Methods, apparatus, electronic devices, and storage media for determining fairness

Abstract

Embodiments of this disclosure provide a method, apparatus, electronic device, and storage medium for determining fairness. The method for determining fairness described herein includes determining a predicted classification for a target dataset based on the target dataset, and determining a fairness metric for a target model based on the target dataset and the predicted classification. The method further includes correcting the fairness metric of the target model based on the probability distribution of the predicted classification relative to the true classification. The method for determining fairness according to embodiments of this disclosure, for target datasets without sensitive variables, can effectively improve the accuracy of fairness metric calculation for the target model without the need to train an auxiliary model using auxiliary data.

Description

Technical Field

[0001] Embodiments of this disclosure relate to the field of computers, and more specifically, to methods, apparatus, electronic devices, and storage media for determining fairness. Background Technology

[0002] In recent years, with the abundance of data and the improvement of computing power, artificial intelligence (AI) technology has developed rapidly, and machine learning technology has gradually been integrated into various fields of society. After being trained on training data, AI models based on machine learning can predict specific data results based on the data to be processed, thereby assisting people in making decisions based on the predicted data results.

[0003] For example, in personalized recommendations in the media field, similar case searches in the legal field, and credit rating in the financial field, artificial intelligence models can determine which category to classify the data to be processed, based on the predicted event. Thus, various fields can train corresponding models tailored to their specific characteristics, facilitating the assessment of the overall data and subsequently predicting the classification of specific data. Summary of the Invention

[0004] Embodiments of this disclosure provide a scheme for determining fairness.

[0005] In a first aspect of this disclosure, a method for determining fairness is provided. The method includes: determining a predicted classification for a target dataset; determining a fairness metric for a target model based on the target dataset and the predicted classification; and correcting the fairness metric of the target model based on the probability distribution of the predicted classification relative to the true classification. According to the first aspect of this disclosure, for target datasets without sensitive variables, the accuracy of the fairness metric calculation for the target model can be effectively improved without using auxiliary data to train an auxiliary model for classification.

[0006] In some embodiments, where the target dataset lacks true sensitive variables, the predicted classification includes predicted sensitive variables, and determining the predicted classification includes: determining the predicted sensitive variables of the target dataset using an auxiliary model based on at least one feature of the target dataset, wherein the auxiliary model is not trained on training data labeled with sensitive variables. In this way, it is possible to label the target dataset with predicted sensitive variables that are comparable to noise by using the target dataset to perform preliminary predictions of sensitive variables on the auxiliary model, thus eliminating the need for true or manually labeled sensitive variables, and significantly saving resources and costs.

[0007] In some embodiments, correcting the fairness metric of the target model includes: for the auxiliary model, determining a transition matrix representing a probability distribution through noisy label learning, wherein the transition matrix includes a probability distribution consisting of the probabilities of each predicted classification relative to each true classification; and correcting the fairness metric of the target model based on the transition matrix. In this way, fairness determined based on a preliminary noisy assessment can be corrected based on the transition matrix, thereby further improving the accuracy of the fairness metric determination for the target model.

[0008] In some embodiments, correcting the fairness metric of the target model further includes: determining a first prior probability based on the transition matrix; determining a second prior probability based on the predicted classification; and correcting fairness based on the first prior probability, the second prior probability, and the transition matrix. In some embodiments, correcting fairness includes updating the fairness by multiplying the inverse of the diagonal matrix of the first prior probability, the diagonal matrix of the second prior probability, the inverse of the transpose of the transition matrix, and the fairness. In some embodiments, the first prior probability represents a statistical prior probability, and the second prior probability represents a noisy prior probability. In this way, the probabilities of the auxiliary model and the prediction of sensitive variables equivalent to noise can be considered, thereby ensuring the accuracy of the determination of the fairness metric of the target model.

[0009] In some embodiments, determining the transition matrix representing the probability distribution through noise label learning includes: using a high-order consensus algorithm based on distance differencing of each noise label and an unconstrained optimization algorithm to determine the transition matrix. In this way, the difference between the predicted and true classifications can be reduced by utilizing the differencing of the predicted sensitive variables, which are equivalent to the noise, thereby enabling accurate determination of the transition matrix used to correct for fairness.

[0010] In some embodiments, correcting the fairness metric of the target model includes determining the row and column structure of the transition matrix representing the probability distribution based on the range of values ​​for the predicted and true classifications. In this way, the specific method of fairness correction can be determined corresponding to the type of classifier, thereby effectively improving the accuracy of determining the fairness metric of the target model.

[0011] In some embodiments, determining the predicted classification for a target dataset includes: determining the predicted classification by an auxiliary model, wherein the auxiliary model includes a binary classifier or a multi-class classifier. In some embodiments, determining the fairness metric of the target model includes: determining the fairness metric of the target model by the target model, wherein the target model includes a binary classifier or a multi-class classifier. In some embodiments, the binary classifier includes a fully connected layer, an activation function layer, and a binary logistic regression layer, and the multi-class classifier includes a fully connected layer, an activation function layer, and a softmax function layer. In this way, the improved method for determining the fairness of artificial intelligence models can be applied to classifiers and classification algorithms in various fields, thereby increasing its applicability.

[0012] In a second aspect of this disclosure, a fairness determination apparatus is provided. The apparatus includes: a prediction classification determination module configured to determine a predicted classification for a target dataset based on the target dataset; a fairness index determination module configured to determine a fairness index of a target model based on the target dataset and the predicted classification; and a fairness index correction module configured to correct the fairness index of the target model based on the probability distribution of the predicted classification relative to the true classification.

[0013] In a third aspect of this disclosure, an electronic device is provided. The electronic device includes: at least one processor; and at least one memory storing computer-executable instructions, the at least one memory and the computer-executable instructions being configured, together with the at least one processor, to cause the electronic device to perform the method according to a first aspect of this disclosure.

[0014] In a fourth aspect of this disclosure, a computer-readable storage medium is provided. The computer-readable storage medium includes computer-executable instructions stored thereon, which, when executed, cause one or more devices to perform the method described according to a first aspect of this disclosure.

[0015] The summary section is provided to present the chosen concepts in a simplified form, which will be further described in the detailed description below. The summary section is not intended to identify key or principal features of this disclosure, nor is it intended to limit the scope of this disclosure. Attached Figure Description

[0016] The above and other objects, features and advantages of this disclosure will become more apparent from the accompanying drawings, in which like reference numerals generally denote like parts.

[0017] Figure 1 This is a schematic diagram illustrating an example system architecture according to some embodiments of the present disclosure;

[0018] Figure 2A A flowchart of a method for determining fairness according to some embodiments of the present disclosure is shown;

[0019] Figure 2B A schematic diagram of the correction architecture of the target model and the auxiliary model according to some embodiments of the present disclosure is shown;

[0020] Figure 3 A flowchart is shown of a method for correcting fairness metrics of a model according to some embodiments of the present disclosure;

[0021] Figure 4 A schematic diagram of a correction architecture for a target model and an auxiliary model according to another embodiment of the present disclosure is shown;

[0022] Figure 5 A flowchart of a method for correcting a fairness metric of a model according to another embodiment of the present disclosure is shown;

[0023] Figure 6 A schematic block diagram of a fairness determination apparatus according to some embodiments of the present disclosure is shown; and

[0024] Figure 7 A schematic block diagram of an example device that can be used to implement embodiments of the present disclosure is shown. Detailed Implementation

[0025] The principles of this disclosure will now be described with reference to several exemplary embodiments illustrated in the accompanying drawings. While preferred embodiments of this disclosure are shown in the drawings, it should be understood that these embodiments are described only to enable those skilled in the art to better understand and implement this disclosure, and are not intended to limit the scope of this disclosure in any way.

[0026] The term "comprising" and its variations as used herein signify open inclusion, i.e., "including but not limited to". Unless otherwise stated, the term "or" means "and / or". The term "based on" means "at least partially based on". The terms "one example embodiment" and "one embodiment" mean "at least one example embodiment". The term "another embodiment" means "at least one additional embodiment". The terms "first", "second", "third", etc., may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0027] It is understood that before using the technical solutions disclosed in the various embodiments of this disclosure, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this disclosure in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0028] It is understood that the data involved in this technical solution (including but not limited to the data itself, the acquisition or use of the data) shall comply with the requirements of relevant laws, regulations and related provisions.

[0029] As mentioned above, classification results predicted by artificial intelligence models can be used to assist people in making decisions. However, the inventors of this application have noted that, due to the inherent characteristics and technical features of machine learning, the predicted classification results may contain a certain degree of unfairness. For example, the training dataset used to train the AI ​​model for recommendation often does not label sensitive variables such as education level, which involve personal privacy restrictions and cannot be publicly accessed. Training the AI ​​model for recommendation in a training dataset lacking sensitive variables will result in the AI ​​model's prediction results not taking into account some personal privacy restrictions, potentially leading to biased unfairness. For example, because the target model has an education level bias, low-end ads may be recommended to people with lower education levels, while high-end ads may be recommended to people with higher education levels, resulting in unfairness in the model's output.

[0030] Therefore, if the training dataset is not labeled with sensitive variables or is incorrectly labeled with them, the predictions of the trained AI model may be biased towards a particular individual or group, leading to unfair treatment and harm to their interests. Therefore, it is necessary to identify and evaluate the fairness of AI models, and update and correct such unfairness to make AI models more trustworthy and achieve fairness and impartiality.

[0031] Regarding how to determine the fairness of an artificial intelligence (AI) model, the inventors of this disclosure have discovered through research that one method involves manually labeling a portion of the training dataset with sensitive variables, and then using this labeled portion to calculate the fairness index of the AI ​​model. However, this method requires significant human resources for labeling sensitive variables, incurring high costs both economically and in terms of time. Furthermore, due to sampling errors and human labeling mistakes, the accuracy of the calculated fairness index is unsatisfactory. On the other hand, the inventors of this disclosure have further researched and considered using publicly available data containing sensitive variables (e.g., publicly available government statistics or open-source data) as auxiliary data. An auxiliary model trained on this auxiliary data is used to label the sensitive variables of the target dataset, thereby calculating the fairness index of the AI ​​model. However, this method requires that the auxiliary data and the target data be independently and identically distributed; otherwise, the accuracy of the calculated fairness index remains unsatisfactory, and it is often difficult to obtain complete or accurate data containing sensitive variables from publicly available sources, making it unsuitable for calculating fairness indices for various types of AI models.

[0032] To at least address the aforementioned problems, exemplary embodiments of this disclosure propose a scheme for determining fairness. This scheme determines a predicted classification for a target dataset and determines a fairness metric for a target model based on the target dataset and the predicted classification. The method further corrects the fairness metric of the target model based on the probability distribution of the predicted classification relative to the true classification. According to embodiments of this disclosure, the method for determining fairness, for target datasets without sensitive variables, can effectively improve the accuracy of the calculated fairness metric of the target model without using auxiliary data to train an auxiliary model. That is, based on the predicted classification of the target dataset lacking sensitive variables by the auxiliary model, a noisy fairness metric of the target model is initially determined, and this fairness metric is further corrected based on the probability distribution of the predicted classification relative to the true classification, thereby processing data with predicted sensitive variables in a manner such as noisy label learning. Therefore, embodiments of this disclosure can effectively improve the accuracy of the calculated fairness metric of the target model without a target dataset with sensitive variables or without using auxiliary data to train the auxiliary model for fairness, thus enabling a more objective evaluation of the fairness of the target model.

[0033] Figure 1 This is a schematic diagram illustrating a system architecture 100 according to some embodiments of the present disclosure. For example... Figure 1 As shown, system architecture 100 may include an auxiliary model 110 for assisting in the prediction of sensitive variables, a target model 120 for fairness assessment, and a target dataset 130. It should be understood that... Figure 1 The system architecture 100 may include more or fewer functional devices, and is not limited thereto.

[0034] In some embodiments, the auxiliary model 110 and the target model 120 can be deployed integratedly or separately on a node, which can be a processing resource capable of implementing the target model and / or the auxiliary model. For example, it can include any form of computing resource for data processing, such as a graphics processing unit (GPU) or a central processing unit (CPU). In some embodiments, the target model and / or the auxiliary model can be a neural network-based artificial intelligence model, such as a model based on any neural network, including convolutional neural networks (CNN), fully convolutional neural networks (FCN), generative adversarial networks (GAN), recurrent neural networks (RNN), or long short-term memory networks (LSTM).

[0035] In some embodiments, the target dataset 130 may provide target data to the auxiliary model 110 and the target model 120. In embodiments of this disclosure, the target dataset 130 does not include sensitive variables and serves as model input to the target model 120 and / or the auxiliary model 110 without sensitive variables. In this document, the term "sensitive variable" refers to a variable that affects the fairness of the artificial intelligence model, such as data items that are difficult to publicly obtain, difficult to fully obtain, or difficult to guarantee accuracy due to factors such as privacy, restrictions, ethics, policies, laws, and structural deficiencies. In embodiments of this disclosure, the target dataset may include various structured and unstructured data (e.g., image data, text data, voice data), etc. Embodiments of this disclosure do not limit this, as long as the data is suitable for training, prediction, and other processing by the artificial intelligence model.

[0036] like Figure 1 As shown, in some embodiments of this disclosure, based on the target dataset 130, the predicted classification 115 of each target data in the target dataset 130 is determined by the auxiliary model 110. Then, based on the target dataset 130 and the predicted classification 115, a noise fairness index 125 of the target model 120 is determined. Since the auxiliary model 110 is an untrained model, the noise fairness index 125 is inaccurate. According to embodiments of this disclosure, based on the probability distribution 135 of the predicted classification relative to the true classification, the fairness index of the target model 120 is corrected, thereby obtaining a corrected fairness index 145, which can characterize the fairness status of the target model 120. The following references... Figures 2A to 7 Various embodiments of this disclosure are further described below.

[0037] Figure 2A A flowchart of a method 200 for determining fairness according to some embodiments of the present disclosure is shown, wherein Figure 2A The process can be adapted for later discussion. Figure 2B and Figure 4 The correction architecture for the auxiliary model and the target model is shown.

[0038] Figure 2B A schematic diagram of a correction architecture 250 for a target model 120 and an auxiliary model 110 according to some embodiments of the present disclosure is shown. In embodiments of the present disclosure, the target model 120 may include a binary classifier or a multi-class classifier, and the auxiliary model 110 may also include a binary classifier or a multi-class classifier. In one example of the present disclosure, for instance, in a scenario where consumption intention is predicted from a dataset including data items such as nickname and browsing time, where the sensitive variable can be gender (male / female), and the predicted classification is male (e.g., a value of 1) or female (e.g., a value of 0), the target model 120 is a binary classifier that classifies consumption intention as high (e.g., a value of 1) or low (e.g., a value of 0) based on the target dataset, and the auxiliary model 110 is a binary classifier that classifies the sensitive variable as male (e.g., a value of 1) or female (e.g., a value of 0) based on the target dataset. It should be understood that, in the embodiments of this disclosure, the binary classifier supports algorithms including logistic regression-based binary classification algorithms, support vector machine-based binary classification algorithms, decision tree-based binary classification algorithms, etc. The embodiments of this disclosure are not limited to these; any binary classifier capable of implementing the embodiments of this disclosure is acceptable. In this way, the fairness index calculation method of artificial intelligence models can be applied to classifiers and classification algorithms in various fields, thereby increasing the scope of applicability.

[0039] refer to Figure 2A At point 201, based on the target dataset, a predicted classification is determined for the target dataset. For example, as... Figure 2B As shown, the auxiliary model 110 sequentially includes a fully connected layer 211, an activation function layer 212, a fully connected layer 213, and a binary logistic regression layer 214. In some embodiments, for example, the target dataset is fed into the auxiliary model 110 (e.g., g(.)). The system provides a target dataset X (e.g., including nicknames and browsing duration). The auxiliary model 110 uses at least a portion of the target dataset X as input t, and sequentially extracts and maps features from the target dataset X through a fully connected layer 211, an activation function layer 212, and a fully connected layer 213. Then, it predicts the classification of the target dataset through a binary logistic regression layer 214. Male (for example, The value can be 1) or female (for example, (The value can be 0). It should be understood that the input t can be all features in the target dataset X, or it can be a subset of features, as long as it can be used to assist the model in predicting classification.

[0040] At point 203, based on the target dataset and the predicted classification, a fairness metric for the target model, i.e., a noise fairness metric, is determined. For example, ... Figure 2B As shown, the target model 120 (e.g., f(.)) sequentially includes a fully connected layer 221, an activation function layer 222, a fully connected layer 223, and a binary logistic regression layer 224. In some embodiments, for example, the target model 120 obtains the predicted classification from the auxiliary model 110. Obtain target dataset X from target dataset 130. Target model 120 is based on target dataset X and predicts classification. The noise fairness index of the target model 120 is determined using the following formula 1.

[0041]

[0042] In some embodiments, the target model 120 combines the target dataset X with the predicted classification. As input, the target dataset X and the predicted classification are extracted sequentially through fully connected layer 221, activation function layer 222, and fully connected layer 223. The features in the model are mapped and then the noise fairness index of the target model 120 is predicted via a binary logistic regression layer 224. Among them, the noise fairness index This indicates the degree of fairness of the target model 120, but it is inaccurate because the auxiliary model 110 is untrained. In some embodiments, the fairness index... Examples include equal odds, equal opportunity, and demographic disparity.

[0043] At point 205, the fairness metric of the target model is corrected based on the probability distribution of the predicted classification relative to the true classification. For example, refer to... Figure 2B Based on the predicted classification determined at point 201 The probability distribution relative to the true classification A is used to determine the noise fairness index at position 203. Make corrections to calculate the fairness index after correction. In some embodiments, it can be based on corrected fairness metrics. To assess the fairness of the target model.

[0044] Therefore, the fairness determination method 200 according to the embodiments of this disclosure can effectively improve the accuracy of the fairness index calculation of the target model for a target dataset without sensitive variables, without the need to train the auxiliary model with auxiliary data.

[0045] In addition, such as Figure 2B As shown, during the process of correcting the fairness index, a corresponding transition matrix can be generated based on the auxiliary model 110, thereby determining the first prior probability (i.e., the statistical prior probability). Based on the predicted classification, the second prior probability (i.e., the noisy prior probability) can be determined. Next, the corrected fairness index can be calculated based on the transition matrix, the first prior probability, the second prior probability, and the noisy fairness index.

[0046] Figure 3 A flowchart of a method 300 for correcting the fairness of a model according to some embodiments of the present disclosure is shown. In embodiments of the present disclosure, the target dataset X lacks true sensitive variables, and the predicted classification... This includes sensitive variables for prediction. For example... Figure 3 As shown in 301, based on at least one feature of the target dataset X, a prediction sensitivity variable for the target dataset X is determined by an auxiliary model 110, wherein the auxiliary model is not trained on training data labeled with the sensitivity variable. In some embodiments, for example, the target dataset X is not labeled with the true sensitivity variable A (e.g., the true gender, male or female), the predicted classification includes the predicted sensitivity variable. (e.g., gender male or female). In a further embodiment, based on at least one feature t (e.g., nickname) in the target dataset X (e.g., including nickname, browsing time), through the auxiliary model 110 (e.g., ... Identify the sensitive variables for predicting the target dataset X. In this embodiment, the auxiliary model 110 is not trained on training data labeled with sensitive variables. That is, in this disclosure, the auxiliary model 110 does not need to use manually labeled data or publicly available data containing sensitive variables (e.g., publicly available government statistics or open-source data) as auxiliary data to label the sensitive variables of the target dataset. Instead, it can directly use the target dataset for classification prediction. In this way, the auxiliary model can make preliminary predictions of sensitive variables using the target dataset to label the target dataset with predicted sensitive variables that are comparable to noise, thus eliminating the need for real or manually labeled sensitive variables and significantly saving resources and costs.

[0047] In embodiments of this disclosure, such as Figure 3As shown in 303, a fairness metric for the target model is determined based on the target dataset and the predicted classification. In some embodiments, for example, the target model 120 obtains the predicted classification from the auxiliary model 110. Based on the target dataset X and the predicted classification Use Formula 1 to determine the fairness index of target model 120.

[0048] like Figure 3 As shown in 304, based on the value ranges of the predicted classification and the true classification, the row and column structure (2×2) of the transition matrix used to represent the probability distribution is determined. In embodiments of this disclosure, as... Figure 3 As shown in 305, for the auxiliary model 110, a transition matrix representing the probability distribution is determined through noise label learning, wherein the transition matrix includes a probability distribution consisting of the probabilities of each predicted classification relative to each true classification. In some embodiments, for example, for the auxiliary model 110, a transition matrix representing the predicted classification is determined through noise label learning. The following transition matrix represents the probability distribution relative to the true classification A:

[0049]

[0050] This transition matrix includes, for example, a probability distribution consisting of the probabilities of predicted male relative to true male, predicted male relative to true female, predicted female relative to true male, and predicted female relative to true female. For example, in the row-column structure (2×2) determined by 304, due to the predicted classification... The value range of is (0, 1) and the value range of the true classification A is (0, 1), determining a 2×2 probability distribution. In this way, the specific method of fairness correction can be determined corresponding to the type of classifier, effectively improving the accuracy of the fairness index determination for the target model. In some embodiments, noisy label learning can be based on confidence learning, forgetting events, sampling loss values, and sample similarity. In this way, the fairness initially determined based on noise can be corrected based on the transition matrix, effectively improving the accuracy of the fairness index determination for the target model. For example, noisy label learning can also be performed using a high-order consensus algorithm based on the distance difference of each noise label and an unconstrained optimization algorithm, where the unconstrained optimization algorithm can include gradient descent, Newton's method, conjugate gradient method, etc. In this way, the difference between the predicted sensitive variables, which are equivalent to noise, can be used to reduce the difference between the predicted classification and the true classification, thereby accurately determining the transition matrix used to correct fairness.

[0051] At position 306, a first prior probability is determined based on the transition matrix, and a second prior probability is determined based on the predicted classification. In some embodiments, for example, based on the transition matrix determined at position 305, the first prior probability is determined using Equation 3 as follows, and the second prior probability is determined based on the predicted classification. The second prior probability is determined using the following formula 4.

[0052]

[0053]

[0054] In some embodiments, the first prior probability represents a statistical prior probability, and the second prior probability represents a noisy prior probability learned based on noisy labels. It should be understood that the method for determining the prior probabilities can be determined in any way based on a pre-estimated probability distribution, and in some cases, posterior probabilities can also be used.

[0055] At point 307, the fairness index of the target model is corrected based on the first prior probability, the second prior probability, and the transition matrix. In some embodiments, the fairness index of the target model is updated by multiplying the inverse of the diagonal matrix of the first prior probability, the diagonal matrix of the second prior probability, the inverse of the transpose of the transition matrix, and the fairness index of the target model, thus determining the fairness of the updated target model 120. In some embodiments, for example, the inverse matrix of the diagonal matrix of the first prior probability is obtained using the following formula 5. The diagonal matrix of the second prior probability The inverse of the transpose of the transition matrix and the noise fairness index of target model 120 Perform matrix multiplication to obtain the corrected fairness index of the updated target model 120.

[0056]

[0057] In some embodiments, the corrected fairness index In a further embodiment, it can be used to evaluate the fairness of the target model 120. In a further embodiment, it can be based on a corrected fairness metric. The fairness of the target model 120 is determined by the range of values ​​for the sensitive variables (e.g., 1 or 0). For example, if the fairness of the target model 120 is characterized by sociodemographic differences, the fairness of the target model 120 can be expressed as: In this way, the probability distributions of the auxiliary model and the sensitive variables of the prediction, which are equivalent to noise, can be taken into account, thereby effectively improving the accuracy of determining the fairness index of the target model.

[0058] According to embodiments of this disclosure, a fairness index for the target model is initially determined based on the predicted classification of a target dataset lacking sensitive variables by an auxiliary model. The fairness index is then corrected based on the probability distribution of the predicted classification relative to the true classification. This approach processes data with predicted sensitive variables using a noise label learning method, effectively improving the accuracy of the fairness index determination without requiring target data with sensitive variables or training the auxiliary model with auxiliary data. For example, by processing the target dataset with the fairness-corrected target model, it can predict the relative levels of gender-equal consumption intentions, mitigating gender bias in the original target dataset and achieving fairer information recommendation.

[0059] In the above Figure 2B and Figure 3 In the embodiments described, the case where the target model and the auxiliary model are binary classifiers is used as an example. The method for determining fairness in the embodiments of this disclosure is also applicable to the case where the target model and the auxiliary model are multi-class classifiers. Figure 4A schematic diagram of the architecture 400 of a target model and an auxiliary model according to another embodiment of this disclosure is shown. In another embodiment of this disclosure, for example, in the case of predicting consumption intention from a dataset including data items such as nickname and browsing time, the sensitive variable is age group (elderly / middle-aged / youth / child), and the predicted classification is elderly (e.g., value 3), middle-aged (e.g., value 2), youth (e.g., value 1), and child (e.g., value 0). The target model 120 is a multi-class classifier that classifies consumption intention into high (e.g., value 3), medium (e.g., value 2), low (e.g., value 1), and none (e.g., value 0) based on the target dataset. The auxiliary model 110 is a multi-class classifier that classifies the sensitive variable into elderly (e.g., value 3), middle-aged (e.g., value 2), youth (e.g., value 1), and child (e.g., value 0) based on the target dataset. In the above example, the value K of the sensitive variable and the predicted classification range is 4. It should be understood that, in the embodiments of this disclosure, the algorithms supported by the multi-class classifier may include multi-class algorithms based on the softmax function, multi-class algorithms based on K-Nearest Neighbor (KNN), multi-class algorithms based on Random Forest, multi-class algorithms based on decision trees, etc., but are not limited to these; any multi-class classifier capable of implementing the embodiments of this disclosure is acceptable. In this way, methods for improving the fairness determination of artificial intelligence models can be applied to classifiers and classification algorithms in various fields, thereby increasing the scope of applicability.

[0060] refer to Figure 4 The auxiliary model 110 sequentially includes a fully connected layer 411, an activation function layer 412, a fully connected layer 413, and a softmax function layer 414. In some embodiments, such as the auxiliary model (e.g., g(.)) deployed on the target model 120 based on the target dataset, The system provides a target dataset X (e.g., including nicknames and browsing duration). The auxiliary model 110 uses at least a portion of the target dataset X as input t, and sequentially extracts and maps features from the target dataset X through a fully connected layer 411, an activation function layer 412, and a fully connected layer 413. Then, it predicts the classification of the target dataset through a softmax function layer 414. Older (e.g., Values ​​are 3), middle-aged (for example, Values ​​are 2), youth (for example, Values ​​are 1), children (e.g., (The value can be 0). It should be understood that the input t can be all features in the target dataset X, or it can be a subset of features, as long as it can be used to assist the model in predicting classification.

[0061] Continue to refer to Figure 4 The target model 120 (e.g., f(.)) sequentially includes a fully connected layer 421, an activation function layer 422, a fully connected layer 423, and a softmax function layer 424. In some embodiments, for example, the target model 120 obtains the predicted classification from the auxiliary model 110. Obtain target dataset X from target dataset 130. Target model 120 is based on target dataset X and predicts classification. Determine the fairness index of target model 120 For example, target model 120 combines the target dataset X with the predicted classification. As input, the target dataset X and the predicted classification are extracted sequentially through fully connected layer 421, activation function layer 422, and fully connected layer 423. The features in the model are mapped and then the noise fairness index of the target model 120 is predicted via a softmax function layer 424. Among them, noise fairness indicators This indicates the degree of fairness in the predictions made by the target model 120.

[0062] Continue to refer to Figure 4 Based on the determined predicted classification Relative to the probability distribution of the true classification A, the determined noise fairness index... Make corrections and determine the fairness indicators after corrections. In some embodiments, it can be based on corrected fairness metrics. To evaluate the fairness of the target model, the transition matrix can be generated based on the auxiliary model 110 to determine the first prior probability (i.e., the statistical prior probability). Based on the predicted classification, the second prior probability (i.e., the noisy prior probability) can be determined. Next, the corrected fairness index can be calculated based on the transition matrix, the first prior probability, the second prior probability, and the noisy fairness index.

[0063] Figure 5 A flowchart of a method 500 for correcting the fairness of a model according to another embodiment of this disclosure is shown. In an embodiment of this disclosure, the target dataset lacks true sensitive variables, and the predicted classification includes predicted sensitive variables. Figure 5 As shown in Figure 501, based on at least one feature of the target dataset, a predictive sensitive variable for the target dataset is determined by an auxiliary model, wherein the auxiliary model is not trained on training data labeled with the sensitive variable. In some embodiments, for example, the target dataset X is not labeled with the true sensitive variable A (e.g., the true age group is elderly / middle-aged / youth / child), and the predicted classification includes the predicted sensitive variable. (For example, the age group is elderly / middle-aged / youth / child). In a further embodiment, based on at least one feature t (e.g., nickname) in the target dataset X (e.g., including nickname, browsing time), through the auxiliary model 110 (e.g., ... Identify the sensitive variables for predicting the target dataset X. In this embodiment, the auxiliary model 110 is not trained on training data labeled with sensitive variables. That is, in this disclosure, the auxiliary model 110 does not need to use manually labeled data or publicly available data containing sensitive variables (e.g., publicly available government statistics or open-source data) as auxiliary data to label the sensitive variables of the target dataset. Instead, it can directly use the target dataset for classification prediction. In this way, the auxiliary model can make preliminary predictions of sensitive variables using the target dataset to label the target dataset with predicted sensitive variables that are comparable to noise, thus eliminating the need for real or manually labeled sensitive variables and significantly saving resources and costs.

[0064] In embodiments of this disclosure, such as Figure 5 As shown in Figure 503, a fairness metric for the target model is determined based on the target dataset and the predicted classification. In some embodiments, for example, the target model 120 obtains the predicted classification from the auxiliary model 110. Based on the target dataset X and the predicted classification Determine the noise fairness index for target model 120.

[0065] In embodiments of this disclosure, such as Figure 5 As shown in Figure 504, based on the range of values ​​for the predicted and true classifications, the row and column structure (K×K) of the transition matrix used to represent the probability distribution is determined. Figure 5 As shown in Figure 505, for the auxiliary model, the transition matrix used to represent the probability distribution is determined through noise label learning. Where the transition matrix This includes a probability distribution consisting of the probabilities of each predicted classification relative to each true classification. For example, in embodiments of this disclosure, due to the predicted classification... The value range of is (0, 1, 2, 3) and the value range of the true classification A is (0, 1, 2, 3), thus determining a 4×4 probability distribution. In this way, the specific method of fairness correction can be determined corresponding to the type of classifier, effectively improving the accuracy of determining the fairness index of the target model. In some embodiments, for example, noise label learning is used for the auxiliary model 110 to determine the method used to represent the predicted classification. The transition matrix is ​​the probability distribution relative to the true classification A. In some embodiments, noisy label learning can be based on confidence learning, forgetting events, sampling loss values, sample similarity, or, for example, a high-order consensus algorithm. In this way, fairness based on a preliminary determination with noise can be corrected based on the transition matrix, effectively improving the accuracy of the fairness index determination for the target model. For example, noisy label learning can also be performed using a high-order consensus algorithm based on the distance difference of each noisy label and an unconstrained optimization algorithm, where the unconstrained optimization algorithm can include gradient descent, Newton's method, conjugate gradient method, etc. In this way, the difference between the predicted and true classifications can be reduced by utilizing the difference of the predicted sensitive variables equivalent to the noise, thereby accurately determining the transition matrix used to correct fairness.

[0066] At position 506, a first prior probability is determined based on the transition matrix, and a second prior probability is determined based on the predicted classification. In some embodiments, for example, the first prior probability is determined based on the transition matrix determined at position 505. Based on predictive classification Determine the second prior probability In some embodiments, the first prior probability represents a statistical prior probability, and the second prior probability represents a noisy prior probability learned based on noisy labels. It should be understood that the method for determining the prior probabilities can be determined in any way based on a pre-estimated probability distribution, and in some cases, posterior probabilities can also be used.

[0067] At point 507, the fairness index of the target model is corrected based on the first prior probability, the second prior probability, and the transition matrix. In some embodiments, the fairness index of the target model is updated by multiplying the inverse of the diagonal matrix of the first prior probability, the diagonal matrix of the second prior probability, the inverse of the transpose of the transition matrix, and the fairness index of the target model, thus determining the fairness of the updated target model 120. In some embodiments, for example, using Formula 5 above, the inverse matrix of the diagonal matrix of the first prior probability is... The diagonal matrix of the second prior probability The inverse of the transpose of the transition matrix And the fairness of target model 120 Perform matrix multiplication to obtain the fairness of the updated target model 120. In some embodiments, it can be based on corrected fairness metrics. To evaluate the fairness of the target model 120. In a further embodiment, this can be based on a corrected fairness metric. The fairness of the target model 120 is determined by the range of values ​​for the sensitive variables (e.g., 3, 2, 1, or 0). For example, if the fairness of the target model 120 is characterized by sociodemographic differences, the fairness of the target model 120 can be expressed as: In this way, the probabilities of sensitive variables in the auxiliary model and the predictions equivalent to noise can be taken into account, thereby effectively improving the accuracy of determining the fairness index of the target model.

[0068] According to embodiments of this disclosure, a fairness index for the target model is initially determined based on the predicted classification of a target dataset lacking sensitive variables by an auxiliary model. The fairness index is then corrected based on the probability distribution of the predicted classification relative to the true classification. This approach processes data with predicted sensitive variables using a noise label learning method. It effectively improves the accuracy of determining the fairness index of the target model without requiring target data and / or auxiliary data with sensitive variables, and without needing to train the auxiliary model on fairness, thus further optimizing the target model's fairness index. For example, by processing the target dataset with the fairness-corrected target model, it can predict a more equitable classification of consumption intentions across age groups, mitigating age bias present in the original target dataset and achieving fairer big data prediction.

[0069] It should be understood that in the embodiments of this disclosure, the number of auxiliary models and target models is not limited to one, but can be a combination of multiple models, and the classifier types of the auxiliary models and target models can also be arbitrarily combined. It should also be understood that in some embodiments, the number of variables included in the predicted sensitive variables and the true sensitive variables is not limited to one type, but can be an arbitrary combination of multiple types.

[0070] According to further research by the inventors, the degree to which corrected fairness approaches true fairness compared to uncorrected fairness, as expressed in terms of equal odds, equal opportunity, and demographic disparity, according to embodiments of this disclosure, is shown in Table 1 below:

[0071] Table 1. Accuracy Improvement Experiment of the Correction Scheme in the Embodiments of this Disclosure

[0072]

[0073] As shown in Table 1 above, the method for determining fairness according to this disclosure can improve the accuracy of corrected fairness by 47% to 89% compared to uncorrected fairness in terms of its approximation of true fairness.

[0074] Figure 6A schematic block diagram of a fairness determination apparatus 600 according to some embodiments of the present disclosure is shown. Figure 6 As shown, the fairness determination device 600 includes a prediction classification determination module 610, a fairness index determination module 620, and a fairness index correction module 630. The prediction classification determination module 610 is configured to determine a predicted classification for the target dataset. The fairness index determination module 620 is configured to determine a fairness index for the target model based on the target dataset and the predicted classification. The fairness index correction module 630 is configured to correct the fairness index of the target model based on the probability distribution of the predicted classification relative to the true classification. According to the fairness determination device 600 of this disclosure, for target datasets without sensitive variables, the accuracy of determining the fairness index of the target model can be effectively improved without using auxiliary data to train the auxiliary model for classification.

[0075] In some embodiments, where the target dataset lacks true sensitive variables, the predicted classification includes predicted sensitive variables, and the predicted classification determination module 610 includes a sensitive variable determination module configured to determine the predicted sensitive variables of the target dataset using an auxiliary model based on at least one feature of the target dataset, wherein the auxiliary model is not trained on training data labeled with sensitive variables. In this way, it is possible to label the target dataset with predicted sensitive variables that are comparable to noise by using the target dataset to perform preliminary predictions of sensitive variables on the auxiliary model, thereby eliminating the need for real or manually labeled sensitive variables and significantly saving resources and costs.

[0076] In some embodiments, the fairness index correction module 630 includes a transition matrix determination module. In some embodiments, the transition matrix determination module is configured to determine a transition matrix representing a probability distribution for an auxiliary model through noisy label learning, wherein the transition matrix includes a probability distribution consisting of the probabilities of each predicted classification relative to each true classification. In some embodiments, the fairness index correction module 630 is configured to correct the fairness index of the target model based on the transition matrix. In this way, fairness based on a preliminary determination with noise can be corrected based on the transition matrix, effectively improving the accuracy of the fairness index determination for the target model.

[0077] In some embodiments, the fairness index correction module 630 includes a prior probability determination module. In some embodiments, the prior probability determination module is configured to determine a first prior probability based on a transition matrix, determine a second prior probability based on a predicted classification, and correct fairness based on the first prior probability, the second prior probability, and the transition matrix. In some embodiments, the fairness index correction module 630 updates fairness by multiplying the inverse of the diagonal matrix of the first prior probability, the diagonal matrix of the second prior probability, the inverse of the transpose of the transition matrix, and fairness. In some embodiments, the first prior probability represents a statistical prior probability, and the second prior probability represents a noisy prior probability. In this way, the probabilities of the auxiliary model and the prediction of sensitive variables equivalent to noise can be considered, thereby effectively improving the accuracy of the fairness index determination for the target model.

[0078] In some embodiments, the transition matrix determination module includes a high-order consensus computation module configured to determine the transition matrix using a high-order consensus algorithm based on distance differences between each noise label and an unconstrained optimization algorithm. In this way, the difference between the predicted and true classifications can be reduced by utilizing the differences of the predicted sensitive variables, which are equivalent to the noise, thereby enabling accurate determination of the transition matrix used to correct for fairness.

[0079] In some embodiments, the transition matrix determination module includes a row and column structure determination module configured to determine the row and column structure of the transition matrix representing the probability distribution based on the range of values ​​for the predicted classification and the true classification. In this way, the specific method of fairness correction can be determined corresponding to the type of classifier, effectively improving the accuracy of determining the fairness index of the target model.

[0080] In some embodiments, the prediction classification determination module 610 determines the predicted classification through an auxiliary model, wherein the auxiliary model includes a binary classifier or a multi-class classifier. In some embodiments, the fairness index determination module 620 determines the fairness index of the target model through the target model, wherein the target model includes a binary classifier or a multi-class classifier. In some embodiments, the binary classifier includes a fully connected layer, an activation function layer, and a binary logistic regression layer, and the multi-class classifier includes a fully connected layer, an activation function layer, and a softmax function layer. In this way, the method for improving the fairness determination of artificial intelligence models can be applied to classifiers and classification algorithms in various fields, thereby increasing the scope of application.

[0081] Figure 7 A schematic block diagram of an example device 700 that can be used to implement embodiments of the present disclosure is shown. Device 700 can be used for the aforementioned process for determining fairness. Figure 7As shown, device 700 includes a central processing unit (CPU) 701, which can perform various appropriate actions and processes according to computer program instructions stored in read-only memory (ROM) 702 or loaded from storage unit 708 into random access memory (RAM) 703. RAM 703 may also store various programs and data required for the operation of device 700. CPU 701, ROM 702, and RAM 703 are interconnected via bus 704. Input / output (I / O) interface 705 is also connected to bus 704.

[0082] Multiple components in device 700 are connected to I / O interface 705, including: input unit 706, such as keyboard, mouse, etc.; output unit 707, such as various types of monitors, speakers, etc.; storage unit 708, such as disk, optical disk, etc.; and communication unit 709, such as network card, modem, wireless transceiver, etc. Communication unit 709 allows device 700 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0083] The various processes and procedures described above, such as the region determination process, region partitioning process, and object configuration process for determining fairness, can be executed by the processing unit 701. For example, in some embodiments, the region determination process, region partitioning process, and object configuration process can be implemented as computer software programs tangibly contained in a machine-readable medium, such as storage unit 708. In some embodiments, part or all of the computer program can be loaded and / or installed on device 700 via ROM 702 and / or communication unit 709. When the computer program is loaded into RAM 703 and executed by CPU 701, one or more actions of the region determination process, region partitioning process, and object configuration process described above can be performed.

[0084] This disclosure can be a method, apparatus, system, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of this disclosure.

[0085] Computer-readable storage media can be tangible devices capable of holding and storing instructions for use by an instruction execution device. Computer-readable storage media can be, for example—but not limited to—electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of computer-readable storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination of the foregoing. The computer-readable storage media used herein are not to be construed as transient signals themselves, such as radio waves or other freely propagating electromagnetic waves, electromagnetic waves propagating through waveguides or other transmission media (e.g., light pulses through fiber optic cables), or electrical signals transmitted through wires.

[0086] The computer-readable program instructions described herein can be downloaded from computer-readable storage media to various computing / processing devices, or downloaded via a network, such as the Internet, local area network, wide area network, and / or wireless network, to an external computer or external storage device. The network may include copper transmission cables, fiber optic transmission, wireless transmission, routers, firewalls, switches, gateway computers, and / or edge servers. A network adapter card or network interface in each computing / processing device receives the computer-readable program instructions from the network and forwards them to the computer-readable storage media in the respective computing / processing device.

[0087] Computer program instructions used to perform the operations of this disclosure may be assembly instructions, instruction set architecture (ISA) instructions, machine instructions, machine-dependent instructions, microcode, firmware instructions, status setting data, or source code or object code written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Smalltalk, C++, etc., and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The computer-readable program instructions may execute entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer may be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or may be connected to an external computer (e.g., via the Internet using an Internet service provider). In some embodiments, electronic circuitry, such as programmable logic circuitry, field-programmable gate arrays (FPGAs), or programmable logic arrays (PLAs), is personalized by utilizing the status information of the computer-readable program instructions to implement various aspects of this disclosure.

[0088] Various aspects of this disclosure are described herein with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer-readable program instructions.

[0089] These computer-readable program instructions can be provided to a processing unit of a general-purpose computer, a special-purpose computer, or other programmable data processing apparatus to produce a machine such that, when executed by the processing unit of the computer or other programmable data processing apparatus, they create means for implementing the functions / actions specified in one or more blocks of the flowchart and / or block diagram. These computer-readable program instructions can also be stored in a computer-readable storage medium that causes a computer, programmable data processing apparatus, and / or other device to operate in a particular manner. Thus, the computer-readable medium storing the instructions comprises an article of manufacture that includes instructions for implementing aspects of the functions / actions specified in one or more blocks of the flowchart and / or block diagram.

[0090] Computer-readable program instructions may also be loaded onto a computer, other programmable data processing apparatus, or other device to cause a series of operational steps to be performed on the computer, other programmable data processing apparatus, or other device to produce a computer-implemented process, thereby causing the instructions executed on the computer, other programmable data processing apparatus, or other device to perform the functions / actions specified in one or more boxes of a flowchart and / or block diagram.

[0091] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of an instruction containing one or more executable instructions for implementing a specified logical function. In some alternative implementations, the functions marked in the blocks may occur in a different order than those shown in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0092] The various embodiments of this disclosure have been described above. These descriptions are exemplary and not exhaustive, and are not limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for information recommendation, comprising: Based on the target dataset, determine a predicted classification for the target dataset, which includes at least one of the user's nickname and browsing duration; Based on the target dataset and the predicted classification, determine the fairness index of the target model; The fairness metric of the target model is corrected based on a transition matrix representing the probability distribution of the predicted classification relative to the true classification, wherein the transition matrix comprises the probability distribution of each predicted classification relative to each true classification. as well as Information recommendation is performed based on the target dataset and the target model corrected by the fairness index.

2. The method of claim 1, wherein the target dataset lacks true sensitive variables, the predicted classification includes predicted sensitive variables, and determining the predicted classification includes: Based on at least one feature of the target dataset, a sensitive variable for the prediction of the target dataset is determined by an auxiliary model, wherein the auxiliary model is not trained on training data labeled with the sensitive variable.

3. The method of claim 2, wherein correcting the fairness metric of the target model comprises: For the auxiliary model, the transition matrix is ​​determined through noise label learning; as well as Based on the transition matrix, the fairness index of the target model is corrected.

4. The method according to claim 3, wherein correcting the fairness index of the target model further includes: Based on the transition matrix, determine the first prior probability; Based on the predicted classification, a second prior probability is determined; as well as The fairness is corrected based on the first prior probability, the second prior probability, and the transition matrix.

5. The method of claim 4, wherein correcting the fairness comprises: The fairness is updated by multiplying the inverse of the diagonal matrix of the first prior probability, the diagonal matrix of the second prior probability, the inverse of the transpose of the transition matrix, and the fairness.

6. The method of claim 4, wherein the first prior probability represents a statistical prior probability, and the second prior probability represents a noisy prior probability.

7. The method of claim 3, wherein determining the transition matrix representing the probability distribution through noise label learning comprises: The transition matrix is ​​determined using a high-order consensus algorithm based on distance difference and unconstrained optimization based on each noise label.

8. The method of claim 1, wherein correcting the fairness metric of the target model comprises: Based on the range of values ​​for the predicted classification and the true classification, the row and column structure of the transition matrix used to represent the probability distribution is determined.

9. The method of claim 1, wherein determining the predicted classification for the target dataset comprises: The predicted classification is determined by an auxiliary model, wherein the auxiliary model includes a binary classifier or a multi-class classifier.

10. The method of claim 1, wherein determining the fairness metric of the target model comprises: The fairness index is generated from the target model, wherein the target model includes a binary classifier or a multi-class classifier.

11. The method according to claim 9 or 10, wherein the binary classifier comprises a fully connected layer, an activation function layer, and a binary logistic regression layer, and the multi-class classifier comprises a fully connected layer, an activation function layer, and a softmax function layer.

12. An information recommendation device, comprising: The prediction classification determination module is configured to determine a predicted classification for a target dataset, which includes at least one of a user's nickname and browsing duration. The fairness index determination module is configured to determine the fairness index of the target model based on the target dataset and the predicted classification. A fairness index correction module is configured to correct the fairness index of the target model based on a transition matrix representing the probability distribution of the predicted classification relative to the true classification, wherein the transition matrix includes the probability distribution of each predicted classification relative to each true classification. as well as The information recommendation module is configured to recommend information based on the target dataset and using the target model corrected by the fairness index.

13. An electronic device, characterized in that, include: processor; as well as A memory storing instructions that, when executed by the processor, cause the electronic device to perform the method according to any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed by an electronic device, cause the electronic device to perform the method according to any one of claims 1 to 11.

Images