CN115345248A - Deep learning-oriented data depolarization method and device - Google Patents
Deep learning-oriented data depolarization method and device Download PDFInfo
- Publication number
- CN115345248A CN115345248A CN202211016753.2A CN202211016753A CN115345248A CN 115345248 A CN115345248 A CN 115345248A CN 202211016753 A CN202211016753 A CN 202211016753A CN 115345248 A CN115345248 A CN 115345248A
- Authority
- CN
- China
- Prior art keywords
- data set
- model
- sample
- sensitive
- training
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- General Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a data depolarization method and a data depolarization device for deep learning, which comprises the steps of firstly selecting an original data set, and extracting class labels and sensitive attribute labels in the original data set; structure discrimination model M 1 Inputting all samples in the original data set into the discriminant model M 1 Searching for a sensitive sample, screening and removing the sensitive sample, and expanding a data set by using a SHAP interpreter to obtain an unbiased data set; construction of a prediction model M 2 Inputting the unbiased data set into the prediction model M 2 Training is carried out; for the prediction model M obtained by training 2 Testing, if the fairness evaluation index chance is equal, considering that the model is fair after training, and finishing the deskewing; and if the fairness evaluation index is not met, the fairness evaluation index is reached.
Description
Technical Field
The invention relates to the field of depolarization, in particular to a data depolarization method and device for deep learning.
Background
The application of deep learning in many industries raises a number of ethical and legal issues, including fairness and prejudice of prediction. Deep learning is not critical to produce unfair decisions, as people believe that the system can help or decide loan approval, criminal decisions, and even healthcare. Deep learning utilizes data to build models that can evaluate tags and attributes of new data. However, available training data often contains deviations about what researchers are unwilling to use for decision making. Deep learning builds models that rely on training data and may result in permanence of these decision biases.
Research shows that the bias of the data set is an important reason for the bias of deep learning decision, and when the classification model is trained on unbalanced data, the trained model shows the bias of a specific population. Most of the existing data depolarization methods achieve a depolarization effect by modifying a data set label, so that some information of a data set is lost, and the model prediction precision is greatly reduced. Therefore, the invention designs a data depolarization method and device facing deep learning, which can ensure prediction accuracy while eliminating bias.
Disclosure of Invention
The invention aims to provide a data depolarization method and device for deep learning, aiming at the defects of the prior art. Sensitive samples in the data set are rebalanced, and then the model is trained, so that the depolarization effect is achieved.
The technical scheme adopted by the invention for realizing the aim is as follows: the first aspect of the embodiment of the present invention provides a data depolarization method for deep learning, which specifically includes the following substeps:
(1) Selecting an original data set, and extracting category labels and sensitive attribute labels in the original data set;
(2) Structure discrimination model M 1 Inputting all samples in the original data set into a discriminant model M 1 Searching for a sensitive sample, screening and removing the sensitive sample, and expanding a data set by using a SHAP interpreter to obtain an unbiased data set;
all samples in the original data set are input into a discrimination model M 1 The process of searching for the sensitive sample specifically comprises the following steps: from the original data setRandomly taking out a sample x, constructing a corresponding sensitive attribute detection sample x ', wherein the two samples x and x' meet the condition that the sensitive attributes are different S x ≠S x′ And other attributes are the same as Z x ＝Z x′ Then the model M is discriminated 1 The difference does not exist between the discrimination results of x and x', and if the difference exists, x is considered as a sensitive sample;
(3) Constructing a prediction model M 2 Inputting the unbiased data set obtained in the step (3) into a prediction model M 2 Training is carried out;
(4) The prediction model M obtained by the training in the step (3) is subjected to 2 Testing is carried out, if the fairness evaluation index chance is equal, the prediction model M is considered 2 After training, the device reaches fairness and completes deviation removal; and (4) if the fairness evaluation index is not met, repeating the step (3) until the fairness evaluation index is reached.
Further, the original data set is a COMPAS data set.
Further, the process of expanding the data set by using the SHAP interpreter is specifically as follows:
measuring the importance of each feature vector to the model prediction result by using SHAP; suppose the ith sample is x i The jth feature of the ith sample is x i_j Discrimination model M 1 The predicted value for this sample is y i Whole discrimination model M 1 Is y _ base, then the SHAP value obeys the following equation:
y i ＝y base +f(x i1 )+f(x i2 )+…+f(x ik )
wherein f (x) i ) Is x i The SHAP value of (1); f (x) i1 ) Is the contribution of the 1 st feature in the ith sample to the final predicted value yi when f (x) i1 )>0, the characteristic is shown to improve the predicted value; otherwise, the feature is shown to reduce the predicted value;
for f (x) ik ) And sequencing, setting a ranking threshold value in a self-defined manner, selecting a characteristic expansion sample larger than the ranking threshold value, expanding the data set without the sensitive sample into the size of the original data set, and obtaining an unbiased data set.
Further, the discriminant model M 1 Is composed of a fully connected neural network; the prediction model M 2 The neural network consists of 6 layers of fully-connected neural networks, and the number of neurons in each layer is 64, 32, 16, 8, 4 and 2 respectively.
Further, the model M in the step (3) 2 The training process specifically comprises the following steps: the full connection layer uses SoftMax as an activation function to classify and identify data, loss functions in the training process all use cross entropy loss functions, and an optimizer uses Adam to optimize; the loss function is formulated as follows:
L 1 ＝-[y·log(p)+(1-y)·log(1-p)]
wherein y represents the label of the specimen; p represents the probability that the prediction result is 1.
Further, the fairness assessment index in step (4) is formulated as follows:
wherein the content of the first and second substances,
representing model prediction, and S is a sensitive attribute.
A second aspect of the embodiments of the present invention provides a deep learning oriented data depolarization device, including one or more processors, configured to implement the above deep learning oriented data depolarization method.
A third aspect of the embodiments of the present invention provides a computer-readable storage medium, on which a program is stored, which, when being executed by a processor, is configured to implement the deep learning oriented data depolarization method described above.
The invention has the beneficial effects that: the invention discloses a data depolarization method and device for deep learning, which are used for obtaining an unbiased data set by screening and removing sensitive samples; and training the prediction model by utilizing an unbiased data set, so that the prediction model meets fairness evaluation indexes, and the unbiased effect is achieved.
Drawings
In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only some embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to these drawings without creative efforts.
FIG. 1 is a block diagram of a model in an embodiment of the invention;
FIG. 2 is a flow chart of a rethread prediction oriented depolarization method provided by an embodiment of the invention;
fig. 3 is a schematic diagram of a rethreading prediction oriented depolarization device provided by an embodiment of the invention.
Detailed Description
Reference will now be made in detail to the exemplary embodiments, examples of which are illustrated in the accompanying drawings. When the following description refers to the accompanying drawings, like numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatus and methods consistent with certain aspects of the invention, as detailed in the appended claims.
The terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. As used in this specification and the appended claims, the singular forms "a", "an", and "the" are intended to include the plural forms as well, unless the context clearly indicates otherwise. It should also be understood that the term "and/or" as used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.
It is to be understood that although the terms first, second, third, etc. may be used herein to describe various information, these information should not be limited to these terms. These terms are only used to distinguish one type of information from another. For example, first information may also be referred to as second information, and similarly, second information may also be referred to as first information, without departing from the scope of the present invention. The word "if" as used herein may be interpreted as "at" \8230; "or" when 8230; \8230; "or" in response to a determination ", depending on the context.
The following detailed description of embodiments of the invention is provided in connection with the accompanying drawings. The features of the following examples and embodiments may be combined with each other without conflict.
Referring to fig. 1 to 2, the present invention provides a rethrough prediction oriented depolarization method, comprising the following steps:
(1) Selecting an original data set, extracting category labels and sensitive attribute labels in the original data set
In the embodiment of the invention, a COMPAS data set, an Adult data set or a Bank data set is selected as an original data set. In the embodiment of the invention, the ratio of 5: a ratio of 1 divides the training set and the test set.
The process of extracting the category label and the sensitive attribute label from the training set specifically comprises the following steps: each sample in the COMPAS dataset has 18 category attributes, among which are multiple potentially sensitive attributes such as age, gender, etc. Particularly, one-hot encoding is required to be performed on the non-binary data categories in the training set, and the corresponding category labels are obtained.
(2) Structure discrimination model M 1 Inputting all samples in the original data set into a discriminant model M 1 Searching for a sensitive sample, screening and removing the sensitive sample, and expanding a data set by using a SHAP interpreter to obtain an unbiased data set;
(2.1) construction of the discrimination model M 1 The method specifically comprises the following steps: the discriminant model M in the embodiment of the invention 1 Is composed of a fully connected neural network.
(2.2) inputting all samples in the original data set into a discriminant model M 1 The process of searching for the sensitive sample specifically comprises the following steps: randomly taking out a sample x from the original data set, constructing a corresponding sensitive attribute detection sample x ', wherein the two samples x and x' meet the condition that the sensitive attributes are different S x ≠S x′ And other attributes are the same as Z x ＝Z x′ Then the model M is discriminated 1 And the discrimination results of x and x' should not be different, and if the difference exists, x is considered as a sensitive sample.
<xnotran> , x = [1,0,1,0,0,0,1,0,0,0,1,0,1,1,0,0,0,0], 9 "", 0, "", 9 , 1, x' = [1,0,1,0,0,0,1,0,1,0,1,0,1,1,0,0,0,0]. </xnotran>
(2.3) screening for desensitized samples
If x is a sensitive sample, deleting x from the training set; if x is not a sensitive sample, x is retained. Repeating the above operations, traversing all samples in the COMPAS data set to obtain a data set X without sensitive samples 1 。
(2.4) augmenting the data set with a SHAP interpreter, specifically:
after screening the data set by the method described in step (2.2), the remaining data set X is 1 The data set does not contain sensitive samples and is a fair data set, but the data set loses a large number of samples, the information contained in the samples can greatly contribute to the prediction task, and the data set X is used for predicting the data set 1 The trained model may have low classification accuracy, and therefore the data set needs to be expanded.
The importance of each feature vector to the model prediction results is first measured using SHAP. Suppose the ith sample is x i The jth feature of the ith sample is x i_j Discrimination model M 1 The predicted value for this sample is y i Whole discrimination model M 1 Is y _ base, then the SHAP value obeys the following equation:
y i ＝y base +f(x i1 )+f(x i2 )+…+f(x ik )
wherein f (x) i ) Is x i The SHAP value of (1). Viewed directly, f (x) i1 ) That is, the 1 st feature pair in the ith sample is the most significantThe contribution of the final predicted value yi when f (x) i1 )>0, the characteristic is shown to improve the predicted value and also acts positively; conversely, the feature is shown to reduce the predicted value, and has adverse effect.
Finding out the characteristics greatly contributing to the model prediction result through SHAP, and determining the f (x) ik ) And (4) sequencing, then expanding the sample based on the characteristics with the top rank, selecting the characteristics with the top 30% of the top rank to expand the sample in the embodiment of the invention, and expanding the data set without the sensitive sample in the step (2.3) into the original data set size, wherein the data set is an unbiased data set X'.
(3) Construction of a prediction model M 2 Inputting the unbiased data set obtained in the step (3) into a prediction model M 2 Training is carried out;
prediction model M 2 The neural network consists of 6 layers of fully-connected neural networks, and the number of neurons in each layer is 64, 32, 16, 8, 4 and 2. And the full connection layer uses SoftMax as an activation function to classify and identify the data, the loss functions in the training process all use cross entropy loss functions, and the optimizer uses Adam. Inputting the unbiased data set X' synthesized in the step (2) into a prediction model M 2 In the embodiment of the present invention, the size of the training batch is set to 100, a preheating learning rate strategy is adopted in the training stage, and Adam is adopted for optimization by the optimizer. The loss function takes the form of cross entropy, and the formula is as follows:
L 1 ＝-[y·log(p)+(1-y)·log(1-p)]
where y represents the cable of the sample. P represents the probability that the prediction result is 1.
(4) The prediction model M obtained by the training in the step (3) is subjected to 2 Testing, if the fairness evaluation index chance is equal, considering that the model is fair after training, and finishing the deskewing; and (4) if the fairness evaluation index is not met, repeating the step (3) until the fairness evaluation index is reached.
Inputting the test set divided in the step (1.2) into a trained prediction model M 2 When the test result of the test set meets the fairness evaluation index chance equality, the prediction model M is considered to be 2 After training, fairness is achieved.
The chance equality formula is:
wherein
And (3) representing model prediction, wherein S is a sensitive attribute, and when the results on the left side and the right side of the formula are equal or similar, the model is considered to achieve the depolarization effect.
Example 1
(1) And selecting the COMPAS data set as an original data set, and extracting category labels and sensitive attribute labels in the original data set.
In embodiment 1 of the present invention, the "sex" of the sample is used as the sensitive attribute, and the attribute is "male" or "female", and the binarization processing is performed on the attribute, so that "male" is encoded as 1 and "female" is encoded as 0.
(2) Structure discrimination model M 1 Inputting all samples in the original data set into a discriminant model M 1 Searching for a sensitive sample, screening and removing the sensitive sample, and expanding the data set by using a SHAP interpreter to obtain an unbiased data set.
(2.1) construction of the discrimination model M 1 The method specifically comprises the following steps: the discriminant model M in the embodiment of the invention 1 Is composed of a fully-connected neural network.
(2.2) input all samples in the original dataset into the discriminant model M 1 The process of searching for the sensitive sample specifically comprises the following steps: randomly taking out a sample x from the original data set, constructing a corresponding sensitive attribute detection sample x ', wherein the two samples x and x' meet the condition that the sensitive attributes are different S x ≠S x′ And other attributes are the same as Z x ＝Z x′ Then the discriminant model M 1 And (4) judging that the discrimination results of x and x' are different, and if the discrimination results are different, considering that x is a sensitive sample.
(2.3) screening for desensitized samples
If x is a sensitive sample, x is concentrated from the training setDeleting the line; if x is not a sensitive sample, x is retained. Repeating the above operations, traversing all samples in the COMPAS data set to obtain a data set X without sensitive samples 1 。
(2.4) augmenting the dataset with a SHAP interpreter
Finding features that contribute greatly to model prediction results by SHAP, pair f (x) ik ) And (4) sequencing, then expanding the sample based on the characteristics with the top rank, selecting the characteristics with the top 30% of the top rank to expand the sample in the embodiment of the invention, and expanding the data set without the sensitive sample in the step (2.3) into the original data set size, wherein the data set is an unbiased data set X'.
(3) Construction of a prediction model M 2 Inputting the unbiased data set obtained in the step (3) into a prediction model M 2 And (5) training.
(4) The prediction model M obtained by the training in the step (3) is subjected to 2 Testing, if the fairness evaluation index chance is equal, considering that the model is fair after training, and finishing the depolarization; and (4) if the fairness evaluation index is not met, repeating the step (3) until the fairness evaluation index is reached.
Inputting the test set divided in the step (1.2) into a trained prediction model M 2 The method comprises the steps of (1) testing, predicting whether an individual crimes again, wherein the output result is only 0 or 1, wherein 0 represents that crimes cannot be crimes again, and 1 represents that crimes can be crimes again. Calculating the fairness evaluation index through calculation, and when the test result of the test set meets the fairness evaluation index chance equality, considering the prediction model M 2 After training, fairness is achieved.
Example 2
(1) Selecting an Adult data set as an original data set, and extracting category labels and sensitive attribute labels in the original data set.
In embodiment 2 of the present invention, the "sex" of the sample is used as the sensitive attribute, and the attribute is "male" or "female", and the binarization processing is performed on the attribute, so that "male" is encoded as 1 and "female" is encoded as 0.
(2) Structure discrimination model M 1 Inputting all samples in the original data set into a discriminant modelType M 1 Searching for a sensitive sample, screening and removing the sensitive sample, and expanding the data set by using a SHAP interpreter to obtain an unbiased data set.
(2.1) construction of the discrimination model M 1 The method comprises the following specific steps: the discriminant model M in the embodiment of the invention 1 Is composed of a fully-connected neural network.
(2.2) inputting all samples in the original data set into a discriminant model M 1 The process of searching for the sensitive sample specifically comprises the following steps: randomly taking out a sample x from an original data set, constructing a sensitive attribute detection sample x ' corresponding to the sample x ', wherein the two samples x and x ' meet the condition that the sensitive attributes are different S x ≠S x′ And other attributes are the same as Z x ＝Z x′ Then the model M is discriminated 1 And the discrimination results of x and x' should not be different, and if the difference exists, x is considered as a sensitive sample.
(2.3) screening for desensitized samples
If x is a sensitive sample, deleting x from the training set; if x is not a sensitive sample, x is retained. Repeating the above operations, traversing all samples in the COMPAS data set to obtain a data set X without sensitive samples 1 。
(2.4) augmenting the dataset with a SHAP interpreter
Finding features that contribute greatly to model prediction results by SHAP, pair f (x) ik ) Sorting is carried out, then the samples are expanded on the basis of the characteristics with the top rank, the characteristics with the top 30% of the ranks are selected to expand the samples in the embodiment of the invention, and the data set after the sensitive samples are removed in the step (2.3) is expanded into the size of the original data set, wherein the data set is an unbiased data set X'.
(3) Constructing a prediction model M 2 Inputting the unbiased data set obtained in the step (3) into a prediction model M 2 And (5) training.
(4) For the prediction model M obtained by training in the step (3) 2 Testing, if the fairness evaluation index chance is equal, considering that the model is fair after training, and finishing the deskewing; and (4) if the fairness evaluation index is not met, repeating the step (3) until the fairness evaluation index is reached.
Inputting the test set divided in the step (1.2) into a trained prediction model M 2 When the test result of the test set meets the fairness evaluation index chance equality, the prediction model M is considered to be 2 After training, fairness is achieved.
Constructed prediction model M fair The output predicted value is a binary variable, and the prediction model M in embodiment 2 of the present invention fair The method is used for predicting the annual income of an individual, and the output result is only 0 or 1,0 which represents that the annual income is less than 50K, and 1 which represents that the annual income is more than 50K.
The formula for fairness assessment index chance equality is as follows:
wherein
And (3) representing a model prediction result, wherein S is a sensitive attribute, and when the results on the left side and the right side of the formula are equal or similar, the model is considered to achieve a depolarization effect.
Corresponding to the embodiment of the deep learning-oriented data depolarization method, the invention also provides an embodiment of a deep learning-oriented data depolarization device.
Referring to fig. 3, an apparatus for data depolarization facing deep learning according to an embodiment of the present invention includes one or more processors, configured to implement the method for data depolarization facing deep learning in the foregoing embodiment.
The embodiment of the data depolarization device oriented to deep learning of the invention can be applied to any equipment with data processing capability, such as computers and other equipment or devices. The device embodiments may be implemented by software, or by hardware, or by a combination of hardware and software. The software implementation is taken as an example, and as a logical device, the device is formed by reading corresponding computer program instructions in the nonvolatile memory into the memory for running through the processor of any device with data processing capability. In terms of hardware, as shown in fig. 3, a hardware structure diagram of any device with data processing capability where the deep learning-oriented data deskewing apparatus is located according to the present invention is a hardware structure diagram, where in addition to the processor, the memory, the network interface, and the nonvolatile memory shown in fig. 3, any device with data processing capability where the apparatus is located in the embodiment may generally include other hardware according to the actual function of the any device with data processing capability, and details thereof are not repeated.
The specific details of the implementation process of the functions and actions of each unit in the above device are the implementation processes of the corresponding steps in the above method, and are not described herein again.
For the device embodiment, since it basically corresponds to the method embodiment, reference may be made to the partial description of the method embodiment for relevant points. The above-described embodiments of the apparatus are merely illustrative, and the units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of the present invention. One of ordinary skill in the art can understand and implement without inventive effort.
An embodiment of the present invention further provides a computer-readable storage medium, on which a program is stored, where the program, when executed by a processor, implements the deep learning oriented data depolarization method in the foregoing embodiment.
The computer readable storage medium may be an internal storage unit, such as a hard disk or a memory, of any data processing device described in any previous embodiment. The computer readable storage medium can be any device with data processing capability, such as a plug-in hard disk, a Smart Media Card (SMC), an SD Card, a Flash memory Card (Flash Card), etc. provided on the device. Further, the computer readable storage medium may include both an internal storage unit and an external storage device of any data processing capable device. The computer-readable storage medium is used for storing the computer program and other programs and data required by the arbitrary data processing-capable device, and may also be used for temporarily storing data that has been output or is to be output.
The embodiments described in this specification are merely illustrative of implementations of the inventive concept and the scope of the present invention should not be considered limited to the specific forms set forth in the embodiments but rather by the equivalents thereof as may occur to those skilled in the art upon consideration of the present inventive concept.
Claims (8)
1. A deep learning-oriented data depolarization method is characterized by specifically comprising the following substeps:
(1) Selecting an original data set, and extracting category labels and sensitive attribute labels in the original data set;
(2) Structure discrimination model M 1 Inputting all samples in the original data set into a discriminant model M 1 Searching for a sensitive sample, screening and removing the sensitive sample, and expanding a data set by using a SHAP interpreter to obtain an unbiased data set;
all samples in the original data set are input into a discrimination model M 1 The process of searching for the sensitive sample specifically comprises the following steps: randomly taking out a sample x from the original data set, constructing a corresponding sensitive attribute detection sample x ', wherein the two samples x and x' meet the condition that the sensitive attributes are different S x ≠S x′ And other attributes are the same as Z x ＝Z x′ Then the model M is discriminated 1 The discrimination results of x and x 'should not be different, if the discrimination results of x and x' are different, x is considered as a sensitive sample;
(3) Construction of a prediction model M 2 Inputting the unbiased data set obtained in the step (3) into a prediction model M 2 Training is carried out;
(4) For the prediction model M obtained by training in the step (3) 2 Testing is carried out, if the fairness evaluation index chance is met, the prediction model M is considered to be equal 2 After training, the health care product is obtainedCompleting the depolarization when the balance is fair; and (4) if the fairness evaluation index is not met, repeating the step (3) until the fairness evaluation index is reached.
2. The method of claim 1, wherein the raw data set is a COMPAS data set.
3. The method of claim 1, wherein augmenting the data set using the SHAP interpreter is by:
measuring the importance of each feature vector to the model prediction result by using SHAP; suppose the ith sample is x i The jth feature of the ith sample is x i_j Discrimination model M 1 The predicted value for this sample is y i Whole discrimination model M 1 Is y _ base, then the SHAP value obeys the following equation:
y i ＝y base +f(x i1 )+f(x i2 )+…+f(x ik )
wherein f (x) i ) Is x i The SHAP value of (1); f (x) i1 ) Is the contribution of the 1 st feature in the ith sample to the final predicted value yi when f (x) i1 )>0, the characteristic is shown to improve the predicted value; otherwise, the feature is shown to reduce the predicted value;
for f (x) ik ) And sequencing, setting a ranking threshold value in a self-defined manner, selecting a characteristic expansion sample larger than the ranking threshold value, expanding the data set without the sensitive sample into the size of the original data set, and obtaining an unbiased data set.
4. Method according to claim 1, characterized in that the discriminant model M is 1 Is composed of a fully connected neural network; the prediction model M 2 The neural network consists of 6 layers of fully-connected neural networks, and the number of neurons in each layer is 64, 32, 16, 8, 4 and 2.
5. The method of claim 1, wherein the model M in step (3) 2 The training process specifically comprises the following steps: is totally connected withThe interface uses SoftMax as an activation function to classify and identify data, loss functions in the training process all use cross entropy loss functions, and an optimizer uses Adam to optimize; the loss function is formulated as follows:
L 1 ＝-[y·log(p)+(1-y)·log(1-p)]
wherein y represents the label of the sample; p represents the probability that the prediction result is 1.
6. The method of claim 1, wherein the fairness assessment indicator in step (4) is formulated as follows:
wherein the content of the first and second substances,
representing model prediction, and S is a sensitive attribute.
7. A deep learning oriented data depolarization apparatus comprising one or more processors configured to implement the deep learning oriented data depolarization method of any one of claims 1 to 6.
8. A computer-readable storage medium, on which a program is stored, which, when being executed by a processor, is adapted to carry out the deep learning oriented data depolarization method of any one of claims 1 to 6.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211016753.2A CN115345248A (en) | 2022-08-24 | 2022-08-24 | Deep learning-oriented data depolarization method and device |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211016753.2A CN115345248A (en) | 2022-08-24 | 2022-08-24 | Deep learning-oriented data depolarization method and device |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115345248A true CN115345248A (en) | 2022-11-15 |
Family
ID=83953162
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211016753.2A Pending CN115345248A (en) | 2022-08-24 | 2022-08-24 | Deep learning-oriented data depolarization method and device |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115345248A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117315379A (en) * | 2023-11-29 | 2023-12-29 | 中电科大数据研究院有限公司 | Deep learning-oriented medical image classification model fairness evaluation method and device |
-
2022
- 2022-08-24 CN CN202211016753.2A patent/CN115345248A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117315379A (en) * | 2023-11-29 | 2023-12-29 | 中电科大数据研究院有限公司 | Deep learning-oriented medical image classification model fairness evaluation method and device |
| CN117315379B (en) * | 2023-11-29 | 2024-03-12 | 中电科大数据研究院有限公司 | Deep learning-oriented medical image classification model fairness evaluation method and device |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US10303737B2 (en) | Data analysis computer system and method for fast discovery of multiple Markov boundaries | |
| Charpentier et al. | Econometrics and machine learning | |
| CN110852881B (en) | Risk account identification method and device, electronic equipment and medium | |
| EP4322031A1 (en) | Recommendation method, recommendation model training method, and related product | |
| CN113887580B (en) | Contrast type open set image recognition method and device considering multi-granularity correlation | |
| Mohammadi et al. | Improving linear discriminant analysis with artificial immune system-based evolutionary algorithms | |
| CN110222733B (en) | High-precision multi-order neural network classification method and system | |
| CN112016313A (en) | Spoken language element identification method and device and alarm situation analysis system | |
| CN115688024A (en) | Network abnormal user prediction method based on user content characteristics and behavior characteristics | |
| CN112749737A (en) | Image classification method and device, electronic equipment and storage medium | |
| CN116610831A (en) | Semanteme subdivision and modal alignment reasoning learning cross-modal retrieval method and retrieval system | |
| CN115345248A (en) | Deep learning-oriented data depolarization method and device | |
| CN113837836A (en) | Model recommendation method, device, equipment and storage medium | |
| CN113298254A (en) | Deskewing method and device for deep migration learning | |
| CN114119191A (en) | Wind control method, overdue prediction method, model training method and related equipment | |
| CN115329872A (en) | Sensitive attribute identification method and device based on comparison learning | |
| CN115310606A (en) | Deep learning model depolarization method and device based on data set sensitive attribute reconstruction | |
| CN115344794A (en) | Scenic spot recommendation method based on knowledge map semantic embedding | |
| CN111400413B (en) | Method and system for determining category of knowledge points in knowledge base | |
| CN114170000A (en) | Credit card user risk category identification method, device, computer equipment and medium | |
| CN110609961A (en) | Collaborative filtering recommendation method based on word embedding | |
| CN111291829A (en) | Automatic determination method and system for selected pictures | |
| CN113704409B (en) | False recruitment information detection method based on cascading forests | |
| CN115329883B (en) | Semantic similarity processing method, device and system and storage medium | |
| WO2012029152A2 (en) | Device and method for calculation of distance between datasets |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |