Correction data determination device, correction data determination method, and program

Abstract

In order to enable suitable error correction corresponding to an analysis task, an information processing device (1) comprises: an acquisition unit (11) for acquiring target data; a calculation unit (12) that calculates, for each error or each of the attributes of the error, the degree of impact a plurality of errors included in the target data have on an evaluation index of a machine learning model; and a determination unit (13) that determines data to be corrected from the target data on the basis of the degrees of impact calculated by the calculation unit (12).

Description

[Technical Field]

[0001] The present invention relates to techniques for analyzing data. [Background technology]

[0002] Data quality is an issue in data analysis. Examples of cases where data quality can be an issue include "variations in notation," "missing values," "abnormal values," and "formatting discrepancies." Patent Document 1, for example, is an example of a document disclosing a so-called data cleansing technique for correcting errors and the like contained in data. Patent Document 1 describes a technique that appropriately addresses data inconsistencies between business systems and enables highly accurate data analysis by identifying the content of the data cleansing process based on discrepancies in business data related to objects between each business system, and performing the data cleansing process with the identified content. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] International Publication No. 2018 / 207506 Summary of the Invention [Problem to be solved by the invention]

[0004] However, it is known that the errors to be corrected in data cleansing vary depending on the type of analytical task in machine learning. The technology described in Patent Document 1 has a problem in that it is not possible to correct errors taking into account the analytical task.

[0005] One aspect of the present invention has been made in view of the above-mentioned problems, and one example of a purpose thereof is to enable appropriate error correction according to the analysis task. [Means for solving the problem]

[0006] According to one aspect of the present invention Correction data determination The device includes an acquisition means for acquiring target data, a calculation means for calculating the degree of influence that multiple errors contained in the target data have on the evaluation index of the machine learning model for each error or for each attribute of the error, and a determination means for determining data to be corrected from the target data based on the degree of influence calculated by the calculation means.

[0007] According to one aspect of the present invention Correction data determination The method includes at least one processor acquiring target data, calculating the degree of influence that multiple errors contained in the target data have on the evaluation index of a machine learning model for each of the errors or for each attribute of the errors, and determining data to be corrected from the target data based on the calculated degree of influence.

[0008] A program according to one aspect of the present invention causes a computer to function as an acquisition means for acquiring target data, a calculation means for calculating the degree of influence that multiple errors contained in the target data have on the evaluation index of a machine learning model for each of the errors or for each attribute of the errors, and a determination means for determining data to be corrected from the target data based on the degree of influence calculated by the calculation means. [Effects of the Invention]

[0009] According to one aspect of the present invention, it is possible to perform appropriate error correction according to the analysis task. [Brief explanation of the drawings]

[0010] [Figure 1] 1 is a block diagram showing a configuration of an information processing device according to a first exemplary embodiment. [Figure 2] 1 is a flowchart showing the flow of an information processing method according to the first exemplary embodiment. [Figure 3] FIG. 10 is a block diagram showing the configuration of an information processing device according to a second exemplary embodiment. [Figure 4] 10 is a flowchart showing the flow of an information processing method according to the second exemplary embodiment. [Figure 5] 10A to 10C are diagrams illustrating specific examples of errors detected by an error detection unit according to the second exemplary embodiment. [Figure 6] FIG. 10 is a diagram showing a specific example of grouping of errors by a grouping unit according to the second exemplary embodiment. [Figure 7] FIG. 10 is a diagram showing a specific example of evaluation data generated by an evaluation data generating unit according to the second exemplary embodiment. [Figure 8] FIG. 10 is a diagram showing a specific example of an influence calculated by an influence calculation unit according to the second exemplary embodiment. [Figure 9] 10 is a diagram illustrating a specific example of a determination process performed by a determination unit according to the second exemplary embodiment. FIG. [Figure 10] FIG. 10 is a diagram showing a specific example of a data correction process performed by a data cleansing unit according to the second exemplary embodiment. [Figure 11] FIG. 1 is a diagram illustrating an example of a computer that executes instructions of a program, which is software that realizes the functions of each device according to each exemplary embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0011] Exemplary Embodiment 1 A first exemplary embodiment of the present invention will be described in detail with reference to the drawings. This exemplary embodiment is a basic form of the exemplary embodiments described below.

[0012] (Configuration of information processing device) The configuration of an information processing device 1 according to this exemplary embodiment will be described with reference to Fig. 1. Fig. 1 is a block diagram showing the configuration of the information processing device 1. The information processing device 1 includes an acquisition unit 11, a calculation unit 12, and a determination unit 13.

[0013] The acquisition unit 11 acquires target data. The calculation unit 12 calculates the degree of influence that multiple errors contained in the target data have on the evaluation index of the machine learning model for each of the errors or for each attribute of the errors. The determination unit 13 determines data to be corrected from the target data based on the degree of influence calculated by the calculation unit 12.

[0014] As described above, the information processing device 1 according to this exemplary embodiment has a configuration including an acquisition unit 11 that acquires target data, a calculation unit 12 that calculates the degree of influence that multiple errors contained in the target data have on the evaluation index of the machine learning model for each error or for each attribute of the error, and a determination unit 13 that determines data to correct from the target data based on the influence calculated by the calculation unit 12. Therefore, the information processing device 1 according to this exemplary embodiment has the effect of being able to perform appropriate error correction according to the analysis task.

[0015] (Information Processing Program) The functions of the information processing device 1 described above can also be realized by a program. The information processing program according to this exemplary embodiment causes a computer to function as an acquisition means for acquiring target data, a calculation means for calculating the influence of multiple errors included in the target data on the evaluation index of a machine learning model for each of the errors or for each attribute of the errors, and a determination means for determining data to be corrected from the target data based on the influence calculated by the calculation means.

[0016] (Flow of information processing method) The flow of information processing method S1 according to this exemplary embodiment will be described with reference to Fig. 2. Fig. 2 is a flow diagram showing the flow of information processing method S1. The execution entity of each step in information processing method S1 may be a processor provided in information processing device 1, or may be a processor provided in another device, or each step may be executed by a processor provided in a different device.

[0017] In step S11, at least one processor acquires target data. In step S12, at least one processor calculates the influence of multiple errors contained in the target data on the evaluation index of the machine learning model for each of the errors or for each attribute of the errors. In step S13, at least one processor determines data to be corrected from the target data based on the influence calculated in step S12.

[0018] As described above, the information processing method S1 according to this exemplary embodiment employs a configuration in which at least one processor acquires target data to be evaluated, calculates the influence of multiple errors contained in the target data on the evaluation index of the machine learning model for each error or for each attribute type of the error, and determines data to be corrected from the target data based on the calculated influence. Therefore, the information processing method S1 according to this exemplary embodiment has the effect of enabling appropriate error correction according to the analysis task.

[0019] Exemplary Embodiment 2 A second exemplary embodiment of the present invention will be described in detail with reference to the drawings. Note that components having the same functions as those described in the first exemplary embodiment are given the same reference numerals, and their description will not be repeated.

[0020] <Configuration of information processing device> 3 is a block diagram showing the configuration of an information processing device 1A according to exemplary embodiment 2. The information processing device 1A includes a control unit 10A, a storage unit 20A, an input / output unit 30A, and a communication unit 40A.

[0021] (input / output section) Input / output devices such as a keyboard, a mouse, a display, a printer, and a touch panel are connected to the input / output unit 30A. The input / output unit 30A receives various types of information input to the information processing device 1A from the connected input devices. Under the control of the control unit 10A, the input / output unit 30A outputs various types of information to the connected output devices. An example of the input / output unit 30A is an interface such as a USB (Universal Serial Bus). The input / output unit 30A may also include a display panel, a speaker, a keyboard, a mouse, a touch panel, and the like.

[0022] (Communications Department) The communication unit 40A communicates with devices external to the information processing device 1A via a communication line. While the specific configuration of the communication line does not limit the present exemplary embodiment, examples of the communication line include a wireless LAN (Local Area Network), a wired LAN, a WAN (Wide Area Network), a public line network, a mobile data communication network, or a combination thereof. The communication unit 40A transmits data supplied from the control unit 10A to other devices, and supplies data received from other devices to the control unit 10A.

[0023] (Control unit) The control unit 10A includes an acquisition unit 11, a calculation unit 12, a determination unit 13, an error detection unit 14, a data cleansing unit 18, an evaluation unit 19, and an analysis result output unit 20. The calculation unit 12 also includes a grouping unit 15, an evaluation data generation unit 16, and an impact calculation unit 17.

[0024] (Acquisition Department) The acquisition unit 11 acquires target data D. The target data D is the target of data analysis, and as an example, is data including multiple records. Examples of data including multiple records include structured data such as table data, semi-structured data written in a data description language such as JSON (JavaScript Object Notation: registered trademark) or XML (Extensible Markup Language), and unstructured data representing a document written in a natural language. As an example, a record is a row of a table, and includes one or more sets of attribute names and attribute values ​​corresponding to the columns of the table.

[0025] In this exemplary embodiment, the target data D contains multiple errors. Errors occur due to various factors, such as tabulation errors and variations in the notation of different data. Examples of errors include attribute values ​​contained in records having different data types (numeric, text, date, etc.), duplicate records being included in the target data D, records containing missing values, and records containing incorrect data.

[0026] If the target data D containing such errors is analyzed as is, the accuracy of the data analysis will be poor or the correct data analysis results will not be obtained. Therefore, if the target data D contains errors, the accuracy of the analysis can be improved by performing data cleansing.

[0027] (Error detection section) The error detection unit 14 detects multiple errors contained in the target data D. The method used by the error detection unit 14 to detect errors is arbitrary, but as an example, the error detection unit 14 may detect errors contained in the target data D using a rule-based detection method, or may detect errors by inference using a trained model generated by machine learning.

[0028] When detecting an error using a rule-based detection method, the error detection unit 14 may determine that an error has occurred in the following cases: (i) an attribute value is missing; (ii) the attribute value is not within a predetermined range; (iii) the attribute value of the first attribute name is inconsistent with the attribute value of the second attribute name; or (iv) the format of the attribute value is incorrect.

[0029] Furthermore, when detecting errors by inference using a trained model, the machine learning method of the trained model is not limited, and, for example, a decision tree-based, linear regression, or neural network method may be used, or two or more of these methods may be used. For example, the input of the trained model includes records contained in the target data D, and for example, the output of the trained model includes labels indicating the presence or absence of errors or the type of errors contained in the records.

[0030] (Calculation section) The calculation unit 12 calculates the degree of influence that an error contained in the target data D has on the evaluation index of the analytical model for each error or for each attribute of the error. Here, the analytical model is a machine learning model corresponding to an analytical task. Examples of analytical tasks include, but are not limited to, annual income prediction, sales prediction, and morbidity prediction.

[0031] An error attribute is an index for classifying errors or information indicating the classification result of errors, such as the type of error, information identifying each group into which errors are grouped, etc. When errors are grouped into multiple groups, they may be grouped by error type, or multiple types of errors may be included in one group. In other words, multiple types may be associated with one attribute.

[0032] (Analysis model) The analytical model is a model for analyzing the target data D, and is generated by machine learning, for example. iFor example, ' may be a linear model that performs regression analysis on the prediction of annual income. The machine learning method of the analytical model is not limited, and for example, a decision tree-based, linear regression, or neural network method may be used, or two or more of these methods may be used.

[0033] As an example, the input of the analytical model includes target data D. As an example, the output of the analytical model includes information indicating the estimated annual income. However, the input and output of the analytical model are not limited to the above examples and may include other information.

[0034] (Group Division) The grouping unit 15 groups the multiple errors detected by the error detection unit 14 according to the characteristics of the errors. The grouping method used by the grouping unit 15 is arbitrary, but as an example, the errors may be grouped by type of error, or multiple types of errors may be grouped together. More specifically, as an example, the grouping unit 15 may group the errors by type of detection method used by the error detection unit 14 (for example, by rule). As an example, the grouping unit 15 may cluster the multiple errors using a clustering method such as spectral clustering.

[0035] (Evaluation data generation unit) The evaluation data generating unit 16 generates evaluation data D that includes pseudo errors in the target data D. i ´ (i=1, 2, ..., n) is generated for each error or each attribute of the error, where n is the evaluation data D i The number of errors or error attributes is the number of errors or error attributes. i ' corresponds to the evaluation data D', the evaluation data generating unit 16 generates pseudo errors for each attribute, for example, i ´. Also, the error and evaluation data D i ' correspond one-to-one, the evaluation data generating unit 16 generates pseudo errors similar to each error, for example, to generate evaluation data D i Include in ´.

[0036] Evaluation data D i The method for generating the evaluation data D' is arbitrary. For example, the evaluation data generating unit 16 generates the evaluation data D' by a rule-based generation method such as deleting existing data or removing hyphens. i As another example, the evaluation data generation unit 16 may generate the evaluation data D′ using a generative model such as an autoencoder or a generative adversarial network (GAN). i In this case, the input of the generative model includes, for example, the target data D, and the output includes, for example, the evaluation data D i Contains ´.

[0037] (Impact calculation department) The impact calculation unit 17 calculates the impact for each error or for each attribute of the error. More specifically, as an example, the impact calculation unit 17 calculates the impact for each attribute corresponding to each group divided by the grouping unit 15. In this case, more specifically, as an example, the impact calculation unit 17 calculates the impact for each attribute corresponding to each group divided by the grouping unit 15. i Using ´, the impact s i Calculate.

[0038] Evaluation data D i When using the analysis model MD ′, the influence calculation unit 17 uses the analysis model MD ′ generated using the target data D, for example. init Performance and evaluation data D i The analytical model MD generated using i Based on the comparison results with each performance of i Calculate the impact s i is, for example, a value that represents the degree of change (for example, the rate of change) in the performance of the analytical model. i For each of the ´, the impact s i By calculating the n influences s i In the following, the influence S={s1, s2, ..., s n}

[0039] (Decision section) The determining unit 13 determines the influence S={s1, s2, . . . , s n}, the determination unit 13 determines the data to be corrected from the target data D. More specifically, as an example, the determination unit 13 uses the impact S calculated by the calculation unit 12 to calculate a second impact on the evaluation index for each of multiple partial data included in the target data D, and determines the partial data to be corrected based on the calculated second impact for each partial data. Here, the partial data is data included in the target data D, and as an example, is a record included in table data including multiple records. In other words, when the target data D is table data including multiple records, the determination unit 13 determines the record to be corrected based on the impact S calculated for each type of error, as an example.

[0040] (Data Cleansing Department) The data cleansing unit 18 corrects the data determined by the determination unit 13. As an example, the data cleansing unit 18 may correct the data in response to a user operation. More specifically, the data cleansing unit 18 may output the data to be corrected to an output device such as a display panel, and correct the data based on information input by an input device operated by the user.

[0041] Furthermore, as an example, the data cleansing unit 18 may perform data correction by inference using a trained model that has been machine-learned. In this case, the machine learning method of the trained model is not limited, and as an example, a decision tree-based, linear regression, or neural network method may be used, or two or more of these methods may be used. Here, as an example, the input of the trained model includes a set of attribute names and attribute values ​​of a record that includes an error. As an example, the output of the trained model includes corrected attribute values. However, the data cleansing method performed by the data cleansing unit 18 is not limited to the above example, and other methods may also be used. For example, the data cleansing unit 18 may perform rule-based data correction.

[0042] (Evaluation Department) The evaluation unit 19 evaluates the corrected data D after the errors have been corrected by the data cleansing unit 18. clean Analytical model MD clean Generate the analytical model MD clean Here, the evaluation unit 19 evaluates the performance of the corrected data D after the errors have been corrected by the data cleansing unit 18. clean The sequential determination process is stopped when the result of evaluating the analysis model MD satisfies a predetermined condition. clean The condition is that the mean square error (MSE) of the predicted value indicating the prediction result by the decision unit 13 is less than a predetermined threshold. The decision unit 13 and the evaluation unit 19 are an example of a decision means according to the present specification.

[0043] (Analysis result output section) The analysis result output unit 20 outputs information indicating the analysis result. The information indicating the analysis result is, for example, the corrected data D clean , and analytical model MD clean The information indicating the analysis result may include at least one of the influence degree S calculated by the calculation unit 12 and the second influence degree for each partial data. The analysis result output unit 20 may output the information indicating the analysis result by transmitting the information to another device connected via the communication unit 40A, or may output the information to an output device connected via the input / output unit 30A. The analysis result output unit 20 may output the information by writing the information to the storage unit 20A or another external storage device.

[0044] (Storage part) The storage unit 20A stores target data D, evaluation data D1', D2', . . . , D n ´, corrected data D clean , analytical model MD init ,Analysis models MD1´,MD2´,…,MD n ´, analytical model MD clean , are stored. In the following, analytical model MD, analytical models MD1', MD2', ..., MDn ´, analytical model MD clean , and , when it is not necessary to distinguish between them, they are also simply referred to as "analysis model MD." Here, "the analysis model MD is stored in the storage unit 20A" means that parameters that define the analysis model MD are stored in the storage unit 20A.

[0045] <Flow of information processing method> The flow of an information processing method S1A, which is an example of an information processing method according to exemplary embodiment 2, will be described with reference to Fig. 4. Fig. 4 is a flow diagram showing the flow of the information processing method S1A.

[0046] (Step S101) In step S101, the acquisition unit 11 acquires target data D and an analysis task. In this example, the target data D is training data D used to generate an analytical model. train and test data D for evaluating the performance of the analytical model. test The acquisition unit 11 may receive the target data D and the analysis task from another device via the communication unit 40A, or may acquire the target data D and the analysis task from an input device connected via the input / output unit 30A. The acquisition unit 11 may also acquire the target data D and the analysis task by reading the target data D and the analysis task from the storage unit 20A or another external storage device.

[0047] (Step S102) In step S102, the error detection unit 14 detects multiple errors contained in the target data D and outputs an error index indicating the location of the error. As an example, the error detection unit 14 detects errors using a rule-based detection method. Alternatively, the error detection unit 14 may detect errors by inference using a trained model generated by machine learning.

[0048] Fig. 5 is a diagram showing specific examples of errors detected by error detection unit 14. In the example of Fig. 5, error detection unit 14 determines that an error has occurred when, for example, an attribute value is missing, the attribute value of a predetermined attribute name is not within a predetermined range, the attribute value of a first attribute name and the attribute value of a second attribute name are inconsistent, or the format of the attribute value of a predetermined attribute name is incorrect. In the example of Fig. 5, error detection unit 14 detects errors E1 to E5 in target data D.

[0049] (Step S103) In step S103, the grouping unit 15 divides the errors detected by the error detection unit 14 into a plurality of groups, and defines a set of error groups G={g1, g2, . . . , g n} is output.

[0050] Fig. 6 is a diagram showing a specific example of grouping by the grouping unit 15. In the example of Fig. 6, the grouping unit 15 classifies a plurality of errors E1 to E5 into four groups: a group g1 of missing values, a group g2 of format errors, a group g3 of inconsistencies, and a group g4 of outliers.

[0051] (Step S104) In step S104, the evaluation data generating unit 16 generates the evaluation data for the groups g1, g2, . . . , g n For each group, we add errors similar to the errors in that group and create new evaluation data D i ´ is generated.

[0052] Figure 7 shows the evaluation data D i 7 is a diagram showing a specific example of a missing value group g1. In the example of FIG. 7, the evaluation data generation unit 16 generates evaluation data D1' corresponding to the missing value group g1 by replacing some of the attribute values ​​of the records included in the target data D with missing values ​​E11. In addition, the evaluation data generation unit 16 generates evaluation data D2' corresponding to the format error group g2 by replacing the attribute value of "zip code" of the records included in the target data D with attribute value E12 with the hyphen deleted.

[0053] (Step S105) In step S105, the influence calculation unit 17 calculates n pieces of evaluation data D i Each of the data is used as training data for the analytical model MD i ´ and generate the analytical model MD i In this example, the analytical model MD i ´ and analytical model MD init are models corresponding to the analysis tasks acquired by the acquisition unit 11 in step S101, and are generated by a common generation method corresponding to the analysis tasks.

[0054] As an example, the influence calculation unit 17 uses a function eval() to evaluate the analytical model to calculate the generated analytical model MD i Here, the function eval() is a function that takes an analytical model as input and outputs a score that evaluates its performance. In other words, the impact s i teeth, s i =eval(MD i ´) The performance evaluation index for the analysis is arbitrary, but as an example, MSE (mean square error) may be calculated during the above regression analysis, or the difference from the MSE calculated for the target data D, which is the original data, may be calculated.

[0055] FIG. 8 shows the influence s calculated by the influence calculation unit 17. i 8 is a diagram for explaining a specific example of the above. In the example of FIG. 8, the horizontal axis indicates the number of increased errors, and the vertical axis indicates the analytical performance of the analytical model. In the example of FIG. 8, the analytical model MD4' generated using the evaluation data D4' is init The performance of the analytical model MD3' generated using the evaluation data D3' is 0.1 lower than that of the analytical model MD init In addition, the analytical model MD1' generated using the evaluation data D1' is initThe performance of the analytical model MD2' generated using the evaluation data D2' is 0.3 times lower than that of the analytical model MD init In the example of FIG. 8, the impact calculation unit 17 calculates the performance of the analysis model MD init Analysis model MD for i The degree of performance degradation of ´ is calculated as the impact.

[0056] (Step S106) In step S106, the determination unit 13 determines the n evaluation results (influence s i ) is a set of influences S={s1,s2,…,s n} to determine the data to be modified. In this example, the input of the determination unit 13 includes the target data D and the impact S. The output of the determination unit 13 includes the priority I of the modification of the data record. In other words, in this operation example, the determination unit 13 determines the priority of the data to be modified based on the impact S.

[0057] The method for selecting the data to be corrected is arbitrary, but as an example, the determination unit 13 uses the influence S calculated by the calculation unit 12 to calculate a second influence on the above evaluation index for each of the multiple records included in the target data D, and determines the record to be corrected based on the calculated second influence for each record.

[0058] 9 is a diagram showing a specific example of the determination process of the determination unit 13. In the example of FIG. 9, the target data D includes records r1 to r3. In the example of FIG. 9, the impact s corresponding to the attribute of the error included in each record is i The sum of these is calculated as the second influence of each record.

[0059] In the example of Figure 9, the influence of group g1 is "0.3", the influence of group g2 is "0.2", and the influence of group g3 is "0. 5" and the influence of group g4 is "0.1", the second influences of records r1 to r3 will be as follows: Record r1 contains two errors in group g2, so the second influence of record r1 is 0.5+0.5=1. Record r2 contains one error in group g1 and one error in group g4, so the second influence of record r2 is 0.3+0.1=0.4. Record r3 contains one error in group g3, so the second influence of record r3 is 0.2. In the example of FIG. 9, the determination unit 13 determines the record with the highest second influence as the record to be corrected.

[0060] (Step S107) In step S107, the data cleansing unit 18 corrects the data determined in step S106. Here, the input to the data cleansing unit 18 includes, for example, the target data D and the order priority I of the record to be corrected. The output of the data cleansing unit 18 includes, for example, corrected data D obtained by correcting the record to be corrected in the target data D. clean Includes.

[0061] In step S107, the number of records to be modified at one time by the data cleansing unit 18 may be set in advance. In this case, the data cleansing unit 18 selects the preset number of records from among the multiple records to be modified based on priority I, and modifies the selected records.

[0062] The data cleansing unit 18 may modify the data in any manner, but as an example, the data cleansing unit 18 may output a screen on a display for the user to modify the data and modify the data in accordance with the user's operation. Alternatively, the data cleansing unit 18 may modify the data to be modified using a rule-based modification method. Alternatively, the data cleansing unit 18 may modify the data by inference using a trained model generated by machine learning.

[0063] 10 is a diagram showing a specific example of data correction processing by the data cleansing unit 18. In the example of FIG. 10, the data cleansing unit 18 corrects the attribute values ​​of "age" and "annual income" of record r1 included in target data D. Corrected data D clean is the record r1 after the modification. clean Includes.

[0064] (Step S108) In step S108, the evaluation unit 19 evaluates the corrected data D clean Analytical model MD clean Generate the analytical model MD clean The evaluation method used by the evaluation unit 19 is arbitrary, but as an example, the evaluation unit 19 may perform a regression analysis on the prediction of annual income using a linear model for the annual income prediction task and evaluate the results of the analysis using the mean square error (MSE) of the predicted value.

[0065] (Step S109) In step S109, the evaluation unit 19 determines whether the evaluation result satisfies a predetermined stopping condition. An example of the presentation condition is that the MSE (prediction error) is less than 0.2. If the evaluation result satisfies the stopping condition (YES in step S109), the evaluation unit 19 ends the processing. On the other hand, if the evaluation result does not satisfy the stopping condition (NO in step S109), the evaluation unit 19 returns to the processing in step S106 and continues the data correction processing.

[0066] In other words, in steps S106 to S109, the determination unit 13 sequentially determines the data to be corrected with reference to the above priority order, and the evaluation unit 19 generates corrected data D in which the data determined by the determination unit 13 in the target data D is corrected. clean If the evaluation result satisfies a predetermined target value, the sequential determination process is stopped.

[0067] <Effects of information processing devices> However, in large-scale data, it is not realistic to correct all errors before performing data analysis. This is because correcting all errors contained in large-scale data requires a huge amount of time and cost. In contrast, in this exemplary embodiment, by prioritizing cleansing of errors that have a large impact on the analysis task, it is possible to achieve more accurate data analysis while reducing costs.

[0068] Furthermore, conventional data cleansing techniques have the problem that the errors that can be corrected and the machine learning models that can be applied are limited. Furthermore, it is known that the errors to be corrected differ depending on the type of analysis task in machine learning, and it is not possible to perform error correction that takes the analysis task into consideration. In contrast, the information processing device 1A according to this exemplary embodiment calculates the impact based on the machine learning model (i.e., the analysis task) for each type of error and determines the data to be corrected. Therefore, the present invention has the effect of enabling error correction that takes the analysis task into consideration for any machine learning model, regardless of the type of error.

[0069] In the information processing device 1A according to this exemplary embodiment, the calculation unit 12 calculates an influence degree s for each attribute corresponding to each group obtained by grouping the plurality of errors according to the characteristics of the errors. i Therefore, the information processing device 1A according to this exemplary embodiment can determine data to be corrected by taking into account the degree of influence of each group divided according to the characteristics of the error.

[0070] In the information processing device 1A according to this exemplary embodiment, the calculation unit 12 generates evaluation data D1′, D2′, . . . , D n ' is generated for each error or for each attribute of the error, and the generated evaluation data D1', D2', ..., D nTherefore, according to the information processing device 1A according to this exemplary embodiment, the impact is calculated using evaluation data generated for each error or for each attribute of the error, thereby making it possible to more accurately determine the data to be corrected.

[0071] In the information processing device 1A according to this exemplary embodiment, the calculation unit 12 calculates the performance of the analytical model MD generated using the target data D and the performance of the evaluation data D1′, D2′, . . . , D n The analytical models MD1', MD2', ..., MD n Based on the comparison results with the performance of each of the n Therefore, according to the information processing device 1A according to this exemplary embodiment, the degree of influence is calculated based on a change in performance of the analytical model generated using evaluation data including pseudo errors, thereby making it possible to more accurately determine the data to be corrected.

[0072] Furthermore, in the information processing device 1A according to this exemplary embodiment, the determining unit 13 uses the influence S calculated by the calculating unit 12 to calculate a second influence on the evaluation index for each of the multiple records included in the target data D, and determines the record to be corrected based on the calculated second influence for each record. Therefore, the information processing device 1A according to this exemplary embodiment can more preferably select the record to be corrected from the multiple records.

[0073] Furthermore, in the information processing device 1A according to this exemplary embodiment, the determination unit 13 is configured to determine the priority of data to be corrected based on the impact S. Therefore, according to the information processing device 1A according to this exemplary embodiment, the priority of data to be corrected is determined based on the impact of an error, thereby making it possible to more appropriately determine the priority.

[0074] Furthermore, in the information processing device 1A according to this exemplary embodiment, the determination unit 13 is configured to sequentially determine the data to be corrected by referring to the priority order. Therefore, according to the information processing device 1A according to this exemplary embodiment, the process of sequentially determining the data to be corrected can be performed with higher accuracy.

[0075] In the information processing device 1A according to this exemplary embodiment, the determination unit 13 generates corrected data D clean The present invention employs a configuration in which the sequential determination process is stopped when the evaluation result of the above condition satisfies a predetermined target value. By repeatedly performing cleansing until the stopping condition is met, it is possible to achieve higher accuracy of data analysis at a certain cost than before, and also to reduce the cost required to achieve a certain accuracy target than before. In this way, according to this exemplary embodiment, it is possible to achieve data cleansing in which the quality of the target data meets a predetermined target value while reducing the processing load related to data cleansing.

[0076] [Software implementation example] Some or all of the functions of the information processing devices 1 and 1A may be realized by hardware such as an integrated circuit (IC chip), or may be realized by software.

[0077] In the latter case, the information processing devices 1 and 1A are realized, for example, by a computer that executes instructions of a program, which is software that realizes each function. An example of such a computer (hereinafter referred to as computer C) is shown in FIG. 11. The computer C includes at least one processor C1 and at least one memory C2. The memory C2 stores a program P for operating the computer C as the information processing device 1 and 1A. In the computer C, the processor C1 reads and executes the program P from the memory C2, thereby realizing each function of the information processing device 1 and 1A.

[0078] The processor C1 may be, for example, a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), a micro processing unit (MPU), a floating point number processing unit (FPU), a physics processing unit (PPU), a microcontroller, or a combination thereof. The memory C2 may be, for example, a flash memory, a hard disk drive (HDD), a solid state drive (SSD), or a combination thereof.

[0079] The computer C may further include a RAM (Random Access Memory) for expanding the program P during execution and for temporarily storing various data. The computer C may also include a communication interface for transmitting and receiving data to and from other devices. The computer C may also include an input / output interface for connecting input / output devices such as a keyboard, mouse, display, and printer.

[0080] Furthermore, the program P can be recorded on a non-transitory tangible recording medium M that can be read by the computer C. Such a recording medium M can be, for example, a tape, a disk, a card, a semiconductor memory, or a programmable logic circuit. The computer C can acquire the program P via such a recording medium M. The program P can also be transmitted via a transmission medium. Such a transmission medium can be, for example, a communication network or broadcast waves. The computer C can also acquire the program P via such a transmission medium.

[0081] [Appendix 1] The present invention is not limited to the above-described embodiments, and various modifications are possible within the scope of the claims. For example, embodiments obtained by appropriately combining the technical means disclosed in the above-described embodiments are also included in the technical scope of the present invention.

[0082] [Appendix 2] Some or all of the above-described embodiments can also be described as follows: However, the present invention is not limited to the following described aspects. (Appendix 1) An information processing device comprising: an acquisition means for acquiring target data; a calculation means for calculating the degree of influence that multiple errors contained in the target data have on an evaluation index of a machine learning model for each of the errors or for each attribute of the errors; and a determination means for determining data to be corrected from the target data based on the degree of influence calculated by the calculation means.

[0083] (Appendix 2) The information processing device according to claim 1, wherein the calculation means calculates the degree of impact for each attribute corresponding to each group obtained by dividing the plurality of errors into groups according to the characteristics of the errors.

[0084] (Appendix 3) The information processing device described in Appendix 1 or 2, wherein the calculation means generates evaluation data including a pseudo-error in the target data for each error or for each attribute of the error, and calculates the impact using the generated evaluation data.

[0085] (Appendix 4) The information processing device described in Appendix 3, wherein the calculation means calculates the influence based on a comparison result between the performance of a machine learning model generated using the target data and the performance of each machine learning model generated using the evaluation data.

[0086] (Appendix 5) An information processing device described in any one of Appendices 1 to 4, wherein the determination means uses the influence calculated by the calculation means to calculate a second influence on the evaluation index for each of multiple partial data included in the target data, and determines the partial data to be modified based on the calculated second influence for each partial data.

[0087] (Appendix 6) 6. The information processing device according to claim 1, wherein the determining means determines a priority order of data to be modified based on the degree of impact.

[0088] (Appendix 7) 7. The information processing device according to claim 6, wherein the determining means sequentially determines the data to be modified by referring to the priority order.

[0089] (Appendix 8) 8. The information processing device according to claim 7, wherein the determination means stops the sequential determination process when an evaluation result of corrected data obtained by correcting the determined data satisfies a predetermined target value.

[0090] (Appendix 9) An information processing method including the steps of: acquiring target data; calculating, by at least one processor, the degree of influence that multiple errors contained in the target data have on an evaluation index of a machine learning model for each of the errors or for each attribute of the errors; and determining data to be corrected from the target data based on the calculated degree of influence.

[0091] (Appendix 10) A program that causes a computer to function as an acquisition means for acquiring target data, a calculation means for calculating the degree of influence that multiple errors contained in the target data have on the evaluation index of a machine learning model for each of the errors or for each attribute of the errors, and a determination means for determining data to be corrected from the target data based on the degree of influence calculated by the calculation means.

[0092] [Appendix 3] Some or all of the above-described embodiments can also be expressed as follows. An information processing device comprising at least one processor that executes an acquisition process for acquiring target data, a calculation process for calculating the degree of influence that multiple errors contained in the target data have on the evaluation index of a machine learning model for each of the errors or for each attribute of the errors, and a determination process for determining data to be corrected from the target data based on the degree of influence calculated in the calculation process.

[0093] The information processing device may further include a memory that stores a program for causing the processor to execute the acquisition process, the calculation process, and the determination process. The program may be recorded on a computer-readable, non-transitory, tangible recording medium. [Explanation of symbols]

[0094] 1, 1A Information processing equipment 11 Acquisition Department 12 Calculation section 13 Decision Section

Claims

1. An acquisition means for acquiring target data; a calculation means for calculating the degree of influence that a plurality of errors contained in the target data has on an evaluation index of a machine learning model for each of the errors or for each attribute of the errors; a determining means for determining data to be corrected from the target data based on the degree of influence calculated by the calculating means; A correction data determination device comprising:

2. the calculation means calculates the degree of influence for each attribute corresponding to each group obtained by dividing the plurality of errors into groups according to the characteristics of the errors; 2. The apparatus for determining correction data according to claim 1.

3. the calculation means generates evaluation data including a pseudo error in the target data for each error or for each attribute of the error, and calculates the impact using the generated evaluation data.

3. The apparatus for determining correction data according to claim 1 or 2.

4. the calculation means calculates the influence level based on a comparison result between the performance of a machine learning model generated using the target data and the performance of each machine learning model generated using the evaluation data; 4. The apparatus for determining correction data according to claim 3.

5. The determining means using the influence calculated by the calculation means, calculate a second influence on the evaluation index for each of a plurality of partial data included in the target data; determining the partial data to be corrected based on the calculated second influence degree for each partial data; 3. The apparatus for determining correction data according to claim 1 or 2.

6. the determining means determines the priority of the data to be corrected based on the degree of influence.

3. The apparatus for determining correction data according to claim 1 or 2.

7. the determining means sequentially determines the data to be corrected by referring to the priority order.

7. The apparatus for determining correction data according to claim 6.

8. the determining means stops the sequential determining process when an evaluation result of corrected data obtained by correcting the determined data satisfies a predetermined target value; 8. The apparatus for determining correction data according to claim 7.

9. At least one processor Obtaining target data; Calculating the degree of influence that a plurality of errors contained in the target data have on the evaluation index of the machine learning model for each of the errors or for each attribute of the errors; determining data to be corrected from the target data based on the calculated degree of influence; A method for determining correction data, comprising:

10. Computer, An acquisition means for acquiring target data; a calculation means for calculating the degree of influence that a plurality of errors contained in the target data has on an evaluation index of a machine learning model for each of the errors or for each attribute of the errors; a determining means for determining data to be corrected from the target data based on the degree of influence calculated by the calculating means; A program that functions as a

Images