0051] It should be noted that the present invention is not limited to the embodiments described above, but includes various modifications and equivalent configurations within the spirit of the attached claims. For example, the embodiments described above are described in detail for the purpose of clearly illustrating the present invention, and the present invention is not necessarily limited to having all the configurations described. Also, each element described in parallel in this embodiment may be configured such that at least one of the elements is connected in series with respect to the other elements. [Industrial applicability]

[0052] This invention can be applied to a missing data completion device relating to a technology for completing missing data. [Explanation of Symbols]

[0053] 10...Data table, 20...Parameter definition unit, 30...Missing data prediction unit, 40...Extraction unit, 50...Priority flag assignment unit, 60...Priority ranking determination unit, 100...Missing data completion device

Claims

1. a data table in which a plurality of column data items are defined in the column direction and a plurality of entries each consisting of data of the plurality of column data items in the row direction, wherein missing data exists in a specific column data item in some of the entries; a parameter definition section for defining an interpolation amount adjustment parameter used to adjust the proportion of data in the specific column data item that needs to be interpolated; a missing data prediction unit that predicts the missing data of the specific column data item based on data of other column data items other than the specific column data item in the plurality of entries that make up the data table, and the data of the specific column data item; an extracting unit that extracts at least one entry included in a predetermined ratio defined by the complement amount adjustment parameter from each data item of the specific column data item; a complementary priority flag assigning unit that assigns a complementary priority flag to at least one of the entries included in the predetermined ratio; a priority ranking determination unit that counts the number of entries to which the complementary priority flags are assigned for each of the entries included in the data table, and determines an integrated complementary priority ranking in descending order of the number of complementary priority flags; A missing data compensation device comprising:

2. a display unit that displays, for each entry including the missing data, the data for each specific column data item and the list data of the integration and completion priority order; 2. The missing data compensation device according to claim 1.

3. The complementary priority flag assigning unit assigning the complementation priority flag to each of the plurality of predetermined ratios defined as the complementation amount adjustment parameter; The display unit The amount of data supplementation measurement work based on the list data and the prediction accuracy of the missing data prediction unit for the data of the specific column data item are displayed.

3. The missing data compensation device according to claim 2.

4. The missing data prediction unit Selecting one predetermined ratio from the plurality of predetermined ratios defined as the complement amount adjustment parameter so that the overall prediction accuracy of the training data including the missing data satisfies the target accuracy.

4. The missing data compensation device according to claim 3.

5. The missing data prediction unit A model is trained using data of column data items other than the specific column data item in the plurality of entries constituting the data table as explanatory variables and data of the specific column data item as a target variable, and the missing data of the specific column data item is predicted by machine learning using the model.

2. The missing data compensation device according to claim 1.

6. a parameter definition step of defining in a parameter definition section an interpolation amount adjustment parameter used to adjust the proportion of data in a specific column data item in a data table in which a plurality of column data items are defined in a column direction and a plurality of entries are formed from each piece of data in the plurality of column data items in a row direction, and in which missing data exists in a specific column data item in some of the entries; a missing data prediction step in which a missing data prediction unit predicts the missing data of the specific column data item based on data of column data items other than the specific column data item and the data of the specific column data item in the plurality of entries constituting the data table; an extraction step in which an extraction unit extracts at least one entry included in a predetermined ratio defined by the complement amount adjustment parameter from each data item of the specific column data item; a complementary priority flag assigning step in which a complementary priority flag assigning unit assigns a complementary priority flag to at least one of the entries included in the predetermined ratio; a priority order determination step in which a priority order determination unit counts the number of entries to which the complementary priority flags are assigned for each of the entries included in the data table, and determines an integrated complementary priority order in descending order of the number of complementary priority flags; A missing data imputation method comprising:

7. a display step of displaying, on a display unit, the data for each specific column data item and the list data of the integration and completion priority order for each entry including the missing data; 7. The missing data filling method according to claim 6.

8. The complementary priority flag assigning unit assigning the complementation priority flag to each of a plurality of values ​​set as the complementation amount adjustment parameter; In the display step, The display unit displays the amount of data complementation measurement work based on the list data and the prediction accuracy of the missing data prediction unit for the data of the specific column data item.

8. The missing data filling method according to claim 7.

9. In the missing data prediction step, The missing data prediction unit selects one predetermined ratio from the plurality of predetermined ratios defined as the complement amount adjustment parameter so that the overall prediction accuracy of the training data including the missing data satisfies a target accuracy.

9. The missing data filling method according to claim 8.

10. In the missing data prediction step, The missing data prediction unit learns a model using data of column data items other than the specific column data item in the plurality of entries constituting the data table as explanatory variables and data of the specific column data item as a target variable, and predicts the missing data of the specific column data item by machine learning using the model.

7. The missing data filling method according to claim 6.

Images