Identification and Quantification of Crosstalk Bias Based on Expert Knowledge

Abstract

To provide a method, system and storage medium for identifying and quantizing a crossconnection bias on the basis of expert knowledge.SOLUTION: A method comprises: acquiring a machine learning model based upon a plurality of factors corresponding to an object, and having been trained about the object; acquiring expert information provided by humans about the object, and representing relations among the plurality of factors as to how the plurality of factors affect one another; generating a structural cause and effect model representing the relations among the plurality of factors on the basis of the expert information; identifying, as crossconnection factors, factors which may cause a crossconnection bias at the machine learning model among the plurality of factors on the basis of the structural cause and effect model; and estimating the crossconnection bias on the basis of the identified crossconnection factors.SELECTED DRAWING: Figure 1A

Description

Technical Field

[0001] The present disclosure generally relates to identifying and quantifying confounding bias based on expertise.

Background Art

[0002] Generally, bias is a favorable or adverse prejudice towards one entity, person, or group compared to others. In data, bias creates variables that do not properly capture the object being represented. Thus, when the data is skewed, conclusions based on that data may not be representative of the object being represented.

[0003] The claimed subject matter of the present disclosure is not limited to embodiments that operate only in the environments as described above, or to embodiments that solve any of the disadvantages as described above. Rather, this background is provided only to explain one exemplary technical field in which some of the embodiments described in the present disclosure may be implemented.

Summary of the Invention

[0004] In an exemplary embodiment, the method may include obtaining a machine learning model trained with respect to an object. The machine learning model may be based on a plurality of factors corresponding to the object. The method may include obtaining expert information provided by a human with respect to the object. The expert information may indicate the relationships among the plurality of factors regarding how the plurality of factors affect each other. The method may include generating a structural causal model representing the relationships among the plurality of factors based on the expert information. The method may include identifying, as confounding factors, factors among the plurality of factors that cause confounding bias in the machine learning model based on the structural causal model. The method may include estimating the confounding bias based on the identified confounding factors.

[0005] In some embodiments, the method may include adjusting a machine learning model based on an estimated confounding bias to reduce the confounding bias. In some aspects, the subject may be artwork, and the expert information provided by a human may be obtained from an art expert. Identifying a factor as a confounding factor may include performing a causal inference analysis with respect to a structural causal model.

[0006] In some aspects, generating a structural causal model may include identifying input factors and output factors from a plurality of factors, where the input factors affect the result of the output factors, and determining the dependencies between the input factors and the output factors.

[0007] In some aspects, estimating a confounding bias may include determining a perceived causal effect associated with the subject, determining an observed causal effect using an objective analysis architecture, and determining the difference between the perceived causal effect and the observed causal effect.

[0008] In some embodiments, the subject is image classification. Determining the perceived causal effect may include matching an image based on a confounding factor. Determining the observed causal effect may include matching an image using a matching algorithm. The difference between the perceived causal effect and the observed causal effect may include the difference between the matching based on the confounding factor and the matching using the matching algorithm.

[0009] In another exemplary embodiment, one or more computer-readable storage media may store instructions that, in response to being executed by one or more processors, cause the system to perform the above methods and any suitable variations thereof.

[0010] In yet another example embodiment, the system may include one or more processors and one or more computer-readable storage media storing instructions that, responsive to being executed by the one or more processors, cause the system to perform the methods and any suitable variations described above.

[0011] The objectives and advantages of the embodiments will be realized and achieved at least by the elements, features, and combinations particularly pointed out in the claims. The above summary and the following detailed description are both provided by way of example and are descriptive, not limiting, of the claimed invention.

[0012] Example embodiments will be described and explained in further particularity and detail through the use of the accompanying drawings.

Brief Description of the Drawings

[0013]

Figure 1A

Figure 1B

Figure 2

Figure 3

Modes for Carrying Out the Invention

[0014] Some embodiments described in this disclosure are related to quantifying crosstalk bias based on expertise. Generally, bias is a favorable or adverse prejudice towards one entity, person, or group compared to others. In modeling or analyzing an object, bias gives rise to variables that do not appropriately capture the object being represented. Thus, when data or a model is biased, it means that the conclusions drawn based on that data or model do not represent the object being represented.

[0015] A confounder (also called a confounding variable, confounding factor, or lurking variable) may be a variable that acts on both the dependent and independent variables, causing an apparent association. Confounding bias may similarly be a bias that causes spurious correlations between input and output variables.

[0016] The disclosed embodiments include ways to identify and quantify confounding bias by using expert knowledge. The disclosed embodiments may estimate the bias based on expert knowledge, which may be incorporated into a structural causal model, and then the structural causal model is evaluated for the bias using the rules of causal inference. According to the disclosed embodiments, the bias may be quantified as the difference between the perceived causal effect (e.g., based on expert knowledge) and the observed causal effect (based on data).

[0017] As described above, confounding bias can result in models that do not represent the population being represented. This can be a problem for applications that implement such biased models. Machine learning is implemented in various fields, especially in healthcare, justice, finance, etc. In such situations, models may be used to represent people and expected outcomes based on various factors. For example, machine learning may be implemented to help determine who is likely to be eligible for a credit card, who should be treated with a particular drug, who should be released on bail, etc. However, if the models used in machine learning are biased against certain attributes (e.g., ethnicity, gender, age, etc.), certain people may be treated unfairly or inaccurately as a result of the bias.

[0018] In a further example, machine learning may be used to create generative art. Generative art refers to artworks that are created, in whole or in part, through the use of autonomous systems. In such applications, machine learning may implement a model to determine features of an artwork that would otherwise require decisions directly made by an artist. However, if the models used for generative art contain biases, the resulting generative art may be skewed, inaccurate, or unrepresentative. Additionally, or alternatively, machine learning may be used to analyze art, but the models and / or data used to train the corresponding machine learning may also contain biases that can distort the analysis.

[0019] So far, biases have been estimated in various applications such as text, categorical data, voice data, images, and others. However, such bias estimation has been implemented in situations where the bias is clearly defined and the ground truth regarding the bias is known. As used herein, ground truth refers to information resulting from direct observation (e.g., empirical evidence) rather than information resulting from inference. However, it is not always possible to find the ground truth regarding bias through direct observation, and previous bias estimation techniques may not be effective in estimating bias when the ground truth regarding the bias is unknown. For example, some subjects such as the generation and / or analysis of art may be relatively subjective such that the data and models used in machine learning applications related to such subjects may embed biases (e.g., hidden or subconscious biases in information about subjects obtained from humans).

[0020] Accordingly, the disclosed embodiments may be used to identify, quantify, and adjust biases in a model, for example, when the ground truth regarding the bias is not necessarily known or obvious before the model is generated. Such embodiments may be used, for example, in machine learning applications to improve the models being used.

[0021] In the disclosed embodiments, an expert may provide expertise. The expertise may include information based on their field-specific knowledge regarding dependencies between various factors or variables of interest. Thus, the expert may provide information identifying dependencies that may exist in the data. This expert information may then be used, for example, in machine learning, to identify biases in the data or models that can be used. The expertise may be used to generate a causal model, and the structure of the causal model may be used to identify biases in either the data or the data-based model. Based on the structural causal model, a particular type of bias (whether it is data-specific or model-specific) may be determined. One particular type of bias that may be determined is confounding bias, which is a model-specific bias. Once the confounding bias is determined, the bias may be reduced or eliminated so that the results based on the model do not include the confounding bias.

[0022] Embodiments of the present disclosure are described with reference to the accompanying drawings.

[0023] Figure 1A is a diagram depicting an example environment 100 related to inferring events that occur in relation to a software program, arranged according to at least one embodiment described in the present disclosure. Environment 100 may include a causal model module 104. The causal model module 104 may obtain a model trained with respect to a target, such as a machine learning model 101. The machine learning model 101 may be based on a plurality of factors corresponding to the target. The causal model module 104 may also obtain specialized information provided by a human with respect to the target, such as specialized knowledge information 102. The specialized knowledge information 102 may indicate the relationships among a plurality of factors regarding how the factors influence each other. The causal model module 104 may identify input factors and output factors from among the plurality of factors. The input factors may have the potential to affect the result of the output factors, and the causal model module 104 may identify the input factors and output factors based on this relationship.

[0024] The causal model module 104 may be configured to generate a structural causal model 106. The structural causal model 106 may represent the relationships among a plurality of factors regarding how the factors influence each other. Thus, the structural causal model 106 may be based on the specialized knowledge information 102 and may indicate the relationships among a plurality of factors corresponding to the target. To generate the structural causal model 106, the causal model module 104 may determine the dependencies between the input factors and the output factors and / or the dependencies among the plurality of factors. In one example, the causal model module 104 may receive a text description of the specialized knowledge information 102 and may use the text description to determine the relationships among a plurality of factors corresponding to the target and / or the relationships among the factors.

[0025] Environment 100 may further include a confounding factor identification module 108. The confounding factor identification module 108 may identify a confounding factor 110 based on the structural causal model 106. The confounding factor 110 may be one of a plurality of factors from the expert knowledge information 102 (represented in the structural causal model 106 and representing a plurality of factors in the machine learning model 101). That factor can be identified as the confounding factor 110 because it causes confounding bias in the machine learning model 101.

[0026] Environment 100 may further include a bias estimation module 112. The bias estimation module 112 may be trained or configured to estimate an estimated confounding bias 114 based on the confounding factor 110 identified by the confounding factor identification module 108. The estimated confounding bias 114 may be an estimate of the confounding caused by the confounding factor 110.

[0027] The expert knowledge information 102 may be any suitable information obtained from an expert regarding a subject the expert is knowledgeable about. In some embodiments, the expert knowledge information 102 may be a textual representation of information obtained from an expert, although other configurations may be implemented. In some aspects, the expert knowledge information 102 may be based on the opinions of one or more experts. For example, if the subject is a work of art, the expert knowledge information 102 may be information regarding factors that affect the work of art. In such aspects, the factors may include geography, the style of the era, the style of the artist, and any other suitable factors that may affect a particular work of art or art in general.

[0028] In other examples, the expert may be a financial expert, and the subject may be the suitability for receiving a loan. In such a scenario, the factors may include income, credit information, existing debt, and any other appropriate factors that may affect the suitability for receiving a loan. In a further example, the expert may be a medical expert, and the subject may be the treatment of a patient with a drug. In such a scenario, the factors may include the patient's allergies, medical history, existing medication plan, pharmacological interactions, desired medical outcome, and any other appropriate factors that may affect the treatment of the patient with the drug.

[0029] The expert knowledge information 102 may include observed factors and unobserved factors corresponding to the subject. The observed factors may be factors that can be determined by direct observation (e.g., empirical evidence), and the unobserved factors may be factors that cannot be determined by direct observation. The expert knowledge information 102 may include any factors identified as relevant by the expert, and such an evaluation may be subjectively determined by the expert.

[0030] As described above, the structural causal model 106 may be generated by the causal model module 104. The structural causal model 106 may show the dependencies among the factors corresponding to the subject, which in turn may be determined based on the expert knowledge information 102. The expert knowledge information 102 may determine whether an edge exists between any two variables in the structural causal model 106 and, if so, what the direction of that edge is. Thus, the expert knowledge information 102 may specify or determine the structural causal model 106 based on expert knowledge.

[0031] Figure 1B represents an example of a structural causal model 106, although other representations of structural causal models may be implemented in accordance with the concepts described herein. In Figure 1B, the structural causal model 106 is represented by a directed acyclic graph (DAG) that represents the relationships between factors or variables (sometimes called nodes). In the example shown, the structural causal model 106 includes factors G, P, A, and L. The relationships between factors G, P, A, and L are represented by lines that extend between the factors. As shown, the lines are directed, that is, they have a single arrow indicating their influence. Thus, in the configuration shown in Figure 1B, factor G has an influence on factors P and A, and factor P has an influence on factors A and L.

[0032] The structural causal model 106 includes four factors G, P, A, and L and four relationships between them, although the structural causal model 106 may include any suitable number of factors and relationships between those factors. Further, the factors and relationships may be based on expert knowledge information 102 obtained from experts.

[0033] The structural causal model 106 may include factors that are input factors, output factors, and confounding factors. An input factor may be a factor whose influence on other factors can be determined. An output factor may be a factor whose causal effect of an input factor can be determined. A confounding factor may be a factor that causes a spurious correlation between an input factor and an output factor. As described above, in some situations, some of the factors may be observable (e.g., by empirical evidence), and some of the factors may not be observable.

[0034] In one example, the structural causal model 106 of FIG. 1B represents the relationships between factors corresponding to artworks. Thus, the factors may be based on information obtained from experts in the artwork (e.g., the object of the causal model is the artwork). In this example, factor G may represent geography, P may represent the style of the era, A may represent the style of the artist, and L may represent landscape painting (e.g., the object of the artwork). Thus, geography G gives rise to the style A of the artist, geography G also gives rise to the style P of the era, then the style P of the era gives rise to both the style A of the artist and the landscape painting L, and the style A of the artist gives rise to the landscape painting L. Specifically, the style P of the era (e.g., an art movement such as the Renaissance or Impressionism) may affect both the style A of the artist and the landscape painting L (e.g., the artwork). Next, the style P of the era may be affected by geography G (e.g., the region / nationality of the artist). Geography G may also independently affect the style A of the artist. These relationships may be encoded in the structural causal model 106 to describe artwork creation.

[0035] In this example, the structural causal model 106 may model the style of a particular artist, which is represented by the quantified causal effect of the artist's style A on the artwork landscape painting L. However, there may be a number of factors that affect the artwork, and the structure of the structural causal model 106 may vary based on expert opinion.

[0036] As described above, the confounding factor identification module 108 of FIG. 1A may identify the confounding factor 110 based on the structural causal model 106. In some aspects, causal inference analysis may be performed with respect to the structural causal model to identify factors as confounding factor 110. The confounding factor 110 may be identified based on the rules of causal inference. In the example of FIG. 1B, the rules of causal inference may indicate that the style P of the era is a confounding factor that causes confounding bias in the landscape painting L (e.g., the artwork).

[0037] As described above, the rules of causal inference may be used to determine confounding factor 110. In one example, a set of criteria called "d-separation" may be used to determine confounding factor 110. Considering a causal graph, the d-separation criterion may be used to determine the backdoor paths that need to be blocked to exclude the confounding effect. A backdoor path is a path from the output variable to the input variable. Such a path may need to be blocked because it introduces a spurious correlation between the input and the output. Variable Z is said to block the path between X and Y when conditioning on Z renders both X and Y independent. Thus, X is the input, Y is the output, and Z is the confounding factor that affects both X and Y, and in that case, conditioning on Z removes the backdoor path. D-separation determines which variables need to be conditioned on to block the backdoor path, and these are based on the direction of the arrows in the causal graph.

[0038] Referring again to FIG. 1A, bias estimation module 112 will be described in further detail. Once confounding factor 110 is identified, bias estimation module 112 may estimate an estimated confounding bias 114 caused by confounding factor 110 (e.g., in an artwork or other object).

[0039] Bias estimation module 112 may identify which parts of machine learning model 101 may be biased, i.e., which confounding factors (e.g., confounding factor 110) are implicitly or even unconsciously biasing the model (e.g., the analysis of an artwork or other object). Once the bias is identified, the bias may then be quantified using confounding factor 110.

[0040] In some aspects, estimating confounding bias may include determining one or more adjustment factors using a causal analysis tool. Estimating confounding bias may further include learning a representation of an image (e.g., artwork) using a residual network or other suitable neural network. Estimating confounding bias may further include determining similar images using a matching algorithm and calculating a difference between a perceived causal effect and an observed causal effect.

[0041] In some aspects, estimating bias may include determining a perceived causal effect using a structural causal model trained with respect to confounding factors, determining an observed causal effect using an objective analysis architecture, and estimating confounding bias based on a difference between the perceived causal effect and the observed causal effect.

[0042] In some situations, the subject may be image classification. In such an aspect, determining a perceived causal effect may include collating an image based on confounding factors (e.g., using a classifier model), and determining an observed causal effect may include collating an image using a matching algorithm. In such an aspect, the difference between the perceived causal effect and the observed causal effect may include a difference between collation based on confounding factors and collation using a matching algorithm.

[0043] Referring to an example where the subject is artwork, two sets of images belonging to two different artists may be compared. The two sets of images may be collated based on a confounding variable. Thus, images from the two sets may be paired based on the confounding variable, such that pairs of images have the same or similar confounding variables. Any remaining deviation between pairs of images may be determined to be caused by differences in the artists' styles.

[0044] One or more domain experts (e.g., experts in art or other subjects) may provide information for handling and control groups corresponding to two artist styles to identify images. Information from the domain experts may then be used to identify the factors of interest in each image. For example, the factor of interest may be a particular era style, and the expert may identify images corresponding to that particular era style. Using a classifier model trained to predict the style of an era, the best match between one of the two artist styles may be determined. The matching algorithm may be implemented to obtain a match using features extracted from the image. In some embodiments, the features may be extracted by using a residual network or other suitable neural network. In some embodiments, the matching algorithm may use an optimum or greedy search, which may be parameterized by a threshold for the match and may be able to have multiple matches. The values of the matched pairs may then be used to estimate counterfactuals. If there are more than one match, the average of all the matches may be used as the counterfactual.

[0045] The confounding bias in modeling the artist's style may be determined. The artist's style may be modeled, for example, using a machine learning model. The set of images of the artist's actual artworks may be represented by A, and the set of computer-generated images (e.g., generative art) in the artist's style may be represented by B. Thus, A corresponds to the artist's actual artworks, and B corresponds to the artist's generative artworks. The two artists in this scenario are the actual artist and the machine learning model, and the confounding bias between A and B can be determined.

[0046] Causal models may be obtained from the knowledge of art experts. Using the rules of causal inference, we can obtain a set of variables that we need to adjust to remove confounding bias. For example, the variables may include the styles of the times. Thus, the machine learning model may be trained to match the images in set A with the images in set B based on the styles of the times. When we do this, any remaining differences between the images in set A and the images in set B should be caused by confounding bias, as all other factors are the same between set A and set B. To determine the styles of the times, a neural network for image recognition may be implemented. Accordingly, the images of set A and set B may be represented by vectors of the styles of the times.

[0047] Once the feature representations of the images of the two sets A and B are obtained, the nearest neighbors of the elements in set B to the elements in set A may be determined using a matching procedure. If there is more than one match, the average of all the matches may be used.

[0048] In some aspects, the bias may be quantified as the difference between the perceived causal effect and the observed causal effect. For example, the bias may be given by the formula:

Number

[0049] Here, the observed effect y corresponds to the concatenated features of the images without adjustment of the confounding factors, and the perceived causal effect is the concatenated feature representation of the counterfactual obtained later by incorporating expert knowledge and adjusting the confounding factors (e.g., by matching with similar time styles among the styles of other artists). In the above formula, for each image i in set B, the match may be determined as the average of all the nearest neighbors in set A. Thus, M i corresponds to the set of all the nearest neighbors of the image i belonging to set B with respect to set A. j represents the image corresponding to the nearest neighbor of i in set A. y i is, for example, the feature representation of the image i regarding the style of the times obtained from a neural network for image recognition.

[0050] As described above, the cross bias may be calculated as the difference between the feature representation of an image and its nearest neighbor in the other set. This is because, assuming that all other factors are the same between the two sets A and B, any difference between the two images should correspond to the bias in modeling the artist's style.

[0051] In some embodiments, quantifying the cross bias may enable the machine learning model to be adjusted to reduce or eliminate the cross bias within the model. Quantifying the cross bias may also enable the identification of bias in generative artwork or the identification of the influence of a particular past artwork on generative artwork. Quantifying the cross bias may also assist in setting the value and composition of generative artwork, or assist art historians specializing in art in studying and understanding the style of new artworks.

[0052] Furthermore, quantifying the cross bias may also enable the machine learning model to be adjusted to reduce or eliminate the cross bias. Accordingly, the bias estimation module 112 may be configured to adjust the machine learning model based on the estimated cross bias to reduce or eliminate the cross bias.

[0053] Without departing from the scope of the present disclosure, changes, additions, or deletions may be made to FIG. 1A. For example, the environment 100 may include more or fewer elements than those illustrated and described in the present disclosure.

[0054] Furthermore, the various modules described with respect to environment 100 may include code and routines configured to enable a computer device to perform one or more operations described with respect to the corresponding module. Additionally, or alternatively, one or more modules may be implemented using hardware including a processor, a microprocessor (e.g., for performing or controlling one or more operations), a field programmable gate array (FPGA), or an application specific integrated circuit (ASIC). In some other instances, one or more modules may be implemented using a combination of hardware and software. In the present disclosure, the operations described as being performed by a particular module include operations that the particular module may direct the corresponding system to perform.

[0055] Furthermore, in some embodiments, at least a portion of one or more routines, one or more instructions, or code of two or more of the described modules may be combined such that they may be considered the same element or may have a common section that may be a contemplated portion of two or more modules. The depiction of the modules in the specification and FIG. 1A is thus intended to simplify the description of the operations being performed and is not given as a limiting implementation or configuration.

[0056] FIG. 2 represents a block diagram of an example computer system 202 according to at least one embodiment of the present disclosure. The computer system 202 may be configured to implement or direct one or more operations associated with a causal model module, an interference factor identification module, and / or a bias estimation module (e.g., causal model module 104, structural causal model, and / or bias estimation module 112 of FIG. 1A). The computer system 202 may include a processor 250, a memory 252, and a data storage 254. The processor 250, the memory 252, and the data storage 254 may be communicatively coupled.

[0057] In general, processor 250 may include any suitable dedicated or general-purpose computer, computing entity, or processing device that includes various computer hardware or software modules, and may be configured to execute instructions stored on any applicable computer-readable storage medium. For example, processor 250 may include a microprocessor, a microcontroller, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or any other digital or analog circuit configured to interpret and / or execute program instructions and / or process data. Although shown as a single processor in FIG. 2, processor 250 may include any number of processors configured to individually or collectively execute or direct the execution of any number of the operations described in this disclosure. Additionally, one or more of the processors may be present in one or more different electronic devices, such as different servers.

[0058] In some embodiments, processor 250 may be configured to interpret and / or execute program instructions stored in memory 252, data storage 254, or both memory 252 and data storage 254, and / or process data stored therein. In some embodiments, processor 250 may fetch program instructions from data storage 254 and load the program instructions into memory 252. After the program instructions are loaded into memory 252, processor 250 may execute the program instructions.

[0059] For example, in some embodiments, the above-described modules (the causal model module 104, the structural causal model 106, and / or the bias estimation module 112 of FIG. 1A) may be included in the data storage 254 as program instructions. The processor 250 may fetch the program instructions of the corresponding module from the data storage 254 and may load the program instructions of the corresponding module into the memory 252. After the program instructions of the corresponding module are loaded into the memory 252, the processor 250 may execute the program instructions so that the computer system 202 can implement the operations associated with the corresponding module as desired by the instructions.

[0060] The memory 252 and the data storage 254 may include a computer-readable storage medium that carries or stores computer-executable instructions or data structures. Such a computer-readable storage medium may include any available medium that can be accessed by a general-purpose or special-purpose computer such as the processor 250. By way of example, and not limitation, such computer-readable storage media may include random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disk storage, magnetic disk storage or other magnetic storage devices, flash memory devices (e.g., solid state memory devices), or any other tangible or non-transitory computer-readable storage medium that may be used to carry or store particular program code in the form of computer-executable instructions or data structures and that can be accessed by a general-purpose or special-purpose computer. Combinations of the above may also be included within the scope of computer-readable storage media. Computer-executable instructions may include, for example, instructions and data configured to cause the processor 250 to perform a particular operation or group of operations.

[0061] Changes, additions, or deletions may be made to the computer system 202 without departing from the scope of the present disclosure. For example, in some embodiments, the computer system 202 may include any number of other components that may not be explicitly illustrated or described.

[0062] FIG. 3 is a flowchart of a method 300 that serves as an example of quantifying crosstalk bias based on expertise according to at least one embodiment described in the present disclosure. The method 300 may be executed by any suitable system, apparatus, or device. For example, one or more of the causal model module 104, the structural causal model 106, and / or the bias estimation module 112 of FIG. 1A, or the computer system 202 of FIG. 2 (e.g., directed by one or more modules), may perform one or more operations related to the method 300. Although represented as separate blocks, the steps and operations associated with one or more of the blocks of the method 300 may be divided into further blocks, combined into fewer blocks, or deleted, depending on a particular implementation.

[0063] The method 300 may begin at block 301, where a machine learning model may be obtained. The machine learning model may be trained with respect to a target. The machine learning model may be based on a plurality of factors corresponding to the target.

[0064] At block 302, expert information provided by a human with respect to the target may be obtained. In some embodiments, the information may each include a plurality of factors corresponding to the target. The expert information may indicate the relationships among the plurality of factors with respect to how the plurality of factors affect each other.

[0065] In block 304, a structural causal model may be generated. In some embodiments, the structural causal model may show the relationships between multiple factors regarding how the multiple factors affect each other. The structural causal model may be based on expert information and may represent the relationships between multiple factors. In some aspects, generating the structural causal model includes determining the dependencies between input factors and output factors identified from the multiple factors.

[0066] In some aspects, generating the structural causal model may include identifying input factors and output factors from the multiple factors. Since an input factor may affect the result of an output factor, one of the multiple factors may be identified as an input factor if it affects the result of other factors. Further, one of the multiple factors may be identified as an output factor if it is affected by other factors (e.g., input factors).

[0067] In block 306, confounding factors may be identified. In some embodiments, a factor among the multiple factors may be identified as a confounding factor in that it causes a confounding bias with respect to the information about the subject. That factor among the multiple factors may be identified as a confounding factor based on the structural causal model. In some aspects, identifying a factor as a confounding factor may include performing a causal inference analysis on the structural causal model.

[0068] In block 308, the confounding bias may be estimated. Additionally, or alternatively, the confounding bias may be quantified. In some embodiments, the confounding bias may be estimated based on the identified confounding factors. In some aspects, estimating the confounding bias may include determining a perceived causal effect related to the subject. Estimating the confounding bias may include using an objective analysis architecture to determine the observed causal effect. Estimating the confounding bias may include determining the difference between the perceived causal effect and the observed causal effect.

[0069] In some embodiments of method 300, the object may be image classification. In such a configuration, determining the perceived causal effect may include matching images based on confounding factors. Additionally, or alternatively, determining the observed causal effect may include matching images using a matching algorithm. The difference between the perceived causal effect and the observed causal effect may include the difference between matching based on confounding factors and the matching using the matching algorithm.

[0070] In some embodiments, method 300 may further include adjusting a machine learning model based on an estimated confounding bias to reduce and / or eliminate the confounding bias.

[0071] In some aspects of method 300, the object may be artwork, and the expert information provided by humans may be obtained from art experts.

[0072] The disclosed configurations are advantageous because they do not require ground truth regarding bias. Rather, the bias may be tested based on assumptions about the structure in the causal model. Such configurations may be applicable in various settings and may be generalizable across data modalities. Further, the disclosed configurations incorporate domain-specific knowledge to improve the identification and quantification of bias.

[0073] Without departing from the scope of the present disclosure, changes, additions, or deletions may be made to method 300. For example, the operations of method 300 may be implemented in a different order. Additionally, or alternatively, two or more operations may be executed simultaneously. Further, the operations and actions described are provided merely as examples, and some of the operations and actions may be optional, combined into fewer operations and actions, or extended to additional operations and actions without departing from the essence of the disclosed embodiments.

[0074] As described above, embodiments described in this disclosure may involve the use of a special-purpose or general-purpose computer (e.g., the processor 250 of FIG. 2) that includes various computer hardware or software modules, as will be discussed in more detail below. Further, as described above, embodiments described in this disclosure may be implemented using a computer-readable medium (e.g., the memory 252 or data storage 254 of FIG. 2) that conveys or stores computer-executable instructions or data.

[0075] As used in this disclosure, the terms "module" or "component" can refer to a particular hardware implementation configured to perform the operations of the module or component, and / or a software object or software routine that can be stored in and / or executed by the general-purpose hardware of a computer system (e.g., a computer-readable medium, a processing device, etc.). In some embodiments, the different components, modules, engines, and services described in this disclosure may be implemented as objects or processes that execute in a computer system (e.g., as separate threads). Although some of the systems and methods described in this disclosure are generally described as being implemented in software (stored in and / or executed by general-purpose hardware), specific hardware implementations, or combinations of software and specific hardware implementations are also possible and contemplated. As used herein, a "computing entity" can be any computer system as previously defined in this disclosure, or any module or combination of modules that execute in a computer system.

[0076] In this disclosure, terms used especially in the appended claims (e.g., the body of the appended claims) are generally intended as "open" terms (e.g., the term "including" should be construed to mean "including, but not limited to", the term "having" should be construed to mean "having at least", the term "includes" should be construed to mean "includes, but not limited to", etc.).

[0077] Furthermore, when a specific number is intended in an introduced claim recitation, such intent is clearly recited in the claim, and in the absence of such a recitation, there is no such intent. For example, for purposes of illustration, in the following appended claims, introductory phrases such as "at least one" and "one or more" may be used to introduce claim recitations. However, just because such phrases are used, it should not be construed that when a claim recitation is introduced by an indefinite article such as "a" or "an", even if both an introductory phrase such as "one or more" or "at least one" and an indefinite article such as "a" or "an" are included within the same claim, a particular claim containing the introduced claim recitation is limited to examples that contain only one of the recited matters (e.g., "a" and / or "an" should be construed to mean "at least one" or "one or more"). The same applies when a claim recitation is introduced using a definite article.

[0078] Furthermore, even if a specific number is specified in the introduced claim description, it will be understood by those skilled in the art that such a description should generally be interpreted to mean at least the number described (for example, if there is a description of simply "two items" without any other modifiers, this description means at least two items, or two or more items). Further, when expressions such as "at least one of A, B, and C" or "one or more of A, B, and C" are used, generally, such a construction is intended to include only A, only B, only C, both A and B, both A and C, both B and C, and / or all of A, B, and C, etc.).

[0079] Furthermore, any disjunctive and / or disjunctive clause representing two or more alternative terms, whether in the specification, the claims, or the drawings, should be understood to be intended to include one of those terms, any of those terms, or both of those terms. For example, the clause "A or B" should be understood to include the possibilities of "A or B" or "A and B".

[0080] All examples and conditional language given in this disclosure are intended for the educational purpose of assisting the reader in understanding the concepts contributed by the inventor to the advancement of the art and the present invention, and should be interpreted as not being limited to such specifically given examples and conditions. Although embodiments of the present disclosure have been described in detail, various changes, substitutions, and alternatives may be made without departing from the spirit and scope of the present disclosure.

[0081] In addition to the above embodiments, the following appendices are disclosed. (Appendix 1) Obtaining a machine learning model trained with respect to an object, the machine learning model being based on a plurality of factors corresponding to the object, and Obtaining specialized information provided by humans regarding the object, where the specialized information indicates the relationship between the plurality of factors regarding how the plurality of factors affect each other, Generating a structural causal model representing the relationship between the plurality of factors based on the specialized information, Identifying, as confounding factors, factors among the plurality of factors that cause confounding bias in the machine learning model based on the structural causal model, Estimating the confounding bias based on the identified confounding factors A method having the above. (Appendix 2) Further comprising adjusting the machine learning model based on the estimated confounding bias so as to reduce the confounding bias, The method according to Appendix 1. (Appendix 3) The object is an artwork, The specialized information provided by the human is obtained from art experts, The method according to Appendix 1. (Appendix 4) Identifying the factors as the confounding factors includes performing causal inference analysis on the structural causal model, The method according to Appendix 1. (Appendix 5) Generating the structural causal model includes Identifying input factors and output factors from the plurality of factors, where the input factors affect the result of the output factors, Determining the dependency between the input factors and the output factors And including The method according to Appendix 1. (Appendix 6) Estimating the confounding bias includes Determining the perceived causal effect related to the object, Determining the observed causal effect using an objective analysis architecture, Determining the difference between the perceived causal effect and the observed causal effect including the method according to Appendix 1. (Appendix 7) The object is image classification, determining the perceived causal effect includes collating images based on the confounding factor, determining the observed causal effect includes collating images using a matching algorithm, the difference between the perceived causal effect and the observed causal effect includes the difference between the collation based on the confounding factor and the collation using the matching algorithm, the method according to Appendix 6. (Appendix 8) A system having one or more processors and one or more computer-readable storage media storing instructions wherein, in response to being executed by the one or more processors, the instructions cause the system to acquire a machine learning model trained with respect to an object, the machine learning model being based on a plurality of factors corresponding to the object, acquire expert information provided by a human with respect to the object, the expert information indicating the relationship between the plurality of factors regarding how the plurality of factors affect each other, generate a structural causal model representing the relationship between the plurality of factors based on the expert information, identify, as a confounding factor, a factor among the plurality of factors that causes a confounding bias in the machine learning model based on the structural causal model, estimate the confounding bias based on the identified confounding factor and cause the system to perform operations having. (Appendix 9) The operations further include adjusting the machine learning model based on the estimated confounding bias so as to reduce the confounding bias. The system described in Supplementary Note 8. (Supplementary Note 10) The object is an artwork, and the specialized information provided by the human is obtained from an expert in art. The system described in Supplementary Note 8. (Supplementary Note 11) Identifying the factor as the confounding factor includes performing causal inference analysis on the structural causal model. The system described in Supplementary Note 8. (Supplementary Note 12) Generating the structural causal model is to identify input factors and output factors from the plurality of factors, where the input factors affect the result of the output factors, and to determine the dependency between the input factors and the output factors and includes The system described in Supplementary Note 8. (Supplementary Note 13) Estimating the confounding bias includes determining the perceived causal effect related to the object, determining the observed causal effect using an objective analysis architecture, and determining the difference between the perceived causal effect and the observed causal effect and includes The system described in Supplementary Note 8. (Supplementary Note 14) The object is image classification, determining the perceived causal effect includes collating images based on the confounding factor, determining the observed causal effect includes collating images using a matching algorithm, the difference between the perceived causal effect and the observed causal effect includes the difference between the collation based on the confounding factor and the collation using the matching algorithm. The system described in Supplementary Note 13. (Supplementary Note 15) In response to being executed by one or more processors of the system, cause the system to obtain a machine learning model trained with respect to a target, the machine learning model being based on a plurality of factors corresponding to the target, and obtain expert information provided by a human with respect to the target, the expert information indicating relationships among the plurality of factors regarding how the plurality of factors influence each other, and generate a structural causal model representing the relationships among the plurality of factors based on the expert information, and as confounding factors, identify, based on the structural causal model, factors among the plurality of factors that cause confounding bias in the machine learning model, and estimate the confounding bias based on the identified confounding factors One or more computer-readable storage media storing instructions that cause an operation having the above. (Appendix 16) The operation further includes adjusting the machine learning model based on the estimated confounding bias so as to reduce the confounding bias. One or more computer-readable storage media according to Appendix 15. (Appendix 17) Identifying the factors as the confounding factors includes performing causal inference analysis on the structural causal model. One or more computer-readable storage media according to Appendix 15. (Appendix 18) Generating the structural causal model includes identifying input factors and output factors from the plurality of factors, the input factors affecting the result of the output factors, and determining the dependencies between the input factors and the output factors including One or more computer-readable storage media according to Appendix 15. (Appendix 19) Estimating the confounding bias includes Determining the perceived causal effect related to the object Using an objective analysis architecture to determine the observed causal effect Determining the difference between the perceived causal effect and the observed causal effect Including One or more computer-readable storage media described in Appendix 15 (Appendix 20) The object is image classification Determining the perceived causal effect includes collating images based on the confounding factor Determining the observed causal effect includes collating images using a matching algorithm The difference between the perceived causal effect and the observed causal effect includes the difference between the collation based on the confounding factor and the collation using the matching algorithm One or more computer-readable storage media described in Appendix 19

Explanation of symbols

[0082] 100 Environment 101 Machine learning model 102 Expert knowledge information 104 Causal model module 106 Structural causal model 108 Confounding factor identification module 110 Confounding factor 112 Bias estimation module 114 Estimated confounding bias 202 Computer system 250 Processor 252 Memory 254 Data storage

Claims

1. A method executed by a processor, comprising: obtaining a machine learning model trained with respect to a target, the machine learning model being based on a plurality of factors corresponding to the target; obtaining expert information provided by a human with respect to the target, the expert information indicating the relationships among the plurality of factors regarding how the plurality of factors affect each other; generating a structural causal model representing the relationships among the plurality of factors based on the expert information; identifying, as confounding factors, factors among the plurality of factors that cause confounding bias in the machine learning model based on the structural causal model; estimating the confounding bias based on the identified confounding factors; wherein estimating the confounding bias comprises: obtaining real data corresponding to the target; obtaining generated data corresponding to the target by using the machine learning model; determining a perceived causal effect related to the target by performing a comparison based on the confounding factors between the real data and the generated data; determining an observed causal effect related to the target by performing a comparison using a feature matching algorithm between the real data and the generated data; determining a difference between the perceived causal effect and the observed causal effect, and quantifying the confounding bias as the difference; wherein the difference between the perceived causal effect and the observed causal effect includes the difference between the comparison based on the confounding factors regarding which parts of the real data match which parts of the generated data and the comparison using the feature matching algorithm; a method.

2. The method according to claim 1, further comprising adjusting the machine learning model based on the estimated confounding bias to reduce the confounding bias. The method according to claim 1.

3. The target is a work of art, and the expert information provided by the human is obtained from an art expert. The method according to claim 1.

4. Identifying the factors as the confounding factors includes performing causal inference analysis with respect to the structural causal model. The method according to claim 1.

5. Generating the structural causal model comprises: Identifying input factors and output factors from the plurality of factors, wherein the input factors affect the result of the output factors Determining the dependency between the input factor and the output factor Including The method according to claim 1

6. The object is image classification, and the real data and the generated data each include an image Determining the perceived causal effect includes collating images based on the confounding factor Determining the observed causal effect includes collating images using the feature matching algorithm The method according to claim 1

7. A system having one or more processors and one or more computer-readable storage media storing instructions Wherein In response to being executed by the one or more processors, the instructions cause the system to Obtain a machine learning model trained with respect to an object, the machine learning model being based on a plurality of factors corresponding to the object Obtain expert information provided by a human with respect to the object, the expert information indicating the relationship between the plurality of factors regarding how the plurality of factors affect each other Generating a structural causal model representing the relationship between the plurality of factors based on the expert information As a confounding factor, based on the structural causal model, identifying a factor among the plurality of factors that causes a confounding bias in the machine learning model Estimating the confounding bias based on the identified confounding factor Execute an operation having Estimating the confounding bias includes Obtaining real data corresponding to the object Obtaining generated data corresponding to the object by using the machine learning model Determining a perceived causal effect related to the object by performing a collation based on the confounding factor between the real data and the generated data Determining an observed causal effect related to the object by performing a collation using a feature matching algorithm between the real data and the generated data Determining the difference between the perceived causal effect and the observed causal effect, and quantifying the confounding bias as the difference The difference between the perceived causal effect and the observed causal effect includes the difference between matching based on confounding factors regarding which part of the real data coincides with which part of the generated data and matching using the feature matching algorithm. System. Claim 8 The operation further comprises adjusting the machine learning model based on the estimated confounding bias to reduce the confounding bias. The system according to claim 7. Claim 9 The object is artworks, The expert information provided by the human is obtained from art experts. The system according to claim 7. Claim 10 Identifying the factor as a confounding factor includes performing causal inference analysis on the structural causal model. The system according to claim 7. Claim 11 Generating the structural causal model includes identifying input factors and output factors from the plurality of factors, where the input factors affect the result of the output factors, and determining the dependency between the input factors and the output factors including. The system according to claim 7. Claim 12 The object is image classification, the real data and the generated data each include images, Determining the perceived causal effect includes matching images based on the confounding factors. Determining the observed causal effect includes matching images using the feature matching algorithm. The system according to claim 7. Claim 13 In response to being executed by one or more processors of the system, the system obtains a machine learning model trained with respect to an object, the machine learning model being based on a plurality of factors corresponding to the object, and obtains expert information provided by a human with respect to the object, the expert information indicating the relationship between the plurality of factors regarding how the plurality of factors affect each other, and generates a structural causal model representing the relationship between the plurality of factors based on the expert information, and identifies, as confounding factors, factors among the plurality of factors that cause confounding bias in the machine learning model based on the structural causal model, and estimates the confounding bias based on the identified confounding factors Store an instruction to execute an operation having Estimating the confounding bias involves Obtaining real data corresponding to the target, Obtaining generated data corresponding to the target by using the machine learning model, Determining a recognized causal effect related to the target by performing a comparison based on the confounding factor between the real data and the generated data, Determining an observed causal effect related to the target by performing a comparison using a feature matching algorithm between the real data and the generated data, Determining the difference between the recognized causal effect and the observed causal effect, and quantifying the confounding bias as the difference, The difference between the recognized causal effect and the observed causal effect includes the difference between the comparison based on the confounding factor regarding which part of the real data matches which part of the generated data and the comparison using the feature matching algorithm, One or more computer-readable storage media.

14. The operation further includes adjusting the machine learning model based on the estimated confounding bias to reduce the confounding bias. One or more computer-readable storage media according to claim 13.

15. Identifying the factor as the confounding factor includes performing causal inference analysis regarding the structural causal model. One or more computer-readable storage media according to claim 13.

16. Generating the structural causal model includes Identifying input factors and output factors from the plurality of factors, where the input factors affect the result of the output factors, Determining the dependency between the input factor and the output factor including One or more computer-readable storage media according to claim 13.

17. The target is image classification, and the real data and the generated data each include an image. Determining the recognized causal effect includes comparing images based on the confounding factor. Determining the observed causal effect includes comparing images using the feature matching algorithm. One or more computer-readable storage media according to claim 13.

Images