Training methods and systems to mitigate model bias caused by unmeasured confounding factors

Abstract

The present invention discloses a training method and system for alleviating model bias caused by unmeasurable confounding factors. The relevant method includes: combining user-item pair data, through sensitivity analysis, estimating the uncertainty set of sensitivity coefficients when unmeasurable confounding factors exist, establishing a first loss function in combination with the uncertainty set of sensitivity coefficients, inputting a data set consisting of multiple user-item pair data into a machine learning model to be trained, and optimizing the first loss function using adversarial training; or establishing a second loss function in combination with the uncertainty set of sensitivity coefficients and a pre-trained machine learning model, inputting a data set consisting of multiple user-item pair data into a machine learning model to be trained, and optimizing the second loss function using adversarial training. The solution provided by the present invention does not require human intervention and specific expert domain knowledge, and can effectively alleviate model bias caused by unmeasurable confounding factors.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a training method and system for alleviating model deviation caused by unmeasurable confounding factors. Background Art

[0002] In recent years, machine learning models, which provide intelligent services by fitting vast amounts of data, have been widely applied across various fields. However, due to the inherent presence of numerous confounders in observed data, the resulting models are susceptible to bias. Model bias refers to discrepancies between the output of a machine learning model and the designer's desired output. For example, recommendation models tend to repeatedly recommend repetitive, highly popular short videos. As the biased nature of machine learning models has gained increasing attention, a growing number of methods have been proposed to mitigate this bias.

[0003] Some studies have attempted to mitigate bias in machine learning models using structural causal modeling techniques. These methods require two key prerequisites: 1) leveraging the prior knowledge of domain experts to construct a causal graph describing the causal relationships between key factors in the data generation phase; 2) using domain experts to identify all confounding factors and measure their specific values. With the help of the causal graph and knowledge of confounding factors, bias correction training is used to mitigate the model's bias. Another group of studies has attempted to mitigate bias in machine learning models using methods based on the propensity score. These methods calculate the propensity score from the measured confounding factor values and then use this propensity score to construct a bias correction training method. This allows the model training process to mask the influence of reverberation factors, ultimately mitigating the model's bias.

[0004] However, potential confounding factors in data are ever-changing and often difficult to measure, making them difficult to meet the requirements of existing model bias mitigation methods. For example, the training data used by machine learning models for recommendation systems contains a variety of potential confounding factors, such as user income, social status, and product popularity, rendering existing methods ineffective. Summary of the Invention

[0005] The purpose of the present invention is to provide a training method and system for alleviating model bias caused by unmeasurable confounding factors. By estimating the uncertainty set of sensitivity coefficients through sensitivity analysis and performing model training, the robustness of the model to potential confounding factors is increased, thereby alleviating the bias characteristics of the machine learning model.

[0006] The purpose of the present invention is achieved through the following technical solutions:

[0007] A training method for alleviating model bias caused by unmeasured confounding factors, including:

[0008] Combined with the user-item pair data, through sensitivity analysis, the uncertainty set of the sensitivity coefficient in the presence of unmeasured confounding factors is estimated. The sensitivity coefficient is the inverse of the propensity coefficient.

[0009] For the loss function of the currently selected bias correction method based on the tendency coefficient, a first loss function is established in combination with an uncertain set of sensitivity coefficients, a data set consisting of multiple user-item pair data is input into the machine learning model to be trained, and the first loss function is optimized using adversarial training; or a second loss function is established on the basis of the first loss function in combination with a pre-trained machine learning model, a data set consisting of multiple user-item pair data is input into the machine learning model to be trained, and the second loss function is optimized using adversarial training.

[0010] A training system for mitigating model bias caused by unmeasured confounding factors, comprising:

[0011] A sensitivity analysis unit, configured to combine user-item pair data and, through sensitivity analysis, estimate an uncertainty set of sensitivity coefficients when unmeasured confounding factors exist, wherein the sensitivity coefficient is the inverse of the propensity coefficient;

[0012] A training unit is configured to establish a first loss function for the loss function of the currently selected bias correction method based on the propensity coefficient in combination with an uncertain set of sensitivity coefficients, input a data set consisting of multiple user-item pair data into the machine learning model to be trained, and optimize the first loss function using adversarial training; or establish a second loss function based on the first loss function in combination with a pre-trained machine learning model, input a data set consisting of multiple user-item pair data into the machine learning model to be trained, and optimize the second loss function using adversarial training.

[0013] A processing device comprising: one or more processors; a memory for storing one or more programs;

[0014] When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0015] A readable storage medium stores a computer program, which implements the aforementioned method when the computer program is executed by a processor.

[0016] It can be seen from the technical solution provided by the present invention that, under the premise that the number and values of confounding factors are unpredictable and no additional assumptions are made, the uncertainty set of sensitivity coefficients is estimated through sensitivity analysis, and the loss function is constructed in combination with the uncertainty set of sensitivity coefficients for model training. This can increase the robustness of the model to potential confounding factors, and through adversarial strategies, ensure that the machine learning model has good performance for any possible tendency coefficient in the uncertainty set, effectively alleviating the bias phenomenon of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0018] Figure 1 A flowchart of a training method for alleviating model bias caused by unmeasured confounding factors provided by an embodiment of the present invention;

[0019] Figure 2 A schematic diagram of a training system for alleviating model bias caused by unmeasured confounding factors provided by an embodiment of the present invention;

[0020] Figure 3 A schematic diagram of a processing device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] The following is a clear and complete description of the technical solutions in the embodiments of the present invention, in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0022] First, the following terms may be used in this article:

[0023] The terms "include," "comprises," "contains," "has," or other similar expressions should be interpreted as non-exclusive. For example, "including certain technical features (such as raw materials, components, ingredients, carriers, dosage forms, materials, dimensions, parts, components, mechanisms, devices, steps, procedures, methods, reaction conditions, processing conditions, parameters, algorithms, signals, data, products, or manufactured articles, etc.) should be interpreted as including not only the technical features explicitly listed, but also other technical features known in the art that are not explicitly listed.

[0024] The following is a detailed description of a training scheme for alleviating model bias caused by unmeasured confounding factors provided by the present invention. Any content not described in detail in the embodiments of the present invention belongs to the prior art known to professionals in the field. Where specific conditions are not specified in the embodiments of the present invention, the conditions are carried out according to conventional conditions in the art or the conditions recommended by the manufacturer. Where the manufacturer of the reagents or instruments used in the embodiments of the present invention is not specified, they are all conventional products that can be purchased commercially.

[0025] Example 1

[0026] The embodiment of the present invention provides a training method for alleviating the model deviation caused by unmeasured confounding factors, such as Figure 1 As shown, it mainly includes:

[0027] 1. Based on the user-item pair data, through sensitivity analysis, the uncertainty set of the sensitivity coefficient in the presence of unmeasured confounding factors is estimated. The sensitivity coefficient is the inverse of the propensity coefficient.

[0028] The preferred implementation of this step is as follows:

[0029] (1) Combine the user-item pair data and calculate the propensity coefficient when there are no unmeasured confounding factors.

[0030] (2) Through sensitivity analysis, a structural equation of the propensity coefficient in the presence of unmeasured confounding factors was constructed. Combined with the influence of unmeasured confounding factors on the propensity coefficient, the relationship between the propensity coefficient in the presence of unmeasured confounding factors and the propensity coefficient in the absence of unmeasured confounding factors was determined, and the uncertainty set of the sensitivity coefficient in the presence of unmeasured confounding factors was estimated.

[0031] In the embodiment of the present invention, a sensitivity analysis was performed to construct a structural equation for the propensity coefficient in the presence of unmeasured confounding factors, which is expressed as:

[0032]

[0033] in, represents the propensity coefficient in the presence of unmeasured confounding factors, m(·) refers to any type of function, is also a function of any type, the function m(·) is related to There is no constraint between them, they can be the same function or different functions; (u,i) represents a user-item pair data, u represents the user index, i represents the item index; x u,i Represents the characteristics of a user-item pair data, h u,i Features that represent unmeasurable confusion factors, such as the mood of the user when browsing the product corresponding to a user-item pair (u,i), the quality of the product in a specific environment, and other unmeasurable features. represents the influence of unmeasured confounding factors on the propensity coefficient.

[0034] Range constraints on the effect of a given unmeasured confounder on the propensity coefficient Determining the propensity coefficient in the presence of unmeasured confounders Compared with the propensity coefficient p when there is no unmeasured confounding factor u,i The relationship between them is expressed as:

[0035]

[0036] After simplification, we get:

[0037]

[0038] Where Γ is the set value,

[0039] Estimate the uncertainty set W of the sensitivity coefficient in the presence of unmeasured confounding factors: in, represents a single sensitivity coefficient, a u,i with b u,i Represents the lower and upper bounds of the uncertainty set of the sensitivity coefficients.

[0040] 2. For the loss function of the currently selected bias correction method based on the propensity coefficient, a first loss function is established in combination with an uncertain set of sensitivity coefficients, a data set consisting of multiple user-item pair data is input into the machine learning model to be trained, and the first loss function is optimized using adversarial training; or a second loss function is established based on the first loss function in combination with a pre-trained machine learning model, a data set consisting of multiple user-item pair data is input into the machine learning model to be trained, and the second loss function is optimized using adversarial training.

[0041] (1) For the loss function of the currently selected inverse bias correction method, the first loss function is established in combination with the uncertainty set of the sensitivity coefficient, which is expressed as:

[0042]

[0043] Among them, L RD-IPS (φ) is the first loss function, φ represents the parameters of the machine learning model to be trained, W represents the uncertainty set of sensitivity coefficients, represents a single sensitivity coefficient, |D| represents the number of user-item pairs in the dataset D, (u,i) represents the data for a user-item pair, u represents the user index, and i represents the item index; o u,i e u,irepresents the original loss function without bias correction, o u,i Indicates whether a user-item pair (u,i) is observed. If observed, the value is 1; if not observed, the value is 0. u,i It represents the difference between the model prediction result and the true label corresponding to a user-item pair (u,i).

[0044] The core idea here is to optimize the loss function L against the adversarial RD-IPs (φ), that is, by selecting the sensitivity coefficient in the uncertainty set of the sensitivity coefficient, the loss value of the first loss function continues to increase, and at the same time, the parameters of the machine learning model to be trained are optimized so that the loss value of the first loss function continues to decrease.

[0045] (2) For the loss function of the currently selected anti-propensity coefficient bias correction method, a second loss function is established by combining the uncertainty set of sensitivity coefficients with the pre-trained machine learning model, which is expressed as:

[0046]

[0047] Among them, L BRD-IPS (φ) represents the second loss function, φ represents the parameters of the machine learning model to be trained, W represents the uncertainty set of sensitivity coefficients, represents a single sensitivity coefficient, |D| represents the number of user-item pairs in the dataset D, (u,i) represents the data for a user-item pair, u represents the user index, and i represents the item index; Represents the parameters of the pre-trained machine learning model. u,i e u,i (φ) has the same meaning as the first type of loss function formula, indicating that when the parameter is φ, o u,i e u,i The value of This means that the model uses pre-trained machine learning model parameters The difference between the corresponding prediction result and the true label. u,i e u,i (φ) and The essential difference is that the latter value will not change with model optimization and acts as an "anchor point", while the former value will continue to change with model optimization.

[0048] Similarly, in this case, by selecting a sensitivity coefficient from the uncertainty set of the sensitivity coefficients, the loss value of the second loss function continues to increase, and at the same time, the parameters of the machine learning model to be trained are optimized so that the loss value of the second loss function continues to decrease.

[0049] In order to more clearly demonstrate the technical solution and technical effects provided by the present invention, the two core parts of the present invention are introduced in detail below.

[0050] 1. Sensitivity analysis framework.

[0051] Sensitivity analysis refers to analyzing how a target variable changes when a variable is perturbed, based on the interrelationships between variables. The sensitivity analysis method established in this proposal is primarily used to analyze the impact of potential unmeasured confounding factors on the propensity score, an important parameter used to mitigate model bias. The structural equation form of the propensity score, which is widely used in statistics, is:

[0052]

[0053] Among them, (u,i) represents the data for a user-item pair, u represents the user index, i represents the item index; p u,i represents the propensity coefficient of a user-item pair to the data, m(·) represents any type of function, which ensures the universality of the structural equation, x u,i is the input feature of the user-item pair used to calculate the propensity coefficient. When there are no unmeasured confounding factors in the data, the above formula can calculate an accurate propensity coefficient, thereby achieving accurate model bias correction. However, unmeasured confounding factors are ubiquitous, so after considering the influence of unmeasured confounding factors, the actual propensity coefficient structural equation should be:

[0054]

[0055] The main difference from the previous formula is that the term describing the effect of unmeasured confounding factors on the propensity coefficient is added Among them, h u,i is the characteristic of the unmeasured confounding factor, Represents any type of function, ensuring the universality of the calculation method. Since the unmeasurable confusion factor cannot be measured, its characteristic h cannot be directly obtained by any technology. u,i , leading to an accurate propensity coefficient It cannot be calculated directly, and only the inaccurate propensity coefficient p can be calculated. u,i .

[0056] The sensitivity analysis method established in the present invention adopts an indirect approach based on the inaccurate tendency coefficient p u,i Estimate the exact propensity coefficient Given the influence range of the unmeasured confounding factor, After that, we can get p u,i and The relationship between the two is:

[0057]

[0058] By further simplifying the above formula, we can get:

[0059]

[0060] in,

[0061] So far, this method has obtained an accurate propensity coefficient through the sensitivity analysis framework. Range The present invention refers to the range W as the uncertainty set of accurate sensitivity coefficients. Although the nature of unmeasured confounding factors makes it impossible to calculate a specific accurate propensity coefficient, the sensitivity analysis framework established by the present invention can accurately calculate the range of propensity coefficients.

[0062] 2. Adversarial training framework.

[0063] Among the existing model bias correction methods, the most widely used one is the bias correction method based on the propensity coefficient. Among these methods, the inverse propensity coefficient bias correction (IPS) method is the most classic. The loss function of the inverse propensity coefficient bias correction method is:

[0064]

[0065] Among them, φ represents the parameters of the machine learning model, |D| represents the size of the dataset, and o u,i e u,i Represents the original loss function without bias correction.

[0066] Due to the presence of unmeasured confounding factors in the data, p u,i The bias coefficient is inaccurate, which makes it impossible to effectively mitigate the model bias using the above method. Therefore, the present invention designs a new bias correction method based on adversarial training for the above traditional methods. RD-IPS (Robust Deobfuscation Technology - Anti-Bias Correction Method for Bias Coefficients) is established for the anti-bias correction method of bias coefficients:

[0067]

[0068] The core idea is to optimize the loss function L RD-IPS (φ), that is, by selecting This causes the loss value to continue to increase, while optimizing the model parameter φ to keep the loss value continuously decreasing. Ultimately, through the min-max (minimization-maximization) adversarial strategy, the machine learning model is guaranteed to have good performance for any possible tendency coefficient in the uncertainty set, effectively alleviating the model's bias.

[0069] At the same time, in order to deal with the risks of deploying correction methods that may exist in real scenarios, this method further establishes the BRD-IPS method (Robust Deobfuscation Technology with Anchor Points - Anti-Bias Correction Method) with "no side effects". The loss function of BRD-IPS is:

[0070]

[0071] The difference between it and RD-IPS is that it achieves the “no side effect” feature by introducing a benchmark model into the loss function. The benchmark model is a pre-trained common machine learning model with parameters of This method theoretically ensures that even in the worst case, BRD-IPS can better mitigate model bias than the original IPS method, ensuring that the training strategy established by this invention can be safely and harmlessly applied to various machine learning models.

[0072] For other bias correction methods, such as DR and AutoDebias, a similar method is used to construct the first loss function, that is, p in the original loss function u,i Replace with And according to Will use Bring in, thus Introduced into the first loss function. Similarly, based on the construction of the first loss function, the benchmark model is introduced to construct the second loss function.

[0073] In the embodiment of the present invention, the machine learning model can be a recommendation model of any structure, and therefore, no limitation is made to a specific model structure.

[0074] The above-mentioned solution of the embodiment of the present invention can be applied to the machine learning model used for recommendation of large e-commerce or short video platforms. On the basis of not changing the original recommendation model, it participates in model training as an additional module, so that the trained model has good robustness to the deviation caused by potential confounding factors, and achieves better recommendation effect when it goes online.

[0075] The above-mentioned scheme of the embodiment of the present invention mainly has the following advantages: a training framework for alleviating model bias phenomenon is established. The existing methods for alleviating model bias require a large amount of domain knowledge and expert intervention as prerequisites. The bias mitigation framework proposed by this method does not require human intervention and specific expert domain knowledge. While effectively alleviating the model bias caused by unmeasurable confusion factors, the proposed "side effect-free" adversarial training framework further ensures the universality of the method, so that the present invention can be widely used in various types of deployed machine learning models.

[0076] To demonstrate the performance of the above-mentioned solution, we used three of the most widely used machine learning models in recommendation fields—IPS, DR, and AutoDebias—as the basis. DR is a dual-robust bias correction method, and AutoDebias is an adaptive bias correction method. We also used two different datasets, Yahoo! R3 and Coat, to verify the effectiveness of the present invention in mitigating bias. The degree of bias mitigation was measured using the model's UAUC and NDCG@5 metrics on an unbiased test set. Higher values indicate greater bias mitigation.

[0077] Table 1: Performance comparison of different methods on two datasets

[0078]

[0079] In Table 1, RD-* indicates training using the first loss function established using the aforementioned solution of the present invention, while BRD-* indicates training using the second loss function established using the aforementioned solution of the present invention. The symbol * denotes IPS, DR, and AutoDebias. As can be seen from Table 1, both the RD and BRD frameworks of this method improve the performance of the three base models on two different datasets, fully validating the effectiveness of this invention.

[0080] Example 2

[0081] The present invention also provides a training system for alleviating model deviation caused by unmeasured confounding factors, which is mainly implemented based on the method provided in the above embodiment, such as Figure 2 As shown, the system mainly includes:

[0082] A sensitivity analysis unit, configured to combine user-item pair data and, through sensitivity analysis, estimate an uncertainty set of sensitivity coefficients when unmeasured confounding factors exist, wherein the sensitivity coefficient is the inverse of the propensity coefficient;

[0083] A training unit is configured to establish a first loss function for the loss function of the currently selected bias correction method based on the propensity coefficient in combination with an uncertain set of sensitivity coefficients, input a data set consisting of multiple user-item pair data into the machine learning model to be trained, and optimize the first loss function using adversarial training; or establish a second loss function based on the first loss function in combination with a pre-trained machine learning model, input a data set consisting of multiple user-item pair data into the machine learning model to be trained, and optimize the second loss function using adversarial training.

[0084] Those skilled in the art will clearly understand that for the convenience and brevity of description, only the division of the above-mentioned functional modules is used as an example. In actual applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above.

[0085] Example 3

[0086] The present invention also provides a processing device, such as Figure 3 As shown, it mainly includes: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method provided by the aforementioned embodiment.

[0087] Furthermore, the processing device further includes at least one input device and at least one output device; in the processing device, the processor, memory, input device, and output device are connected via a bus.

[0088] In the embodiment of the present invention, the specific types of the memory, input device, and output device are not limited; for example:

[0089] The input device can be a touch screen, image acquisition device, physical button or mouse;

[0090] The output device may be a display terminal;

[0091] The memory may be a random access memory (RAM) or a non-volatile memory, such as a disk memory.

[0092] Example 4

[0093] The present invention also provides a readable storage medium storing a computer program, which implements the method provided in the above embodiment when the computer program is executed by a processor.

[0094] In the embodiments of the present invention, the computer-readable storage medium may be provided in the aforementioned processing device, for example, as a memory in the processing device. Alternatively, the computer-readable storage medium may be a USB flash drive, a removable hard drive, a read-only memory (ROM), a magnetic disk, or an optical disk, among other media capable of storing program code.

[0095] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present invention should be included in the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be based on the scope of protection of the claims.

Claims

1. A training method for alleviating model bias caused by unmeasured confounding factors, characterized in that: include: Combined with the user-item pair data, through sensitivity analysis, the uncertainty set of the sensitivity coefficient in the presence of unmeasured confounding factors is estimated. The sensitivity coefficient is the inverse of the propensity coefficient. For the loss function of the currently selected bias correction method based on the propensity coefficient, a first loss function is established in combination with an uncertain set of sensitivity coefficients, a data set consisting of multiple user-item pairs is input into the machine learning model to be trained, and the first loss function is optimized using adversarial training; or a second loss function is established based on the first loss function and combined with a pre-trained machine learning model, a data set consisting of multiple user-item pairs is input into the machine learning model to be trained, and the second loss function is optimized using adversarial training; The step of estimating the uncertainty set of the sensitivity coefficient when the unmeasured confounding factor exists by combining the user-item pair data with the sensitivity analysis includes: calculating the propensity coefficient when the unmeasured confounding factor does not exist by combining the user-item pair data; constructing the propensity coefficient structural equation when the unmeasured confounding factor exists by the sensitivity analysis, determining the relationship between the propensity coefficient when the unmeasured confounding factor exists and the propensity coefficient when the unmeasured confounding factor does not exist by combining the influence of the unmeasured confounding factor on the propensity coefficient, and estimating the uncertainty set of the sensitivity coefficient when the unmeasured confounding factor exists; The sensitivity analysis is used to construct a structural equation for the propensity coefficient when unmeasured confounding factors are present, and the influence of unmeasured confounding factors on the propensity coefficient is combined to determine the relationship between the propensity coefficient when unmeasured confounding factors are present and the propensity coefficient when unmeasured confounding factors are not present. The uncertainty set of the sensitivity coefficient when unmeasured confounding factors are present is estimated to include: Through sensitivity analysis, the structural equation of the propensity coefficient in the presence of unmeasured confounding factors was constructed, which is expressed as: in, represents the propensity coefficient when there is an unmeasured confounding factor, m(·) represents any type of function, Represents a function of any type, (u,i) represents a user-item pair data, u represents the user index, i represents the item index; x u,i Represents the characteristics of a user-item pair data, h u,i represents the characteristics of the unmeasured confounding factor, represents the influence of unmeasured confounding factors on the propensity coefficient; Range constraints on the effect of a given unmeasured confounder on the propensity coefficient Determining the propensity coefficient in the presence of unmeasured confounders Compared with the propensity coefficient p when there is no unmeasured confounding factor u,i The relationship between them is expressed as: After simplification, we get: Where Γ is the set value, Estimate the uncertainty set W of the sensitivity coefficient in the presence of unmeasured confounding factors: in, represents a single sensitivity coefficient, a u,i with b u,i Represents the lower and upper bounds of the uncertainty set of the sensitivity coefficients.

2. A training method for alleviating model bias caused by unmeasured confounding factors according to claim 1, characterized in that: Optimizing the first loss function by using adversarial training, or optimizing the second loss function by using adversarial training includes: By selecting a sensitivity coefficient in the uncertainty set of the sensitivity coefficients, the loss value of the first loss function is continuously increased, and at the same time, the parameters of the machine learning model to be trained are optimized so that the loss value of the first loss function is continuously reduced; Alternatively, the loss value of the second loss function is continuously increased by selecting a sensitivity coefficient from the uncertain set of sensitivity coefficients, while optimizing the parameters of the machine learning model to be trained so that the loss value of the second loss function is continuously decreased.

3. A training method for alleviating model bias caused by unmeasured confounding factors according to claim 1, characterized in that: The loss function of the currently selected bias correction method based on the propensity coefficient is combined with the uncertainty set of the sensitivity coefficient to establish the first loss function, which includes: For the loss function of the currently selected inverse tendency coefficient deviation correction method, the first loss function is established in combination with the uncertainty set of the sensitivity coefficient and expressed as: Among them, L RD-IPS (φ) is the first loss function, φ represents the parameters of the machine learning model to be trained, W represents the uncertainty set of sensitivity coefficients, represents a single sensitivity coefficient, |D| represents the number of user-item pairs in the dataset D, (u,i) represents the data for a user-item pair, u represents the user index, and i represents the item index; o u,i Indicates whether the user-item pair data (u,i) is observed, with a value of 1 if observed and 0 if not; e u,i Represents the difference between the prediction result of the machine learning model to be trained and the true label corresponding to the user-item pair data (u,i).

4. A training method for alleviating model bias caused by unmeasured confounding factors according to claim 1 or 3, characterized in that: Establishing a second loss function based on the first loss function and combining it with a pre-trained machine learning model includes: For the loss function of the currently selected inverse bias correction method, the second loss function established is expressed as: Among them, L BRD-IPS (φ) represents the second loss function, φ represents the parameters of the machine learning model to be trained, W represents the uncertainty set of sensitivity coefficients, represents a single sensitivity coefficient, |D| represents the number of user-item pairs in the dataset D, (u,i) represents the data for a user-item pair, u represents the user index, and i represents the item index; Represents the parameters of a pre-trained machine learning model; o u,i Indicates whether the user-item pair data (u,i) is observed, with a value of 1 if observed and 0 if not; e u,i (φ) represents the difference between the prediction result of the machine learning model to be trained with the parameter φ corresponding to the user-item pair data (u,i) and the actual label. The parameters corresponding to the user-item pair data (u,i) are The difference between the prediction results of the pre-trained machine learning model and the true label.

5. A training system for alleviating model bias caused by unmeasured confounding factors, characterized in that: The method according to any one of claims 1 to 4 is implemented, and the system comprises: A sensitivity analysis unit, configured to combine user-item pair data and, through sensitivity analysis, estimate an uncertainty set of sensitivity coefficients when unmeasured confounding factors exist, wherein the sensitivity coefficient is the inverse of the propensity coefficient; A training unit is configured to establish a first loss function for the loss function of the currently selected bias correction method based on the propensity coefficient in combination with an uncertain set of sensitivity coefficients, input a data set consisting of multiple user-item pair data into the machine learning model to be trained, and optimize the first loss function using adversarial training; or establish a second loss function based on the first loss function in combination with a pre-trained machine learning model, input a data set consisting of multiple user-item pair data into the machine learning model to be trained, and optimize the second loss function using adversarial training.

6. A processing device, characterized in that include: one or more processors; a memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the method according to any one of claims 1 to 4.

7. A readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 4 is implemented.

Images