Ethics diagnostic device and ethics diagnostic method

Abstract

To appropriately diagnose the ethicality of an AI model prediction result.SOLUTION: An ethicality diagnosis device stores: sensitive feature data which is data obtained by associating a value for a sensitive feature amount that is a feature amount for which a certain amount of care is required in handling from an ethical standpoint and a value for a selected feature amount that is one or more feature amounts selected from feature amounts constituting an AI model; a sensitive feature coefficient which is a value indicating a degree of impact imparted by each selected feature amount on the sensitive feature amount by analyzing a relation between the value for the sensitive feature amount and the value for the selected feature amount; and an importance level per feature amount which is a value indicating a degree of impact imparted by each selected feature amount on an AI model prediction result. The ethicality diagnosis device obtains a non-ethical degree which is a value indicating a degree of ethicality of a prediction result output by the AI model on the basis of the sensitive feature coefficient and the importance level per feature amount.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The present invention relates to an ethical diagnostic device and an ethical diagnostic method. [Background technology]

[0002] In recent years, systems that utilize AI models (AI: Artificial Intelligence) have been used in a variety of fields. However, ensuring the ethics and fairness of AI models has become an issue. For example, if the training data used to train an AI model is influenced by prejudices or gaps due to gender, age, race, ethnicity, etc. (if there is bias in the training data), the output of the AI ​​model will also be influenced by these prejudices and gaps.

[0003] Regarding the ethics of AI models, for example, Patent Document 1 describes an evaluation device configured to efficiently and reliably evaluate the risks of models installed in white-box AI systems and analytical engines. The evaluation device acquires one or more explainable predictive models, determines the risks of the one or more models based on the one or more models and ethical risk factor information, which is information that becomes ethical risk factors, and selects and outputs a model based on the risk assessment results. The evaluation device generates a statement describing the one or more models in a language for each of the one or more models based on the relationships between the elements of the one or more models, and determines the risks of the one or more models using the statement, at least one of the elements of the statement, and the ethical risk factor information.

[0004] Furthermore, Non-Patent Document 1 describes a tool that was created based on the premise that if AI is trained using training data based on achievements and trends that are biased (existent bias) due to customs and historical background, the direction of learning may change significantly.The document aims to use the above tool to investigate, report, and mitigate biases that are included in the results derived by AI and stem from attributes such as race, gender, region, and age. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] International Publication No. 2021 / 199201 [Non-patent literature]

[0006] [Non-Patent Document 1] IBM Research Trusted AI, “AI Fairness 360”, [online], March 24, 2020, IBM Research & IBM, [searched on March 24, 2020], Internet,<URL:http: / / aif360.mybluemix.net / > Summary of the Invention [Problem to be solved by the invention]

[0007] The risk assessment device described in Patent Document 1 generates sentences that represent the relationship between explanatory variables and objective variables of a model, calculates the similarity between the features of the generated sentences and the features of ethical risk factor information, and assesses ethical risk based on the frequency of occurrences with a predetermined similarity. Furthermore, the technology described in Non-Patent Document 1 provides tools for reducing bias in training data, models during training, and predicted labels. However, the technologies described in both documents are mechanisms for assessing the ethics of a model before or after it is applied to an actual usage scenario, and do not immediately assess the ethics of the prediction results output by the model in an actual usage scenario.

[0008] Furthermore, the tool described in Non-Patent Document 1 has a function that arbitrarily changes the prediction results output by the AI ​​model, and using this function may lead to a decline in the quality of the model. Furthermore, it is difficult to completely eliminate the effects of bias, and the tool described in Non-Patent Document 1 does not guarantee that the prediction results output by the AI ​​model will not contain ethical issues.

[0009] The present invention has been made in consideration of the above background, and aims to provide an ethics diagnosis device and an ethics diagnosis method that can appropriately diagnose the ethics of the prediction results output by an AI model. [Means for solving the problem]

[0010] One aspect of the present invention for achieving the above object is an ethics diagnosis device that diagnoses the ethics of prediction results output by an AI model, the device being configured using an information processing device having a processor and a storage device, and storing sensitive feature data, which is data that associates values ​​of sensitive features, which are features that require a certain degree of consideration in handling from an ethical perspective, with values ​​of selected features, which are one or more features selected from the features of an AI model, and analyzing the relationship between the values ​​of the sensitive features and the values ​​of the selected features, to store sensitive feature coefficients, which are values ​​indicating the degree of influence that each of the selected features has on the sensitive feature, and feature importance, which is a value indicating the degree of influence that each of the selected features has on the prediction result of the AI ​​model, and then calculating a level of unethics, which is a value indicating the degree of ethics of the prediction result output by the AI ​​model, based on the sensitive feature coefficients and the feature importance.

[0011] Other problems and solutions disclosed in the present application will be made clear in the detailed description and drawings. [Effects of the Invention]

[0012] According to the present invention, it is possible to appropriately diagnose the ethics of the prediction results output by an AI model. [Brief explanation of the drawings]

[0013] [Figure 1] FIG. 2 is a diagram illustrating an example of main functions of the ethical diagnostic device. [Figure 2] FIG. 2 is a system flow diagram illustrating an example of the main functions of the ethical diagnostic device. [Figure 3] FIG. 10 is a diagram illustrating an example of S feature data. [Figure 4] FIG. 10 is a diagram showing an example of the results of logistic regression analysis. [Figure 5] 10 is an example of a prediction / diagnosis result display screen. [Figure 6] 10 is an example of a detailed diagnosis screen for each S feature amount. [Figure 7] 1 is an example of an information processing device used to configure an ethical diagnostic device. DETAILED DESCRIPTION OF THE INVENTION

[0014] Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings. In the following description, various data may be described using expressions such as "information" and "data," but the various data may be expressed or managed in a manner other than the data structure shown as an example.

[0015] FIG. 1 is a block diagram showing the main functions of a system (hereinafter referred to as "ethics diagnostic device 100") that diagnoses the ethics of prediction results output by an AI model (machine learning model, hereinafter referred to as "model"), shown as one embodiment. FIG. 2 is a system flow diagram explaining the main functions of ethics diagnostic device 100. Ethics diagnostic device 100 is configured using one or more information processing devices (computers). Below, the main functions of ethics diagnostic device 100 will be explained with reference to these figures.

[0016] In this embodiment, an example will be described in which the model to be diagnosed is a model that outputs information (hereinafter referred to as "evaluation information") regarding the evaluation of the skills of an applicant for a job opening at a company (hereinafter referred to as "interviewee") by inputting features (voice pitch, voice volume, gaze direction, facial expression, number of nods, heart rate, etc.) extracted from video data of an interview conducted with the interviewee.

[0017] The ethical assessment device 100 is configured to assess features that require certain consideration in handling from an ethical perspective (e.g., race, sex, nationality, age, employment status, place of origin, place of residence, gender minority, religion, physical / intellectual disability, ideology, etc., hereinafter referred to as "sensitive features"). The ethicality of the prediction results output by the model is diagnosed by focusing on the "S feature" (or "S feature").

[0018] As shown in FIG. 1 or 2, the ethics diagnostic device 100 includes the functions of a memory unit 110, an information acquisition management unit 130, a feature extraction unit 135, a learning data generation unit 140, a model learning unit 145, a prediction unit 150, a feature-by-feature importance calculation unit 155, a feature selection unit 160, an S feature data generation unit 165, an S feature data analysis unit 170, an ethics diagnostic unit 175, and a prediction / diagnosis result output unit 180.

[0019] Of the above functions, the storage unit 110 stores information (data) such as input data 111, features 112, correct labels 113, training data 114, models 115, prediction results 116, importance of each feature 117, selected features 118, S features 119, S feature data 120, S feature coefficients 121, and prediction / diagnosis results 122.

[0020] Among these, the input data 111 is data from which feature quantities 112 to be input to the model 115 are extracted. In this embodiment, as an example, the input data 111 is video data of a captured image of an interviewee.

[0021] The feature 112 is a feature 112 extracted by the feature extraction unit 135 from the input data 111. In this embodiment, the feature 112 is, for example, the interviewee's "voice pitch," "voice volume," "number of gaze deviations," "average heart rate," "variance of nods," "minimum value of surprise emotion," etc. The feature 112 is not only provided to the model 115 in an actual application scenario of the model 115, but is also used to generate the training data 114. In the former case, the feature 112 is, for example, a feature extracted from video data capturing the interviewee's behavior, and in the latter case, the feature 112 is, for example, a feature extracted from video data of another interviewee captured in the past.

[0022] The correct label 113 is a correct label of evaluation information that is assigned to the feature 112 when generating the training data 114. In this embodiment, the correct label 113 is, for example, a numerical value that indicates the level of skill of the interviewee.

[0023] The training data 114 is data (teaching data) used for training the model 115. The training data 114 is generated by assigning a correct label 113 to sample data of the feature 112 (a value for each feature 112 generated based on the input data 111).

[0024] The model 115 is a machine learning model that outputs the results of learning using training data 114 for input feature quantities 112 as prediction results 116. In this embodiment, the model 115 outputs, as evaluation information, evaluation scores (e.g., evaluation scores on a five-point scale) of the interviewee for each of the preset evaluation items (level of listening, volume of voice, ability to understand questions, gaze, facial expression, etc.). The type of model 115 is not limited, but may be, for example, regression (linear regression, logistic regression, etc.). These include neural networks (e.g., convolutional neural networks), trees (e.g., decision trees, random forests, gradient boosting), and neural networks (e.g., convolutional neural networks).

[0025] The prediction result 116 is information output by the model 115 in response to the input value of the feature quantity 112. In this embodiment, the prediction result 116 is, for example, the evaluation score of the interviewee for each of the above evaluation items.

[0026] The feature amount importance 117 is a value indicating the degree of influence that each feature amount 112 has on the prediction result 116. A method for calculating the feature amount importance 117 will be described later.

[0027] The selected feature 118 is a feature selected by the feature selection unit 160 from the feature 112 provided to the model 115. The selected feature 118 is used to generate S feature data 120.

[0028] The S feature 119 is the S feature described above.

[0029] The S feature data 120 is data in which the values ​​of one or more selected feature quantities 118 are associated with the S feature quantity.

[0030] The S feature coefficient 121 is a value indicating the degree of influence that each of the selected features 118 has on the S feature.

[0031] The prediction / diagnosis result 122 is information about the result of the ethical diagnosis by the ethicality diagnosis unit 175 of the prediction result output by the model 115. As will be described later, the ethicality diagnosis unit 175 calculates a value (index) indicating the degree of ethicality of the prediction result output by the model 115 (hereinafter referred to as "unethicality") based on the importance of each feature 117 and the S feature coefficient 121, and outputs the calculated unethicality and information based on the unethicality as the prediction / diagnosis result 122.

[0032] The information acquisition and management unit 130 shown in FIG. 1 acquires various information (input data 111, correct answer label 113, designation (or selection criteria) of selected features 118, S features 119, etc.) used to diagnose the ethics of the prediction results output by the model 115 via a user interface, a communication network, etc., and manages the acquired information in the storage unit 110.

[0033] The feature extraction unit 135 extracts feature quantities 112 from the input data 111. The method for extracting the feature quantities 112 is not necessarily limited. In this embodiment, the feature extraction unit 135 extracts the feature quantities 112 by, for example, performing principal component analysis on optical flows acquired from video data and identifying representative features from the eigenvalues.

[0034] The training data generation unit 140 generates training data 114 by assigning a correct answer label 113 to the feature 112. The correct answer label 113 is set by the user via a user interface, for example.

[0035] The model learning unit 145 learns the model 115 based on the training data 114. For example, the model learning unit 145 inputs the values ​​of the feature quantities 112 in the training data 114 to the model 115, compares the values ​​output by the model 115 with the labels of the training data 114, and adjusts the parameters constituting the model 115 based on the difference (feedback of the difference), thereby learning the model 115.

[0036] The prediction unit 150 predicts the input data 111 (video) in an actual application scenario of the model 115. The feature quantities 112 extracted from the data (data) are input to a model 115, and the information output by the model 115 is obtained as a prediction result 116. The prediction result 116 is provided to a user, such as a human resources officer who screens the interviewee, via a user interface, for example.

[0037] The feature amount importance calculation unit 155 calculates the feature amount importance 117. The method for calculating the feature amount importance 117 is not necessarily limited, but the feature amount importance calculation unit 155 calculates the feature amount importance 117 using a method such as "SHAP (Shapley Additive exPlanations)", "Shapley Value", "Cohort Shapley Value", or "Local Permutation Importance", for example.

[0038] The feature selection unit 160 selects a predetermined number of selected features 118 from the features 112 extracted by the feature extraction unit 135. Note that the feature selection unit 160 may not only select some of the features 112 extracted by the feature extraction unit 135 as the selected features 118, but may also select all of the features 112 as the selected features 118.

[0039] The S feature data generation unit 165 generates the S feature data 120 by associating the values ​​of one or more selected features 118 with the values ​​of the S feature. The S feature data generation unit 165 receives settings of the S feature to be associated with the selected feature 118 and the respective values ​​from the user, for example, via a user interface.

[0040] 3 shows an example of the S feature data 120. The illustrated S feature data 120 is made up of multiple records each having the following items: data ID 1191, interviewee ID 1192, S feature amount 1193, and selected feature amount 1194. One record of the S feature data 120 corresponds to one piece of sample data (combination of values ​​of each selected feature amount) extracted from the input data 111 (video data).

[0041] Of the above items, the data ID 1191 stores a data ID that is an identifier of the sample data. The interviewee ID 1192 stores an interviewee ID that is an identifier of the interviewee. The S feature 1193 stores the value of the S feature described above. The selected feature 1194 stores the value of each of one or more selected features 118 associated with the S feature.

[0042] 3 may be generated and displayed via a user interface. Also, a user interface for editing the contents of the screen may be provided so that the user can edit the contents of the S feature data 120.

[0043] 1 or 2 analyzes the S feature data 120 to determine the S feature coefficients 121. In this embodiment, the S feature data analysis unit 170 performs logistic regression analysis using the S feature quantities as the objective variables and the selected feature quantities (for example, the selected feature quantities normalized to Z values ​​(mean "0" and variance "1")) as the explanatory variables, and normalizes the obtained regression coefficients so that the sum of their absolute values ​​becomes "1.0" to determine the S feature coefficients 121.

[0044] The number of selected features (explanatory variables) used in the logistic regression analysis is, for example, 1 / 10 of the number of sample data for the smaller of the numbers of sample data for each value that the S feature can take. For example, if the S feature is "gender" and the number of sample data for "male" is 40 and the number of sample data for "female" is 60, the number of selected features (explanatory variables) is set to 4, which is the number of sample data for males with the smaller number of sample data, "40", multiplied by 1 / 10.

[0045] In addition, for example, when multicollinearity is recognized between selected features (explanatory variables), one of the selected features that is correlated may be excluded. A regression analysis is performed using a feature selection algorithm on the data, and if the VIF (Variance Inflation Factor) calculated from the following formula (hereinafter referred to as "Formula 1") exceeds a preset threshold, In this case, one of the selected features is excluded. i is the multiple correlation coefficient (i is a natural number assigned to each combination of explanatory variables).

number

[0046] Furthermore, when performing logistic regression analysis (S feature data 120) on multiple combinations of S features (objective variables) and selected features (explanatory variables) with different selected features (varied selected features) for comparison, the Matthews Correlation Coefficient (MCC) may be calculated by cross-validation, and the combination with the largest MCC may be selected from the combinations. In this case, the S feature coefficient 121 may be calculated by multiplying the normalized regression coefficient by the MCC, for example, so that the comparison results of the multiple combinations are reflected in the S feature coefficient 121.

[0047] In this embodiment, the degree of influence of the selected feature (explanatory variable) on the S feature (objective variable) is calculated by logistic regression analysis. However, the degree of influence may be calculated by other methods.

[0048] Figure 4 shows an example of the results of logistic regression analysis. This figure shows the analysis results when the value of the S feature (objective variable) "gender" is set to "male." In this example, the regression coefficient values ​​for each selected feature, such as "voice pitch," "average voice volume," and "variance of gaze deviation frequency," obtained by logistic regression analysis, are normalized so that the sum of their absolute values ​​is "1.0." The S characteristic coefficient is set to 121.

[0049] For example, a screen showing the contents of FIG. 4 may be displayed via a user interface so that the user can check the results of the logistic regression analysis.

[0050] 1 or 2 calculates the degree of unethics based on the importance of each feature 117 and the S feature coefficient 121, and outputs the calculated degree of unethics as a prediction / diagnosis result 122. For example, the ethical diagnosis unit 175 calculates the degree of unethics as follows.

[0051] First, normalize the importance of each feature so that the sum of its absolute values ​​is "1.0". Using the formula (hereinafter referred to as "Formula 2"), the sum of the values ​​obtained by multiplying the importance of each feature and the S feature coefficient is calculated as the degree of unethicality for each prediction result.

number

[0052] The prediction / diagnosis result output unit 180 shown in FIG. 1 or 2 outputs, via a user interface, an image showing the contents of the prediction result 116 and the contents of the prediction / diagnosis result 122 (the ethical diagnosis result). A screen (hereinafter referred to as "prediction / diagnosis result presentation screen 500") is generated and output.

[0053] 5 is an example of a prediction / diagnosis result presentation screen 500. As shown in the figure, the prediction / diagnosis result presentation screen 500 has an evaluation item selection field 511, an interview theme selection field 512, a video display field 513, an interviewee evaluation result confirmation field 514, and an unethicality level display field 515.

[0054] A user such as a human resources person can select an evaluation item using a pull-down menu in the evaluation item selection field 511. In this example, the user selects "listening level."

[0055] In the interview theme selection field 512, the user can select an interview theme by operating the mouse, keyboard, etc. In this example, the user selects "Theme 2."

[0056] In the video display field 513, a playback video of video data taken when an interviewee is interviewed on the interview theme selected by the user in the interview theme selection field 512 is displayed.

[0057] The interviewee evaluation result confirmation field 514 displays the evaluation result of the interviewee predicted by the prediction unit 150 using the model 115. As shown in the figure, the interviewee evaluation result confirmation field 514 is provided with a pull-down menu for correcting the evaluation result, and the user can correct the evaluation result as appropriate.

[0058] The unethicality level display field 515 displays the results (unethicality level for each S feature) of the diagnosis made by the ethics diagnosis unit 175 on the ethics of the prediction result 116 when the model 115 makes a prediction using the video data displayed in the video display field 513 as input data 111. In this example, the unethicality levels of the S features "gender," "age," "place of origin," and "orientation" are displayed in the form of a bar graph.

[0059] When the user selects any one of the S features in the unethicality display field 515, the prediction / diagnosis result output unit 180 generates and outputs a screen (hereinafter referred to as the "S feature-by-feature diagnostic detail screen 600") that displays information such as the ethicality judgment result for the selected S feature, the S feature coefficient used to calculate the unethicality of the selected S feature, and the importance of each feature.

[0060] 6 shows an example of an S feature-specific diagnostic details screen 600 that is displayed when the user selects the S feature "gender" in the unethicality level display field 515 on the prediction / diagnosis result presentation screen 500. As shown in the figure, the prediction / diagnosis result presentation screen 500 has an ethicality diagnostic result display field 611, an S feature coefficient display field 612, an importance level display field 613 for each feature, and an unethicality level display field 614.

[0061] The ethical diagnosis result display field 611 displays information indicating the result of the ethical diagnosis performed by the ethical diagnosis unit 175 on the prediction result 116 output by the model 115 based on the unethicality level. For example, if the unethicality level exceeds a preset threshold (50% (0.5) in this example), the ethical diagnosis unit 175 determines that there is a problem with the ethicality of the model 115 for the corresponding S feature. If the unethicality level is equal to or less than the threshold, the ethical diagnosis unit 175 determines that there is no problem with the ethicality of the prediction result 116 for the S feature. In this example, the unethicality level is "0.67," which exceeds the threshold, and therefore, content indicating that there is a problem with the ethicality of the prediction result 116 for the S feature "gender" is displayed in the ethical diagnosis result display field 611.

[0062] The S feature coefficient display field 612 displays the value of the S feature coefficient 121 used to calculate the unethical level. The feature-by-feature importance display field 613 displays the value of the feature-by-feature importance 117 used to calculate the unethical level. In this example, the ethics diagnosis unit 175 calculates the S feature coefficients 121 of the S feature quantities "maximum voice pitch," "average voice volume," and "variance of gaze deviation frequency" by The unethical level (0.67=0.81×0.79+0.16×0.19+0.03×0.02) is calculated by substituting the value of the feature value and the value of the feature value importance 117 into Equation 2. The value is displayed.

[0063] As described above, the ethics diagnosis device 100 of this embodiment calculates the degree of unethics, which is a value indicating the ethics of the prediction result 116 output by the model, based on the S feature coefficient 121, which is a value indicating the degree of influence that each of the selected features 118 has on the S feature 119, and the feature importance 117, which is a value indicating the degree of influence that each of the selected features 118 has on the prediction result 116 of the model 115. Therefore, the ethics diagnosis device 100 can appropriately diagnose the ethics of the prediction result 116 output by the model 115.

[0064] Furthermore, according to the ethical diagnostic device 100 of this embodiment, it is possible to provide the user with an index for determining whether or not there is an ethical problem with the prediction result 116 output by the model 115. Furthermore, even if the prediction result 116 contains bias, it is possible to provide the user with information indicating whether or not there is an ethical problem.

[0065] Furthermore, since the ethical diagnostic device 100 judges the ethicality of the prediction result 116 and does not involve any arbitrary changes to the prediction result 116, it is possible to prevent the quality of the model 115 from deteriorating.

[0066] Furthermore, if the prediction result 116 of the model 115 has an ethical problem, a warning is output, so that the user can be reliably informed (made aware) that the prediction result 116 of the model 115 has an ethical problem.

[0067] FIG. 7 shows an example of the configuration of an information processing device that constitutes the ethical diagnostic device 100. The illustrated information processing device 10 includes a processor 11, a main memory device 12, an auxiliary memory device 13, an input device 14, an output device 15, and a communication device 16. Note that the illustrated information processing device 10 may be realized, in whole or in part, by using virtual information processing resources provided using virtualization technology, process space separation technology, or the like, such as a virtual server provided by a cloud system. Also, all or in part of the functions provided by the information processing device 10 may be realized, for example, by using a virtual server provided by a cloud system via an API (Application Program Interface), or the like. Alternatively, the ethics diagnostic device 100 may be configured using a plurality of information processing devices 10 connected to each other so as to be able to communicate with each other.

[0068] In the figure, the processor 11 may be, for example, a CPU (Central Processing Unit), PU (Micro Processing Unit), GPU (Graphics Processing Unit), FPGA (Field Programmable Gate Array), ASIC (Application Specific Integrated Circuit) ), AI (Artificial Intelligence) chips, etc.

[0069] The main memory device 12 is a device for storing programs and data, and is, for example, a ROM (Read Only Memory). These include non-volatile memory (NVRAM (Non Volatile RAM)), RAM (Random Access Memory), and non-volatile memory (NVRAM).

[0070] The auxiliary storage device 13 is, for example, an SSD (Solid State Drive), a hard disk drive, The auxiliary storage device 13 may be a hard disk, an optical storage device (e.g., a CD (Compact Disc), a DVD (Digital Versatile Disc)), a storage system, a reading / writing device for a recording medium such as an IC card, an SD card, or an optical recording medium, or a storage area of ​​a cloud server. Programs and data can be read into the auxiliary storage device 13 via a recording medium reading device or a communication device 16. The programs and data stored in the auxiliary storage device 13 are read into the main storage device 12 as needed.

[0071] The input device 14 is an interface that accepts input from the outside, and is, for example, a keyboard, a mouse, a touch panel, a card reader, a pen-input tablet, a voice input device, or the like.

[0072] The output device 15 is an interface that outputs various information such as the progress of processing and the results of processing. The output device 15 is, for example, a display device (liquid crystal monitor, LCD (Liquid Crystal Display), graphic card, etc.) that visualizes the various information described above, a device that converts the various information described above into audio (audio output device (speaker, etc.)), or a device that converts the various information described above into text (printer, etc.). Note that, for example, the information processing device 10 may be configured to input and output information to and from other devices via the communication device 16.

[0073] The input device 14 and the output device 15 constitute a user interface that receives information from the user and presents information to the user.

[0074] The communication device 16 is a device that realizes communication with other devices. The communication device 16 is a wired or wireless communication interface that realizes communication with other devices via a communication medium such as a communication network, and is, for example, a NIC (Network Interface Card), a wireless communication module, a USB module, or the like.

[0075] The information processing device 10 may be equipped with, for example, an operating system, a file system, a DBMS (DataBase Management System) (relational database, NoSQL, etc.), a KVS (Key-Value Store), etc.

[0076] Each function of ethics diagnostic device 100 is realized by processor 11 reading and executing a program stored in main memory device 12, or by hardware (FPGA, ASIC, AI chip, etc.) constituting ethics diagnostic device 100. Ethics diagnostic device 100 stores the various pieces of information (data) described above, for example, as a database table or a file managed by a file system.

[0077] Although one embodiment of the present invention has been described above, it goes without saying that the present invention is not limited to the above embodiment and can be modified in various ways without departing from the spirit of the present invention. For example, the above embodiment has been described in detail to clearly explain the present invention, and the present invention is not necessarily limited to an embodiment having all of the described configurations. Furthermore, it is possible to add, delete, or replace part of the configuration of the above embodiment with other configurations.

[0078] For example, the present invention is not limited to the case where the model 115 is a model that learns through supervised learning, but can also be applied to the case where the model 115 is a model that learns through unsupervised learning.

[0079] Furthermore, the above-mentioned configurations, functional units, processing units, processing means, etc. may be partly or entirely implemented in hardware by, for example, designing them as integrated circuits. Furthermore, the above-mentioned configurations, functions, etc. may be implemented in software by a processor interpreting and executing a program that implements each function. Information such as the programs, tables, and files that implement each function may be stored in a memory, a hard disk, a recording device such as an SSD (Solid State Drive), or an I / O device. It can be placed on a recording medium such as a C card, SD card, or DVD.

[0080] Furthermore, the layout of the various functional units, processing units, and databases of each information processing device described above is merely an example, and the layout of the various functional units, processing units, and databases can be changed to an optimal layout in terms of the performance, processing efficiency, communication efficiency, etc. of the hardware and software that these devices are equipped with.

[0081] Furthermore, the configuration (schema, etc.) of the database that stores the various types of data described above can be flexibly changed from the perspective of efficient use of resources, improved processing efficiency, improved access efficiency, improved search efficiency, and the like. [Explanation of symbols]

[0082] 100 Ethics diagnosis device, 110 Memory unit, 111 Input data, 112 Features, 113 Correct labels, 114 Training data, 115 Model, 116 Prediction results, 117 Importance of each feature, 118 Selected features, 119 S features, 120 S feature data, 121 S feature coefficient, 122 prediction / diagnosis result, 130 information acquisition management unit, 135 feature extraction unit, 140 learning data generation unit, 145 model learning unit, 150 prediction unit, 155 feature importance calculation unit, 160 feature selection unit, 165 S feature data generation unit, 170 S feature data analysis unit, 175 ethical diagnosis unit, 180 prediction / diagnosis result output unit

Claims

1. An ethical diagnostic device that diagnoses the ethicality of prediction results of an AI model, The information processing device includes a processor and a storage device, Sensitive feature data is data that associates values ​​of sensitive features, which are features that require a certain degree of consideration in handling from an ethical perspective, with values ​​of selected features, which are one or more features selected from the features of the AI ​​model; a sensitive feature coefficient, which is a value indicating the degree of influence of each of the selected feature quantities on the sensitive feature quantities by analyzing the relationship between the value of the sensitive feature quantity and the value of the selected feature quantity; The importance of each feature is a value indicating the degree of influence that each of the selected features has on the prediction result of the AI ​​model; and Remember, calculating a degree of unethicality, which is a value indicating the degree of ethics of the prediction result output by the AI ​​model, based on the sensitive feature coefficient and the importance of each feature; Ethics diagnostic device.

2. The ethical diagnostic device according to claim 1, L ｉ The normalized importance of each feature, s ｉ is the S feature coefficient, i is a natural number that identifies the S feature coefficient, and n is the number of selected features, and the unethical degree is calculated by the following formula: Ethics diagnostic device.

3. The ethical diagnostic device according to claim 1, performing a logistic regression analysis on the sensitive feature data, using the sensitive feature quantity as a response variable and the selected feature quantity as an explanatory variable, to obtain a regression variable as the sensitive feature coefficient; Ethics diagnostic device.

4. The ethical diagnostic device according to claim 3, generating a plurality of pieces of the sensitive feature data having different combinations of the selected feature quantities; performing a logistic regression analysis on each of the sensitive feature data; Calculating a Matthews Correlation Coefficient (MCC) for each of the sensitive feature data by cross-validation; selecting the regression coefficients calculated using the sensitive feature data that maximize the MCC as the sensitive feature coefficients; Ethics diagnostic device.

5. The ethical diagnostic device according to claim 3, excluding one of the selected features that is correlated when multicollinearity exists between the selected features; Ethics diagnostic device.

6. The ethical diagnostic device according to claim 5, Variance Inflation Factor (VIF) is used as an indicator of whether multicollinearity exists. Using determining that multicollinearity exists between the selected feature quantities when the VIF between the selected feature quantities exceeds a predetermined threshold; Ethics diagnostic device.

7. The ethical diagnostic device according to claim 1, The importance of each feature is calculated using one of "SHAP" (Shapley Additive exPlanations), "Shapley Value", "Cohort Shapley Value", and "Local Permutation Importance". Ethics diagnostic device.

8. The ethical diagnostic device according to claim 1, a user interface that accepts settings of the sensitive feature; Ethics diagnostic device.

9. The ethical diagnostic device according to claim 1, a user interface for accepting the setting of the sensitive feature data; Ethics diagnostic device.

10. The ethical diagnostic device according to claim 1, a user interface for outputting the determined unethical level or information based on the unethical level; Ethics diagnostic device.

11. The ethical diagnostic device according to claim 1, a user interface for outputting the sensitive feature coefficient and the importance of each feature used in calculating the degree of unethicality; Ethics diagnostic device.

12. The ethical diagnostic device according to claim 1, a user interface that outputs a warning when the unethical level value exceeds a preset threshold; Ethics diagnostic device.

13. An ethical diagnostic method for diagnosing the ethics of a prediction result of an AI model, An information processing device having a processor and a storage device, Sensitive feature data is data that associates values ​​of sensitive features, which are features that require a certain degree of consideration in handling from an ethical perspective, with values ​​of selected features, which are one or more features selected from the features of the AI ​​model; a sensitive feature coefficient, which is a value indicating the degree of influence of each of the selected feature quantities on the sensitive feature quantities by analyzing the relationship between the value of the sensitive feature quantity and the value of the selected feature quantity; The importance of each feature is a value indicating the degree of influence that each of the selected features has on the prediction result of the AI ​​model; and and A step of calculating a degree of unethicality, which is a value indicating the degree of ethics of the prediction result output by the AI ​​model, based on the sensitive feature coefficient and the importance of each feature. An ethical diagnostic method.

14. The ethical diagnostic method according to claim 13, The information processing device is ｉ The normalized importance of each feature, s ｉ is an S feature coefficient, i is a natural number that identifies the S feature coefficient, and n is the number of selected features, and the unethical degree is calculated using the following formula: Further implementation of this ethical diagnostic method.

15. The ethical diagnostic method according to claim 13, a step in which the information processing device performs a logistic regression analysis on the sensitive feature data, using the sensitive feature quantity as a response variable and the selected feature quantity as an explanatory variable, to obtain a regression variable as the sensitive feature coefficient; Further implementation of this ethical diagnostic method.

Images