Data set fairness

Abstract

The embodiment of the invention relates to data set fairness. The present application relates to a method of obtaining a machine learning model and a data set for training the machine learning model to estimate a fairness-accuracy curve of the machine learning model and determine a confidence interval of the fairness-accuracy model, and a system and software for performing the method.

Description

Technical Field

[0001] The present application relates to a method for obtaining a machine learning model and a data set, wherein the data set is used to train the machine learning model to estimate the fairness-accuracy curve of the trained machine learning model and determine the confidence interval of the fairness-accuracy model; and a system and software for executing the method. Background Art

[0002] In terms of machine learning fairness, training models that minimize differences across different sensitive groups in a dataset often results in reduced accuracy of the trained machine learning model. The relationship between fairness and accuracy for machine learning models and datasets is often referred to as the fairness-accuracy tradeoff. The severity of this tradeoff fundamentally depends on the characteristics of the dataset, such as the imbalance or bias of the dataset, and thus using a universal fairness requirement across different datasets may lead to unreliable fairness-accuracy determinations.

[0003] This poses significant challenges for deploying these models in real-world settings. One application of determining and tuning fairness-accuracy in machine learning is training machine learning models to minimize disparities across various sensitive groups such as race or gender. For example, consider two crime datasets: dataset A contains records from communities where crime rates are evenly distributed across all racial groups, while dataset B is from communities where historical factors have led to disproportionate crime rates among specific racial groups. Intuitively, training a race-agnostic model is more challenging for the latter due to the uneven distribution of crime rates across racial groups in the former. Therefore, applying the same fairness-safety trade-off to both datasets may not be appropriate to account for the different characteristics and potential biases of the two datasets.

[0004] This example highlights one of the major challenges in fairness for machine learning models. More specifically, it is impractical to set uniform requirements for fairness across different datasets while adhering to a basic accuracy benchmark. Current approaches seek to determine the accuracy of a model for a given fairness value. However, in order to maintain a basic model accuracy benchmark, it is necessary to determine the range of fairness violations that can be achieved to achieve the desired model accuracy. Specifically, it is desirable to determine the range of allowable fairness violations for a given dataset, model class, and desired accuracy. Attempts have been made to determine the accuracy of a model for a given fairness value; however, current approaches are computationally inefficient and only determine model accuracy for a single fairness value.

[0005] Therefore, a computationally efficient method is needed to approximate the fairness-accuracy trade-off tailored to individual models and datasets over a wide range of fairness and accuracy values.

[0006] The embodiments described below are not limited to implementations that solve any or all of the disadvantages of the known approaches described above. Summary of the invention

[0007] The present invention content is provided to introduce a series of concepts in a simplified form, which are further described in the following detailed description. The present invention content is not intended to identify the key features or essential features of the subject matter claimed for protection, nor is it intended to be used to determine the scope of the subject matter claimed for protection; variations and alternative features that facilitate the work of the present invention and / or are used to achieve substantially similar technical effects should be considered to fall within the scope of the present invention disclosed herein.

[0008] According to a first aspect, a computer-implemented method is provided, the method comprising: obtaining a machine learning model; obtaining a data set for training the machine learning model to produce a trained machine learning model; estimating a fairness-accuracy curve of the trained machine learning model by training the machine learning model on the data set, wherein the fairness-accuracy curve comprises a plurality of fairness values ​​corresponding to a plurality of accuracy values; and determining confidence intervals for the plurality of fairness values ​​of the fairness-accuracy curve, wherein the confidence intervals are a range of expected fairness values ​​for the fairness values ​​corresponding to the accuracy values.

[0009] According to one embodiment, the method discloses that determining the confidence interval includes using a calibration data set, wherein the calibration data set is disjoint from the data set.

[0010] According to one embodiment, the method discloses that the probability that the fairness value is within the range of expected fairness values ​​is defined as:

[0011]

[0012] in: is the fairness-accuracy curve, and is a confidence interval, where α is the significance level, where α∈(0,1); and ψ is a random variable.

[0013] According to one embodiment, the method discloses a base truth fairness-accuracy curve that is within a confidence interval of the fairness-accuracy curve.

[0014] According to one embodiment, the method discloses a base truth fairness-accuracy curve that is within a confidence interval of the fairness-accuracy curve.

[0015] According to one embodiment, the method discloses estimating a fairness-accuracy curve includes utilizing a one-shot training framework.

[0016] According to one embodiment, the method discloses that the method includes dynamically estimating a fairness-accuracy curve during inference of a machine learning model.

[0017] According to one embodiment, the method discloses that each of the plurality of fairness values ​​represents a compromised fairness value depending on its corresponding plurality of accuracy values.

[0018] According to one embodiment, the method discloses that the fairness-accuracy curve to be estimated is:

[0019]

[0020] in: is the fairness-accuracy curve, and Ψ is an accuracy value from a plurality of accuracy values, and Ψ∈[0,1]; h is a model from a set of models H; Φ fair () is a function for determining fairness; and Acc() is a function for determining accuracy.

[0021] According to one embodiment, the method discloses that training the model on the dataset includes training the model on a regularization loss defined as:

[0022] L λ (θ)=E[L CE (h θ (X), Y)]+λL fair (h θ )

[0023] Where: L λ (θ) is the regularized loss function; θ is the weight of the model; E[] is the expected value function; L CE is the cross entropy loss function; h θ is the model; X is the feature vector; Y is the label feature; λ is the regularization parameter; L fair is a fairness violationΦ fair (h) Fairness violation loss function.

[0024] According to one embodiment, the method discloses that training the model on the dataset includes training the model on a regularization loss defined as:

[0025]

[0026] Where: h θ is the model; X is the feature vector; Λ is the space of regularization parameter λ; λ is the regularization parameter; P λ is the distribution of the regularization parameter; L CE is the cross entropy loss function; Y is the label feature; and L fairis a fairness violation fair (h) Fairness violation loss function.

[0027] According to one embodiment, the method discloses that the fairness-accuracy curve is estimated as

[0028]

[0029] in: is the fairness-accuracy curve, and ψ is the accuracy value from multiple accuracy values, and ψ∈[0,1]; i is the index of the dataset sample; k is the total number of samples in the dataset; is a function that determines fairness; is the function used to determine accuracy; and It's a model.

[0030] According to one embodiment, the method discloses that the degree of fairness is at least one of: group fairness, equal opportunity, and equal probability.

[0031] According to a second aspect, there is a data processing system comprising means for performing any of the above methods.

[0032] According to a third aspect, there is a computer program comprising instructions which, when the program is executed by a computer, cause the computer to perform any one of the above methods.

[0033] A computer-readable medium comprising instructions which, when executed by a computer, cause the computer to perform any of the above methods.

[0034] The methods described herein may be performed by software in a machine-readable form on a tangible storage medium, for example in the form of a computer program comprising computer program code means adapted to perform all the steps of any of the methods described herein when the program is run on a computer, and wherein the computer program may be embodied on a computer-readable medium. Examples of tangible (or non-transitory) storage media include disks, thumb drives, memory cards, etc., and do not include propagating signals. The software may be suitable for execution on a parallel processor or a serial processor, so that the method steps may be performed in any suitable order or simultaneously.

[0035] This application recognizes that firmware and software can be valuable, separately tradable commodities. It is intended to cover software that runs on or controls "dumb" or standard hardware to perform desired functions. It is also intended to cover software that "describes" or defines the configuration of hardware, such as HDL (Hardware Description Language) software used to design silicon chips or configure general-purpose programmable chips to perform desired functions.

[0036] Preferred features may be combined as appropriate, as will be apparent to those skilled in the art, and may be combined with any aspect of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Embodiments of the present invention will be described by way of example with reference to the following drawings, in which:

[0038] Figure 1 A flow chart of a computer-implemented method for estimating a fairness-accuracy curve for a trained machine learning model and determining a confidence interval for the fairness-accuracy curve is shown;

[0039] Figure 2 showing a graph including a fairness-accuracy curve and corresponding confidence intervals for a trained machine learning model; and

[0040] Figure 3 A computing system 2000 is shown on which any of the above-described methods may be performed.

[0041] Common reference numbers are used throughout the drawings to indicate similar features. DETAILED DESCRIPTION

[0042] Embodiments of the present invention are described below by way of example only. These examples represent the best modes currently known to the applicant for putting the present invention into practice, although they are not the only way to achieve this goal. This description sets forth the functions of the examples and the sequence of steps for constructing and operating the examples. However, the same or equivalent functions and sequences may be implemented by different examples.

[0043] Figure 1 A flow chart of a computer-implemented method for estimating a fairness-accuracy curve for a trained machine learning model and determining a confidence interval for the fairness-accuracy curve is shown.

[0044] The method provides a computationally efficient way to estimate the fairness-accuracy curve and provides a range of expected fairness values ​​for each accuracy value.

[0045] In step 10, the method includes obtaining a data set for training a machine learning model. The data set may include multiple training samples. All methods and systems described herein may be used for any type of data set. Example data sets include, but are not limited to, at least one of the following data sets: a tabular data set, a computer vision data set, a time series data set, a graphical data set, an audio data set, a geospatial data set, and a natural language data set.

[0046] Each training sample may include the following triple (X, A, Y). X∈X represents a feature vector, where X is a feature vector from feature vector set X. A∈A represents a sensitive attribute, where A is a sensitive attribute from feature vector set A. Y∈Y:={0,1} represents a label, where Y is a label from label set Y. In the illustrated embodiment, a possible label set may be defined as Y:={0,1}; however, those skilled in the art will appreciate that other label sets may be defined.

[0047] The feature vector may include, but is not limited to, at least one of the following: group characteristics; gender; race; ethnicity; salary; height; any personal characteristics; and any other characteristics depending on the technical problem to be solved.

[0048] Sensitive attributes may include, but are not limited to, at least one of the following: gender; race; ethnicity; salary; wealth; and any other characteristics depending on the technical problem being solved.

[0049] The labels may include, but are not limited to, at least one of the following: binary classification; category classification; and any other labels depending on the technical problem to be solved.

[0050] In a specific example, feature vector X includes a person's income level, the person's loan amount, the person's previous payment record; sensitive attributes A include the person's racial identity; and labels Y include loan default predictions.

[0051] In step 20, the method includes obtaining a machine learning model. The machine learning model is denoted as "h" in this article. The machine learning model h is taken from the model class H. For example, the model class can be a neural network. This type of model can be a family of parameterized classifiers For example, a neural network, whose prediction value is R is a real number. θ is the weight of model h. is the parameter space of model h. For each parameter space There may be R d space, where the weight θ is R d A vector in space.

[0052] The machine learning model can be any model. Example machine learning models include, but are not limited to, neural networks and any other gradient-based models.

[0053] In step 30, the method includes estimating a fairness-accuracy curve for the trained machine learning model.

[0054] The accuracy of a machine learning model can be defined as the percentage of correct predictions. The accuracy of a model h is denoted as acc(h). The accuracy threshold, or the expected minimum accuracy of a model h, is denoted as ψ, where ψ∈[0,1].

[0055] The fairness value may be any measure of fairness of a model trained on a data set (hereinafter referred to as a trained model). In one embodiment, the fairness value may be a fairness violation value, where the fairness violation value is an amount of fairness that is weighed depending on an accuracy value. Thus, each of the plurality of fairness values ​​may represent a compromised fairness value depending on its corresponding plurality of accuracy values. Examples include, but are not limited to, at least one of: population equality, equal opportunity, and equal probability.

[0056] Population equality refers to when all sensitive attributes A have equal selection rates. For any a∈A, ideal population equality is defined as P(h(X)=1|A=a)=P(h(X)=1), where A is an example sensitive attribute. The fairness violation value of population equality is defined as

[0057] φ DP (h):=|P(h(X)=1|A=1)-P(h(X)=1|A=0)|:=|P(.

[0058] Equal opportunity is when the true positive rates of all sensitive groups are equal. For any a∈A, ideal equal opportunity is defined as P(h(X)=1|A=a,Y=1)=P(h(X)=1|Y=1). The fairness violation value of equal opportunity is defined as

[0059] φ EOP (h)=|P(h(X)=1|A=1, Y=1)-P(h(X)=1|A=0, Y=1)|.

[0060] :=|P(When Y=A={0,1}, the fairness violation value can be expressed as φ fair (h)∈[0,1]. For example, for an accuracy value of 0.8, the fairness violation value may be 0.65, and for an accuracy value of 0.99, the fairness violation value may be 0.95. The estimation of the fairness-accuracy curve aims to minimize the fairness violation value so as to optimize the fairness of the model at a given model accuracy. Therefore, for a trained machine learning model, for an accuracy above the accuracy threshold, the fairness-accuracy curve may be expressed as:

[0061]

[0062] in:

[0063] is the fairness-accuracy curve, and

[0064] ψ is an accuracy value from the plurality of accuracy values, and

[0065] ψ is a plurality of accuracy values ​​from a plurality of accuracy values, and ψ∈[0,1];

[0066] h is a model in a set of models H;

[0067] is the function used to determine fairness; and

[0068] is the function used to determine accuracy.

[0069] Estimating the fairness-accuracy curve includes mapping the accuracy value and fairness value of the fairness-accuracy curve into

[0070] As described above, it is desirable to determine a range of allowable fairness violation values ​​for a model relative to the model's accuracy for a given dataset. One solution is to determine a baseline truth fairness-accuracy curve for the model when training on the dataset. The baseline truth fairness-accuracy curve shows the minimum fairness violation relative to the accuracy of the trained machine learning model. The baseline truth fairness-accuracy curve can be used as a lookup table to determine fairness violations given a desired accuracy. However, the baseline truth fairness-accuracy curve is often not available, and therefore various optimization techniques have been proposed to estimate the baseline truth fairness-accuracy curve.

[0071] Obtaining the exact ground truth fairness-accuracy curve defined in Equation 1 is important for several reasons. First, the dataset is limited, which restricts us to calculate the accuracy acc(h) and fairness violation φ fair (h) can be accurately calculated. Second, Eq. The constrained optimization problem required to solve the value is non-trivial to solve exactly and may require training several separate machine learning models across different accuracy constraints. For example, see the empirical fairness-accuracy curve 420, where each point on the empirical fairness-accuracy curve 420 corresponds to a single trained machine learning model.

[0072] In an embodiment, the fairness-accuracy curve of the trained machine learning model may be estimated by training the machine learning model using a regularized loss defined as:

[0073] L λ (θ)=E[L CE (h θ (X), Y)]+λL fair (h θ ) Equation 2

[0074] in:

[0075] h λ (θ) is the regularization loss function;

[0076] θ is the weight of the model;

[0077] E[] is the expected value function;

[0078] L CE is the cross entropy loss function;

[0079] h θ It is a model;

[0080] X is the eigenvector;

[0081] Y is the label feature;

[0082] λ is the regularization parameter; and

[0083] L fair is a fairness violation fair (h) Fairness violation loss function.

[0084] The parameter λ∈R≥0 in Lλ modulates the accuracy-fairness trade-off, where lower values ​​of λ result in higher accuracy and fewer fairness violations, and vice versa. Therefore, given an accuracy threshold ψ, there exists a value of λ≥0 such that the loss L in Eq. 2 is λ (θ) provides a smooth surrogate loss for the constrained optimization problem in Equation 1. The surrogate loss can be optimized using a gradient-based optimization method. Obtaining the estimated fairness-accuracy curve involves optimizing on a grid of λ, where is the minimum value of the loss Lλ.

[0085] In an embodiment, the fairness-accuracy curve of a trained machine learning model may be estimated as:

[0086]

[0087] in:

[0088] is the fairness-accuracy curve, and

[0089] is multiple ψ is the accuracy value from multiple accuracy values, ψ∈[0,1];

[0090] i is the index of the dataset sample;

[0091] k is the total number of samples in the dataset;

[0092] is the function used to determine fairness;

[0093] yes is the function used to determine accuracy;

[0094] is a model; and

[0095]

[0096] and table and Representation classifier Empirical fairness violations and accuracy on the held-out data. However, training multiple models can be computationally expensive, especially when the model class H is a large-scale model (e.g., neural network). In addition, accuracy and fairness violations may not change continuously with changes in the value of λ, and a small increase in the value of λ may lead to a large change in the relationship between accuracy and fairness violation. Therefore, it is difficult to find a model with accuracy close to a given value of ψ, because the λ parameter has a large impact on the model accuracy. There is almost no control effect.

[0097] Thus, in an embodiment, the method includes estimating a fairness-accuracy curve by utilizing a one-shot training (YOTO) framework. The YOTO framework allows training a single model to determine a fairness-accuracy curve for the model, the curve comprising a plurality of fairness values ​​corresponding to a plurality of accuracy values.

[0098] Advantageously, the YOTO framework avoids the computational overhead associated with training multiple models when estimating the fairness-accuracy curve. The model to be trained when using the YOTO framework can be defined as hθ:X×Λ→R, which as an additional input and is trained to recover the classifier obtained by minimizing Lλ′ at inference time hθ(·,λ′). Λ is the space of regularization parameters λ.

[0099] As described above, obtaining the estimated fairness-accuracy curve involves optimizing over a grid of λ, where ∈ is the minimum value of the loss Lλ. In the YOTO framework, the YOTO framework solves the optimization problem where the parameter λ is sampled from the distribution Pλ instead of fixing λ. Therefore, during training, the model observes many different values ​​of λ and learns to optimize the loss Lλ for all values ​​of λ simultaneously.

[0100] In an embodiment, the fairness-accuracy curve may be estimated by training a machine learning model using an updated regularization loss defined as:

[0101]

[0102] in:

[0103] h θ It is a model;

[0104] X is the eigenvector;

[0105] Λ is the space of regularization parameters λ;

[0106] λ is the regularization parameter;

[0107] P λ is the distribution of the regularization parameter;

[0108] L CE is the cross entropy loss function;

[0109] Y is the label feature; and

[0110] L fair is a fairness violation fair (h) Fairness violation loss function.

[0111] In an embodiment, the fairness-accuracy curve may be estimated as:

[0112]

[0113] in:

[0114] is the fairness-accuracy curve, and

[0115] is a plurality of ψ is an accuracy value from a plurality of accuracy values, and ψ∈[0,1];

[0116] i is the index of the dataset sample;

[0117] k is the total number of samples in the dataset;

[0118] yes is the function used to determine fairness;

[0119] yes is the function used to determine accuracy;

[0120] yes is a model; and

[0121]

[0122] Advantageously, the regularized loss of the YOTO framework alleviates the computational burden of having to train multiple models when approximating the trade-off curve.

[0123] In an embodiment, the method includes dynamically estimating a fairness-accuracy curve during inference of a machine learning model. At inference time, the model can be conditioned on a selected parameter value λ′ and recover a model trained to optimize Lλ′.

[0124] Advantageously, the fairness-accuracy curve Helps characterize the fairness of a dataset.

[0125] In step 40, the method includes determining confidence intervals for a plurality of fairness values ​​of the fairness-accuracy curve. The confidence interval is an expected fairness value range for fairness values ​​corresponding to accuracy values.

[0126] In an embodiment, the probability that the fairness value is within the expected fairness value range is defined as:

[0127]

[0128] ,in:

[0129] yes is the fairness-accuracy curve, and

[0130] yes is a confidence interval, where

[0131] is the significance level, where α∈(0,1); and

[0132] is a random variable, where ψ is a real number in [0,1].

[0133] In an embodiment, the confidence interval is determined using the calibration data set Dcal The calibration data set Dcal does not intersect with the data set. A random variable ψ is obtained from the calibration data set Dcal, where Dcal:={(Xi, Ai, Yi)}i. In an embodiment, the confidence interval includes an upper bound and a lower bound.

[0134] In an embodiment, the method comprises applying the confidence interval to Previously, Dcal was used to determine the accuracy acc(h) and fairness violation φ for a given model h fair (h), respectively expressed as and Notice, and One possible way to construct these confidence intervals is to use an assumed light concentration inequality, such as the Hoefding inequality. However, it is noted that one skilled in the art will appreciate that other assumed light concentration inequalities or other suitable methods may be used to determine the confidence intervals.

[0135] Advantageously, the confidence interval provides an expected range of fairness values ​​for a particular fairness value given a dataset, model class, and expected accuracy.

[0136] Advantageously, the confidence interval provides an acceptable range of fairness violations for any given accuracy value.

[0137] Advantageously, confidence intervals provide the ability to adjust the fairness of machine learning models while maintaining their accuracy. Confidence intervals allow machine learning practitioners to make fairness decisions with limited accuracy.

[0138] Figure 2 A graph is shown that includes fairness-accuracy curves and corresponding confidence intervals for trained machine learning models.

[0139] The graph shows that the ground truth fairness-accuracy curve 430 is within the confidence interval of the estimated fairness-accuracy curve 410. The area bounded by the confidence interval represents models h whose accuracy is above a threshold accuracy and within the expected range of fairness values ​​for a given accuracy value.

[0140] In an embodiment, the confidence interval represents the range of permissible fairness violations for a given accuracy value. In an embodiment, the region above the confidence interval 440 corresponds to a suboptimal accuracy-fairness tradeoff region. If acc(h0) ≥ ψ for classifier h0, the fairness violation value Φfair(h0) is within the confidence interval Above , then h0 may be suboptimal in terms of fairness violation, i.e., Φfair(h0) can be reduced while keeping the accuracy fixed.

[0141] In an embodiment, the region below the confidence interval 450 corresponds to a region where the accuracy-fairness tradeoff is unlikely to be achieved. A model with acc(h) ≥ ψ is unlikely to achieve the confidence interval The following fairness violations.

[0142] Figure 3 A computing system 2000 is shown on which any of the above methods can be performed. In particular, the computing system 2000 can include a single computing device or component, such as a laptop computer, tablet computer, desktop computer, or other computing device. Alternatively, the functionality of the system 2000 can be distributed across multiple computing devices.

[0143] The computing system 2000 may include one or more controllers, such as a controller 2005, which may be, for example, a central processing unit processor (CPU), a graphics processing unit (GPU), a chip, or any suitable processor or computing or operating device, such as the mentioned FPGA, an operating system 2015, a memory 2020 storing executable code 2025, a storage device 2030 which may be external to the system or embedded in the memory 2020, one or more input devices 2035, and one or more output devices 2040.

[0144] One or more processors in one or more controllers (such as controller 2005) can be configured to perform any of the methods described herein. For example, one or more processors within controller 2005 can be connected to a memory 2020 storing software or instructions that, when executed by one or more processors, cause the one or more processors to perform methods according to some embodiments of the present invention. Controller 2005 or a central processing unit within controller 2005 can be configured to perform the methods described above, for example, using instructions stored in memory 2020.

[0145] Input device 2035 can be or can include a mouse, keyboard, touch screen or keypad or any suitable input device. It should be appreciated that, as shown in box 2035, any suitable number of input devices can be operably connected to computing system 2000. Output device 2040 can include one or more displays, speakers and / or any other suitable output devices. It should be appreciated that, as shown in box 2040, any suitable number of output devices can be operably connected to computing system 2000. Input and output devices can be used, for example, to enable a user to select information to be displayed, such as images and graphics shown here.

[0146] In the above embodiments, all or part of the method may be performed by a server. The server may include a single server or a network of servers. In some examples, the functionality of the server may be provided by a network of servers distributed across geographic regions, such as a globally distributed server network, and a user / operator of the method may be connected to an appropriate network server in the network server based on, for example, the user's location.

[0147] The above embodiments are fully automatic. In some examples, a user or operator of the system may manually instruct some steps of the method to be performed.

[0148] In the described embodiments of the invention, portions of the system may be implemented in the form of computing and / or electronic devices. Such devices may include one or more processors, which may be microprocessors, controllers, or any other suitable type of processor for processing computer executable instructions to control the operation of the device so as to collect and record routing information. In some examples, for example, where a system-on-chip architecture is used, the processor may include one or more fixed function blocks (also referred to as accelerators) that implement a portion of the method in hardware (rather than software or firmware). Platform software including an operating system or any other suitable platform software may be provided at a computing-based device to enable application software to be executed on the device.

[0149] The various functions described herein can be implemented in hardware, software, or any combination thereof. If implemented in software, the function can be stored on a computer-readable medium or transmitted via a computer-readable medium as one or more instructions or codes. Computer-readable media may include, for example, computer-readable storage media. Computer-readable storage media may include volatile or non-volatile, removable or non-removable media implemented in any method or technology for storing information such as computer-readable instructions, data structures, program modules, or other data. Computer-readable storage media may be any available storage media that a computer can access. As an example and not limitation, such computer-readable storage media may include RAM, ROM, EEPROM, flash memory or other memory devices, CD-ROM or other optical disk storage devices, magnetic disk storage devices or other magnetic storage devices, or any other medium that can be used to carry or store desired program codes in the form of instructions or data structures and can be accessed by a computer. Optical disks and disks as used herein include compact discs (CDs), laser optical disks, optical disks, digital versatile discs (DVDs), floppy disks, and blue-ray discs (BDs). In addition, propagated signals are not included in the scope of computer-readable storage media. Computer-readable media also include communication media, which include any media that helps to transmit a computer program from one place to another. For example, a connection can be a communications medium. For example, if the software is transmitted from a website, server or other remote source using coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, radio and microwave, these technologies are included in the definition of communications medium. Combinations of the above should also be included within the scope of computer-readable media.

[0150] Alternatively or additionally, the functions described herein may be at least partially performed by one or more hardware logic components. For example, but not limited to, the hardware logic components that may be used may include field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), program specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), etc.

[0151] Although shown as a single system, it should be understood that the computing device can be a distributed system. Thus, for example, several devices can communicate via a network connection and can collectively perform tasks described as being performed by the computing device. Although shown as a local device, it should be understood that the computing device can be remotely located and accessed via a network or other communication link (e.g., using a communication interface).

[0152] The term "computer" is used herein to refer to any device having processing capabilities such that it can execute instructions. Those skilled in the art will recognize that such processing capabilities are incorporated into many different devices, and thus the term "computer" includes PCs, servers, mobile phones, personal digital assistants, and many other devices.

[0153] Those skilled in the art will recognize that the storage device for storing program instructions can be distributed across a network. For example, a remote computer can store a process example described as software. A local or terminal computer can access the remote computer and download a part or all of the software to run the program. Alternatively, the local computer can download software fragments as needed, or execute some software instructions at the local terminal, and execute some software instructions at the remote computer (or computer network). Those skilled in the art will also recognize that all or part of the software instructions can be executed by a dedicated circuit (such as a DSP, a programmable logic array, etc.) by utilizing conventional techniques known to those skilled in the art.

[0154] It should be understood that the benefits and advantages described above may relate to one embodiment, or may relate to several embodiments. The embodiments are not limited to those that solve any or all of the stated problems, nor are they limited to those that have any or all of the stated benefits and advantages. Variations should be considered to be included within the scope of the present invention.

[0155] Any reference to "an" item refers to one or more of those items. The term "comprising" is used herein to mean including the identified method steps or elements, but such steps or elements do not comprise an exclusive list and the method or apparatus may include additional steps or elements.

[0156] As used herein, the terms "component" and "system" are intended to encompass a computer-readable data storage device configured with computer-executable instructions that, when executed by a processor, cause certain functions to be performed. Computer-executable instructions may include routines, functions, etc. It should also be understood that a component or system may be located on a single device or distributed across several devices.

[0157] Additionally, as used herein, the term "exemplary" is intended to mean "serving as an illustration or example of something."

[0158] Furthermore, where the detailed description or claims use the term "includes," such term is intended to be inclusive in a manner similar to the term "comprising," in that "comprising" is interpreted as "comprising" when used as a transitional term in a claim.

[0159] The accompanying drawings illustrate exemplary methods. Although these methods are shown and described as a series of actions performed in a particular order, it should be understood and appreciated that these methods are not limited by order. For example, some actions may occur in a different order than described herein. In addition, an action may occur simultaneously with another action. In addition, in some cases, all actions may not be required to implement the methods described herein.

[0160] In addition, the actions described herein may include computer executable instructions, which may be implemented by one or more processors and / or stored on a computer readable medium. Computer executable instructions may include routines, subroutines, programs, execution threads, etc. In addition, the results of the actions of the method may be stored in a computer readable medium, displayed on a display device, etc.

[0161] The order of steps of the methods described herein is exemplary, but the steps may be performed in any suitable order, or simultaneously where appropriate. In addition, steps may be added or replaced in any of the methods, or individual steps may be deleted from any of the methods, without departing from the scope of the subject matter described herein. Aspects of any of the above examples may be combined with aspects of any of the other examples to form further examples without losing the effect sought.

[0162] It should be understood that the above description of the preferred embodiment is given by way of example only, and various modifications may be made by those skilled in the art. The foregoing includes examples of one or more embodiments. Of course, in order to describe the above aspects, it is impossible to describe every conceivable modification and change of the above-mentioned device or method, but those skilled in the art will recognize that many further modifications and arrangements of various aspects are possible. Therefore, the described aspects are intended to cover all such changes, modifications and variations that fall within the scope of the appended claims.

Claims

1. A computer-implemented method, the method comprising: Obtain a machine learning model; Obtaining a data set for training the machine learning model to generate a trained machine learning model; estimating a fairness-accuracy curve of the trained machine learning model by training the machine learning model on the dataset, wherein the fairness-accuracy curve includes a plurality of fairness values ​​corresponding to a plurality of accuracy values; as well as Confidence intervals for the plurality of fairness values ​​for the fairness-accuracy curve are determined, wherein a confidence interval is a range of expected fairness values ​​for fairness values ​​corresponding to accuracy values. 2 . The method of claim 1 , wherein determining the confidence interval comprises using a calibration data set, wherein the calibration data set is disjoint from the data set.

3. A method according to any preceding claim, wherein the probability that the fairness value is within the range of expected fairness values ​​is defined as: in: is the fairness-accuracy curve, and is the confidence interval, where α is the significance level, where α∈(0,1); and ψ is a random variable.

4. The method of any preceding claim, wherein a reference truth fairness-accuracy curve is within the confidence interval of the fairness-accuracy curve.

5. The method of any preceding claim, wherein estimating the fairness-accuracy curve comprises: Leverage a one-shot training framework.

6. A method according to any preceding claim, wherein the method comprises: The fairness-accuracy curve is dynamically estimated during inference of the machine learning model.

7. A method according to any preceding claim, wherein the plurality of fairness values ​​each represent a compromised fairness value depending on its corresponding plurality of accuracy values.

8. A method according to any preceding claim, wherein the fairness-accuracy curve to be estimated is: in: is the fairness-accuracy curve, and Ψ is an accuracy value from the plurality of accuracy values, and ψ∈[0,1]; h is a model in a set of models H; Φ fair () is the function used to determine fairness; and Acc() is a function used to determine the accuracy.

9. The method of any preceding claim, wherein training the model on the dataset comprises training the model on a regularization loss defined as: THE λ (θ)=E[L CE (h θ (X),Y)]+λL fair (h θ ), in: L λ (θ) is the regularization loss function; θ is the weight of the model; is that E[] is the expected value function; L CE is the cross entropy loss function; h θ is the model; X is the eigenvector; Y is the label feature; λ is the regularization parameter; L fair is a fairness violationΦ fair (h) Fairness violation loss function.

10. The method of any preceding claim, wherein training the model on the dataset comprises training the model on a regularization loss, the regularization loss being defined as: in: h θ is the model; X is the eigenvector; Λ is the space of regularization parameters λ; λ is the regularization parameter; P λ is the distribution of the regularization parameter; L CE is the cross entropy loss function; Y is the tag feature; and L fair is a fairness violation fair (h) Fairness violation loss function.

11. A method according to any preceding claim, wherein the fairness-accuracy curve is estimated as in: yes is the fairness-accuracy curve, and is an accuracy value from the plurality of accuracy values, and ψ∈[0,1]; i is the index of the sample in the dataset; k is the total number of samples in the dataset; yes is the function used to determine fairness; yes is the function used to determine accuracy; as well as yes is the model.

12. The method of claim 1, wherein the fairness is at least one of: group fairness, equal opportunity, and equal probability.

13. A data processing system comprising means for executing the method according to any one of claims 1 to 12.

14. A computer program comprising instructions which, when executed by a computer, cause the computer to perform the method according to any one of claims 1 to 12.

15. A computer-readable medium comprising instructions which, when executed by a computer, cause the computer to perform the method according to any one of claims 1 to 12.

Images