Fair improvement method and equipment for recommendation system click rate prediction model and medium

Abstract

The invention discloses a fairness improvement method and device for a click rate prediction model of a recommendation system and a medium, and the method comprises the following steps: S1, training the click rate prediction model of the recommendation system by using a sample data set, and extracting the representation information of the last linear layer of the trained click rate prediction model as the input of an agent model; s2, constructing an influence function, calculating an accuracy influence value and a fairness influence value, and constructing an influence value matrix; and S3, calculating a weighted value of each piece of sample data, and carrying out reweighting on the sample data to obtain a new sample data set. The invention further provides corresponding equipment and a storage medium, the problems that recommendation accuracy needs to be sacrificed for improving recommendation fairness in a click rate prediction model of an existing recommendation system, the data processing amount is large, the model training cost is high and the like are solved, feature representation of a final linear layer of an original model is extracted to serve as an agent of a sample, and the recommendation efficiency is improved. The recommendation fairness can be improved under the condition that the recommendation accuracy is not influenced.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a method, device and medium for improving the fairness of a click-through rate prediction model of a recommendation system. Background Art

[0002] In the application of information technology and artificial intelligence technology, the recommendation system is the key link between users and the information world. Especially in application scenarios such as e-commerce and online entertainment, it will directly affect the system's responsiveness and user experience, thereby affecting the economic benefits of the industry. In the recommendation system, click-through rate prediction is one of the important tasks of the recommendation system. Its accuracy is directly related to user experience and merchant interests, and the fairness of the recommendation mechanism is also a very important consideration indicator.

[0003] Traditional recommendation algorithms, such as collaborative filtering and content-based recommendation, have achieved certain results in improving the accuracy of recommendations, but often ignore the fairness of recommendations. The main risk point of these recommendation algorithm models is that they often have significant differences in the accuracy of algorithm recommendations for user groups with different wealth, so algorithm discrimination occurs. This unfair recommendation mechanism may not only aggravate market segmentation, but also cause greater social risks. Therefore, while ensuring high recommendation accuracy, it is also necessary to ensure the fairness of the recommendation system to avoid discriminatory effects on different user groups. The current fairness improvement method is usually to set additional constraints on the algorithm to ensure that different user groups are not affected by their sensitive attributes such as wealth, race, and gender when obtaining services or information, so as to achieve fairer results. However, such operations usually limit the freedom of the algorithm in optimizing individual-level predictions or recommendations, which may lead to a decline in overall accuracy. Therefore, how to ensure the fairness of recommendations while ensuring the accuracy of click-through rate prediction is an important research direction. Chinese patent announcement number CN116701950B, announcement date January 5, 2024, named "A method, device and medium for debiasing a click-through rate prediction model for a recommendation system" discloses a method for improving a click-through rate prediction model, including a deviation generation process and a root cause analysis step, a first-order weight reduction operation step or a first-order weight reconstruction operation step, to bring the model recommendation results closer to the user's real interests, and to improve the fairness of the item side of biased testing and the accuracy of unbiased test recommendations with a small sacrifice of accuracy. However, the debiasing method needs to take into account the impact of multiple layers of data on the recommendation accuracy and fairness, and needs to process a large amount of data. The traditional sample influence value estimation method and weight adjustment method need to store the gradient of each sample. In the actual application of click-through rate prediction, the high complexity of the model and the huge amount of data often make it difficult to directly save the sample gradient. Summary of the invention

[0004] The present invention overcomes the problems in the existing recommendation system click rate prediction model that improving recommendation fairness requires sacrificing recommendation accuracy, large data processing volume, high model training cost, etc., and provides a fairness improvement method, device and medium for the recommendation system click rate prediction model. The method extracts the feature representation of the last linear layer of the original model as a proxy for the sample, simplifies the data processing process of model training, saves computing resource costs and training time costs, and can improve recommendation fairness without affecting the recommendation accuracy.

[0005] In order to achieve the above object, the present invention adopts the following scheme:

[0006] The fairness improvement method of the click rate prediction model of the recommendation system includes the following steps:

[0007] S1: Use the sample data set to train a click-through rate prediction model of the recommendation system, and extract the representation information of the last linear layer of the trained click-through rate prediction model as the input of the proxy model, where the proxy model is a linear model;

[0008] S2: Construct an influence function and use it to calculate the accuracy impact value and fairness impact value of each sample data in the sample data set on the proxy model, and construct an influence value matrix using the accuracy impact value and fairness impact value;

[0009] S3: Calculate the weighted value of each sample data using the linear programming method according to the influence value matrix, and re-weight the sample data according to the weighted value to obtain a new sample data set.

[0010] Preferably, the influence function is constructed in step S2 as follows:

[0011] In the proxy model, given a training set It contains n pairs of input features x i and its corresponding label y i , and then a proxy classifier is given, the proxy classifier uses the loss function l to perform empirical risk minimization training, and the optimal parameters of the proxy classifier are saved as:

[0012]

[0013] When one of the samples x is down-scaled by a small margin ∈ j When the weights are becomes When the loss function l is strictly convex and quadratically differentiable, if the accuracy or fairness-based action function f is differentiable, then the sample x is adjusted downward j The influence function of the weight on fairness and accuracy is:

[0014]

[0015] in, By removing the sample x j Then retrain the model to get is the Hessian matrix of the loss function l.

[0016] Preferably, the method for calculating the accuracy impact value in step S2 is:

[0017] Function f is used as an accuracy-based action function, then sample x is removed from the training data set. i The impact function on accuracy is:

[0018]

[0019] Where V is the validation set.

[0020] Preferably, the fairness impact value is calculated in step S2 as follows:

[0021] According to the equal opportunity index, we construct a binary classification problem Y={0,1}, set the label to y and the prediction result to Sensitive attribute a∈{0,1}, the fairness formula is:

[0022]

[0023] Using 0 / 1 to represent zero-one loss, the quantitative index of fairness is expressed as:

[0024]

[0025] Using training loss instead of zero-one loss, the fairness-based action function f is:

[0026]

[0027] According to the action function f, the removed sample x is obtained i The fairness impact function of is:

[0028]

[0029] Where V is the validation set.

[0030] Preferably, the linear programming method used in step S3 is:

[0031] While minimizing fairness and maintaining accuracy, the reduction in the corresponding weight is controlled to the minimum. The formula is expressed as:

[0032]

[0033] where ωi is the weight, I util (e i ) and I fair (e i ) are the accuracy impact value and fairness impact value of the sample respectively.

[0034] Preferably, the linear programming method used in step S3 is:

[0035] Under the condition that the unfairness and accuracy loss are not greater than the set threshold, the reduction of the corresponding weight is controlled to the minimum, and the formula is expressed as:

[0036]

[0037] where ω i is the weight, I util (e i ) and I fair (e i ) are the accuracy impact value and fairness impact value of the sample, respectively, where parameters β and γ are used to adjust the weight.

[0038] Preferably, in step S3, while re-weighting the sample data according to the weighted value, a data enhancement operation is performed on the sample data to obtain a new sample data set, and the data enhancement operation includes:

[0039] Mark the accuracy impact value and fairness impact value of the sample data in the coordinate system of the impact value distribution diagram;

[0040] The data of the labeled points in the first quadrant of the influence value distribution map are enhanced by using mixed perturbation;

[0041] The data of the marked points exceeding the threshold in the third quadrant are deleted by threshold control.

[0042] Preferably, in step S1, when there are two or more linear layers at the end of the click-through rate prediction model, the results of the last two or more linear layers are added and channel-concatenated and output as representation information of the linear layers.

[0043] The present application also provides an electronic device, comprising: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor executes the method described in any one of claims 1 to 8.

[0044] The present application also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the method described in any one of claims 1 to 8 is implemented.

[0045] The present invention includes at least the following beneficial effects: (1) by accurately calculating the impact value of each training sample on the accuracy and fairness of the model, and adjusting the sample weights according to these impact values, the fairness of the recommendation system can be effectively improved without compromising the accuracy; (2) by extracting the feature representation of the last linear layer of the original model as a proxy for the sample, the data processing process of the model training is simplified, the simulation efficiency is improved, and the computing resource cost and training time cost are saved; (3) because the feature of the last linear layer of the original model is extracted as a proxy, there is no need to consider the differences in the structural design of the front and middle layers of different models and the differences in the feature processing methods, this method The method can be applied to various click-through rate prediction models, and has high flexibility and wide applicability; (4) It can clearly calculate the specific impact values ​​of accuracy and fairness in the model, providing an intuitive basis for operations such as reweighting, data enhancement or deletion, which is conducive to the further development of the model and the integration of other modules to achieve more functional expansion; (5) In addition to the basic weights, it also combines strategies such as data enhancement and data deletion to further enrich the training data and improve the generalization ability of the model. Data deletion is performed by deleting the third quadrant points through threshold control, which helps to remove samples that have a negative impact on model performance and further improve the use effect of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 A schematic diagram of a fairness improvement method of the present invention;

[0047] Figure 2 This is an example diagram of the accuracy and fairness impact value distribution of training sample data of the present invention. DETAILED DESCRIPTION

[0048] The present invention is further described in detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0049] like Figure 1 As shown, the fairness improvement method of the click rate prediction model of the recommendation system provided by the present invention includes the following steps:

[0050] S1: Use the sample data set to train the click-through rate prediction model of the recommendation system, and extract the representation information of the last linear layer of the trained click-through rate prediction model as the input of the proxy model, which is a linear model; the click-through rate model has a multi-layer feature extraction and processing structure, and the output result of its last linear layer is representative of the feature extraction of the input data of the entire model. By extracting the representation information of the last linear layer as the input of the proxy model, the use and processing flow of the data can be simplified. The proxy model is set according to the linear layer of the click-through rate prediction model. When simulating the original model, a simpler network architecture and parameter settings are used. Only the representation information of the last linear layer is used as input during the training process to simplify the computational complexity and improve the simulation efficiency.

[0051] S2: Construct an influence function and use it to calculate the accuracy impact value and fairness impact value of each sample data in the sample data set on the proxy model, and use the accuracy impact value and fairness impact value to construct an influence value matrix; the influence function represents the degree of influence of each sample data on the overall performance of the click-through rate model, and the influence indicators include the accuracy and fairness of the recommendation. By calculating the accuracy impact value and fairness impact value of each sample data, the influence value matrix can be constructed by combining the elements of the matrix.

[0052] S3: According to the influence value matrix, a linear programming method is used to calculate the weighted value of each sample data, and the sample data is re-weighted according to the weighted value to obtain a new sample data set. The click-through rate prediction model trained with the updated sample data can achieve improved fairness.

[0053] This method can effectively improve the fairness of the recommendation system without compromising the accuracy by accurately calculating the impact of each training sample on the accuracy and fairness of the model, and adjusting the sample weights based on these impact values. By extracting the feature representation of the last linear layer of the original model as a proxy for the sample, the data processing process of model training is simplified, the simulation efficiency is improved, and the computing resource cost and training time cost are saved. Since the features of the last linear layer of the original model are extracted as a proxy, there is no need to consider the differences in the structural design of the front and middle layers of different models and the differences in feature processing methods. This method can be applied to various click-through rate prediction models and has high flexibility and wide applicability.

[0054] In another technical solution, the influence function is constructed in step S2 as follows:

[0055] In the proxy model, given a training set It contains n pairs of input features x i and its corresponding label y i, and then a proxy classifier is given, the proxy classifier uses the loss function l to perform empirical risk minimization training, and the optimal parameters of the proxy classifier are saved as:

[0056]

[0057] When one of the samples x is down-scaled by a small margin ∈ j When the weights are becomes When the loss function l is strictly convex and quadratically differentiable, if the accuracy or fairness-based action function f is differentiable, then the sample x is adjusted downward j The influence function of the weight on fairness and accuracy is:

[0058]

[0059] in, By removing the sample x j Then retrain the model to get is the Hessian matrix of the loss function l.

[0060] The method for calculating the accuracy impact value in step S2 is:

[0061] Function f is used as an accuracy-based action function, then sample x is removed from the training data set. i The impact function on accuracy is:

[0062]

[0063] Where V is the validation set.

[0064] The method for calculating the fairness impact value in step S2 is:

[0065] According to the equal opportunity index, we construct a binary classification problem Y={0,1}, set the label to y and the prediction result to Sensitive attribute a∈{0,1}, considering group fairness, choose equal opportunity as the fairness indicator. The equal opportunity indicator focuses on the true positive rate between different groups. Different groups are divided based on some sensitive attributes, such as gender, race, etc., and the fairness formula is:

[0066]

[0067] Using 0 / 1 to represent zero-one loss, the quantitative index of fairness is expressed as:

[0068]

[0069] Using training loss instead of zero-one loss, the fairness-based action function f is:

[0070]

[0071] According to the action function f, the removed sample x is obtained i The fairness impact function of is:

[0072]

[0073] Where V is the validation set.

[0074] In another technical solution, the linear programming method used in step S3 is:

[0075] While minimizing fairness and maintaining accuracy, the reduction in the corresponding weight is controlled to the minimum. The formula is expressed as:

[0076]

[0077] where ω i is the weight, I util (e i ) and I fair (e i ) are the accuracy impact value and fairness impact value of the sample respectively.

[0078] As a comparative example, the comparison results of the model before and after training using this linear programming method are shown in Table 1 below:

[0079] Table 1

[0080]

[0081]

[0082] According to the data in Table 1, the smaller the fairness value is, the closer it is to a completely fair recommendation method. The fairness of the trained model is significantly better than that before training, with little impact on the accuracy.

[0083] The linear programming method used in step S3 is:

[0084] Under the condition that the unfairness and accuracy loss are not greater than the set threshold, the reduction of the corresponding weight is controlled to the minimum, and the formula is expressed as:

[0085]

[0086] where ω i is the weight, I util (e i ) and I fair (e i) are the accuracy impact value and fairness impact value of the sample, respectively, where parameters β and γ are used to adjust the weight.

[0087] As a comparative example, the comparison results of the model before and after training using this linear programming method are shown in Table 2 below:

[0088] Table 2

[0089]

[0090] From the data in Table 2, we can see that the fairness of the trained model is significantly better than that before training without reducing the accuracy.

[0091] Optional linear programming methods include at least:

[0092] Increase fairness as much as possible while ensuring that the accuracy does not decrease and the total weight reduction does not exceed the set threshold;

[0093] Try to simultaneously ensure minimal weight changes, minimal accuracy loss, and maximum fairness improvement as much as possible.

[0094] In another technical solution, Figure 2 As shown, in step S3, while re-weighting the sample data according to the weighted value, a data enhancement operation is performed on the sample data to obtain a new sample data set, and the data enhancement operation includes:

[0095] Mark the accuracy impact value and fairness impact value of the sample data in the coordinate system of the impact value distribution diagram;

[0096] The data of the labeled points in the first quadrant of the influence value distribution map are enhanced by using mixed perturbation;

[0097] The data of the marked points exceeding the threshold in the third quadrant are deleted by threshold control.

[0098] In step S1, when there are two or more linear layers at the end of the click rate prediction model, the results of the last two or more linear layers are added and channel-concatenated and output as representation information of the linear layers.

[0099] The present application also provides an electronic device, comprising: at least one processor, and a memory connected to the at least one processor in communication, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor performs a method as described in any one of the above schemes. The electronic device can be any terminal device including a mobile phone, a laptop computer, a desktop computer, a tablet computer, a PDA (Personal Digital Assistant), a POS (Point of Sales), a vehicle-mounted computer, etc.

[0100] The present application also provides a storage medium on which a computer program is stored, characterized in that when the program is executed by a processor, a method as described in any one of the above schemes is implemented.

[0101] Through the description of the above implementation mode, the technicians in the relevant field can clearly understand that the present invention can be implemented by means of software plus necessary general hardware, and of course, it can also be implemented by special hardware including special integrated circuits, special CPUs, special memories, special components, etc. In general, all functions completed by computer programs can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be various, such as analog circuits, digital circuits or special circuits. However, for the present invention, software program implementation is a better implementation mode in more cases. Based on such an understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a software product, which is stored in a readable storage medium, such as a computer floppy disk, a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a disk or an optical disk, etc., including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the present invention.

[0102] It should be noted that, although the above describes the various steps in a specific order, it does not mean that the various steps must be performed in the above specific order. In fact, some of these steps can be performed concurrently or even in a different order, as long as the required functions can be achieved. The number of devices and processing scales described here are used to simplify the description of the present invention, and the application, modification and variation of the present invention are obvious to those skilled in the art.

[0103] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation modes, and they can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. A method for improving the fairness of a click-through rate prediction model of a recommendation system, characterized in that: The following steps are involved: S1: Use the sample data set to train a click-through rate prediction model of the recommendation system, and extract the representation information of the last linear layer of the trained click-through rate prediction model as the input of the proxy model, where the proxy model is a linear model; S2: Construct an influence function and use it to calculate the accuracy impact value and fairness impact value of each sample data in the sample data set on the proxy model, and construct an influence value matrix using the accuracy impact value and fairness impact value; S3: Calculate the weighted value of each sample data using the linear programming method according to the influence value matrix, and re-weight the sample data according to the weighted value to obtain a new sample data set.

2. The method for improving fairness of the click rate prediction model of the recommendation system according to claim 1, characterized in that: The way to construct the influence function in step S2 is: In the proxy model, given a training set It contains n pairs of input features x i and its corresponding label y i , and then a proxy classifier is given, the proxy classifier uses the loss function l to perform empirical risk minimization training, and the optimal parameters of the proxy classifier are saved as: When one of the samples x is down-scaled by a small margin ∈ j When the weights are becomes When the loss function l is strictly convex and quadratically differentiable, if the accuracy or fairness-based action function f is differentiable, then the sample x is adjusted downward j The influence function of the weight on fairness and accuracy is: in, By removing the sample x j Then retrain the model to get is the Hessian matrix of the loss function l.

3. The method for improving fairness of the click rate prediction model of the recommendation system according to claim 2, characterized in that: The method for calculating the accuracy impact value in step S2 is: Function f is used as an accuracy-based action function, then sample x is removed from the training data set. i The impact function on accuracy is: Where V is the validation set.

4. The method for improving fairness of the click rate prediction model of the recommendation system according to claim 2, characterized in that: The method for calculating the fairness impact value in step S2 is: According to the equal opportunity index, we construct a binary classification problem Y = {0, 1), set the label to y and the prediction result to Sensitive attribute a∈{0,1}, the fairness formula is: Using 0 / 1 to represent zero-one loss, the quantitative index of fairness is expressed as: Using training loss instead of zero-one loss, the fairness-based action function f is: According to the action function f, the removed sample x is obtained i The fairness impact function of is: Where V is the validation set.

5. The method for improving fairness of the click rate prediction model of the recommendation system according to claim 1, characterized in that: The linear programming method used in step S3 is: While minimizing fairness and maintaining accuracy, the reduction in the corresponding weight is controlled to the minimum. The formula is expressed as: where ω i is the weight, I util (e i ) and I fair (e i ) are the accuracy impact value and fairness impact value of the sample respectively.

6. The method for improving the fairness of the click rate prediction model of the recommendation system according to claim 1, characterized in that: The linear programming method used in step S3 is: Under the condition that the unfairness and accuracy loss are not greater than the set threshold, the reduction of the corresponding weight is controlled to the minimum, and the formula is expressed as: where ω i is the weight, I util (e i ) and I fair (e i ) are the accuracy impact value and fairness impact value of the sample, respectively, where parameters β and γ are used to adjust the weight.

7. The method for improving fairness of the click rate prediction model of the recommendation system according to claim 1, characterized in that: In step S3, while re-weighting the sample data according to the weighted value, a data enhancement operation is performed on the sample data to obtain a new sample data set, and the data enhancement operation includes: Mark the accuracy impact value and fairness impact value of the sample data in the coordinate system of the impact value distribution diagram; The data of the labeled points in the first quadrant of the influence value distribution map are enhanced by using mixed perturbation; The data of the marked points exceeding the threshold in the third quadrant are deleted by threshold control.

8. The method for improving fairness of the click rate prediction model of the recommendation system according to claim 7, characterized in that: In step S1, when there are two or more linear layers at the end of the click rate prediction model, the results of the last two or more linear layers are added and channel-concatenated and output as representation information of the linear layers.

9. An electronic device, characterized in that include: At least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor performs the method described in any one of claims 1-8.

10. A storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the method described in any one of claims 1 to 8 is implemented.

Images