KR20240055469A - Apparatus and Method for Fairness Visualization of AI Learning Dataset - Google Patents

Abstract

The present invention relates to an apparatus and method for visualizing the fairness of an artificial intelligence learning dataset, and more specifically, to an uploader that uploads an original dataset in batch form, receives a plurality of sensitive attributes (Sensitive Columns) as input, and A structural analysis module that derives subsets corresponding to each sensitive attribute (Sensitive Column), a sampling module that outputs a correction data set with fairness corrected after sampling according to the size of multiple subsets, and the original data set to facilitate checking data fairness. The present invention relates to an apparatus and method for visualizing the fairness of an artificial intelligence learning dataset, including a visualization module that visualizes at least one of a dataset and a correction dataset in a preset format.

Description

Apparatus and Method for Fairness Visualization of AI Learning Dataset}

The present invention relates to an apparatus and method for visualizing the fairness of an artificial intelligence learning dataset. More specifically, the present invention relates to an artificial intelligence learning dataset that outputs a correction dataset with the fairness corrected from the original dataset and visually provides the degree of bias and fairness through a fairness index. This relates to an apparatus and method for visualizing the fairness of an intelligent learning data set.

Recently, artificial intelligence algorithms have been growing rapidly, and can be used to assist or automate human judgment by predicting risks and impacts before making decisions. Artificial intelligence algorithms can learn decision-making models in various topics such as public safety, policy, finance, medicine, and employment, and these decision-making models can often even learn biases that exist in society. Decisions made by algorithms that can have a significant impact on a subject's life may be biased based on characteristics that should be unrelated to the decision, and may disadvantage individuals belonging to specific groups such as gender, race, or religion.

As artificial intelligence (AI) technology develops rapidly and is applied to various industrial fields, there is a need to discuss the adverse effects of artificial intelligence and its impact on society as a whole. In particular, fairness issues such as bias against gender, race, social groups, etc. or lack of transparency become issues.

For the proper operation and results of artificial intelligence (AI), it is very important to ensure the fairness of the dataset for learning. For example, in the case of an artificial intelligence model that determines whether or not there is a stroke, if a learning dataset with an unfair ratio of stroke patients to normal people (eg, patients: general population = 80,000: 20,000) is used, errors may occur due to the unfairness of the learning data. Results can be derived. In addition, correcting only the ratio of one sensitive attribute (eg. stroke status) in the data (eg. patient: general population = 20,000 items: 20,000 items) may result in bias in other dependent attributes (eg. patient-male: Since patient-female = 15,000: 5,000) may occur, proper results of artificial intelligence learning cannot be expected.

In addition, to date, a method of removing static data sets has been used, and additional correction technology for additionally collected data is insufficient. And even if correction technology is applied, there is no technology that can visually check the degree of correction of the dataset before and after the correction technology is applied, so it is desperately needed in this technology field.

Republic of Korea Patent Publication No. 10-2022-0090359 Republic of Korea Patent Publication No. 10-2021-0028724

The present invention is intended to solve the above problems. Artificial intelligence learning data is sampled according to the size of a plurality of subsets and outputs a correction dataset with fairness corrected to resolve the unfairness of countless data collected indiscriminately. The purpose is to obtain three fairness visualization devices and methods.

The purpose of the present invention is to provide an apparatus and method for visualizing the fairness of an artificial intelligence learning dataset that visualizes at least one of the original dataset and the correction dataset in a preset format to facilitate checking data fairness.

The technical problems to be achieved by the present invention are not limited to the technical problems mentioned above, and other technical problems not mentioned can be clearly understood by those skilled in the art from the description of the present invention.

In order to achieve the above object, the fairness visualization device of the artificial intelligence learning dataset of the present invention includes an uploader that uploads the original dataset in batch form; A structural analysis module that receives a plurality of sensitive attributes (Sensitive Columns) as input and derives a subset corresponding to each sensitive attribute (Sensitive Columns) from the original dataset; A sampling module that samples according to the size of a plurality of subsets and then outputs a correction dataset with corrected fairness; and a visualization module that visualizes at least one of the original dataset and the corrected dataset in a preset format to facilitate checking data fairness.

In order to achieve the above object, the method for visualizing the fairness of an artificial intelligence learning dataset of the present invention includes an original dataset upload step in which an original dataset in batch form is uploaded by an uploader; A subset derivation step in which a plurality of sensitive attributes (Sensitive Columns) are input by a structural analysis module, and a subset corresponding to each sensitive attribute (Sensitive Columns) is derived from the original data set; A correction data set output step in which a correction data set with fairness corrected after sampling according to the size of a plurality of subsets is output by a sampling module; and a visualization step in which at least one of the original dataset and the corrected dataset is visualized in a preset format by a visualization module to facilitate confirmation of data fairness.

As described above, according to the present invention, the unfairness of the acquired dataset can be gradually solved, and the problem that the performance of the artificial intelligence model may deteriorate as the artificial intelligence model is learned with countless data collected indiscriminately in the past can be solved. There is a possible effect.

In addition, the present invention visualizes the degree of bias and fairness of the uploaded original dataset and the correction dataset whose fairness has been corrected, so that the user can visually compare and confirm the bias and fairness of the dataset, and the unfairness in the dataset is visualized. This has the effect of ensuring that the items are confirmed in detail so that the fairness of the items can be improved.

The effects of the present invention are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the detailed description and claims.

Figure 1 is a configuration diagram of a fairness visualization device for an artificial intelligence learning dataset of the present invention.
Figure 2 is a diagram showing a plurality of sensitive attributes (Sensitive Columns) input according to an embodiment of the present invention.
Figure 3 is a diagram visualizing data in matrix form according to sensitive attributes (Sensitive Columns) input according to an embodiment of the present invention.
Figure 4 is a diagram visualizing fairness and bias in graph form according to an embodiment of the present invention.
Figure 5 is a diagram (a) visualizing the numerical characteristics of a numeric attribute (Numeric Column) and a diagram (b) visualizing the categorical characteristics of a categorical attribute (Categorical Column) according to an embodiment of the present invention.
Figure 6 is a diagram visualizing all data in the original dataset (Dataset ₀ ) or the corrected dataset (Dataset ₂ ) according to an embodiment of the present invention.
Figure 7 is a flowchart of the fairness visualization method of the artificial intelligence learning dataset of the present invention.
Figure 8 is a diagram showing a first sampling step according to an embodiment of the present invention.
Figure 9 is a diagram showing a second sampling step and a third sampling step according to an embodiment of the present invention.

The terms used in this specification are general terms that are currently widely used as much as possible while considering the function in the present invention, but this may vary depending on the intention or precedent of a person working in the art, the emergence of new technology, etc. In addition, in certain cases, there are terms arbitrarily selected by the applicant, and in this case, the meaning will be described in detail in the description of the relevant invention. Therefore, the terms used in the present invention should be defined based on the meaning of the term and the overall content of the present invention, rather than simply the name of the term.

Unless otherwise defined, all terms used herein, including technical or scientific terms, have the same meaning as generally understood by a person of ordinary skill in the technical field to which the present invention pertains. Terms defined in commonly used dictionaries should be interpreted as having meanings consistent with the meanings they have in the context of the related technology, and unless clearly defined in the present application, should not be interpreted in an ideal or excessively formal sense. No.

Hereinafter, embodiments according to the present invention will be described in detail with reference to the accompanying drawings. Figure 1 is a configuration diagram of a fairness visualization device for an artificial intelligence learning dataset of the present invention. Figure 2 is a diagram showing a plurality of sensitive attributes (Sensitive Columns) input according to an embodiment of the present invention. Figure 3 is a diagram visualizing data in matrix form according to sensitive attributes (Sensitive Columns) input according to an embodiment of the present invention. Figure 4 is a diagram visualizing fairness and bias in the form of a graph according to an embodiment of the present invention. Figure 5 is a diagram (a) visualizing the numerical characteristics of a numeric attribute (Numeric Column) and a diagram (b) visualizing the categorical characteristics of a categorical attribute (Categorical Column) according to an embodiment of the present invention. Figure 6 is a diagram visualizing all data in the original dataset (Dataset ₀ ) or the corrected dataset (Dataset ₂ ) according to an embodiment of the present invention.

Figure 7 is a flowchart of the fairness visualization method of the artificial intelligence learning dataset of the present invention. Figure 8 is a diagram showing the first sampling step (S320) according to an embodiment of the present invention. Figure 9 is a diagram showing the second sampling step (S340) and the third sampling step (S360) according to an embodiment of the present invention.

Fairness visualization device for artificial intelligence learning dataset

The fairness visualization device of the artificial intelligence learning dataset mentioned in the present invention may be a system equipped with a complex device and module. Alternatively, it is a platform server that implements a platform and provides platform services so that multiple user terminals can access it through an Internet network, correct the fairness of the learning dataset, and visually check the fairness of the corrected learning dataset. You can. Alternatively, it may be a recording medium that can be read by a computer device that records a program that performs a method of correcting the fairness of an artificial intelligence learning dataset and a method of visualizing the fairness of the corrected learning dataset. In other words, the fairness visualization device for the artificial intelligence learning dataset can be implemented in various ways.

Referring to Figure 1, the fairness visualization device of the artificial intelligence learning dataset of the present invention includes an uploader 100, a structural analysis module 200, a sampling module 300, and a visualization module 400.

The uploader 100 uploads the original dataset in batch form.

Batch, as referred to in the present invention, is a type of computer data processing and refers to processing data to be processed over a certain period of time or in a certain amount. In other words, the upload 100 may process data acquired over a certain period of time or in a certain amount in batches, rather than acquiring data in real time.

According to one embodiment of the present invention, the uploader 100 is used in transportation, health and medical services, disaster safety, environment, etc., such as the National Transportation DB, Health and Medical Big Data Open System, National Disaster and Safety Portal, Ministry of Environment and Culture Data Plaza public data portal, etc. Original datasets can be obtained from servers and databases that provide big data in various fields, including culture and tourism. And the original data set may be a batch-type data set stored for a certain period of time in each server and database, or a batch-type data set stored in a certain amount.

Next, the structural analysis module 200 receives a plurality of sensitive attributes (Sensitive Columns) and derives a subset corresponding to each sensitive attribute (Sensitive Columns) from the original dataset.

The sensitive attributes (Sensitive Columns) mentioned in the present invention are items that can affect the output value output from artificial intelligence technology, such as whether there is a stroke or whether it is male or female. The structural analysis module 200 can receive a plurality of sensitive attributes (Sensitive Columns) through the input module 600 included in the fairness visualization device of the artificial intelligence learning dataset of the present invention or through a separate user terminal.

In addition, the structural analysis module 200 may randomly divide the original dataset into a plurality of subsets according to a plurality of sensitive columns. For example, if the gender attribute is input as a sensitive attribute (Sensitive Columns), the structural analysis module 200 can divide the original data set into a male subset and a female subset. Alternatively, if evaluation grade attributes including grades A to D are input as sensitive columns, the structural analysis module 200 divides the original dataset into a grade A subset, a grade B subset, or a grade C subset. and D-class subsets.

Next, the sampling module 300 samples according to the size of a plurality of subsets and then outputs a correction data set with fairness corrected.

The sampling module 300 can gradually correct the fairness of the data set by going through several sampling processes, and output a corrected data set with improved fairness as a result.

In the first sampling process, the sampling module 300 includes an average value calculation module 310 that calculates the average value for a plurality of subsets (Subset ₀ ) derived from the original dataset (Dataset ₀ ), and the average value It is characterized by including a primary sampling module 320 that first samples a plurality of subsets (Subset ₀ ) using .

The primary sampling module 320 may compare the size of each subset based on the average value. If the size of a random subset is larger than the average value, primary sampling can be performed according to the average value. If the size of a random subset is smaller than or equal to the average value, primary sampling can be performed to maintain the size.

The size of the subset mentioned in the present invention refers to the number of data included in each subset. According to one embodiment of the present invention, if four subsets are derived from the structural analysis module 200 and the sizes of each are 200, 600, 100, and 300, the average value calculation module 310 calculates the average value for the four subsets. 300 can be calculated. Then, the first sampling module 320 can compare the size of each subset based on the average value 300. The first sampling model 320 can sample to maintain the size of 200 because subset 1 with a size of 200 is smaller than the average value of 300, and subset 2 with a size of 600 is larger than the average value of 300, so the average value is 300. It can be sampled to have a size of 300 excluding 300. And since subset 3 with a size of 100 is smaller than the average value of 300, it can be sampled to maintain that size of 100. Finally, subset 3 with a size of 300 has the same size as the average value of 300, so it can be sampled to maintain that size of 300. You can sample to As a result, the sizes of the four subsets initially sampled from the primary sampling module 320 may range from 200, 600, 100, and 300 to 200, 300, 100, and 300. Therefore, the primary sampling module 320 of the present invention has the significant effect of primarily reducing bias between subsets.

Next, in the second sampling, the structural analysis module 200 selects the residual data set (Dataset ₁ ) remaining after the first sampling from the first sampling module 320 among the original data set (Dataset ₀ ) and the A plurality of subsets (Subset _A , Subset _B ) are derived from the original data set (Dataset ₀ ), and the sampling module 300 compares the sizes of the corresponding subsets and calculates the difference between the sizes of the corresponding subsets. It is characterized in that it further includes a difference value calculation module 330 that calculates each and a secondary sampling module 340 that secondarily samples a plurality of subsets (Subset _A and Subset _B ) using the difference values.

According to an embodiment of the present invention, a plurality of subsets (Subset _A ) derived from the residual dataset (Dataset ₁ ) are A-1 and A-2 and may have sizes of 200 and 300, and the original dataset ( Multiple subsets (Subset _B ) derived from Dataset ₀ ) are B-1 and B-2 and can have sizes of 200 and 600. At this time, the residual data set (Dataset ₁ ) and the original data set (Dataset ₀ ) are each subset derived from the structural analysis module 200 with the same sensitive attribute, so that the subsets can correspond.

The difference calculation module 330 may compare the sizes of the A-1 subset and the corresponding B-1 subset and calculate the difference between the subsets as 0. Then, the sizes of the A-2 subset and the corresponding B-2 subset can be compared, and the difference between the subsets can be calculated as 300.

The secondary sampling module 340 can maintain the respective sizes at 200 and 200 because the difference value between the A-1 subset and the B-1 subset is 0. And since the difference between the A-2 subset and the B-2 subset is 300, the size of the B-2 subset can be subtracted from 600 by the difference value and sampled to a size of 300. At this time, the subset with a larger size among the corresponding subsets is characterized in that it is sampled using the difference value. Therefore, as the secondary sampling module 340 performs secondary sampling, the size of the A-1 subset is set to 200, the size of the A-2 subset is set to 200, the size of the B-1 subset is set to 300, and the size of the B-2 subset is set to 300. You can do this.

Therefore, by the secondary sampling module 340 of the present invention, the remaining data set (Dataset ₁ ) remaining after the first sampling from the primary sampling module 320 and the original data set received in batch form from the uploader 100 ( There is a notable effect of being able to secondarily sample each subset (Subset _A , Subset _B ) derived from Dataset ₀ ).

Next, in the third sampling, the sampling module 300 includes a minimum value selection module 350 that compares the sizes of the plurality of subsets sampled secondarily from the second sampling module 340 and then selects the minimum value. It further includes a 3rd sampling module 360 that performs 3rd sampling by adding the minimum value to each of the plurality of subsets first sampled from the 1st sampling module 320.

According to one embodiment of the present invention, the size of A-1 secondary sampled through the secondary sampling module 340 is 200, the size of A-2 is 200, the size of B-1 is 300, and the size of B-2 is 200. The size of may be 300. Then, the minimum value selection module 350 can select 200, which is the smallest size value, as the minimum value.

And if the sizes of subsets 1 to 4 first sampled from the first sampling module 320 are 200, 300, 100, and 300, they have a ratio of 2:3:1:3, and the 3rd sampling module 360 The minimum value of 200 can be added to the sizes of subsets 1 to 4, respectively. Therefore, the sizes of subsets 1 to 4 may be 400, 500, 300, and 500, and may have a ratio of 1.33:1.67:1:1.67.

Therefore, the difference between subsets can be significantly reduced by the 3rd sampling module 360, and the fairness of the original dataset (Dataset ₀ ) is gradually corrected through the 1st to 3rd sampling process to produce a correction dataset (Dataset). ₂ ) can be output.

Next, the visualization module 400 visualizes at least one of the original dataset (Dataset ₀ ) and the corrected dataset (Dataset ₂ ) in a preset format to easily check data fairness.

The visualization module 400 can visualize bias and fairness in various formats.

First, looking at the embodiment of FIGS. 2 and 3, the uploader 100 can upload an original dataset (Dataset ₀ ) in a CSV file format. The structural analysis module 200 includes speed limit (A), weather (B), lighting (C), initial collision type (D), road surface condition (E), and damage (F) through a number of sensitive attributes (Sensitive Columns). You can receive input as . Then, the visualization module 400 is characterized in that it visualizes the data included in the original dataset (Dataset ₀ ) in matrix form according to the plurality of sensitive attributes (Sensitive Columns). Therefore, there is a notable effect in that users can check each data value for each attribute in detail.

In addition, looking at the embodiment of FIG. 4, the visualization module 400 uses the original dataset (Dataset ₀ ) and the corrected dataset (Dataset ₂ ) as a series, and calculates statistical parity difference and equal opportunity difference. It is characterized by visualizing the fairness and bias of each fairness index, including Equal Opportunity Difference, Average Odds Difference, Disparate Impact, and theil index, in graph form.

The user can make the following decisions through the graph visualized by the visualization module 400. First, in the graph for any fairness index, if the index of the correction dataset (Dataset ₂ ) (Mitigated) is closer to the fairness baseline than the index of the original dataset (Dataset ₀ ) (Original), it is fair and well-corrected. It can be interpreted as For example, in the Statistical Parity Difference index of FIG. 4, the index of the correction dataset (Dataset ₂ ) (Mitigated) is -0.09, and is fairer than the index of -0.18 of the original dataset (Dataset ₀ ) (Original). (Fair) Since it is closer to the baseline of 0, the user can judge that fairness has been corrected only for the Statistical Parity Difference indicator.

In addition, in the Equal Opportunity Difference indicator and the Average Odds Difference indicator, the indicator of the original dataset (Dataset ₀ ) (Original) is better than the indicator of the calibration dataset (Dataset ₂ ) (Mitigated). It can be confirmed that it is close to the fairness baseline. In other words, bias occurred in other indicators through the sampling module 300. The visualization module 400 uses the original dataset (Dataset 0) or a correction dataset (Dataset ₀ ) according to the purpose of utilizing the dataset by a user who has confirmed that an unfair result has been derived from the sampling module 300. ₂ ) has a notable effect of visually helping users decide whether to use a dataset.

In addition, looking at the embodiment of FIG. 5, the visualization module 400 distinguishes a sensitive column having data in a numeric form among a plurality of sensitive columns and displays a numeric column in the numeric column. A numeric feature display unit 410 analyzes the numeric features and displays them in a chart, and among a plurality of sensitive columns, a sensitive column with character data is distinguished and a categorical feature is displayed. It is characterized by including a categorical feature display unit 420 that analyzes categorical features for columns and displays them in a chart, allowing the data fairness by feature to be confirmed.

Looking at one embodiment of Figure 5 (a), the numerical characteristic display unit 410 includes an original dataset (Dataset ₀ ) and a correction dataset (Dataset ₂ ) for the speed limit attribute among a plurality of sensitive attributes (Sensitive Columns). The number (Count), mean (Mean), standard deviation (Std dev), percentage of data being 0 (Zeros), minimum value (Min), median, and maximum value (Max) of each data are numerical characteristics. Each can be derived as (Numeric Features). And the numerical characteristic display unit 410 can visualize the data of each dataset in a chart format by reflecting the derived characteristics.

Looking at an embodiment of Figure 5 (b), the category characteristic display unit 420 displays the original dataset (Dataset ₀ ) and the correction dataset (Dataset ₂ ) for weather attributes among a plurality of sensitive attributes (Sensitive Columns). The number of each data (Count), the type of weather within the property (Unique), the type of weather with a large number of data (Top), the number of weather data with a large number of data (Freq top), and the average string length (Avg str len) are categorized. Each can be derived as Categorical Features. And the category characteristic display unit 420 can visualize the data of each dataset in a chart format by reflecting the derived characteristics.

Therefore, by the numerical characteristic display unit 410 and the category characteristic display unit 420, the user visually compares the data distribution for sensitive attributes (Sensitive Columns) between the original dataset (Dataset ₀ ) and the corrected dataset (Dataset ₂ ). And there is a remarkable effect that can be confirmed.

In addition, the visualization module 400 is configured to binning at least one of the original dataset (Dataset ₀ ) and the corrected dataset (Dataset ₂ ) using the categorical attribute (Categorical Column). A binning setting unit 430, a scattering setting unit 440 that scatters at least one of the original dataset (Dataset ₀ ) and the corrected dataset (Dataset ₂ ) using the numeric attribute (Numeric Column). , a series setting unit 450 that sets the categorical features within the categorical column to be expressed in different colors, and when arbitrary data is selected, data values for each attribute for the arbitrary data are displayed in detail It is characterized by visualization so that the fairness of the entire dataset can be confirmed by further including a detailed characteristic display unit 460.

In general, binning refers to the function of grouping and managing pixels in an image composed of multiple pixels. That is, the binning setting unit 430 of the present invention configures the binning X-axis and Binning is set on at least one of the Y-axes, and the data is visually displayed by dividing it according to the set category properties.

Conversely, scattering refers to the function of scattering data. In other words, the scattering setting unit 440 of the present invention configures the numeric properties into the scattering Set to at least one of the following and display the data visually by dividing it according to the set numerical properties.

Looking at the example of FIG. 6, the object of visualization may be any original dataset (Dataset ₀ ) with one numeric column and six categorical columns. At this time, the binning setting unit 430 may select a crash type (Crush_type), one of the categorical columns, to be visualized on the binning X-axis. At this time, the accident type (Crush_type) can be divided into Injury and/or Tow due to crash, No Injury, and/or Drive away. Then, the data in the original dataset (Dataset ₀ ) can be visualized separately by being classified by the accident type (Crush_type) selected from the binning setting unit 430. Additionally, the binning setting unit 430 can select Damage, one of the categorical columns, to be visualized on the binning Y-axis. At this time, damage can be divided into 1500 or more and 1500 or less. Then, the data in the original dataset (Dataset ₀ ) can be visualized separately by the damage selected from the binning setting unit 430. Therefore, as in the embodiment of FIG. 6, the data in the original dataset (Dataset ₀ ) can be divided into four quadrants and displayed together.

In addition, the graph visualized in each quadrant can be set from the series setting unit 450 to express category characteristics within the first accident type (First_crash_type) in different colors. And when arbitrary data is selected from the graph visualized in each quadrant, the characteristics of each attribute for the arbitrary data can be displayed in detail from the detailed characteristic display unit 460. In other words, the original dataset (Dataset ₀ ) has a damage of over 1500, and the data with an accident type (Crush_type) of Injury and/or Tow due to crash is relatively small. This can be identified, and it can be confirmed that it is a highly unfair dataset.

By the visualization module 400 of the present invention capable of visualizing data in various ways, the original dataset (Dataset ₀ ) uploaded from the uploader 100 and the correction dataset (Dataset) whose fairness has been corrected from the sampling module 300 ₂ ), the degree of bias and fairness can be visualized respectively according to the fairness index.

In addition, by the visualization module 400 of the present invention, the categorical features of one categorical column of the original dataset (Dataset ₀ ) and the correction dataset (Dataset ₂ ) are visualized together. Users can visually compare and see how much the fairness has improved for the corresponding category attributes. In addition, the Numeric Features for the Numeric Column of the original dataset (Dataset ₀ ) and the correction dataset (Dataset ₂ ) are visualized together, allowing the user to visually see how much the fairness of the corresponding numeric column has improved. You can compare and check.

In addition, by the visualization module 400 of the present invention, the original dataset (Dataset ₀ ) is targeted and has at least one of a categorical column on the binning XY axis and a numeric column on the scattering XY axis. The data distribution according to the selected number of properties is visualized, allowing the user to visually check the data distribution according to the multiple properties of the dataset.

Next, the present invention further includes a storage module 500 for storing the remaining data so that the remaining data after sampling can be included in another original dataset (Dataset ₀₀ ) to be uploaded from the uploader 100. It is characterized by

In other words, the present invention does not simply remove the residual data remaining after the sampling, but stores the residual data from the storage module 500 and later uploads another batch of original data sets from the upload 100. By combining them, the correction dataset (Dataset ₂ ) with corrected fairness can be output again.

Therefore, according to the present invention, the unfairness of the dataset can be solved incrementally, and the problem of the performance of the artificial intelligence model being deteriorated as the artificial intelligence model is learned with countless data collected indiscriminately has a significant effect. There is.

In addition, the present invention can visually compare and confirm the degree of bias and fairness of the uploaded original dataset and the correction dataset whose fairness has been corrected, and can check unfair items in the dataset to supplement the fairness of the items. It has a remarkable effect.

How to visualize the fairness of artificial intelligence learning datasets

Referring to Figure 7, the fairness visualization method of the artificial intelligence learning dataset of the present invention includes an original dataset upload step (S100), a subset derivation step (S200), a correction dataset output step (S300), and a visualization step (S400). do.

First, in the original dataset upload step (S100), the original dataset in batch form is uploaded by the uploader 100.

Batch, as referred to in the present invention, is a type of computer data processing and refers to data to be processed being organized and processed for a certain period of time or a certain amount. In other words, in the original dataset upload step (S100), data may not be acquired in real-time data units, but data acquired over a certain period of time or a certain amount may be processed in batches.

According to one embodiment of the present invention, the original data set upload step (S100) is performed on transportation, health care, and disaster safety, such as the National Transportation DB, Health and Medical Big Data Open System, National Disaster and Safety Portal, Ministry of Environment and Culture Data Plaza public data portal, etc. Original data sets can be obtained from servers and databases that provide big data in various fields such as environment, culture and tourism. And the original data set may be a batch-type data set stored for a certain period of time in each server and database, or a batch-type data set stored in a certain amount.

Next, in the subset derivation step (S200), a plurality of sensitive attributes (Sensitive Columns) are input by the structural analysis module 200, and a subset corresponding to each sensitive attribute (Sensitive Columns) is derived from the original dataset. do.

The sensitive attributes (Sensitive Columns) mentioned in the present invention are items that can affect the output value output from artificial intelligence technology, such as whether there is a stroke or whether it is male or female. In the subset derivation step (S200), a plurality of sensitive attributes (Sensitive Columns) may be input through the input module 600 included in the fairness visualization device of the artificial intelligence learning dataset or a separate user terminal.

In the subset derivation step (S200), the original dataset may be randomly divided into a plurality of subsets according to a plurality of sensitive columns. For example, if the gender attribute is input as a sensitive attribute (Sensitive Columns), in the subset derivation step (S200), the original dataset may be divided into a male subset and a female subset. Alternatively, if an evaluation grade attribute including grades A to D is input as a sensitive attribute (Sensitive Columns), the subset derivation step (S200) determines that the original dataset is a grade A subset, a grade B subset, or a grade C subset. and D-class subsets.

Next, in the correction data set output step (S300), the sampling module 300 samples a plurality of subsets according to their size and then outputs a correction data set whose fairness has been corrected.

The correction data set output step (S300) is characterized in that the fairness of the data set is gradually corrected through several sampling processes, and as a result, a correction data set with improved fairness is output.

Referring to FIG. 8, in the first sampling process, the correction data set output step (S300) generates a plurality of subsets (Subset ₀ ) derived from the original data set (Dataset ₀ ) by the average value calculation module 310. It includes an average value calculation step (S310) in which the _average value for It is characterized by:

The size of the subset mentioned in the present invention refers to the number of data included in each subset. According to one embodiment of the present invention, if four subsets are derived from the subset derivation step (S210) and the sizes of each are 200, 600, 100, and 300, the average value calculation step (S310) is the average value for the four subsets. 300 can be calculated. Then, in the first sampling step (S320), the average value of 300 is used as a standard, and the size of each subset can be compared. In the first sampling step (S320), subset 1 with a size of 200 is smaller than the average value of 300, so it can be sampled to maintain the size of 200, and subset 2 with a size of 600 is larger than the average value of 300, so the average value It can be sampled to have a size of 300 excluding 300. And since subset 3 with a size of 100 is smaller than the average value of 300, it can be sampled to maintain the size of 100, and finally, subset 3 with a size of 300 has the same size as the average value of 300, so it can be sampled to maintain the size of 300. can be sampled to As a result, the sizes of the four subsets initially sampled from the primary sampling module 320 may range from 200, 600, 100, and 300 to 200, 300, 100, and 300. Therefore, there is a significant effect of primarily reducing bias between subsets through the average value calculation step (S310) and the first sampling step (S320) of the present invention.

9, in the secondary sampling process, the correction data set output step (S300) progresses to the subset derivation step (S200) so that secondary sampling can be performed when the bias between subsets is initially completed after primary sampling. It is possible to regress.

The subset derivation step (S220a, S220b) after the first sampling is the residual data set remaining after the first sampling from the first sampling step (S320) among the original data set (Dataset ₀ ) by the structural analysis module 200. Multiple subsets (Subset _A and Subset _B ) can be derived from (Dataset ₁ ) and the original dataset (Dataset ₀ ), respectively.

And the correction data set output step (S300) is a difference value calculation step (S330) in which the sizes of the corresponding subsets are compared by the difference value calculation module 330 and the difference values of the sizes of the corresponding subsets are calculated. It further includes a secondary sampling step (S340) in which a plurality of subsets (Subset _A and Subset _B ) are secondarily sampled by using the difference value by the secondary sampling module 340.

According to an embodiment of the present invention, a plurality of subsets (Subset _A ) derived from the residual dataset (Dataset ₁ ) are A-1 and A-2 and may have sizes of 200 and 300, and the original dataset ( Multiple subsets (Subset _B ) derived from Dataset ₀ ) are B-1 and B-2 and can have sizes of 200 and 600. At this time, the residual data set (Dataset ₁ ) and the original data set (Dataset ₀ ) are each derived with the same sensitive attribute from the subset derivation step (S220a, S220b) after the first sampling, so that the subsets can correspond.

In the difference value calculation step (S330), the sizes of the A-1 subset and the corresponding B-1 subset are compared, and the difference value between the subsets may be calculated as 0. Then, the sizes of the A-2 subset and the corresponding B-2 subset are compared, and the difference between the subsets can be calculated as 300.

In the second sampling step (S340), since the difference value between the A-1 subset and the B-1 subset is 0, the respective sizes can be maintained at 200 and 200. And since the difference value between the A-2 subset and the B-2 subset is 300, the size of the B-2 subset can be subtracted from 600 by the difference value and sampled to a size of 300. At this time, the subset with a larger size among the corresponding subsets is characterized in that it is sampled using the difference value. Therefore, in the second sampling step (S340), the size of the A-1 subset is 200, the size of the A-2 subset is 200, the size of the B-1 subset is 300, and the size of the B-2 subset is 300. It can be.

Therefore, the remaining dataset (Dataset ₁ ) remaining after the first sampling from the first sampling step (S320) by the second sampling step (S340) of the present invention and the data received in batch form from the original dataset upload step (S100) There is a notable effect in that each subset (Subset _A , Subset _B ) derived from the original dataset (Dataset ₀ ) can be sampled secondarily.

9, in the third sampling, the correction data set output step (S300) compares the sizes of a plurality of subsets sampled secondarily from the second sampling step (S340) by the minimum value selection module 350. Then, by the minimum value selection step (S350) and the 3rd sampling module 360 in which the minimum value is selected, the minimum value is added to each of the plurality of subsets first sampled from the 1st sampling step (S320), resulting in 3rd sampling. It is characterized in that it further includes a tea sampling step (S360).

According to one embodiment of the present invention, the size of A-1 secondarily sampled through the second sampling step (S340) is 200, the size of A-2 is 200, the size of B-1 is 300, and the size of B-2 is 200. The size of may be 300. Then, in the minimum value selection step (S350), 200, which is the smallest size value, may be selected as the minimum value.

And if the sizes of subsets 1 to 4 sampled from the first sampling step (S320) are 200, 300, 100, and 300, they have a ratio of 2:3:1:3, and the third sampling step (S360) The minimum value of 200 may be added to the sizes of subsets 1 to 4, respectively. Accordingly, the sizes of subsets 1 to 4 may be 400, 500, 300, and 500, and may have a ratio of 1.33:1.67:1:1.67.

Therefore, in the present invention, the difference between subsets can be significantly reduced by the third sampling step (S360), and the fairness of the original data set (Dataset ₀ ) can be gradually corrected to output a corrected data set (Dataset ₂ ). there is.

Next, in the visualization step (S400), at least one of the original dataset (Dataset ₀ ) and the corrected dataset (Dataset ₂ ) is visualized in a preset format by the visualization module 400 to facilitate confirmation of data fairness. do.

In the visualization step (S400), bias and fairness can be visualized in various formats.

First, looking at the embodiment of FIGS. 2 and 3, in the original dataset upload step (S100), the original dataset (Dataset ₀ ) in CSV file format may be uploaded. The subset derivation step (S200) includes speed limit (A), weather (B), lighting (C), initial collision type (D), road surface condition (E), and damage (F) into a number of sensitive attributes (Sensitive Columns). It can be input as . Then, the visualization step (S400) is characterized in that the data included in the original dataset (Dataset ₀ ) is visualized in matrix form according to the plurality of sensitive attributes (Sensitive Columns). Therefore, there is a notable effect in that users can check each data value for each attribute in detail.

In addition, looking at the embodiment of FIG. 4, the visualization step (S400) uses the original dataset (Dataset ₀ ) and the corrected dataset (Dataset ₂ ) as a series, and calculates the statistical parity difference (Statistical Parity Difference) and the equal opportunity difference ( It features the fairness and bias of each fairness index, including Equal Opportunity Difference, Average Odds Difference, Disparate Impact, and theil index, visualized in graph form.

The user can make the following decisions through the graph visualized from the visualization step (S400). First, in the graph for any fairness index, if the index of the correction dataset (Dataset ₂ ) (Mitigated) is closer to the fairness baseline than the index of the original dataset (Dataset ₀ ) (Original), it is fair and well-corrected. It can be interpreted as For example, in the Statistical Parity Difference index of FIG. 4, the index of the correction dataset (Dataset ₂ ) (Mitigated) is -0.09, and is fairer than the index of -0.18 of the original dataset (Dataset ₀ ) (Original). (Fair) Since it is closer to the baseline of 0, fairness can be judged to have been corrected only for the Statistical Parity Difference indicator.

In addition, in the Equal Opportunity Difference indicator and the Average Odds Difference indicator, the original data set (Dataset ₀ ) (Original) indicator is fairer than the indicator of the calibration data set (Dataset ₂ ) (Mitigated). (Fair) It can be confirmed that it is close to the baseline. In other words, bias occurred in other indicators through the correction data set output step (S300). In the visualization step (S400), the user who confirms that unfair results were derived from the correction data set output step (S300) uses the original dataset (Dataset ₀ ) for the purpose of utilizing the dataset, or uses the correction data Using a set (Dataset ₂ ) has the notable effect of visually helping users decide whether to use a dataset.

In addition, looking at the embodiment of FIG. 5, in the visualization step (S400), a sensitive column having data in the form of a number is distinguished among a plurality of sensitive columns by the numeric characteristic display unit 410. By the numeric feature display step (S410) and the category feature display unit (420) in which the numeric features for the numeric columns are analyzed and displayed in a chart, characters among a plurality of sensitive features (Sensitive Columns) By including a categorical feature display step (S420) in which sensitive columns with data in the form are distinguished and categorical features for the categorical columns are analyzed and displayed in charts, data by feature It is characterized by ensuring fairness.

Looking at one embodiment of Figure 5(a), the numerical characteristic display step (S410) includes an original dataset (Dataset ₀ ) and a correction dataset (Dataset ₂₎ for the speed limit attribute among a plurality of sensitive attributes (Sensitive Columns). ) The number (Count), mean (Mean), standard deviation (Std dev), percentage of data being 0 (Zeros), minimum value (Min), median, and maximum value (Max) of each data are numerical values. Each can be derived as Numeric Features. And in the numerical characteristic display step (S410), the derived characteristics are reflected and the data of each dataset can be visualized in chart format.

Looking at an embodiment of Figure 5 (b), the category characteristic display step (S420) includes an original dataset (Dataset ₀ ) and a correction dataset (Dataset ₂ ) for weather attributes among a plurality of sensitive attributes (Sensitive Columns). The number of each data (Count), the type of weather in the property (Unique), the type of weather with a large number of data (Top), the number of weather data with a large number of data (Freq top), and the average string length (Avg str len) Each can be derived as categorical features. And in the category characteristic display step (S420), the derived characteristics are reflected and the data of each dataset can be visualized in chart format.

Therefore, by the numerical characteristic display step (S410) and the category characteristic display step (S420) of the present invention, the user converts the original dataset (Dataset ₀ ) and the corrected dataset (Dataset ₂ ) into data for sensitive attributes (Sensitive Columns). There is a remarkable effect of being able to visually compare and confirm distributions.

In addition, looking at the embodiment of FIG. 6, the visualization step (S400) uses the categorical attribute (Categorical Column) by the binning setting unit 430 to create the original dataset (Dataset ₀ ) and the correction dataset ( In the binning setting step (S430) in which at least one of Dataset ₂ is binning, the numeric attribute (Numeric Column) is used by the scattering setting unit 440 to combine the original dataset (Dataset ₀ ) and the correction data. In the scattering setting step (S440) in which at least one of the three (Dataset ₂ ) is scattered, the categorical features in the categorical column are colored in different colors by the series setting unit 450. It further includes a series setting step (S450) that is set to be expressed and a detailed characteristic display step (S460) in which, when arbitrary data is selected by the detailed characteristic display unit (460), data values for each attribute for the arbitrary data are displayed in detail. By doing so, it is characterized by visualization so that the fairness of the entire dataset can be confirmed.

In general, binning refers to the function of grouping and managing pixels in an image composed of multiple pixels. In other words, in the binning setting step (S430) of the present invention, in order to confirm that numerous data in the dataset are grouped into one of the categorical columns such as weather, initial collision type, road surface condition, and damage, the category properties are binning Binning is set on at least one of the Y-axes, and data is divided and visually displayed according to the set category properties.

Conversely, scattering refers to the function of scattering data. In other words, in the scattering setting step (S440) of the present invention, in order to confirm that numerous data in the dataset are grouped into one of the numeric attributes (Numeric Column) such as speed limit and lighting, the numeric attribute is the scattering X axis and the scattering Y axis. is set to at least one of the following, and the data is divided according to the set numerical properties and displayed visually.

Looking at the example of FIG. 6, the object of visualization may be any original dataset (Dataset ₀ ) with one numeric column and six categorical columns. At this time, in the binning setting step (S430), the accident type (Crush_type), one of the categorical columns, can be selected to be visualized on the binning X-axis. At this time, the accident type (Crush_type) can be divided into Injury and/or Tow due to crash, No Injury, and/or Drive away. Then, the data in the original dataset (Dataset ₀ ) can be visualized separately by being classified by the accident type (Crush_type) selected in the binning setting step (S430). Additionally, in the binning setting step (S430), damage, one of the categorical columns, may be selected to be visualized on the binning Y-axis. At this time, damage can be divided into 1500 or more and 1500 or less. Then, the data in the original dataset (Dataset ₀ ) can be visualized separately by the damage selected in the binning setting step (S430).

Therefore, as in the embodiment of FIG. 6, the data in the original dataset (Dataset ₀ ) can be divided into four quadrants and displayed together. The graph visualized in each quadrant can be set so that category characteristics within the first accident type (First_crash_type) are expressed in different colors from the series setting step (S450). And when arbitrary data is selected from the graph visualized in each quadrant, the characteristics of each attribute for the arbitrary data can be displayed in detail from the detailed characteristic display step (S460). In other words, the original dataset (Dataset ₀ ) has damage over 1500, and the data with the accident type (Crush_type) of Injury and/or Tow due to crash is relatively small. It can be visually understood from the , and it can be confirmed that it is a highly unfair dataset.

By the visualization step (S400) of the present invention, which allows data visualization in various ways, fairness is corrected from the original dataset (Dataset ₀ ) uploaded from the original dataset upload step (S100) and the correction dataset output step (S300). The degree of bias and fairness of the corrected dataset (Dataset ₂ ) can be visualized respectively according to the fairness index.

In addition, by the visualization step (S400) of the present invention, the categorical features for one categorical column of the original dataset (Dataset ₀ ) and the correction dataset (Dataset ₂ ) are visualized together. Users can visually compare and see how much the fairness has improved for the corresponding category attributes. In addition, the Numeric Features for the Numeric Column of the original dataset (Dataset ₀ ) and the correction dataset (Dataset ₂ ) are visualized together, allowing the user to visually see how much the fairness of the corresponding numeric column has improved. You can compare and check.

In addition, by the visualization step (S400) of the present invention, the original dataset (Dataset ₀ ) is targeted and at least one of a categorical column on the binning XY axis and a numeric column on the scattering XY axis is included. The data distribution according to the selected number of properties is visualized, allowing the user to visually check the data distribution according to the multiple properties of the dataset.

Next, the present invention stores the residual data so that the residual data remaining after sampling can be included in another original dataset (Dataset ₀₀ ) to be uploaded from the original dataset upload step (S100) by the storage module 500. It is characterized in that it further includes a storage step (S500). Meanwhile, the visualization step (S400) and the storage step (S500) are not limited to the described order, and the storage step (S500) is performed after the visualization step (S400) or the visualization step (S400) is performed after the storage step (S500). Alternatively, the visualization step (S400) and the storage step (500) may be performed simultaneously.

In other words, in the present invention, the residual data remaining after the sampling is not simply removed, but the residual data is stored from the storage step (S500) and later stored as an original data set in another batch form from the original data set upload step (S100). Once uploaded, it can be combined with this so that the correction dataset (Dataset ₂ ) with corrected fairness can be output and visualized again.

Therefore, according to the present invention, the unfairness of the dataset can be solved incrementally, and the problem that the performance of the artificial intelligence model may deteriorate as the artificial intelligence model is learned with countless data collected indiscriminately can be solved. It works.

In addition, the present invention can visually compare and confirm the degree of bias and fairness of the uploaded original dataset and the correction dataset whose fairness has been corrected, and unfair items in the dataset can be confirmed and the fairness of the items can be improved. It has a remarkable effect in enabling

Embodiments may be implemented by hardware, software, firmware, middleware, microcode, hardware description language, or any combination thereof. When implemented as software, firmware, middleware, or microcode, program code or code segments that perform necessary tasks may be stored in a computer-readable storage medium and executed by one or more processors.

And aspects of the subject matter described herein may be described in the general context of computer-executable instructions, such as program modules or components that are executed by a computer. Typically, program modules or components include routines, programs, objects, and data structures that perform specific tasks or implement specific data types. Aspects of the subject matter described herein may be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media, including memory storage devices.

Although the embodiments have been described with limited examples and drawings as described above, various modifications and variations can be made from the above description by those skilled in the art. For example, the described techniques are performed in an order different from the described method, and/or the components of the described system, structure, device, circuit, etc. are combined or combined in a different form than the described method, or in a different configuration. Appropriate results may be achieved by substitution or substitution of elements or equivalents.

Therefore, other implementations, other embodiments, and equivalents to the claims also fall within the scope of the claims described below.

100.. Uploader
200.. Structural analysis module
300.. Sampling module
310.. Average value calculation unit
320.. Primary sampling unit
330.. Difference value calculation unit
340.. Secondary sampling unit
350.. Minimum value selection part
360.. 3rd sampling unit
400.. Visualization module
410.. Numerical characteristic display unit
420.. Category attribute display
430.. Binning setting section
440.. Scattering setting unit
450.. Series setting section
460.. Detailed characteristics display unit
500.. storage module
600.. Input module

Claims (8)

An uploader that uploads the original dataset in batch form;
A structural analysis module that receives a plurality of sensitive attributes (Sensitive Columns) as input and derives a subset corresponding to each sensitive attribute (Sensitive Columns) from the original dataset;
A sampling module that samples according to the size of a plurality of subsets and then outputs a correction dataset with corrected fairness; and
A visualization device for visualizing at least one of the original dataset and the correction dataset in a preset format to facilitate checking data fairness. A device for visualizing fairness of an artificial intelligence learning dataset, including a visualization module. According to clause 1,
The visualization module is,
A numeric feature display unit that distinguishes sensitive columns with numeric data among a plurality of sensitive columns, analyzes the numeric features for the numeric columns, and displays them on a chart; and
A category characteristic display unit that distinguishes sensitive columns having character data among a plurality of sensitive columns, analyzes categorical features for the categorical columns, and displays them on a chart; By including,
A fairness visualization device for artificial intelligence learning datasets that allows checking data fairness by characteristic. According to clause 2,
The visualization module is,
a binning setting unit for binning at least one of the original data set and the corrected data set using the categorical column;
a scattering setting unit that scatters at least one of the original data set and the corrected data set using the numeric attribute (Numeric Column);
A series setting unit that sets categorical features within the categorical column to be expressed in different colors; and
When arbitrary data is selected, by further including a detailed characteristic display unit to display data values for each attribute of the arbitrary data,
A fairness visualization device for artificial intelligence learning datasets, characterized in that it visualizes the fairness of the entire dataset. According to clause 1,
A storage module for storing the remaining data so that the remaining data after sampling can be included in another original data set to be uploaded from the uploader. A fairness visualization device for an artificial intelligence learning dataset, further comprising: According to clause 1,
The sampling module is,
an average value calculation module that calculates an average value for a plurality of subsets derived from the original data set; and
A fairness visualization device for an artificial intelligence learning dataset, comprising: a primary sampling module that first samples a plurality of subsets using the average value. According to clause 5,
The structural analysis module is,
Among the original datasets, a plurality of subsets are derived from the original dataset and a residual dataset remaining after the first sampling from the first sampling module,
The sampling module is,
a difference value calculation module that compares the sizes of corresponding subsets and calculates difference values between the sizes of the corresponding subsets; and
A fairness visualization device for an artificial intelligence learning dataset, further comprising a secondary sampling module that secondary samples a plurality of subsets using the difference values. According to clause 6,
The sampling module is,
a minimum value selection module that selects the minimum value after comparing the sizes of a plurality of secondary sampled subsets from the secondary sampling module; and
A 3rd sampling module that performs 3rd sampling by adding the minimum value to each of the plurality of subsets first sampled from the 1st sampling module. A device for visualizing fairness of an artificial intelligence learning dataset, further comprising: An original dataset upload step in which the original dataset in batch form is uploaded by the uploader;
A subset derivation step in which a plurality of sensitive attributes (Sensitive Columns) are input by a structural analysis module, and a subset corresponding to each sensitive attribute (Sensitive Columns) is derived from the original data set;
A correction data set output step in which a correction data set with fairness corrected after sampling according to the size of a plurality of subsets is output by a sampling module; and
A visualization method of visualizing the fairness of an artificial intelligence learning dataset, including a visualization step of visualizing at least one of the original dataset and the correction dataset in a preset format to facilitate confirmation of data fairness by a visualization module.