Model training method and device, effect prediction method and device, medium and equipment

Abstract

The embodiment of the invention discloses a model training method and device, an effect prediction method and device, a medium and equipment. The scheme can be executed in a trusted execution environment. In the model training stage, user features mapped by user information of different user groups (receiving different processing measures) are used in a cross mode, and effect prediction models associated with different processing measures are trained. Specifically, when an effect prediction model associated with a first processing measure is trained, not only features mapped by user information of a first group of users receiving the first processing measure and labels thereof are adopted, but also features mapped by user information of a second group of users receiving a second processing measure are adopted; when the effect prediction model associated with the second processing measure is trained, not only the features mapped by the user information of the second group of users accepting the second processing measure and the labels thereof are adopted, but also the features mapped by the user information of the first group of users accepting the first processing measure are adopted.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a model training and effect prediction method, device, medium and equipment. Background Art

[0002] The Conditional Average Treatment Effect (CATE) is a key metric in the field of causal inference, designed to measure the heterogeneous effects of a treatment on different user groups. For example, in healthcare, effective CATE estimation for new drugs or treatments can help us tailor treatment plans to different groups, thereby achieving precision medicine. Similarly, in advertising recommendations, making accurate recommendations for different groups can effectively save costs and enhance advertising effectiveness, which also relies on accurately estimating CATE.

[0003] In practical applications, the CATE estimation problem is transformed into two independent effect prediction problems. Effect prediction models corresponding to two different treatment measures are trained, and then CATE estimation is performed based on the difference in treatment effects predicted by these two effect prediction models.

[0004] The present disclosure provides a more effective effect prediction model training solution. Summary of the Invention

[0005] This embodiment of the present invention provides a model training method, including:

[0006] Obtain user information x1 of a first group of users and representation y1 of the actual effect of a first treatment measure on the first group of users; and obtain user information x2 of a second group of users and representation y2 of the actual effect of a second treatment measure on the second group of users;

[0007] Based on the feature mapping model Φ, map x1 to Φ(x1) and map x2 to Φ(x2);

[0008] Input Φ(x1) into the effect prediction model H1 associated with the first treatment measure, output H1(Φ(x1)), input Φ(x2) into H1, output H1(Φ(x2)); and input Φ(x2) into the effect prediction model H2 associated with the second treatment measure, output H2(Φ(x2)), input Φ(x1) into H2, output H2(Φ(x1));

[0009] With the reduction of the objective function as the training goal, the parameters of Φ, H1, and H2 are adjusted;

[0010] Where P represents the probability operation, and the value of the objective function is positively correlated to: the gap between H1(Φ(x1)) and y1, the gap between H2(Φ(x2)) and y2, the gap between P(H1(Φ(x2))|Φ(x2)) and P(y1|Φ(x1)), and the gap between P(H2(Φ(x1))|Φ(x1)) and P(y2|Φ(x2)).

[0011] The embodiments of this specification provide a model training method, including:

[0012] Obtain user information x1 of a first group of users and representation y1 of the actual effect of a first treatment measure on the first group of users; and obtain user information x2 of a second group of users and representation y2 of the actual effect of a second treatment measure on the second group of users;

[0013] Based on the feature mapping model Φ, map x1 to Φ(x1) and map x2 to Φ(x2);

[0014] Input Φ(x1) into the effect prediction model H1 associated with the first treatment measure, output H1(Φ(x1)), input Φ(x2) into H1, output H1(Φ(x2)); and input Φ(x2) into the effect prediction model H2 associated with the second treatment measure, output H2(Φ(x2)), input Φ(x1) into H2, output H2(Φ(x1));

[0015] With the reduction of the objective function as the training goal, the parameters of Φ, H1, and H2 are adjusted;

[0016] Where P represents the probability operation, and the value of the objective function is positively correlated to: the gap between H1(Φ(x1)) and y1, the gap between H2(Φ(x2)) and y2, the gap between P(H1(Φ(x2))|Φ(x2)) and P(H1(Φ(x1))|Φ(x1)), and the gap between P(H2(Φ(x1))|Φ(x1)) and P(H2(Φ(x2))|Φ(x2)).

[0017] The present invention provides an effect prediction method, including:

[0018] Get the target user's user information x0;

[0019] Based on the feature mapping model Φ, map x0 to Φ(x0);

[0020] Input Φ(x0) into the effect prediction model H1, and output H1(Φ(x0)), which is used to represent the predicted effect representation after the target user accepts the first treatment measure;

[0021] And, Φ(x0) is input into the effect prediction model H2, and H2(Φ(x0)) is output, which is used to represent the predicted effect representation after the target user accepts the second treatment measure.

[0022] The present invention also provides a model training device, including:

[0023] an acquisition module, which acquires user information x1 of a first group of users and a representation y1 of an actual effect of a first treatment measure on the first group of users; and acquires user information x2 of a second group of users and a representation y2 of an actual effect of a second treatment measure on the second group of users;

[0024] The mapping module maps x1 to Φ(x1) and x2 to Φ(x2) based on the feature mapping model Φ.

[0025] A prediction module, inputting Φ(x1) into an effect prediction model H1 associated with a first treatment measure and outputting H1(Φ(x1)), inputting Φ(x2) into H1 and outputting H1(Φ(x2)); and inputting Φ(x2) into an effect prediction model H2 associated with a second treatment measure and outputting H2(Φ(x2)), inputting Φ(x1) into H2 and outputting H2(Φ(x1));

[0026] The adjustment module adjusts the parameters of Φ, H1, and H2 with the goal of reducing the value of the objective function as the training goal;

[0027] Where P represents the probability operation, and the value of the objective function is positively correlated to: the gap between H1(Φ(x1)) and y1, the gap between H2(Φ(x2)) and y2, the gap between P(H1(Φ(x2))|Φ(x2)) and P(y1|Φ(x1)), and the gap between P(H2(Φ(x1))|Φ(x1)) and P(y2|Φ(x2)).

[0028] The embodiment of this specification provides a model training device, including:

[0029] an acquisition module, which acquires user information x1 of a first group of users and a representation y1 of an actual effect of a first treatment measure on the first group of users; and acquires user information x2 of a second group of users and a representation y2 of an actual effect of a second treatment measure on the second group of users;

[0030] The mapping module maps x1 to Φ(x1) and x2 to Φ(x2) based on the feature mapping model Φ.

[0031] A prediction module, inputting Φ(x1) into an effect prediction model H1 associated with a first treatment measure and outputting H1(Φ(x1)), inputting Φ(x2) into H1 and outputting H1(Φ(x2)); and inputting Φ(x2) into an effect prediction model H2 associated with a second treatment measure and outputting H2(Φ(x2)), inputting Φ(x1) into H2 and outputting H2(Φ(x1));

[0032] The adjustment module adjusts the parameters of Φ, H1, and H2 with the goal of reducing the value of the objective function as the training goal;

[0033] Where P represents the probability operation, and the value of the objective function is positively correlated to: the gap between H1(Φ(x1)) and y1, the gap between H2(Φ(x2)) and y2, the gap between P(H1(Φ(x2))|Φ(x2)) and P(H1(Φ(x1))|Φ(x1)), and the gap between P(H2(Φ(x1))|Φ(x1)) and P(H2(Φ(x2))|Φ(x2)).

[0034] The embodiment of this specification also provides an effect prediction device, comprising:

[0035] Acquisition module, obtains the user information x0 of the target user;

[0036] The mapping module maps x0 to Φ(x0) based on the feature mapping model Φ;

[0037] The prediction module inputs Φ(x0) into the effect prediction model H1 and outputs H1(Φ(x0)), which is used to represent the predicted effect representation after the target user accepts the first treatment measure; and inputs Φ(x0) into the effect prediction model H2 and outputs H2(Φ(x0)), which is used to represent the predicted effect representation after the target user accepts the second treatment measure.

[0038] An embodiment of this specification also provides a computer program product, which stores at least one instruction, and the at least one instruction is suitable for being loaded by a processor and executing the above method steps.

[0039] An embodiment of this specification further provides a storage medium, wherein the storage medium stores a computer program, and the computer program is suitable for being loaded by a processor and executing the steps of the above method.

[0040] An embodiment of this specification further provides an electronic device, comprising: a processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the steps of the above method.

[0041] In the technical solution of the present specification, during the model training phase, user features mapped by user information of different user groups (who receive different treatment measures) are cross-used to train effect prediction models associated with different treatment measures. Specifically, when training the effect prediction model H1 associated with the first treatment measure, not only the feature Φ(x1) and its label y1 mapped by the user information x1 of the first group of users who receive the first treatment measure are used, but also the feature Φ(x2) mapped by the user information x2 of the second group of users who receive the second treatment measure are used; when training the effect prediction model H2 associated with the second treatment measure, not only the feature Φ(x2) and its label y2 mapped by the user information x2 of the second group of users who receive the second treatment measure are used, but also the feature Φ(x1) mapped by the user information x1 of the first group of users who receive the first treatment measure are used.

[0042] In addition, the feature mapping model Φ, the effect prediction model H1 associated with the first treatment measure, and the effect prediction model H2 associated with the second treatment measure are jointly trained. The objective function used in the training (reducing the value of the objective function is the training goal) is set specifically so that the objective function is not only positively correlated with the training loss of H1 and the training loss of H2, but also positively correlated with the difference between "the probability distribution of the effect representation of the second group of users predicted by H1" and "the probability distribution of the actual effect representation of the first group of users" (or the probability distribution of the effect representation of the first group of users predicted by H1), and also positively correlated with the difference between "the probability distribution of the effect representation of the first group of users predicted by H2" and "the probability distribution of the actual effect representation of the second group of users" (or the probability distribution of the effect representation of the second group of users predicted by H2). This solution can be applied in a trusted execution environment.

[0043] In the model application stage, the feature Φ(x0) mapped by the target user's user information x0 is input into H1 and H2 respectively to obtain the predicted effect representation after the target user accepts the first treatment measure and the predicted effect representation after the target user accepts the second treatment measure.

[0044] Through the above technical solution, the feature mapping model Φ can be enabled to map user information into user features that meet the no-confusion assumption, and the effect prediction models H1 and H2 can be enabled to have the ability to predict the reverse results, overcome the selective bias caused by the data distribution differences of different user groups, and improve the prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of a model training method provided in an embodiment of this specification;

[0046] Figure 2 This is a flow chart of an effect prediction method provided in an embodiment of this specification;

[0047] Figure 3 This specification also provides a schematic structural diagram of a model training device;

[0048] Figure 4 This is a structural diagram of another model training device provided in an embodiment of this specification;

[0049] Figure 5 This embodiment of the present specification also provides a schematic structural diagram of an effect prediction device;

[0050] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this specification. DETAILED DESCRIPTION

[0051] To make the purpose, technical solutions, and advantages of this specification more clear, the technical solutions of this specification will be clearly and completely described below in conjunction with the specific embodiments of this specification and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of this specification, not all of them. This solution can be applied in a trusted execution environment. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this specification.

[0052] Causal inference is an important field in statistics and machine learning that aims to understand and quantify causal relationships, not just observed associations. Simply put, it is about answering the question "If I do this action, what will happen?" rather than just "Do these things tend to happen together?" In other words, in causal inference, we want to infer the causal effect of one variable (usually a "treatment" or "intervention") on another variable (usually an "outcome") from observed data.

[0053] CATE, short for conditional average treatment effect, is a key concept in causal inference. It represents the average causal effect of a treatment on an outcome, given certain specific conditions (such as individual characteristics and environmental factors). CATE not only helps us understand the differences in the effects of different treatments across groups but also helps us match treatments with more effective outcomes for specific groups.

[0054] Let's say we're studying the impact of a new educational program on student achievement. We can calculate the average treatment effect (CATE) for all students—the average impact of the program across all students. However, CATE allows us to delve deeper, asking questions like: Does the program affect boys and girls differently? Does it have the same effect on low-achieving and high-achieving students? Does the program's effect differ between urban and rural schools? By analyzing these CATEs, we can better understand the actual impact of educational programs and implement more targeted improvements and optimizations.

[0055] T-learner is a method for estimating CATE (conditional average treatment effect). It transforms the CATE estimation problem into two independent effect prediction problems. It uses a machine learning model to predict the results of the first group of users and the second group of users respectively, and then estimates the causal effect by comparing the differences between the two.

[0056] In the T-learner method, two effect prediction models need to be trained, including:

[0057] The effect prediction model corresponding to the first user group is used to predict the effect of the first user group when the first user group receives the first treatment measure. The effect prediction model is usually trained only using the data of the first user group.

[0058] The effect prediction model corresponding to the second user group is used to predict the effect that the second user group will experience when the second user group receives the second treatment measure. This effect prediction model is typically trained using only data from the second user group. It should be noted that the first treatment measure may specifically be a set treatment, and the second treatment measure may be no treatment.

[0059] The CATE can be determined based on the difference in the effect representations predicted by the two effect prediction models for the same user. It should be noted that the predicted effect representation here is typically a probability value. For example, if the first and second treatment measures are two different ad recommendation schemes, then the corresponding effect representation is the probability of the user clicking the ad details link.

[0060] Here we give an example to illustrate the training method of the existing effect prediction model.

[0061] Suppose we have a dataset that records changes in student academic performance on an online education platform, as well as whether or not they use a new learning tool. We would like to estimate the impact of this tool on different groups of students. We divide the student population into two groups: one that uses the new tool (the first treatment) and observes whether this group's grades improve; the other group that does not use the new tool (the second treatment) and also observes whether this group's grades improve.

[0062] The student information of the first group of students is denoted as x1, and the actual effect representation of whether the academic performance of the first group of students has improved is denoted as y1 (for example, y1 is 0 for no improvement, and y1 is 1 for improvement). The student information of the second group of students is denoted as x2, and the actual effect representation of whether the academic performance of the second group of students has improved is denoted as y2.

[0063] Perform feature extraction on x1 to form feature m1 corresponding to x1; perform feature extraction on x2 to form feature m2 corresponding to x2.

[0064] The input m1 is associated with the effect prediction model H1 of the first treatment measure (using the new tool), and the output is H1(m1); the parameters of H1 are adjusted with the goal of reducing the gap between H1(m1) and y1.

[0065] The input m2 is associated with the effect prediction model H2 of the second treatment measure (not using the new tool), and the output is H2(m2); the parameters of H2 are adjusted with the goal of reducing the gap between H2(m2) and y2.

[0066] However, the above solution has the following two problems:

[0067] 1. m1 and m2 obtained by feature extraction of x1 and x2 do not necessarily meet the unconfounding assumption required by the causal inference task. The unconfounding assumption requires independence between treatment measures and potential outcomes, that is, the distribution of treatments is independent of potential outcomes. It can be expressed as Y(1), Y(2)⊥T|X, where Y(1) represents acceptance of one treatment measure and Y(2) represents acceptance of another treatment measure. ⊥ represents independence, T represents the treatment measure variable, and X represents the user dataset.

[0068] 2. There is often a significant difference between the data distribution of the first group of users and the data distribution of the second group of users. If only the features of the first group of users are used to train H1, and only the features of the second group of users are used to train H2, it is easy to cause H1 and H2 to lack counterfactual prediction capabilities and poor prediction accuracy.

[0069] Counterfactual prediction ability refers to the ability of the model to accurately predict unobserved counterfactual results (for example, for H1, the lack of accurate prediction ability of the effect of the second group of users receiving the first treatment measure), which is crucial for accurately estimating CATE.

[0070] Here is an example to illustrate how the difference in data distribution between two groups of users affects the accuracy of CATE estimation. Suppose we have an online advertising campaign and we want to evaluate its impact on user purchasing behavior. We divide the users into two groups:

[0071] Group 1: Users who saw the ad.

[0072] The second group: users who did not see the ad.

[0073] Among these two groups of users, there may be the following distribution differences:

[0074] Demographic differences: The average age of users in the first group may be younger than that of users in the second group, and ads may be more likely to be delivered to this younger user group.

[0075] Geographic location: The first group of users may be more from urban areas, while the second group of users may be more from rural areas. Ads may be shown more frequently in urban areas.

[0076] Online activity: Users in the first group may be more active on the website, and ads may be more likely to be displayed to users who frequently visit the website.

[0077] Historical purchasing behavior: Users in the first group may have more purchase records in the past, and ads may be targeted based on their purchase history.

[0078] Interest tags: Users in the first group may be more inclined to certain interest tags, and advertising targeting may be based on users' interest data.

[0079] The differences in data distribution between these user groups may lead to biases in our assessment of ad effectiveness. For example, if users in the first group are more likely to make a purchase based on their own characteristics (even without seeing the ad), then directly comparing the click-through rates of ads between the two groups may overestimate the effectiveness of the ad.

[0080] Therefore, in order to solve the above two problems, in one or more embodiments of this specification, the following solutions are provided:

[0081] During the model training phase, user features mapped to user information of different user groups (who receive different treatment measures) are cross-used to train effect prediction models associated with different treatment measures. Specifically, when training the effect prediction model H1 associated with the first treatment measure, not only the feature Φ(x1) and its label y1 mapped to the user information x1 of the first group of users who receive the first treatment measure are used, but also the feature Φ(x2) mapped to the user information x2 of the second group of users who receive the second treatment measure are used; when training the effect prediction model H2 associated with the second treatment measure, not only the feature Φ(x2) and its label y2 mapped to the user information x2 of the second group of users who receive the second treatment measure are used, but also the feature Φ(x1) mapped to the user information x1 of the first group of users who receive the first treatment measure are used.

[0082] In addition, the feature mapping model Φ, the effect prediction model H1 associated with the first treatment measure, and the effect prediction model H2 associated with the second treatment measure are jointly trained. The objective function used in the training (reducing the value of the objective function is the training goal) is specifically set so that the objective function is not only positively correlated with the training loss of H1 and the training loss of H2, but also positively correlated with the difference between "the probability distribution of the effect representation of the second group of users predicted by H1" and "the probability distribution of the actual effect representation of the first group of users" (or the probability distribution of the effect representation of the first group of users predicted by H1), and also positively correlated with the difference between "the probability distribution of the effect representation of the first group of users predicted by H2" and "the probability distribution of the actual effect representation of the second group of users" (or the probability distribution of the effect representation of the second group of users predicted by H2).

[0083] In the model application stage, the feature Φ(x0) mapped by the target user's user information x0 is input into H1 and H2 respectively to obtain the predicted effect representation after the target user accepts the first treatment measure and the predicted effect representation after the target user accepts the second treatment measure.

[0084] Through one or more embodiments provided by the present disclosure, the feature mapping model Φ can be enabled to map user information into user features that conform to the no-confusion assumption, and the effect prediction models H1 and H2 can be enabled to have the ability to predict reverse results, thereby overcoming the selective bias caused by data distribution differences between different user groups and improving prediction accuracy.

[0085] The following describes the details of the technical solutions provided in the embodiments of this specification in conjunction with the accompanying drawings.

[0086] Figure 1 This is a flow chart of a model training method provided in an embodiment of this specification, which includes the following steps:

[0087] S100: Obtain user information of a first group of users and a representation of actual effects of a first treatment measure on the first group of users; and obtain user information of a second group of users and a representation of actual effects of a second treatment measure on the second group of users.

[0088] During the model training phase, the user group that serves as training samples needs to be divided into two groups, the first group and the second group, each group including multiple users.

[0089] The user information for the first group of users can be understood as the user information for multiple users in the first group. User information can include various descriptive information about the user, such as height, gender, age, hobbies, place of residence (rural or urban), education level, income level, and so on. Similarly, the user information for the second group of users can be understood as the user information for multiple users in the second group. For ease of description, the user information for the first group of users is denoted as x1, and the user information for the second group of users is denoted as x2.

[0090] The actual effect representation after the first group of users accepts the first treatment measure can be understood as the actual effect reflected by each user in the first group of users after accepting the first treatment measure. The actual effect can be represented by whether it is effective (0 or 1). It can be seen that the actual effect representation after the first group of users accepts the first treatment measure indicates whether the first treatment measure is effective for each user in the group. Similarly, the actual effect representation after the second group of users accepts the second treatment measure indicates whether the second treatment measure is effective for each user in the group. For the sake of convenience of description, the actual effect representation after the first group of users accepts the first treatment measure is denoted as y1, and the actual effect representation after the second group of users accepts the second treatment measure is denoted as y2.

[0091] For example, the first processing measure is a first advertisement recommendation plan, and the second processing measure is a second advertisement recommendation plan; accordingly, the actual effect after the user accepts the processing measure is clicking on the advertisement detail link, or not clicking on the advertisement detail link.

[0092] For another example, the first treatment measure is medical intervention, and the second treatment measure is no medical intervention; accordingly, the actual effect after the user accepts the treatment measure is that the condition improves, or the condition does not improve.

[0093] S102: Based on the feature mapping model, feature mapping is performed on the user information of the first group of users, and feature mapping is performed on the user information of the second group of users.

[0094] S104: Cross-using user features corresponding to the first group of users and user features corresponding to the second group of users to train an effect prediction model associated with the first treatment measure and an effect prediction model associated with the second treatment measure.

[0095] For the convenience of description, the feature mapping model is denoted as Φ, the effect prediction model associated with the first treatment measure is denoted as H1, and the effect prediction model associated with the second treatment measure is denoted as H2.

[0096] Φ is used to map user information x into user features Φ(x) that can be input into the effect prediction model. Specifically, Φ maps x1 to Φ(x1) and x2 to Φ(x2).

[0097] H1 is used to predict the effect H1(Φ(x)) of the first treatment measure applied to the user based on the input Φ(x). H2 is used to predict the effect H2(Φ(x)) of the second treatment measure applied to the second group of users based on the input Φ(x).

[0098] exist Figure 1 In the training process shown, Φ, H1, and H2 are jointly trained. The training objectives are to allow Φ to learn to map user information x into user features Φ(x) that conform to the non-confusion assumption, to allow H1 to learn to accurately predict the effect of the first treatment measure applied to users with features Φ(x), and to allow H2 to learn to accurately predict the effect of the second treatment measure applied to users with features Φ(x).

[0099] exist Figure 1 In the training process shown, for step S104, Φ(x1) is specifically input into the effect prediction model H1 associated with the first treatment measure, and H1(Φ(x1)) is output; Φ(x2) is input into H1, and H1(Φ(x2)) is output; and Φ(x2) is input into the effect prediction model H2 associated with the second treatment measure, and H2(Φ(x2)) is output; Φ(x1) is input into H2, and H2(Φ(x1)) is output.

[0100] S106: Taking reducing the value of the objective function as a training goal, adjusting parameters of the feature mapping model, the effect prediction model associated with the first treatment measure, and the effect prediction model associated with the second treatment measure.

[0101] The value of the objective function in the embodiments of this specification has a positive correlation with the following four values:

[0102] 1. Training loss of H1: the difference between H1(Φ(x1)) and y1.

[0103] 2. Training loss of H2: the difference between H2(Φ(x2)) and y2.

[0104] 3. The difference between "the probability distribution of the effect representation predicted by H1 for the second group of users" and "the probability distribution of the actual effect representation of the first group of users": the gap between P(H1(Φ(x2))|Φ(x2)) and P(y1|Φ(x1)).

[0105] 4. The difference between the probability distribution of the effect representation predicted by H2 for the first group of users and the probability distribution of the actual effect representation for the second group of users: the gap between P(H2(Φ(x1))|Φ(x1)) and P(y2|Φ(x2)).

[0106] The formula of the specific objective function can be flexibly set based on the above principles.

[0107] It should be noted that during the model training phase, for Φ, H1, and H2, the following conditions must be met:

[0108] Requirement 1: Y(1), Y(2)⊥T∣Φ(x).

[0109] Requirement 2: E[Y(1)∣X=x]⊥Y∣Φ(x); E[Y(2)∣X=x]⊥Y∣Φ(x).

[0110] Where Y(1) represents acceptance of the first treatment measure, Y(2) represents acceptance of the second treatment measure, ⊥ represents independence, T represents the treatment variable, X represents the user data set (divided into two groups, x1 and x2), E represents the mathematical expectation, and “|Φ(x)” represents the case where Φ(x) is the condition.

[0111] The above requirement 1 can be transformed into:

[0112] P(Y(2)|T=2,Φ(x))=P(Y(2)|T=1,Φ(x));

[0113] P(Y(1)|T=1, Φ(x))=P(Y(1)|T=2, Φ(x)).

[0114] The above-mentioned requirement 1 after transformation can be roughly equivalent to:

[0115] P(H1(Φ(x2))|Φ(x2))=P(y1|Φ(x1));

[0116] P(H2(Φ(x1))|Φ(x1))=P(y2|Φ(x2)).

[0117] Therefore, during model training, we need to not only aim to reduce the training loss, but also to reduce the gap between P(H1(Φ(x2))|Φ(x2)) and P(y1|Φ(x1)), and the gap between P(H2(Φ(x1))|Φ(x1)) and P(y2|Φ(x2)).

[0118] In addition, as another embodiment,

[0119] The value of the objective function can also have a positive correlation with the following four values:

[0120] 1. Training loss of H1: the difference between H1(Φ(x1)) and y1.

[0121] 2. Training loss of H2: the difference between H2(Φ(x2)) and y2.

[0122] 3. The difference between "the probability distribution of the effect representation predicted by H1 for the second group of users" and "the probability distribution of the actual effect representation of the first group of users": the gap between P(H1(Φ(x2))|Φ(x2)) and P(H1(Φ(x1))|Φ(x1)).

[0123] 4. The difference between the probability distribution of the effect representation predicted by H2 for the first group of users and the probability distribution of the actual effect representation for the second group of users: the gap between P(H2(Φ(x1))|Φ(x1)) and P(H2(Φ(x2))|Φ(x2)).

[0124] The model training process includes multiple iterations. In each iteration, at least part of the data is input into the model to obtain the model output. Then, the model parameters are adjusted in the optimization direction to reduce the value of the objective function, and the next iteration is started until the number of iterations reaches a preset number or the value of the objective function is less than a preset threshold. Figure 1 The training process shown is easy to understand. Steps S104-S106 describe the principle of multiple training iterations.

[0125] Figure 2 This is a flow chart of an effect prediction method provided in an embodiment of this specification, which includes the following steps:

[0126] S200: Obtain user information of the target user.

[0127] The user information of the target user is recorded as x0.

[0128] S202: Based on the feature mapping model, feature mapping is performed on the target user information.

[0129] The feature mapping model is denoted as Φ, which maps x0 to Φ(x0).

[0130] S204: Associating the user feature corresponding to the target user with the effect prediction model of the first processing measure, and outputting a first predicted effect representation, which is used to represent the predicted effect representation after the target user accepts the first processing measure.

[0131] The effect prediction model associated with the first treatment measure is denoted as H1, and the first predicted effect representation is denoted as H1(Φ(x0)).

[0132] S206: Associating the user feature corresponding to the target user with the effect prediction model of the second processing measure, and outputting a second predicted effect representation, which is used to represent the predicted effect representation after the target user accepts the second processing measure.

[0133] The effect prediction model associated with the first treatment measure is denoted as H2, and the first predicted effect representation is denoted as H2(Φ(x0)).

[0134] In addition, the conditional average treatment effect CATE estimation can be performed based on the gap between H1(Φ(x0)) and H2(Φ(x0)).

[0135] Furthermore, if H1(Φ(x0)) is greater than H2(Φ(x0)), the target user is determined to be suitable for the first treatment measure; if H1(Φ(x0)) is less than H2(Φ(x0)), the target user is determined to be suitable for the second treatment measure. In this way, the treatment measure suitable for each user in the user group can be clearly identified, and the user groups can be grouped according to the different treatment measures.

[0136] Continuing with the previous example, suppose we have a dataset that records changes in students' academic performance on an online education platform, as well as whether they use a new learning tool. We want to estimate the impact of this tool on different groups of students. We divide the student population into two groups: one group uses the new tool (the first treatment) and observes whether this group's grades improve; the other group does not use the new tool (the second treatment) and also observes whether this group's grades improve.

[0137] The student information of the first group of students is denoted as x1, and the actual effect representation of whether the academic performance of the first group of students has improved is denoted as y1 (for example, y1 is 0 for no improvement, and y1 is 1 for improvement). The student information of the second group of students is denoted as x2, and the actual effect representation of whether the academic performance of the second group of students has improved is denoted as y2.

[0138] Based on Φ, feature mapping is performed on x1 to form the feature Φ(x1) corresponding to x1; feature mapping is performed on x2 to form the feature Φ(x2) corresponding to x2.

[0139] Input Φ(x1) and Φ(x2) into the effect prediction model H1 associated with the first treatment measure (using the new tool), and output H1(Φ(x1)) and H1(Φ(x2)). Input Φ(x1) and Φ(x2) into the effect prediction model H2 associated with the second treatment measure (not using the new tool), and output H2(Φ(x1)) and H2(Φ(x2)).

[0140] The training goal is to reduce the value of the objective function and adjust the parameters of Φ, H1, and H2. The value of the objective function is positively correlated with: the gap between H1(Φ(x1)) and y1, the gap between H2(Φ(x2)) and y2, the gap between P(H1(Φ(x2))|Φ(x2)) and P(y1|Φ(x1)), and the gap between P(H2(Φ(x1))|Φ(x1)) and P(y2|Φ(x2)).

[0141] After training Φ, H1, and H2, suppose we need to group a class of students into groups suitable for using the new tool to improve their grades and unsuitable for it. Therefore, we need to determine for each student whether the new tool is suitable for improving their grades. For example, we can input Zhang San's user information x0 into Φ for feature mapping, obtaining Φ(x0). Then, we input Φ(x0) into H1 and H2, respectively, obtaining H1(Φ(x0)) and H2(Φ(x0)). Assuming H1(Φ(x0)) is 0.8 and H2(Φ(x0)) is 0.5, then Zhang San is suitable for using the new tool to improve his grades and is likely to be effective.

[0142] Figure 3 The embodiment of this specification also provides a structural diagram of a model training device, including:

[0143] The acquisition module 301 acquires user information x1 of a first group of users and a representation y1 of an actual effect of a first treatment measure on the first group of users; and acquires user information x2 of a second group of users and a representation y2 of an actual effect of a second treatment measure on the second group of users;

[0144] A mapping module 302 maps x1 to Φ(x1) and x2 to Φ(x2) based on a feature mapping model Φ;

[0145] Prediction module 303 inputs Φ(x1) into effect prediction model H1 associated with the first treatment measure and outputs H1(Φ(x1)), inputs Φ(x2) into H1 and outputs H1(Φ(x2)); and inputs Φ(x2) into effect prediction model H2 associated with the second treatment measure and outputs H2(Φ(x2)), inputs Φ(x1) into H2 and outputs H2(Φ(x1)).

[0146] Adjustment module 304 adjusts the parameters of Φ, H1, and H2 with the goal of reducing the value of the objective function as the training goal;

[0147] Where P represents the probability operation, and the value of the objective function is positively correlated to: the gap between H1(Φ(x1)) and y1, the gap between H2(Φ(x2)) and y2, the gap between P(H1(Φ(x2))|Φ(x2)) and P(y1|Φ(x1)), and the gap between P(H2(Φ(x1))|Φ(x1)) and P(y2|Φ(x2)).

[0148] Figure 4 This is a schematic diagram of the structure of another model training device provided in an embodiment of this specification, including:

[0149] The acquisition module 401 acquires user information x1 of a first group of users and a representation y1 of an actual effect of a first treatment measure on the first group of users; and acquires user information x2 of a second group of users and a representation y2 of an actual effect of a second treatment measure on the second group of users;

[0150] A mapping module 402 maps x1 to Φ(x1) and x2 to Φ(x2) based on a feature mapping model Φ;

[0151] Prediction module 403 inputs Φ(x1) into effect prediction model H1 associated with the first treatment measure and outputs H1(Φ(x1)), inputs Φ(x2) into H1 and outputs H1(Φ(x2)); and inputs Φ(x2) into effect prediction model H2 associated with the second treatment measure and outputs H2(Φ(x2)), inputs Φ(x1) into H2 and outputs H2(Φ(x1)).

[0152] An adjustment module 404 adjusts the parameters of Φ, H1, and H2 with the goal of reducing the value of the objective function as a training objective;

[0153] Where P represents the probability operation, and the value of the objective function is positively correlated to: the gap between H1(Φ(x1)) and y1, the gap between H2(Φ(x2)) and y2, the gap between P(H1(Φ(x2))|Φ(x2)) and P(H1(Φ(x1))|Φ(x1)), and the gap between P(H2(Φ(x1))|Φ(x1)) and P(H2(Φ(x2))|Φ(x2)).

[0154] Figure 5 The embodiment of this specification also provides a schematic structural diagram of an effect prediction device, including:

[0155] Acquisition module 501, acquires user information x0 of the target user;

[0156] A mapping module 502 maps x0 to Φ(x0) based on a feature mapping model Φ;

[0157] The prediction module 503 inputs Φ(x0) into the effect prediction model H1 and outputs H1(Φ(x0)), which is used to represent the predicted effect representation after the target user accepts the first treatment measure; and inputs Φ(x0) into the effect prediction model H2 and outputs H2(Φ(x0)), which is used to represent the predicted effect representation after the target user accepts the second treatment measure.

[0158] The above-mentioned device embodiments correspond to the method embodiments. For detailed descriptions, please refer to the description of the method embodiments, which will not be repeated here. The device embodiments are obtained based on the corresponding method embodiments and have the same technical effects as the corresponding method embodiments. For detailed descriptions, please refer to the corresponding method embodiments.

[0159] The embodiments of this specification also provide a computer storage medium, which can store multiple instructions, and the instructions are suitable for being loaded by a processor to execute the method of the embodiments of this disclosure.

[0160] This specification also provides a computer program product, which stores at least one instruction. The at least one instruction is loaded by the processor and executes the method of the embodiment of the present disclosure.

[0161] The embodiments of this specification also provide Figure 6 The electronic device shown in the figure is a schematic diagram of its structure. At the hardware level, the electronic device includes a processor, an internal bus, a network interface, memory, and non-volatile storage, and may also include other necessary hardware. The processor reads the corresponding computer program from the non-volatile storage into the memory and then runs it to implement the above-mentioned voice activity detection method.

[0162] Of course, in addition to software implementation, this specification does not exclude other implementation methods, such as logic devices or a combination of software and hardware, etc. That is to say, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0163] In the 1990s, technological improvements could be clearly distinguished as either hardware improvements (for example, improvements to circuit structures like diodes, transistors, and switches) or software improvements (improvements to process flows). However, with the advancement of technology, many process flow improvements today can now be considered direct improvements to hardware circuit structures. Designers almost always create the corresponding hardware circuit structure by programming the improved process flow into the hardware circuit. Therefore, it cannot be said that a process flow improvement cannot be implemented using hardware modules. For example, a programmable logic device (PLD), such as a field programmable gate array (FPGA), is an integrated circuit whose logical function is determined by user programming. Designers can "integrate" a digital system on a PLD through their own programming, without having to hire a chip manufacturer to design and manufacture a dedicated integrated circuit chip. Moreover, nowadays, instead of manually fabricating integrated circuit chips, this programming is mostly done using "logic compiler" software. This is similar to the software compiler used when developing programs. Before compilation, the original code must also be written in a specific programming language, called a hardware description language (HDL). There is not just one HDL, but many, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, RHDL (Ruby Hardware Description Language), etc. The most commonly used ones are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art will also understand that by simply programming the method flow in one of these hardware description languages ​​and then programming it into an integrated circuit, a hardware circuit that implements the logic method flow can be easily obtained.

[0164] The controller can be implemented in any suitable manner. For example, the controller can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, application-specific integrated circuits (ASICs), programmable logic controllers, and embedded microcontrollers. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320. The memory controller can also be implemented as part of the control logic of the memory. Those skilled in the art will also know that in addition to implementing the controller in a purely computer-readable program code format, the controller can be implemented in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers by logically programming the method steps. Therefore, such a controller can be considered a hardware component, and the devices included therein for implementing various functions can also be considered as structures within the hardware component. Or even, the devices for implementing various functions can be considered as both software modules that implement the method and structures within the hardware component.

[0165] The systems, devices, modules, or units described in the above embodiments may be implemented by computer chips or entities, or by products having certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0166] For the convenience of description, the above devices are described as being divided into various units according to their functions. Of course, when implementing this specification, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0167] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems, or computer program products. Therefore, this specification may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0168] This specification is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of this specification. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0169] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0170] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0171] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0172] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0173] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0174] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a series of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0175] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems, or computer program products. Thus, this specification may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0176] This specification may be described in the general context of computer-executable instructions, such as program modules, executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. This specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media, including storage devices.

[0177] The various embodiments in this specification are described in a progressive manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiments are generally similar to the method embodiments, so the description is relatively simple. For relevant parts, refer to the description of the method embodiments.

[0178] The foregoing is merely an example of the present invention and is not intended to limit the present invention. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention are intended to be included within the scope of the claims of the present invention.

Claims

1. A model training method, comprising: Obtain user information x1 of a first group of users and representation y1 of actual effects of the first group of users after receiving a first treatment measure; and obtain user information x2 of a second group of users and representation y2 of actual effects of the second group of users after receiving a second treatment measure; Based on the feature mapping model Φ, x1 is mapped to Φ(x1), and x2 is mapped to Φ(x2); Input Φ(x1) into the effect prediction model H1 associated with the first treatment measure, output H1(Φ(x1)), input Φ(x2) into H1, output H1(Φ(x2)); and input Φ(x2) into the effect prediction model H2 associated with the second treatment measure, output H2(Φ(x2)), input Φ(x1) into H2, output H2(Φ(x1)); Taking reducing the value of the objective function as the training goal, adjust the parameters of Φ, H1, and H2; Among them, P represents the probability operation, and the value of the objective function is positively correlated to: the gap between H1(Φ(x1)) and y1, the gap between H2(Φ(x2)) and y2, the gap between P(H1(Φ(x2))|Φ(x2)) and P(y1|Φ(x1)), and the gap between P(H2(Φ(x1))|Φ(x1)) and P(y2|Φ(x2)).

2. The method of claim 1, wherein: The first processing measure is a first advertisement recommendation plan, and the second processing measure is a second advertisement recommendation plan; Accordingly, the actual effect after the user accepts the processing measures is to click on the advertisement details link, or not to click on the advertisement details link.

3. The method of claim 1, wherein: The first treatment measure is to perform medical intervention, and the second treatment measure is not to perform medical intervention; Accordingly, the actual effect after the user accepts the treatment measures is that the condition improves, or the condition does not improve.

4. A model training method, comprising: Obtain user information x1 of a first group of users and representation y1 of actual effects of the first group of users after receiving a first treatment measure; and obtain user information x2 of a second group of users and representation y2 of actual effects of the second group of users after receiving a second treatment measure; Based on the feature mapping model Φ, x1 is mapped to Φ(x1), and x2 is mapped to Φ(x2); Input Φ(x1) into the effect prediction model H1 associated with the first treatment measure, output H1(Φ(x1)), input Φ(x2) into H1, output H1(Φ(x2)); and input Φ(x2) into the effect prediction model H2 associated with the second treatment measure, output H2(Φ(x2)), input Φ(x1) into H2, output H2(Φ(x1)); Taking reducing the value of the objective function as the training goal, adjust the parameters of Φ, H1, and H2; Among them, P represents the probability operation, and the value of the objective function is positively correlated to: the difference between H1(Φ(x1)) and y1, the difference between H2(Φ(x2)) and y2, the difference between P(H1(Φ(x2))|Φ(x2)) and P(H1(Φ(x1))|Φ(x1)), and the difference between P(H2(Φ(x1))|Φ(x1)) and P(H2(Φ(x2))|Φ(x2)).

5. An effect prediction method, comprising: Get the user information x0 of the target user; Based on the feature mapping model Φ trained by the method according to any one of claims 1 to 4, mapping x0 to Φ(x0); Input Φ(x0) into the effect prediction model H1 trained by the method according to any one of claims 1 to 4, and output H1(Φ(x0)) to represent the predicted effect representation after the target user accepts the first treatment measure; And, Φ(x0) is input into the effect prediction model H2 trained by the method described in any one of claims 1-4, and H2(Φ(x0)) is output to represent the predicted effect representation after the target user accepts the second processing measure.

6. The method of claim 5, further comprising: Based on the gap between H1(Φ(x0)) and H2(Φ(x0)), the conditional average treatment effect CATE estimation is performed.

7. The method of claim 5, further comprising: If it is determined that H1(Φ(x0)) is greater than H2(Φ(x0)), it is determined that the target user is suitable for the first treatment measure; If it is determined that H1(Φ(x0)) is less than H2(Φ(x0)), it is determined that the target user is suitable for the second processing measure.

8. A model training device, comprising: An acquisition module, which acquires user information x1 of a first group of users and an actual effect representation y1 of the first group of users after receiving a first treatment measure; and acquires user information x2 of a second group of users and an actual effect representation y2 of the second group of users after receiving a second treatment measure; The mapping module maps x1 to Φ(x1) and x2 to Φ(x2) based on the feature mapping model Φ. A prediction module, inputting Φ(x1) into an effect prediction model H1 associated with a first treatment measure, outputting H1(Φ(x1)), inputting Φ(x2) into H1, outputting H1(Φ(x2)); and inputting Φ(x2) into an effect prediction model H2 associated with a second treatment measure, outputting H2(Φ(x2)), inputting Φ(x1) into H2, outputting H2(Φ(x1)); The adjustment module adjusts the parameters of Φ, H1, and H2 to reduce the value of the objective function as the training goal; Among them, P represents the probability operation, and the value of the objective function is positively correlated to: the gap between H1(Φ(x1)) and y1, the gap between H2(Φ(x2)) and y2, the gap between P(H1(Φ(x2))|Φ(x2)) and P(y1|Φ(x1)), and the gap between P(H2(Φ(x1))|Φ(x1)) and P(y2|Φ(x2)).

9. A model training device, comprising: An acquisition module, which acquires user information x1 of a first group of users and an actual effect representation y1 of the first group of users after receiving a first treatment measure; and acquires user information x2 of a second group of users and an actual effect representation y2 of the second group of users after receiving a second treatment measure; The mapping module maps x1 to Φ(x1) and x2 to Φ(x2) based on the feature mapping model Φ. A prediction module, inputting Φ(x1) into an effect prediction model H1 associated with a first treatment measure, outputting H1(Φ(x1)), inputting Φ(x2) into H1, outputting H1(Φ(x2)); and inputting Φ(x2) into an effect prediction model H2 associated with a second treatment measure, outputting H2(Φ(x2)), inputting Φ(x1) into H2, outputting H2(Φ(x1)); The adjustment module adjusts the parameters of Φ, H1, and H2 to reduce the value of the objective function as the training goal; Among them, P represents the probability operation, and the value of the objective function is positively correlated to: the difference between H1(Φ(x1)) and y1, the difference between H2(Φ(x2)) and y2, the difference between P(H1(Φ(x2))|Φ(x2)) and P(H1(Φ(x1))|Φ(x1)), and the difference between P(H2(Φ(x1))|Φ(x1)) and P(H2(Φ(x2))|Φ(x2)).

10. An effect prediction device, comprising: Acquisition module, obtains user information x0 of the target user; A mapping module, mapping x0 to Φ(x0) based on a feature mapping model Φ trained by the method according to any one of claims 1 to 4; A prediction module, inputting Φ(x0) into an effect prediction model H1 trained by the method described in any one of claims 1 to 4, and outputting H1(Φ(x0)), which is used to represent the predicted effect representation after the target user accepts the first treatment measure; and inputting Φ(x0) into an effect prediction model H2 trained by the method described in any one of claims 1 to 4, and outputting H2(Φ(x0)), which is used to represent the predicted effect representation after the target user accepts the second treatment measure.

11. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

12. An electronic device, characterized in that: include: A processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the steps of the method as claimed in any one of claims 1 to 7.

13. A computer program product having at least one instruction stored thereon, characterized in that: When the at least one instruction is executed by the processor, the steps of the method described in any one of claims 1 to 7 are implemented.

Images