Data analysis decision system based on artificial intelligence

Abstract

The invention discloses a data analysis decision-making system based on artificial intelligence, and relates to the technical field of administrative management. In data preprocessing, a resampling technology is used for processing dominant prejudice, so that data of different groups are more balanced in sample quantity, a model is prevented from excessively inclining a specific group, a multi-objective optimization algorithm is introduced, and the data analysis decision-making efficiency is improved. An accuracy loss function of the model is combined with a plurality of fairness loss functions, the weight is dynamically adjusted through a gradient descent method, fairness and accuracy can be optimized at the same time in the training process, an automatic responsibility asking mechanism is introduced, a responsibility allocation matrix is defined, responsibility boundaries of data providers, developers and operators are clearly divided, and the training efficiency is improved. And a log and a tracking system are established to record each decision path, responsibility can be automatically backtracked when an error occurs, and responsibility investigation is carried out according to the responsibility proportion of participants.

Description

Technical Field

[0001] The present invention relates to the field of administrative management technology, and in particular to a data analysis and decision-making system based on artificial intelligence. Background Art

[0002] As artificial intelligence (AI) is increasingly used in data analysis and decision-making systems, it is gradually being applied to administrative management systems. This has brought the dual challenges of improving decision-making efficiency and maintaining fairness to the administrative management system. The lack of fairness and accountability mechanisms in AI systems may lead to increased bias, opacity and social injustice.

[0003] Most current data analysis and decision-making systems use artificial intelligence systems that rely on historical data for training, and these data themselves often have biases, such as gender, race, and income level. If the model is trained using these biased data, then its predictions and decisions will continue or even amplify these biases, leading to discrimination against certain groups. Due to the complexity of AI algorithms, especially deep learning models, their decision paths are not transparent, and bias problems may be hidden inside the model, making it difficult to detect and correct. Even developers and regulators find it difficult to identify implicit biases in the system.

[0004] At the same time, some data analysis and decision-making systems lack accountability mechanisms. The development of artificial intelligence systems involves multiple roles, including data providers, algorithm designers, model developers, and end users. When the system makes unfair or wrong decisions, it is often difficult to determine who should be held responsible for these issues, making it difficult to hold people accountable. Many AI systems, especially deep learning models, are difficult to explain the logic and mechanisms behind the decisions due to their "black box" nature. This inexplicability makes external review and accountability extremely difficult.

[0005] To alleviate the above problems, traditional AI data analysis decision systems reduce bias in model decisions by removing biased variables in training data or making data distribution more balanced through resampling techniques in terms of fairness and accountability mechanisms. The goal is to make the model's training data more diverse and avoid the impact of historical bias on model output. When training the model, fairness constraints such as "Demographic Parity" or "Equal Opportunity" are used to ensure that the model treats different groups fairly. In terms of accountability mechanisms, models with strong interpretability such as decision trees or logistic regression are used to track every decision step of the algorithm, providing supervisors with a transparent perspective on the system.

[0006] However, data debiasing technology mainly targets explicit variables, such as gender and race, but has limited effect on implicit biases or complex data relationships. In practical applications, many biases cannot be eliminated through simple data processing, such as the impact of implicit factors such as socioeconomic status and cultural background on model decision-making.

[0007] Although the introduction of fairness constraints can improve the fairness of decision-making, it often leads to a decrease in the accuracy of the model. This trade-off makes it difficult for the system to find the best balance between efficiency and fairness, especially in scenarios such as medical diagnosis that require highly accurate predictions.

[0008] In terms of accountability mechanism, although the transparent model has improved the interpretability of some systems, the interpretability problem is still prominent in complex models such as deep learning. Many efficient deep learning models are difficult to trace their decision-making paths due to their complex structures. This makes it difficult to clearly identify responsibilities after problems occur in actual applications. In AI projects involving multiple parties, the responsibility boundaries of data providers, developers, and users are blurred, resulting in difficulty in enforcing accountability when problems arise in the system, and inability to conduct effective supervision and accountability. Therefore, there is an urgent need for an artificial intelligence-based data analysis and decision-making system to solve such problems. Summary of the invention

[0009] In view of the above existing problems, the present invention is proposed.

[0010] The present invention provides an artificial intelligence-based data analysis decision system to solve the problems that data analysis decision systems are obviously insufficient in dealing with implicit bias, balancing fairness and accuracy, and tracing the responsibility of complex models.

[0011] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0012] In one aspect, the present invention provides a data analysis and decision-making system based on artificial intelligence, which comprises:

[0013] The data collection module collects, reviews, cleans and debias the data, and outputs a data set that has been processed, representative and fair.

[0014] Model selection and construction module, select the appropriate model architecture according to the characteristics of the data and task requirements,

[0015] The model training and optimization module is used to train the model and introduce a multi-dimensional optimization mechanism during the model training process. This module is used to ensure fair treatment of different groups.

[0016] The real-time inference module is responsible for model performance monitoring, dynamic adjustment, and online inference. After the model is trained and optimized, it is deployed to the actual system.

[0017] The automated accountability module clarifies the allocation of responsibilities for deviations or errors that may occur in the decision-making process.

[0018] In another aspect, the present invention provides a data analysis and decision-making method based on artificial intelligence, comprising:

[0019] Step S1, data preprocessing and debiasing,

[0020] Pre-process the collected raw data to identify and process explicit and implicit biases.

[0021] Step S2, model selection and interpretability optimization,

[0022] According to the task requirements and data characteristics, select the corresponding model architecture and embed interpretability tools into the deep neural network model.

[0023] The model selection method in step S2 is:

[0024] Models with strong interpretability, such as logistic regression and decision trees, are suitable for scenarios that require high transparency and interpretability (such as medical diagnosis, judicial decision-making, etc.).

[0025] Deep learning models, such as recurrent neural networks (RNNs) and convolutional neural networks (CNNs), are suitable for high-dimensional and complex data types, including images, text, and time series.

[0026] In step S2, for the deep neural network model, an interpretability tool is embedded, and the LIME (Local Interpretable Model Agnostic Explanations) tool is used to generate a report on the contribution of features in the prediction.

[0027] Step S3: Model training and fairness constraints.

[0028] Based on the model selected in step S2, model training is performed. During the model training process, a fairness constraint mechanism is introduced to ensure fairness for different groups while maintaining the efficiency of the model.

[0029] In step S3, the fairness constraint is:

[0030] Demographic balance ensures that the proportions of different groups in the forecast results remain relatively consistent,

[0031] Equal opportunity, ensuring that all groups have equal opportunities in classification accuracy,

[0032] Equal error rate, balancing the misclassification rates of different groups to avoid unfairly high error rates for specific groups

[0033] Step S4: Deployment of accountability mechanism and automated accountability.

[0034] Embed an automated accountability mechanism in the data analysis and decision-making system, clarify the responsibilities of each participant, and establish a traceable process.

[0035] In step S4, responsibility boundaries are divided, including:

[0036] Data providers are responsible for the accuracy, representativeness and compliance of the data.

[0037] Developers are responsible for model design, training, and fairness verification, and regularly submit fairness and bias assessment reports.

[0038] The operator is responsible for monitoring the actual application of the model.

[0039] Furthermore, in step S1, explicit bias processing refers to: reviewing explicit bias variables including gender and race, using resampling methods (such as undersampling or oversampling) to balance the bias in the data set,

[0040] Implicit bias processing refers to: using adversarial debiasing models to automatically identify and process implicit biases, including complex sources of bias such as socioeconomic status and cultural background.

[0041] During the data processing and model training process in step S1, bias information in the data is continuously monitored, and feature engineering and parameter adaptive adjustment are used to generate a debiased dataset with enhanced fairness.

[0042] Furthermore, in step S1, the data preprocessing and debiasing method is:

[0043] Let the categorical variable be X i , i represents the i-th feature, for a categorical variable X i , calculate the proportion of the target variable Y under each category:

[0044] Where P(Y=1|X i =x) is the conditional probability that the target variable Y is 1 under a specific category x, N is the total number of samples, j represents the sample index, and X ii represents the value of the ith feature of the jth sample. I(·) is an indicator function. It takes the value of 1 when the condition is met and 0 otherwise. The difference in the proportion of gender or race dominant variables under different classifications is calculated. If the difference in the proportion of a certain group is too large, there is an explicit bias.

[0045] Balancing the explicit bias by oversampling or undersampling,

[0046] Suppose there are two groups G 1 and G 2, representing gender and ethnic groups respectively, and resampling was performed so that the number of samples between groups and Keep the proportions balanced:

[0047] in, It is group G 1 The number of samples, It is group G 2 The number of samples, ResampleRatio represents the resampling ratio, which is adjusted to be close to 1 to balance explicit biases such as gender or race.

[0048] The adversarial debiasing model is used to identify and process implicit bias. The model consists of two parts, including the main model and the bias detector. The goal of the main model is to minimize the loss function L M , that is, optimizing the classification accuracy without considering the bias, the goal of the bias detector is to maximize the loss function L D , detect potential implicit bias,

[0049] The loss function of the main model is:

[0050] Where N is the total number of samples, Y i is the true label of the i-th sample, is the predicted value of the ith sample,

[0051] The loss function of the bias detector is:

[0052] Among them, Z i is the potential biased label of the i-th sample, is the biased label predicted by the model,

[0053] The final model optimizes the adversarial loss function to remove bias:

[0054] Lt otal =L M -λL D , where L total is the final loss function, λ controls bias removal and classification accuracy,

[0055] Optimize the adversarial debiasing model to minimize L M And maximize L D ,Right now:

[0056] Among them, θ M is the main model parameter, θ D are the bias detector parameters.

[0057] Furthermore, in step S1, dynamic bias detection and continuous adjustment are performed, and bias detection is monitored by fairness indicators, including:

[0058] Demographic Balance:

[0059] When DP is close to 1, it means that the proportions of the prediction results of the two groups are consistent.

[0060] Equal Opportunity:

[0061] If EO is close to 1, it means that the two groups have equal chances in classification accuracy.

[0062] Perform adaptive parameter adjustment:

[0063] Among them, λ t is the balance parameter for the current training step, α is the learning rate, It is the gradient of the total loss function and outputs a fairness-enhanced debiased dataset.

[0064] Furthermore, in step S3, a multi-objective optimization algorithm is used to dynamically adjust the weight between fairness and model accuracy.

[0065] Furthermore, in step S3, a loss function corresponding to the fairness constraint condition is defined, including:

[0066] Demographic equilibrium loss function:

[0067] in, For group g 1 The positive prediction probability of For group g 2 The positive prediction probability, L DP is the demographic equilibrium loss function, the goal is to make this value as close to 0 as possible,

[0068] This loss function measures the difference in the model's positive prediction results for different groups. If the difference in the positive predictions of two groups is large, it means that there is a statistical imbalance. By minimizing the difference, the two groups are kept consistent in their positive predictions.

[0069] Equal opportunity loss function:

[0070]

[0071] in, It is group g 1 The probability of accurately predicting the positive class under the condition that the true label Y = 1, It is group g 2The probability of accurately predicting the positive class under the condition that the true label Y = 1, L EO is an equal opportunity loss function, the goal is to make the two groups consistent in classification accuracy by minimizing L EO , so that the two groups have the same classification accuracy when the true label is positive,

[0072] Peer error rate loss function:

[0073]

[0074] in, It is group g 1 The false positive rate (FPR), that is, the probability of incorrectly predicting a positive label when the true label is negative, It is group g 2 The false positive rate, It is group g 1 The false negative rate (FNR), that is, the probability of incorrectly predicting a negative label when the true label is positive, It is group g 2 The false negative rate, L ERR is the loss function of the equal error rate. The goal is to keep the error rates of the two groups as consistent as possible by minimizing L ERR , reducing the differences in false positive and false negative rates between groups, thereby reducing the error rate of the model for a specific group.

[0075] Furthermore, in step S3, a multi-objective optimization algorithm is used to dynamically adjust fairness and model accuracy:

[0076] The fairness loss function is combined with the main loss function of the model to construct a multi-objective optimization problem, which is expressed as a cross entropy loss function:

[0077] N is the total number of samples, Y i is the true label of the i-th sample, is the predicted value of the i-th sample, L CE is the cross entropy loss of the model,

[0078] Introduce multi-objective optimization and combine the fairness loss function with the model loss function:

[0079] L total =L CE +λ 1 L DP +λ 2 L EO +λ 3 L ERR , where L CE is the main loss function of the model, L DPis the demographic equilibrium loss, L EO is the equal opportunity loss, L ERR is the peer error rate loss, λ 1 ,λ 2 ,λ 3 is the weight coefficient used to balance model accuracy and fairness constraints;

[0080] The gradient descent method is used to dynamically adjust the weight parameters to balance the accuracy and fairness of the model:

[0081] Among them, λ k is the weight parameter in the kth iteration, α is the learning rate, is the comprehensive loss function L total The gradient of the weight parameter.

[0082] Furthermore, in step S4, an automated accountability process is established, and a log and tracking system is used to record the decision path of each stage of the model. When problems occur, the responsible party can be traced back to improve the efficiency of external supervision.

[0083] Furthermore, in step S4, the responsibilities of each participant are quantified and decision tracking is performed:

[0084] For each participant, a responsibility allocation matrix R is defined to quantify the contribution of each participant to the system risk according to their different tasks in the system.

[0085] R i，j =α j β i，j γ j , where R i，j represents the degree of responsibility of the i-th participant in the j-th task, α j is the contribution of the jth task to the overall risk of the system, ranging from [0, 1], β i，j is the proportion of tasks that the ith participant is directly responsible for in the jth task, reflecting the depth of participation of the participant, with a value range of [0, 1]. j is the importance weight of the jth task, reflecting the criticality of the task, and its value range is [0, 1].

[0086] Quantifying the responsibilities of each participant in each task, the impact of different tasks on system risk, and the overall importance of the task all affect the allocation of responsibilities.

[0087] Take the weighted sum of the responsibilities of all tasks:

[0088] Among them, R iis the total responsibility of the ith participant, N is the total number of tasks in the system, and the responsibilities of all tasks are summed to obtain the total responsibility of each participant in the entire system, quantifying the contribution of each participant to the final decision of the system.

[0089] Automated accountability mechanisms record all steps in the model’s decision-making process:

[0090] Among them, L t is the log record at time t, M is the total number of steps involved in the decision process, and w i is the weight of each step, which is used to reflect the contribution of the step to the final decision, and its value range is [0, 1]. is the prediction result of step i, X i is the input data, θ i is the model parameter of the i-th step,

[0091] Define a deviation detection formula to track errors or deviations in the decision path:

[0092] Among them, E t is the error at time t, Y i is the true output of step i, is the prediction result of step i, δ i is the anomaly detection weight, which indicates the severity of the error in this step. Its value range is [0, 1]. It calculates the difference between the actual output and the predicted output of each step, and monitors whether anomalies occur in the decision-making process in real time. Combined with the anomaly detection weight δ i , determine the severity of an error in a certain step and automatically trigger an alarm when a deviation occurs,

[0093] When an error occurs, the system automatically traces back the responsibility and locates the source of the problem. The system calculates the change in the responsibility of each participant in a task and locates the source of the problem:

[0094] ΔR i，t =R i，t -R i，t-1 , where ΔR i，t is the change in responsibility of participant i at time t, R i，t is the responsibility value of participant i at time t, R i，t-1 is the responsibility value of participant i at the previous moment,

[0095] When problems occur in the system, responsibilities are automatically assigned to different parties, and their responsibilities are quantified based on their actual participation and risk contribution in the decision-making path:

[0096] Among them, A iis the final responsibility value assigned to the i-th participant, ΔR i，t is the change in responsibility of participant i at time t, P is the total number of all participants, and C is the overall cost or penalty of system errors. According to the change in responsibility of each participant, the responsibility and cost after the system has problems are automatically allocated.

[0097] The beneficial effects of the present invention are:

[0098] The present invention uses resampling technology to deal with explicit bias in data preprocessing, making the data of different groups more balanced in sample size, thereby preventing the model from being overly biased towards a specific group. For implicit bias, an adversarial debiasing model is introduced to detect and eliminate implicit bias in the data through an adversarial neural network structure. This allows different groups to be treated fairly when processing complex social background variables, and can automatically handle potential sources of bias when implicit bias is difficult to directly identify.

[0099] The present invention introduces a multi-objective optimization algorithm, combines the accuracy loss function of the model with multiple fairness loss functions, and dynamically adjusts the weights through the gradient descent method, so that fairness and accuracy can be optimized simultaneously during the training process.

[0100] The present invention introduces the LIME interpretability tool to generate feature contribution reports. When processing complex data, such tools are embedded to make the decision path behind each prediction transparent and explainable.

[0101] The present invention introduces an automated accountability mechanism, defines a responsibility allocation matrix, clearly divides the responsibility boundaries of data providers, developers, and operators, and establishes a log and tracking system to record each decision path. When errors occur, it can automatically trace back responsibilities and hold the participants accountable according to their proportion of responsibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0103] Figure 1 This is a schematic diagram of the structure of the data analysis and decision-making system based on artificial intelligence of the present invention;

[0104] Figure 2 The figure is a flow chart of the artificial intelligence-based data analysis and decision-making method of the present invention. DETAILED DESCRIPTION

[0105] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0106] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0107] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0108] Example 1, reference Figure 1 , this embodiment provides a data analysis and decision-making system based on artificial intelligence, including:

[0109] The data collection module collects, reviews, cleans and debias the data, and outputs a data set that has been processed, representative and fair.

[0110] Model selection and construction module, select the appropriate model architecture according to the characteristics of the data and task requirements,

[0111] The model training and optimization module is used to train the model and introduce a multi-dimensional optimization mechanism during the model training process. This module is used to ensure fair treatment of different groups.

[0112] The real-time inference module is responsible for model performance monitoring, dynamic adjustment, and online inference. After the model is trained and optimized, it is deployed to the actual system.

[0113] The automated accountability module clarifies the allocation of responsibilities for deviations or errors that may occur in the decision-making process.

[0114] Example 2, reference Figure 2 This embodiment provides a data analysis and decision-making method based on artificial intelligence, comprising the following steps:

[0115] Step S1, data preprocessing and debiasing,

[0116] Pre-process the collected raw data to identify and process explicit and implicit biases.

[0117] In step S1, explicit bias treatment refers to: reviewing explicit bias variables including gender and race, using resampling methods (such as undersampling or oversampling) to balance the bias in the data set,

[0118] Implicit bias processing refers to: using adversarial debiasing models to automatically identify and process implicit biases, including complex sources of bias such as socioeconomic status and cultural background.

[0119] During the data processing and model training process in step S1, the bias information in the data is continuously monitored, and feature engineering and parameter adaptive adjustment are used to generate a debiased dataset with enhanced fairness.

[0120] In step S1, the data preprocessing and debiasing method is:

[0121] Let the categorical variable be X i , i represents the i-th feature, for a categorical variable X i , calculate the proportion of the target variable Y under each category:

[0122] Where P(Y=1|X i =x) is the conditional probability that the target variable Y is 1 under a specific category x, N is the total number of samples, j represents the sample index, and X ij represents the value of the ith feature of the jth sample. I(·) is an indicator function. It takes the value of 1 when the condition is met and 0 otherwise. The difference in the proportion of gender or race dominant variables under different classifications is calculated. If the difference in the proportion of a certain group is too large, there is an explicit bias.

[0123] Balancing the explicit bias by oversampling or undersampling,

[0124] Suppose there are two groups G 1 and G 2 , representing gender and ethnic groups respectively, and resampling was performed so that the number of samples between groups and Keep the proportions balanced:

[0125] in, It is group G 1 The number of samples, It is group G 2 The number of samples, ResampleRatio represents the resampling ratio, which is adjusted to be close to 1 to balance explicit biases such as gender or race.

[0126] The adversarial debiasing model is used to identify and process implicit bias. The model consists of two parts, including the main model and the bias detector. The goal of the main model is to minimize the loss function L M, that is, optimizing the classification accuracy without considering the bias, the goal of the bias detector is to maximize the loss function L D , detect potential implicit bias,

[0127] The loss function of the main model is:

[0128] Where N is the total number of samples, Y i is the true label of the i-th sample, is the predicted value of the ith sample,

[0129] The loss function of the bias detector is:

[0130] Among them, Z i is the potential biased label of the i-th sample, is the biased label predicted by the model,

[0131] The final model optimizes the adversarial loss function to remove bias:

[0132] Lt otal =L M -λL D , where Lt otal is the final loss function, λ controls bias removal and classification accuracy,

[0133] Optimize the adversarial debiasing model to minimize L M And maximize L D ,Right now:

[0134] Among them, θ M is the main model parameter, θ D are the bias detector parameters,

[0135] In step S1, dynamic bias detection and continuous adjustment are performed, and bias detection is monitored by fairness indicators, including:

[0136] Demographic Balance:

[0137] When DP is close to 1, it means that the proportions of the prediction results of the two groups are consistent.

[0138] Equal Opportunity:

[0139] If EO is close to 1, it means that the two groups have equal chances in classification accuracy.

[0140] Perform adaptive parameter adjustment:

[0141] Among them, λ tis the balance parameter for the current training step, α is the learning rate, It is the gradient of the total loss function, outputting a fairness-enhanced debiased dataset;

[0142] Specifically, explicit biases are resampled in the data preprocessing stage, and implicit biases are dealt with using an adversarial debiasing model to ensure the balance and fairness of the data set, improve the model's inclusiveness to different groups, and significantly reduce unfair decisions caused by historical biases. The introduction of an adversarial debiasing model can automatically identify complex sources of bias such as socioeconomic status.

[0143] Step S2, model selection and interpretability optimization,

[0144] According to the task requirements and data characteristics, select the corresponding model architecture and embed interpretability tools into the deep neural network model.

[0145] The model selection method in step S2 is:

[0146] Models with strong interpretability, such as logistic regression and decision trees, are suitable for scenarios that require high transparency and interpretability (such as medical diagnosis, judicial decision-making, etc.).

[0147] Deep learning models, such as recurrent neural networks (RNNs) and convolutional neural networks (CNNs), are suitable for high-dimensional and complex data types, including images, text, and time series.

[0148] In step S2, for the deep neural network model, an interpretability tool is embedded, and the LIME (Local Interpretable Model Agnostic Explanations) tool is used to generate a report on the contribution of features in the prediction.

[0149] Specifically, a multi-objective optimization algorithm is introduced in the model selection and training process, and multiple fairness constraints are combined to improve the prediction fairness of the model among different groups, avoiding excessively high error rates for specific groups.

[0150] Step S3: Model training and fairness constraints.

[0151] Based on the model selected in step S2, model training is performed. During the model training process, a fairness constraint mechanism is introduced to ensure fairness for different groups while maintaining the efficiency of the model.

[0152] In step S3, the fairness constraint is:

[0153] Demographic balance ensures that the proportions of different groups in the forecast results remain relatively consistent,

[0154] Equal opportunity, ensuring that all groups have equal opportunities in classification accuracy,

[0155] Equal error rates, balancing the misclassification rates of different groups to avoid unfairly high error rates for specific groups,

[0156] In step S3, a multi-objective optimization algorithm is used to dynamically adjust the weight between fairness and model accuracy.

[0157] In step S3, a loss function corresponding to the fairness constraint is defined, including:

[0158] Demographic equilibrium loss function:

[0159] in, For group g 1 The positive prediction probability of For group g 2 The positive prediction probability, L DP is the demographic equilibrium loss function, the goal is to make this value as close to 0 as possible,

[0160] This loss function measures the difference in the model's positive prediction results for different groups. If the difference in the positive predictions of two groups is large, it means that there is a statistical imbalance. By minimizing the difference, the two groups are kept consistent in their positive predictions.

[0161] Equal opportunity loss function:

[0162]

[0163] in, It is group g 1 The probability of accurately predicting the positive class under the condition that the true label Y = 1, It is group g 2 The probability of accurately predicting the positive class under the condition that the true label Y = 1, L EO is an equal opportunity loss function, the goal is to make the two groups consistent in classification accuracy by minimizing L EO , so that the two groups have the same classification accuracy when the true label is positive,

[0164] Peer error rate loss function:

[0165]

[0166] in, is the false positive rate (FPR) of population g1, i.e., the probability of incorrectly predicting a positive label when the true label is negative, is the false positive rate of population g2, is the false negative rate (FNR) of population g1, i.e., the probability of incorrectly predicting a negative label when the true label is positive, It is group g 2 The false negative rate, L ERR is the loss function of the equal error rate. The goal is to keep the error rates of the two groups as consistent as possible by minimizing L ERR , reducing the difference in false positive and false negative rates between groups, thereby reducing the error rate of the model for a specific group,

[0167] In step S3, a multi-objective optimization algorithm is used to dynamically adjust fairness and model accuracy:

[0168] The fairness loss function is combined with the main loss function of the model to construct a multi-objective optimization problem, which is expressed as a cross entropy loss function:

[0169] N is the total number of samples, Y i is the true label of the i-th sample, is the predicted value of the i-th sample, L CE is the cross entropy loss of the model,

[0170] Introduce multi-objective optimization and combine the fairness loss function with the model loss function:

[0171] L total =L CE +λ 1 L DP +λ 2 L Eo +λ 3 L ERR , where L CE is the main loss function of the model, L DP is the demographic equilibrium loss, L EO is the equal opportunity loss, L ERR is the peer error rate loss, λ 1 ,λ 2 ,λ 3 is the weight coefficient used to balance model accuracy and fairness constraints;

[0172] The gradient descent method is used to dynamically adjust the weight parameters to balance the accuracy and fairness of the model:

[0173] Among them, λ k is the weight parameter in the kth iteration, α is the learning rate, is the comprehensive loss function L total Gradients with respect to weight parameters;

[0174] Specifically, a multi-objective optimization algorithm is introduced to combine the main loss function of the model with the fairness loss function. The weight parameters are dynamically adjusted through the gradient descent method without sacrificing fairness. While continuously optimizing the model performance, the fairness and impartiality of the system are maintained.

[0175] Step S4: Deployment of accountability mechanism and automated accountability.

[0176] Embed an automated accountability mechanism in the data analysis and decision-making system, clarify the responsibilities of each participant, and establish a traceable process.

[0177] In step S4, responsibility boundaries are divided, including:

[0178] Data providers are responsible for the accuracy, representativeness and compliance of the data.

[0179] Developers are responsible for model design, training, and fairness verification, and regularly submit fairness and bias assessment reports.

[0180] The operator is responsible for monitoring the actual application of the model.

[0181] In step S4, an automated accountability process is established, using a log and tracking system to record the decision path of each stage of the model. When problems occur, the responsible party can be traced back to improve the efficiency of external supervision.

[0182] In step S4, the responsibilities of each participant are quantified and decision tracking is performed:

[0183] For each participant, a responsibility allocation matrix R is defined to quantify the contribution of each participant to the system risk according to their different tasks in the system.

[0184] R i，j =α j β i，j γ j , where R i，j represents the degree of responsibility of the i-th participant in the j-th task, α j is the contribution of the jth task to the overall risk of the system, ranging from [0, 1], β i，j is the proportion of tasks that the ith participant is directly responsible for in the jth task, reflecting the depth of participation of the participant, with a value range of [0, 1]. j is the importance weight of the jth task, reflecting the criticality of the task, and its value range is [0, 1].

[0185] Quantifying the responsibilities of each participant in each task, the impact of different tasks on system risk, and the overall importance of the task all affect the allocation of responsibilities.

[0186] Take the weighted sum of the responsibilities of all tasks:

[0187] Among them, R i is the total responsibility of the ith participant, N is the total number of tasks in the system, and the responsibilities of all tasks are summed to obtain the total responsibility of each participant in the entire system, quantifying the contribution of each participant to the final decision of the system.

[0188] Automated accountability mechanisms record all steps in the model’s decision-making process:

[0189] Among them, L t is the log record at time t, M is the total number of steps involved in the decision process, and w i is the weight of each step, which is used to reflect the contribution of the step to the final decision, and its value range is [0, 1]. is the prediction result of step i, X i is the input data, θ i is the model parameter of the i-th step,

[0190] Define a deviation detection formula to track errors or deviations in the decision path:

[0191] Among them, E t is the error at time t, Y i is the true output of step i, is the prediction result of step i, δ i is the anomaly detection weight, which indicates the severity of the error in this step. Its value range is [0, 1]. It calculates the difference between the actual output and the predicted output of each step, and monitors whether anomalies occur in the decision-making process in real time. Combined with the anomaly detection weight δ i , determine the severity of an error in a certain step and automatically trigger an alarm when a deviation occurs,

[0192] When an error occurs, the system automatically traces back the responsibility and locates the source of the problem. The system calculates the change in the responsibility of each participant in a task and locates the source of the problem:

[0193] ΔR i，t =R i，t -R i，t-1 , where ΔR i，t is the change in responsibility of participant i at time t, R i，t is the responsibility value of participant i at time t, R i，t-1 is the responsibility value of participant i at the previous moment,

[0194] When problems occur in the system, responsibilities are automatically assigned to different parties, and their responsibilities are quantified based on their actual participation and risk contribution in the decision-making path:

[0195] Among them, A i is the final responsibility value assigned to the i-th participant, ΔR i，t is the change in responsibility of participant i at time t, P is the total number of all participants, C is the overall cost or penalty of system errors, and the responsibility and cost after the system has problems are automatically allocated according to the change in responsibility of each participant;

[0196] Specifically, embedding an automated accountability mechanism greatly improves the transparency of the system and the clarity of responsibility division. By defining a responsibility allocation matrix, automated decision path logging, and anomaly detection mechanisms, the decision path of each step is recorded in real time, and responsibility is automatically traced back when problems arise.

[0197] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A data analysis and decision-making system based on artificial intelligence, characterized in that: include, The data collection module collects, reviews, cleans and debias the data, and outputs a data set that has been processed, representative and fair. Model selection and construction module, select the appropriate model architecture according to the characteristics of the data and task requirements, The model training and optimization module is used to train the model and introduce a multi-dimensional optimization mechanism during the model training process. The real-time inference module is responsible for model performance monitoring, dynamic adjustment, and online inference. After the model is trained and optimized, it is deployed to the actual system. The automated accountability module clarifies the allocation of responsibilities for deviations or errors that may occur in the decision-making process.

2. A data analysis and decision-making method based on artificial intelligence, characterized in that: The data analysis and decision-making system based on artificial intelligence according to claim 1 comprises the following steps: Step S1, data preprocessing and debiasing, Pre-process the collected raw data to identify and process explicit and implicit biases. Step S2, model selection and interpretability optimization, According to the task requirements and data characteristics, select the corresponding model architecture and embed interpretability tools into the deep neural network model. The model selection method in step S2 is: Models with strong interpretability correspond to scenarios that require high transparency and interpretability. Deep learning models for high-dimensional, complex data types, including images, text, and time series. In step S2, for the deep neural network model, an interpretability tool is embedded, and the LIME tool is used to generate a report on the contribution of features in the prediction. Step S3: Model training and fairness constraints. Based on the model selected in step S2, a fairness constraint mechanism is introduced during the model training process. In step S3, the fairness constraint is: Demographic balance ensures that the proportions of different groups in the forecast results remain relatively consistent, Equal opportunity, ensuring that all groups have equal opportunities in classification accuracy, Equal error rate, balancing the misclassification rates of different groups to avoid unfairly high error rates for specific groups Step S4: Deployment of accountability mechanism and automated accountability. Embed an automated accountability mechanism in the data analysis and decision-making system, clarify the responsibilities of each participant, and establish a traceable process. In step S4, responsibility boundaries are divided, including: Data providers are responsible for the accuracy, representativeness and compliance of the data. Developers are responsible for model design, training, and fairness verification, and regularly submit fairness and bias assessment reports. The operator is responsible for monitoring the actual application of the model.

3. The data analysis and decision-making method based on artificial intelligence according to claim 2 is characterized in that: In step S1, explicit bias treatment refers to: reviewing explicit bias variables including gender and race, using resampling methods to balance bias in the data set, Implicit bias processing refers to: using adversarial debiasing models to automatically identify and process implicit biases, including complex sources of bias due to socioeconomic status and cultural background. During the data processing and model training process in step S1, bias information in the data is continuously monitored, and feature engineering and parameter adaptive adjustment are used to generate a debiased dataset with enhanced fairness.

4. The data analysis and decision-making method based on artificial intelligence according to claim 3 is characterized in that: In step S1, the data preprocessing and debiasing method is: Let the categorical variable be X i , i represents the i-th feature, for a categorical variable X i , calculate the proportion of the target variable Y under each category: Where P(Y=1|X i =x) is the conditional probability that the target variable Y is 1 under a specific category x, N is the total number of samples, j represents the sample index, and X ij represents the value of the i-th feature of the j-th sample. I(·) is an indicator function that takes a value of 1 when the condition is met and 0 otherwise. Assume there are two groups G1 and G2, representing gender and ethnic groups respectively, and resample so that the number of samples between groups is and Keep the proportions balanced: in, is the number of samples in population G1, is the number of samples in population G2, ResampleRatio represents the resampling ratio, The adversarial debiasing model is used to identify and process implicit bias. The model consists of two parts, including the main model and the bias detector. The goal of the main model is to minimize the loss function L M , that is, optimizing the classification accuracy without considering the bias, the goal of the bias detector is to maximize the loss function L D , detect potential implicit bias, The loss function of the main model is: Where N is the total number of samples, Y i is the true label of the i-th sample, is the predicted value of the ith sample, The loss function of the bias detector is: Among them, Z i is the potential biased label of the ith sample, is the biased label predicted by the model, The final model optimizes the adversarial loss function to remove bias: L total =L M -λL D , where L total is the final loss function, λ controls bias removal and classification accuracy, Optimize the adversarial debiasing model to minimize L M And maximize L D ,Right now: Among them, θ M is the main model parameter, θ D are the bias detector parameters.

5. The data analysis and decision-making method based on artificial intelligence according to claim 4 is characterized in that: In step S1, dynamic bias detection and continuous adjustment are performed, and bias detection is monitored by fairness indicators, including: Demographic Balance: When DP is close to 1, it means that the proportions of the prediction results of the two groups are consistent. Equal Opportunity: If EO is close to 1, it means that the two groups have equal chances in classification accuracy. Perform adaptive parameter adjustment: Among them, λ t is the balance parameter for the current training step, α is the learning rate, It is the gradient of the total loss function and outputs a fairness-enhanced debiased dataset.

6. The data analysis and decision-making method based on artificial intelligence according to claim 5 is characterized in that: In step S3, a multi-objective optimization algorithm is used to dynamically adjust the weight between fairness and model accuracy.

7. The data analysis and decision-making method based on artificial intelligence according to claim 6 is characterized in that: In step S3, a loss function corresponding to the fairness constraint is defined, including: Demographic equilibrium loss function: in, represents the positive prediction probability for group g1, represents the positive prediction probability for group g2, L DP is the demographic equilibrium loss function, the goal is to make this value as close to 0 as possible, Equal opportunity loss function: in, is the probability that group g1 is accurately predicted as the positive class under the condition that the true label Y = 1, is the probability that group g2 is accurately predicted as the positive class under the condition that the true label Y = 1, L EO is an equal opportunity loss function, the goal is to make the two groups consistent in classification accuracy. Peer error rate loss function: in, is the false positive rate (FPR) of population g1, i.e., the probability of incorrectly predicting a positive label when the true label is negative, is the false positive rate of population g2, is the false negative rate (FNR) of population g1, i.e., the probability of incorrectly predicting a negative label when the true label is positive, is the false negative rate of population g2, L ERR It is a loss function of equal error rate, and the goal is to keep the error rates of the two groups as consistent as possible.

8. The data analysis and decision-making method based on artificial intelligence according to claim 7 is characterized in that: In step S3, a multi-objective optimization algorithm is used to dynamically adjust fairness and model accuracy: The fairness loss function is combined with the main loss function of the model to construct a multi-objective optimization problem, which is expressed as a cross entropy loss function: N is the total number of samples, Y i is the true label of the i-th sample, is the predicted value of the i-th sample, L CE is the cross entropy loss of the model, Introduce multi-objective optimization and combine the fairness loss function with the model loss function: L total =L CE +λ1L DP +λ2L EO +λ3L ERR , where L CE is the main loss function of the model, L DP is the demographic equilibrium loss, L EO is the equal opportunity loss, L ERR is the peer error rate loss, λ1,λ2,λ3 are weight coefficients used to balance model accuracy and fairness constraints; The gradient descent method is used to dynamically adjust the weight parameters: Among them, λ k is the weight parameter in the kth iteration, α is the learning rate, is the comprehensive loss function L total The gradient of the weight parameter.

9. The data analysis and decision-making method based on artificial intelligence according to claim 8 is characterized in that: In step S4, an automated accountability process is established, using a log and tracking system to record the decision path at each stage of the model.

10. The data analysis and decision-making method based on artificial intelligence according to claim 9, characterized in that: In step S4, the responsibilities of each participant are quantified and decision tracking is performed: For each participant, a responsibility allocation matrix R is defined to quantify the contribution of each participant to the system risk according to their different tasks in the system. R i,j =α j β i,j γ j , where R i,j represents the degree of responsibility of the i-th participant in the j-th task, α j is the contribution of the jth task to the overall risk of the system, ranging from [0,1], β i,j is the proportion of tasks that the ith participant is directly responsible for in the jth task, reflecting the depth of participation of the participant, with a value range of [0,1]. j is the importance weight of the jth task, reflecting the criticality of the task, and its value range is [0,1]. Take the weighted sum of the responsibilities of all tasks: Among them, R i is the total responsibility of the ith participant, N is the total number of tasks in the system, Automated accountability mechanisms record all steps in the model’s decision-making process: Among them, L t is the log record at time t, M is the total number of steps involved in the decision process, and w i is the weight of each step, which is used to reflect the contribution of the step to the final decision, and its value range is [0,1]. is the prediction result of step i, X i is the input data, θ i is the model parameter of the i-th step, Define a deviation detection formula to track errors or deviations in the decision path: Among them, E t is the error at time t, Y i is the true output of step i, is the prediction result of step i, δ i is the anomaly detection weight, which indicates the severity of the error in this step. Its value range is [0,1]. It calculates the difference between the actual output and the predicted output of each step, and monitors whether anomalies occur in the decision-making process in real time. Combined with the anomaly detection weight δ i , determine the severity of an error in a certain step and automatically trigger an alarm when a deviation occurs, When an error occurs, the system automatically traces back the responsibility and locates the source of the problem. The system calculates the change in the responsibility of each participant in a task and locates the source of the problem: ΔR i,t =R i,t -R i,t-1 , where ΔR i,t is the change in responsibility of participant i at time t, R i,t is the responsibility value of participant i at time t, R i,t-1 is the responsibility value of participant i at the previous moment, When problems occur in the system, responsibilities are automatically assigned to different parties, and their responsibilities are quantified based on their actual participation and risk contribution in the decision-making path: Among them, A i is the final responsibility value assigned to the i-th participant, ΔR i,t is the change in responsibility of participant i at time t, P is the total number of all participants, and C is the overall cost or penalty of system errors.

Images