Model training or intervention gain prediction method and device, equipment and medium
Through the intervention gain model training method, end-to-end prediction is used to utilize user business information and intervention action sets to perform end-to-end prediction, which solves the problem of user individual incremental impact prediction in the marketing field, reduces costs and errors, adapts to multiple intervention scenarios, and improves prediction efficiency and accuracy.
Patent Information
- Application Number
- CN202411195000.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-28
- Publication Date
- 2024-12-06
AI Technical Summary
In the marketing field, it is difficult for the prior art to effectively predict the incremental impact of individual users on different intervention actions, resulting in low decision-making efficiency, and high sample acquisition cost and serious model error accumulation in multiple intervention actions scenarios.
Provide an intervention gain model training method. By obtaining user business information and intervention action sets, the pre-trained intervention gain model is used to perform end-to-end intervention gain prediction, calculate the difference between the gain response result and the observed data, and adjust the model parameters. , reduce sample acquisition costs and reduce model error.
It realizes efficient intervention gain prediction in the marketing field, reduces sample acquisition costs, improves prediction accuracy, and adapts to multi-intervention action scenarios to reduce model error accumulation.
Smart Images

Figure CN119090543A_ABST
Abstract
Description
Technical Field
[0001] The present disclosure can be used in application scenarios such as generative search, intelligent document editing, intelligent assistants and intelligent e-commerce, and in particular relates to a model training or intervention gain prediction method, device, equipment and medium. Background Art
[0002] In some fields, especially marketing, it is necessary to study the incremental impact of treatment on individual outcomes of users. The incremental impact can also be called uplift or individual treatment effect (ITE).
[0003] Predictive models can now be used to understand the relationships between variables and reveal the impact of interventions on groups, which is crucial for solving problems and making effective decisions. Summary of the invention
[0004] The present disclosure provides a model training or intervention gain prediction method, device, equipment and medium.
[0005] According to one aspect of the present disclosure, a training method for an intervention gain model is provided, comprising:
[0006] Acquire user service information of a sample user, service interaction observation data of the sample user, and an intervention action set; the intervention action set includes a baseline intervention action and at least one other intervention action;
[0007] Determining, by means of an intervention gain model, a predicted response result generated by the sample user for each element in the intervention action set according to the user service information and the intervention action set;
[0008] Determining, by means of the intervention gain model, a gain value corresponding to each of the other intervention actions according to the user service information and the intervention action set;
[0009] Calculating the gain response result corresponding to each of the other intervention actions according to the gain value corresponding to each of the other intervention actions and the predicted response result generated by each element in the intervention action set through the intervention gain model;
[0010] By using the intervention gain model, the difference between the gain response result corresponding to each of the other intervention actions and the corresponding observation response result in the business interaction observation data is calculated to obtain the observation gain loss;
[0011] Parameters of the intervention gain model are adjusted according to the observed gain loss.
[0012] According to one aspect of the present disclosure, there is provided an intervention gain prediction method, comprising:
[0013] Obtain user business information and intervention action set of target users;
[0014] The intervention action set includes a baseline intervention action and at least one other intervention action;
[0015] Inputting the user service information and the intervention action set into a pre-trained intervention gain model; the intervention gain model is trained by the intervention gain model training method as described in any embodiment of the present disclosure;
[0016] Determining, by means of the intervention gain model, a gain value corresponding to each of the other intervention actions according to the user service information and the intervention action set;
[0017] An intervention gain result of the target user is determined according to the gain value corresponding to each of the other intervention actions.
[0018] According to one aspect of the present disclosure, there is provided a training device for an intervention gain model, comprising:
[0019] A sample acquisition module, used to acquire user service information of a sample user, service interaction observation data of the sample user, and an intervention action set; the intervention action set includes a baseline intervention action and at least one other intervention action;
[0020] An intervention gain model, used to determine, based on the user service information and the intervention action set, a predicted response result generated by the sample user for each element in the intervention action set;
[0021] The intervention gain model is used to determine the gain value corresponding to each of the other intervention actions according to the user service information and the intervention action set;
[0022] The intervention gain model is used to calculate the gain response result corresponding to each of the other intervention actions according to the gain value corresponding to each of the other intervention actions and the predicted response result generated by each element in the intervention action set;
[0023] The intervention gain model is used to calculate the difference between the gain response result corresponding to each of the other intervention actions and the corresponding observation response result in the business interaction observation data to obtain the observation gain loss;
[0024] A model parameter adjustment module is used to adjust the parameters of the intervention gain model according to the observed gain loss.
[0025] According to one aspect of the present disclosure, there is provided an intervention gain prediction device, comprising:
[0026] Acquire user service information and an intervention action set of a target user; the intervention action set includes a baseline intervention action and at least one other intervention action;
[0027] Inputting the user service information and the intervention action set into a pre-trained intervention gain model; the intervention gain model is trained by the intervention gain model training method as described in any embodiment of the present disclosure;
[0028] Determining, by means of the intervention gain model, a gain value corresponding to each of the other intervention actions according to the user service information and the intervention action set;
[0029] An intervention gain result of the target user is determined according to the gain value corresponding to each of the other intervention actions.
[0030] According to another aspect of the present disclosure, there is provided an electronic device, comprising:
[0031] at least one processor; and
[0032] a memory communicatively connected to the at least one processor; wherein,
[0033] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the training method or intervention prediction method of the intervention gain model described in any embodiment of the present disclosure.
[0034] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to execute the training method of the intervention gain model or the intervention prediction method described in any embodiment of the present disclosure.
[0035] According to another aspect of the present disclosure, a computer program product is provided, including a computer program, which, when executed by a processor, implements the training method of the intervention gain model or the intervention prediction method described in any embodiment of the present disclosure.
[0036] The disclosed embodiments can achieve end-to-end intervention gain prediction and reduce the accumulated intervention gain prediction error.
[0037] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present disclosure, nor is it intended to limit the scope of the present disclosure. Other features of the present disclosure will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] The accompanying drawings are used to better understand the present solution and do not constitute a limitation of the present disclosure.
[0039] Figure 1 is a flow chart of a training method for an intervention gain model disclosed in an embodiment of the present disclosure;
[0040] Figure 2 is a flow chart of an intervention gain prediction method disclosed in an embodiment of the present disclosure;
[0041] Figure 3 is a scenario diagram of a training method for an intervention gain model disclosed in an embodiment of the present disclosure;
[0042] Figure 4 is a scenario diagram of a training method for an intervention gain model disclosed in an embodiment of the present disclosure;
[0043] Figure 5 is an interface diagram of an intervention gain prediction method disclosed in an embodiment of the present disclosure;
[0044] Figure 6 is an interface diagram of an intervention gain prediction method disclosed in an embodiment of the present disclosure;
[0045] Figure 7 is a scene diagram of an intervention gain prediction method disclosed in an embodiment of the present disclosure;
[0046] Figure 8 is a structural schematic diagram of a training device for an intervention gain model disclosed in an embodiment of the present disclosure;
[0047] Fig. 9 is a structural schematic diagram of a training device for an intervention gain model disclosed in an embodiment of the present disclosure;
[0048] Fig.10 It is a block diagram of an electronic device according to the training method of the intervention gain model or the intervention gain prediction method disclosed in the embodiments of the present disclosure. DETAILED DESCRIPTION
[0049] The following is a description of exemplary embodiments of the present disclosure in conjunction with the accompanying drawings, including various details of the embodiments of the present disclosure to facilitate understanding, which should be considered as merely exemplary. Therefore, it should be recognized by those of ordinary skill in the art that various changes and modifications may be made to the embodiments described herein without departing from the scope and spirit of the present disclosure. Similarly, for the sake of clarity and conciseness, descriptions of well-known functions and structures are omitted in the following description.
[0050] In the disclosed embodiment, the intervention gain model is a model for predicting the causal effect of intervention on user behavior or status. For different application scenarios, intervention and gain may be different. The intervention gain model can be applied to application scenarios such as generative search, document intelligent editing, intelligent assistant and intelligent e-commerce. For example, in the intelligent e-commerce scenario, intervention may refer to the issuance of coupons of different denominations, and gain may be the user's conversion rate. For another example, in the generative search scenario, intervention may be different search strategies, gain may be the accuracy of search results, and accuracy may be expressed by user satisfaction with search results. For another example, in the document intelligent editing scenario, intervention may be different editing options, such as font editing options, paragraph editing options and page layout options, and gain may be document quality. For another example, in the intelligent assistant scenario, intervention may be different input templates, gain may refer to the accuracy of reply results, and accuracy may be expressed by user satisfaction with search results. For another example, in the drug application scenario, intervention may refer to the use of different drugs for treatment, and gain may refer to the user's cure rate. For another example, in the information recommendation scenario, intervention may refer to the use of different recommendation strategies, and gain may refer to the user's retention rate, etc.
[0051] The intervention gain model predicts the user's response results under a certain intervention and the response results without intervention. The difference between the two is the uplift score of the intervention. For example, the difference between the response results of issuing a coupon of a certain denomination and the response results without coupons is the uplift score of the coupon of that denomination. It can be calculated using the following formula:
[0052]
[0053]
[0054]
[0055] Figure 1 This is a flow chart of a method for training an intervention gain model according to an embodiment of the present disclosure. This embodiment can be applied to the case of training an intervention gain model. The method of this embodiment can be executed by a training device for an intervention gain model, which can be implemented in software and / or hardware and is specifically configured in an electronic device with certain data computing capabilities, which can be a server device.
[0056] S101. Obtain user service information of a sample user, service interaction observation data of the sample user, and an intervention action set; the intervention action set includes a benchmark intervention action and at least one intervention action.
[0057] Among them, user service information may refer to information related to intervention actions and expected response results. Response results may refer to the results of the behavior expected to occur by the user, for example, the intervention is to issue red envelopes, the expected user behavior is to use red envelopes, and the result is conversion rate. Exemplarily, user service information may include user information. User information may include: user attribute information, user behavior information, and user interest information. User attribute information may include user age, gender, and place of residence. In an example, user interest information may include electronic products, food, games, and travel. User behavior information may refer to the user's historical operation information. User service information may also include user-service interaction information. In an example, user behavior information may include: operation frequency, operation time period, network environment, and network address. Service may refer to services related to intervention actions and response results. For example, the intervention action is to send takeaway red envelopes, and the response result is the behavior of ordering takeaways, and the service may refer to the takeaway service. Interaction information may refer to the user's behavior information using the service, as well as information fed back to the user by the service. It should be noted that user service information is information collected with the authorization of the user.
[0058] The business interaction observation data is the interaction data between users and businesses that have been collected historically. The intervention action set includes multiple elements, wherein at least one element in the intervention action set is a baseline intervention action, and at least one element is a other intervention action. The intervention actions applied by each element are different, and different intervention actions may refer to different contents involved in the intervention actions. In an example, multiple intervention actions with different incentive methods may include: red envelopes, coupons, redemption codes, points, information push, gifts and payment coupons, etc. The baseline intervention action may be a red envelope, and multiple other intervention actions include coupons, redemption codes and points. For another example, for the same incentive method, multiple actions with different incentive amounts may include: not issuing red envelopes (i.e., the denomination of the red envelope issued is 0), issuing red envelopes with a denomination of 1, and issuing red envelopes with a denomination of 2, etc. The baseline intervention action may be not issuing red envelopes, and multiple other intervention actions include issuing red envelopes with a denomination of 1 and issuing red envelopes with a denomination of 2, etc. For another example, multiple intervention actions with different action contents may be: using drug A, using drug B, using drug C, and using drugs B and C, etc., the baseline intervention action may be using drug A, and multiple other intervention actions include using drug B, using drug C, and using drugs B and C, etc.
[0059] S102: Determine, by means of an intervention gain model, a predicted response result generated by the sample user for each element in the intervention action set according to the user service information and the intervention action set.
[0060] Among them, the intervention gain model is used to input user business information and a set of intervention actions, and output a predicted response result. The predicted response result generated by the sample user for a certain element may refer to the response result generated by the behavior of the sample user predicted to occur under the influence of the element. In fact, the predicted response result will not limit the state of the intervention action originally received by the sample user, and the original state of the sample user may be the state of applying any element. For example, the predicted response result may be the response result generated by the sample user applying any intervention action on the basis of the baseline intervention action. For another example, the predicted response result may be the response result generated by the sample user switching to any intervention action different from the former on the basis of applying any intervention action. For another example, the predicted response result may refer to the response result generated by the sample user switching to the baseline intervention action on the basis of applying any intervention action. Each element in the predicted action set generates at least one predicted response result.
[0061] S103: Determine, by using the intervention gain model, a gain value corresponding to each of the other intervention actions according to the user service information and the intervention action set.
[0062] Among them, the gain value of an intervention action may refer to the difference between the response result generated by the sample user under the influence of the first element and the response result generated when converted to the influence of the second element, wherein the first element or the second element is the intervention action; the first element and the second element are different. If the first element is the intervention action, the second element is different from the intervention action; if the second element is the intervention action, the first element is different from the intervention action. Specifically, the gain value of an intervention action may refer to the difference between the response result generated by the sample user under the influence of a different element and the response result generated when converted to the intervention action. Or the gain value of an intervention action may refer to the difference between the response result generated by the sample user under the influence of the intervention action and the response result generated when converted to a different element. The first element and the second element are elements in the intervention action set.
[0063] In one example, the first element is intervention action A, and the second element is intervention action B.
[0064] In one example, the first element or the second element is an intervention action or a baseline intervention action. If the first element is an intervention action, the second element is a baseline intervention action; if the second element is an intervention action, the first element is a baseline intervention action. The baseline intervention action may be not applying an intervention action.
[0065] S104. Calculate the gain response result corresponding to each of the other intervention actions through the intervention gain model according to the gain value corresponding to each of the other intervention actions and the predicted response result generated by each element in the intervention action set.
[0066] The gain response result of the intervention action may be a response result obtained by performing gain on the intervention action. Specifically, the gain response result of the intervention action may include a first gain response result obtained by applying the intervention action for gain, and a second gain response result obtained by removing the intervention action for gain. The first gain response result obtained by applying the intervention action for gain may be a positive gain, and the second gain response result obtained by removing the intervention action for gain may be a negative gain. Alternatively, the first gain response result obtained by applying the intervention action for gain may be a negative gain, and the second gain response result obtained by removing the intervention action for gain may be a positive gain. The gains of applying the intervention action and removing the intervention action are usually opposite.
[0067] Specifically, the predicted response result of the target element can be understood as the response result generated by the target element affecting the sample user. The gain value of the intervention action can be found as the difference between the response result generated under the influence of the first element and the response result generated under the influence of the second element, and the gain value of the second element can be found as the difference between the response result generated under the influence of the second element and the response result generated under the influence of the first element. Among them, the first element and the second element are different, the target element and the intervention action are different, the first element and the second element are the target element or the intervention action respectively, for example, the first element is the target element, the second element is the intervention action, or the first element is the intervention action, and the second element is the target element.
[0068] If the first element is the target element and the second element is the intervention action, the predicted response result of the target element is fused with the gain value of the intervention action to obtain the response result obtained by converting the target element to the gain value of the intervention action as the first gain response result corresponding to the intervention action.
[0069] If the first element is an intervention action and the second element is a target element, the predicted response result of the target element is fused with the gain value of the intervention action to obtain a response result obtained by converting the intervention action to the target element for gain, that is, a response result obtained by removing the intervention action for gain, as the second gain response result corresponding to the intervention action.
[0070] In one example, the target element is a baseline intervention action, wherein the baseline intervention action may be not applying an intervention action.
[0071] S105. Calculate, by using the intervention gain model, the difference between the gain response result corresponding to each of the other intervention actions and the corresponding observation response result in the business interaction observation data to obtain an observation gain loss.
[0072] Observation data is the collected raw data. Observation data is data obtained by observing a large number of users. Since no specific strategy is adopted to intervene and collect data on a large number of users, the distribution difference of observation data is large, wherein the specific strategy can be an AB experiment. The true value corresponding to the gain response result can be observed. Business interaction observation data may include the observed response results of sample users after applying intervention actions and / or the observed response results of sample users without applying intervention actions. Observation gain loss is used to describe the difference between the response results and the true value of the behavior of the predicted user when switching the applied intervention action. Usually, it is easy to find the observed response results of not applying an intervention action or applying a certain intervention action in business interaction observation data, and it is difficult to find the observed response results of switching from applying an intervention action to applying another intervention action. Therefore, observation gain loss is usually used to describe the difference between the response results and the true value of the behavior of the predicted user when applying or not applying an intervention action.
[0073] Among them, the observed response result of the sample user after applying the intervention action is actually the true value corresponding to the gain response result obtained by applying the intervention action; the observed response result of the sample user without applying the intervention action is actually the true value corresponding to the gain response result obtained by removing the intervention action. The observed response result corresponding to the gain response result corresponding to the intervention action can be queried in the business interaction observation data, the difference between the gain response result and the observed response result is calculated, and the observed gain loss is determined based on the calculated difference.
[0074] S106: Adjust parameters of the intervention gain model according to the observed gain loss.
[0075] The parameters of the intervention gain model can be adjusted with the goal of minimizing or converging the observed gain loss.
[0076] In the prior art, in order to ensure that the input features and interventions are independent of each other, sample data of randomized controlled trials (RCT) are usually obtained by constructing AB experiments. However, the cost of obtaining RCT samples is high, and when there are multiple intervention actions, a set of AB experiments must be created for each intervention action, and high-quality RCT samples must be ensured, which requires higher investment costs.
[0077] According to the technical solution disclosed in the present invention, by predicting the predicted response results generated by sample users for each intervention element in the intervention action set and the gain value of each intervention action, and fusing the gain value of each intervention action with the predicted response result of each element, a gain response result based on the intervention action gain can be obtained, and the observed response result corresponding to the gain response result is queried in the observation data as the true value, and the difference between the gain response result and the true value is calculated to obtain the observed gain loss. The predicted gain response result can be associated with the observation data, and the true value corresponding to the predicted gain response result can be directly obtained from the observation data, thereby reducing the sample acquisition cost. In addition, by implementing intervention gain prediction through an end-to-end model, the accumulation of model errors can be reduced, thereby improving the accuracy of intervention gain prediction. At the same time, multiple intervention actions can be set in the intervention action set to realize an intervention gain prediction scenario that is adapted to multiple interventions.
[0078] Optionally, determining the gain value corresponding to each of the other intervention actions based on the user business information and the intervention action set includes: determining a first gain value corresponding to each of the other intervention actions based on the user business information and the intervention action set; the first gain value corresponding to the other intervention action is a gain value of a response result generated from a baseline intervention action to applying the other intervention action; determining a second gain value corresponding to each of the other intervention actions based on the user business information and the intervention action set; the second gain value corresponding to the other intervention action is a gain value of a response result generated from applying the other intervention action to the baseline intervention action; determining the gain value corresponding to each of the other intervention actions based on the first gain value corresponding to each of the other intervention actions and the second gain value corresponding to each of the other intervention actions.
[0079] Wherein, the sample user originally applies the baseline intervention action, and after changing to apply other intervention actions, the magnitude of the change in the response result is the first gain value of the other intervention action. The sample user originally applies other intervention actions, and after removing the other intervention actions, that is, after changing to apply the baseline intervention action, the magnitude of the change in the response result is the second gain value of the intervention action. The signs of the first gain value and the second gain value are usually opposite. The baseline intervention action may be not applying an intervention action.
[0080] The gain value corresponding to the intervention action includes: a first gain value and a second gain value. In fact, in the real observation data, in particular, the baseline intervention action is not to apply the intervention action, and the response results of applying a certain other intervention action or applying the baseline intervention action are distributed more. Therefore, it is easier to find a large number of corresponding true values in the observation data by selecting any other intervention action and applying the baseline intervention action to switch and predict the gain value after switching, so that a large number of sample data can be formed, and sample data of various types of true values can be obtained in a balanced manner, so as to reduce the imbalance between the sample data of the baseline intervention action and the sample data of the applied intervention action.
[0081] By predicting the gain values of the two situations: changing from applying a baseline intervention action to any other intervention action, or changing from applying any other intervention action to the baseline intervention action, the actual response results in the observed data can be adapted for training, reducing the difficulty of obtaining true values in sample data, reducing the cost of generating sample data, reducing the imbalance between different types of sample data, improving the representativeness of sample data, and improving the prediction accuracy of the intervention gain prediction model.
[0082] Optionally, the step of calculating the gain response result corresponding to each of the other intervention actions according to the gain value corresponding to each of the other intervention actions and the predicted response result generated by each element in the intervention action set includes: for each of the other intervention actions, calculating the difference between the first gain value corresponding to the other intervention action and the predicted response result of the element corresponding to the other intervention action, and determining it as the first response result corresponding to the other intervention action; for each of the other intervention actions, calculating the sum of the second gain value corresponding to the other intervention action and the predicted response result of the element corresponding to the other intervention action, and determining it as the second response result corresponding to the other intervention action. The first response result corresponding to the other intervention action and the second response result corresponding to the other intervention action are determined as the gain response result corresponding to the other intervention action.
[0083] Among them, the first gain value corresponding to other intervention actions refers to the gain value of the response result generated from applying the baseline intervention action to applying the intervention action. At this time, the element corresponding to the intervention action may refer to the intervention action. The predicted response result of the element corresponding to the intervention action is the response result of the sample user applying the intervention action on the basis of the baseline intervention action. The difference between the first gain value corresponding to other intervention actions and the predicted response result of the element corresponding to the intervention action may refer to the response result of the baseline intervention action after the gain of the intervention action is removed, and the first response result is the response result of applying the baseline intervention action. When the baseline intervention action is not applying the intervention action, the first response result is the response result of not applying the intervention action to the sample user.
[0084] Among them, the second gain value corresponding to other intervention actions refers to the gain value of the response result generated from applying the intervention action to the benchmark intervention action. At this time, the element corresponding to the intervention action may refer to the benchmark intervention action. The predicted response result of the element corresponding to the intervention action is the response result of the sample user changing to the benchmark intervention action on the basis of applying the intervention action. The sum of the second gain value corresponding to other intervention actions and the predicted response result of the element corresponding to the intervention action may refer to the response result obtained by adding the gain of applying the intervention action, and the response result is the response result of applying other intervention actions. Accordingly, the second response result is the response result of applying other intervention actions to the sample user.
[0085] The gain response results corresponding to other intervention actions include a first response result and a second response result. As mentioned above, the real observation data includes the response results of applying other intervention actions or benchmark intervention actions to sample users. Therefore, for the gain response results, the corresponding true value can be quickly queried from the business interaction observation data, so that a large amount of correct sample data can be formed.
[0086] In an example, the predicted responses for each element in the intervention action set are as follows:
[0087] μ0(x)=model0(Y 0 ~X 0 )
[0088] μ1(x)=model1(Y 1 ~X 1 )
[0089] …
[0090] μ t (x) = model m (Y m ~X m )
[0091] μ0 represents the predicted response result output by the model when treatment=0.
[0092] μ1 represents the predicted response result output by the model when treatment=1.
[0093] μ t It represents the predicted response result output by the model when treatment = t.
[0094] m: indicates the number of treatments.
[0095] Treatment = 0 is the baseline intervention action, and 0 means no intervention action is applied, that is, the baseline intervention action is no intervention action. Treatment = 1 ... treatment = t is other intervention actions.
[0096] The expected expression of the first gain value corresponding to each other intervention action is as follows:
[0097]
[0098]
[0099] …
[0100]
[0101] The expression on the left side of the identity symbol can be converted to the expression on the right side of the identity symbol. c The above formula is equivalent to the counterfactual expression of the control group. Correspondingly, τ c It means that when the control group is affected by the treatment, a treatment effect is generated. In other words, the gain value τ of the response result generated by the sample user who does not apply the intervention action to the intervention action is c , the expectation is equal to the difference between the observed response of the sample user without intervention and the predicted response of the sample user without intervention.
[0102] represents the i-th sample user who does not perform intervention actions, Represents the observed response result of the i-th sample user without performing any intervention action.
[0103] It represents the gain value of the response result from not applying intervention action to applying intervention action treatment=1 when treatment=1. It represents the predicted response result generated by the intervention action treatment=1 for the i-th sample user who does not perform the intervention action.
[0104] It represents the gain value of the response result from not applying intervention action to applying intervention action treatment=2 when treatment=2. It represents the predicted response result generated by the intervention action treatment=2 for the i-th sample user who does not perform the intervention action.
[0105] It represents the gain value of the response result from not applying any intervention action to applying the intervention action treatment = m when treatment = m. It represents the predicted response result generated by the intervention action treatment=m for the i-th sample user who does not perform the intervention action.
[0106] The expected expression of the second gain value corresponding to each intervention action is as follows:
[0107]
[0108]
[0109] …
[0110]
[0111] Among them, τ t represents the gain (treatment effect) produced by not applying the intervention action (treatment). The above formula is equivalent to the counterfactual expression of the experimental group (treatment group). Correspondingly, τ t It is used to represent the treatment effect when the treatment group is not affected by the treatment. In other words, the gain value τ of the response result from applying the intervention action to not applying the intervention action t , the expectation is equal to the sum of the observed response results of the sample user applying the intervention action and the predicted response results of the sample user applying the intervention action without applying the intervention action.
[0112] It represents the gain value of the response result from applying the intervention action treatment=1 to not applying the intervention action when treatment=1. represents the sample user with the i-th intervention action treatment = 1, It represents the predicted response result of the sample user with intervention action treatment = 1 without intervention action. Represents the observed response result of the intervention action performed by the i-th sample user.
[0113] It represents the gain value of the response result from applying the intervention action treatment=2 to not applying the intervention action when treatment=2. represents the sample user with the intervention action treatment = 2, It represents the predicted response result of the sample user with intervention action treatment = 2 without intervention action. Represents the observed response result of the intervention action performed by the i-th sample user.
[0114] It represents the gain value of the response result from applying the intervention action treatment=m to not applying the intervention action when treatment=m. represents the sample user who has the intervention action treatment=m for the ith time, It represents the predicted response result of the sample user who applies the intervention action treatment = m without applying the intervention action. Represents the observed response result of the intervention action performed by the i-th sample user.
[0115] By calculating the response results of applying other intervention actions to sample users and the response results of applying baseline intervention actions to sample users, the response results of applying other intervention actions to sample users and the response results of applying baseline intervention actions to sample users that can be collected in real observation data can be matched, thereby quickly generating a large amount of correct sample data and improving the utilization rate of observation data. The real response results in the observation data can be adapted for training, and the sample true values can be extracted from the collected observation data to improve the quality of the generated samples and reduce the difficulty of obtaining the true values in the sample data, thereby reducing the training cost of the intervention gain prediction model.
[0116] Optionally, the calculation of the difference between the gain response result corresponding to each of the other intervention actions and the corresponding observed response result in the business interaction observation data to obtain the observed gain loss includes: calculating the first gain loss through the intervention gain model according to the difference between the first response result corresponding to each of the other intervention actions and the corresponding observed response result in the business interaction observation data; calculating the second gain loss through the intervention gain model according to the difference between the second response result corresponding to each of the other intervention actions and the corresponding observed response result in the business interaction observation data; and determining the observed gain loss based on the first gain loss and the second gain loss.
[0117] The first gain loss may be the difference between the predicted result and the true value of the response result of applying the benchmark intervention action to the sample user. The difference calculated by the first response result corresponding to each intervention action may be accumulated to obtain the first gain loss. The second gain loss may be the difference between the predicted result and the true value of the response result of applying other intervention actions to the sample user. The difference calculated by the second response result corresponding to each intervention action may be accumulated to obtain the second gain loss. The sum of the first gain loss and the second gain loss may be calculated to obtain the observed gain loss.
[0118] The first gain loss is equivalent to the loss of the control group, and the second gain loss is equivalent to the loss of the experimental group.
[0119] By calculating the first gain loss according to the difference between the first response result and the corresponding observed response result, calculating the second gain loss according to the difference between the second response result and the corresponding observed response result, and calculating the observed gain loss according to the first gain loss and the second gain loss, the difference between the predicted result and the observed data can be obtained as the loss, and the parameters of the intervention gain model can be adjusted. The more easily obtained observed data can be used to adjust the parameters. In the scenario where observed data is used for modeling, the gain accuracy of the output of the intervention gain model can be improved.
[0120] Optionally, the first gain loss is calculated based on the difference between the first response result corresponding to each of the other intervention actions and the corresponding observed response result in the business interaction observation data, including: for the first response result corresponding to each other intervention action, querying the observed response result corresponding to the other intervention action in the business interaction observation data; for each other intervention action, calculating the difference between the first response result corresponding to the other intervention action and the observed response result corresponding to the other intervention action; summing and averaging the calculated differences of the first response results corresponding to the other intervention actions to calculate the first gain loss.
[0121] For the first response result, the observed response result corresponding to the intervention action is the response result when no intervention action is applied to the sample user. The difference values calculated for the first response results corresponding to each intervention action can be summed and averaged.
[0122] Exemplarily, the business interaction observation data may be first searched for observation data corresponding to the intervention action and the first response result, where the observation data is observation data without applying the intervention action to the sample user, and the observation response result is calculated based on the observation data and the user business information.
[0123] In one example, the observation data is the behavior data of sample users who do not issue red envelopes. The parameter values of service-related parameters are queried or counted from the behavior data and used as the observation response result. For example, query whether the sample user has the behavior of ordering takeout, or count the number of takeouts ordered by the sample user.
[0124] Based on the following formula, the first gain loss is calculated for the i-th sample user:
[0125]
[0126] Among them, if the observed response result corresponding to the first response result is empty, the difference between the first response result corresponding to the intervention action and the observed response result corresponding to the intervention action is 0, that is, when the true value corresponding to the first response result corresponding to the intervention action is not queried, the loss of the first response result corresponding to the intervention action is not counted.
[0127] By querying the observed response results of each first response result, calculating the difference between the two, and accumulating the difference of each first response result to obtain the first gain loss, the gain loss of the control group affected by the intervention action can be converted into the first gain loss, and the intervention gain model can be adjusted based on the gain loss of the control group affected by the intervention action, so that the intervention gain model learns the gain affected by the intervention.
[0128] Optionally, the second gain loss is calculated based on the difference between the second response result and the corresponding observed response result in the business interaction observation data, including: for the second response result corresponding to each other intervention action, querying the observation data corresponding to the intervention action in the business interaction observation data; for each other intervention action, based on the observed response result corresponding to the other intervention action; for each other intervention action, calculating the difference between the second response result corresponding to the other intervention action and the observed response result corresponding to the intervention action; summing and averaging the differences calculated for the second response results corresponding to each of the other intervention actions to calculate the second gain loss.
[0129] For the second response result, the observation data corresponding to the intervention action is the response result of applying the intervention action to the sample user.
[0130] Based on the following formula, the second gain loss is calculated for the i-th sample user:
[0131]
[0132] If the observed response result corresponding to the second response result is empty, the difference between the second response result corresponding to the intervention action and the observed response result corresponding to the intervention action is 0, that is, when the true value corresponding to the second response result corresponding to the intervention action is not queried, the loss of the second response result corresponding to the intervention action is not counted.
[0133] By querying the observed response results of the response for each second response result, calculating the difference between the two, and accumulating the difference of each second response result to obtain the second gain loss, it is possible to convert the gain loss of the experimental group that is not affected by the intervention action into the second gain loss, and adjust the intervention gain model based on the gain loss of the experimental group that is not affected by the intervention action, so that the intervention gain model learns to remove the gain of the intervention.
[0134] Optionally, the training method of the intervention gain model also includes: determining the user's tendency action probability according to the user business information and the intervention action set through the intervention gain model; calculating the tendency score loss according to the user's tendency action probability and each element in the intervention action set through the intervention gain model; adjusting the parameters of the intervention gain model according to the observed gain loss, including: adjusting the intervention gain model parameters according to the observed gain loss and the tendency score loss.
[0135] The user's propensity to act may refer to the probability that the user is likely to be affected by the intervention action. For example, a PSM (Propensity Score Matching) network may be used to evaluate the individual's propensity to the intervention action.
[0136] Loss psm =-∑ j∈m t j log (p j )
[0137] Among them, t j Indicates whether the sample user in the business interaction observation data is affected by the jth element. If the sample user is affected by the jth element, t j is 1, if the sample user is not affected by the jth element, t j is 0. j It represents the probability that the sample user tends to be influenced by the jth element.
[0138] The parameters of the intervention gain model can be adjusted according to the sum of the observed gain loss and the propensity score loss. Specifically, the sum of the observed gain loss and the propensity score loss is calculated to obtain the total loss. The parameters of the intervention gain model are adjusted with the goal of minimizing or converging the total loss.
[0139] It is usually expected that the samples of the experimental group and the control group are matched. For each sample of the experimental group, a user with a similar score is identified in the control group, so that homogeneous experimental and control groups can be obtained. Inter-group comparisons can be made at will, and then individuals in the control group who are as similar as possible to the experimental group in terms of observable characteristics can be found for matching. In this way, under the given observable characteristic variables, the individuals in the treatment group and the control group can be made as similar as possible, thereby alleviating the selection bias problem of the treatment effect.
[0140] By increasing the propensity score loss to adjust the parameters of the intervention gain model, the error caused by selection bias can be reduced, thereby improving the intervention gain prediction accuracy of the intervention gain model.
[0141] Optionally, determining the user's preferred action probability based on the user business information and the intervention action set includes: generating an initial probability based on the user business information and the intervention action set; and fusing the first gain value and the second gain value with the initial probability to obtain the user's preferred action probability.
[0142] The initial probability may refer to the probability that the user tends to be affected by the intervention action. The first gain value and the second gain value are used to add the gain prediction information of each intervention action to the initial probability. The first gain value and the second gain value are fused with the initial probability to obtain the user's tendency action probability, which can optimize the initial probability based on the gain prediction information of multiple intervention actions to obtain the user's tendency action probability, so that the user's tendency action probability is increased with the content of the gain prediction information of multiple intervention actions, thereby adapting to the application scenario of multiple intervention actions.
[0143] In one example, as in the previous example, by calculating the propensity score loss through the PSM Network, the intervention gain model can be expanded to be applicable to multi-treatment business scenarios through a special gate. For example, the gate is used to perform weighted calculations on the sub-probabilities corresponding to each intervention action in the initial probability according to multiple intervention actions, and the new probability corresponding to each intervention action is obtained as the user's propensity action probability.
[0144] By adjusting the initial probability of production according to the first gain value and the second gain value to obtain the user's preferred action probability, the information of multiple intervention actions can be integrated into the user's preferred action probability, which can increase the adaptability of the user's preferred action probability to the application scenarios of multiple intervention actions.
[0145] Optionally, determining the predicted response results of the sample users for each element in the intervention action set based on the user business information includes: determining shared features based on the user business information and the intervention action set; determining the predicted response results of the sample users for each element in the intervention action set based on the shared features; determining the gain value corresponding to each of the other intervention actions based on the user business information and the intervention action set includes: determining the gain value corresponding to each of the other intervention actions based on the shared features.
[0146] Shared features can refer to features extracted from user service information and intervention action sets. As in the previous example, the predicted response results and gain values can be processed by different networks, and the input features of each network are shared features, that is, each network shares input features. In addition, the probability of user-inclined actions can also be calculated based on shared features.
[0147] By using shared features to calculate gain values and predict response results, the gain value prediction network and the response result prediction network share input features, and a common feature representation is learned under limited training data. This feature representation can not only help the model better identify known categories, but also generalize to unseen categories to a certain extent, thereby reducing the risk of overfitting.
[0148] Optionally, determining the shared features according to the user service information and the intervention action set includes: performing data enhancement on the intervention action set, and determining the shared features according to the user service information.
[0149] Performing data enhancement on the intervention action set may refer to increasing the weight of the intervention action set in the shared features so that the shared features have richer information about the intervention actions and are more representative of the intervention actions.
[0150] Specifically, data enhancement of the intervention action set and determination of shared features based on user business information may include: determining initial features based on the intervention action set and the user business information, enhancing the intervention action set in the initial features to obtain shared features; or, enhancing the intervention action set, and determining shared features based on the enhanced intervention action set and user business information.
[0151] In one example, initial features are determined according to the intervention action set and user service information, and the initial features and the intervention action set are fused to obtain shared features.
[0152] In one example, enhancing the intervention action set may include increasing the weight of the intervention action set, weighted fusion of the intervention action set and user service information using the increased weight to obtain fused information, and determining shared features based on the fused information. Increasing the weight of the intervention action set may include weighting each feature extraction layer in the shared network using a gate.
[0153] By enhancing the set of intervention actions in the shared features, the information of the intervention actions can be enriched in the shared features, and the representativeness of the set of intervention actions can be increased. Therefore, subsequent processing based on the shared features can make full use of the information of the intervention actions, so that the intervention gain model can learn the information of the intervention actions well.
[0154] Optionally, data enhancement is performed on the intervention action set, and shared features are determined based on the user business information, including: generating initial features based on the user business information and the intervention action set; and fusing the intervention action set and the initial features to generate shared features.
[0155] The initial feature may refer to a feature directly extracted from the user service information and the intervention action set. The intervention action set and the initial feature fusion may be feature extraction of the intervention action set, and the obtained feature is fused with the initial feature, which may increase the weight ratio of the feature of the intervention action set.
[0156] By fusing the initial features determined according to the intervention action set with the intervention action set again to obtain shared features, it is possible to increase the weight proportion of the intervention action set in the shared features, make full use of the information of the intervention actions, and thus improve the accuracy of the intervention gain model in predicting the gain of the intervention actions.
[0157] Figure 2 This is a flow chart of an intervention gain prediction method disclosed in an embodiment of the present disclosure. This embodiment can be applied to the case where a pre-trained intervention gain model is used to predict intervention gain. The method of this embodiment can be executed by a training device for an intervention gain model, which can be implemented in software and / or hardware and is specifically configured in an electronic device with certain data computing capabilities, which can be a server device.
[0158] S201. Obtain user service information and an intervention action set of a target user; the intervention action set includes a baseline intervention action and at least one other intervention action.
[0159] After authorization by the target user, obtain the target user's user business information. According to the business scenario, configure at least one other intervention action, and combine the baseline intervention action to generate an intervention action set.
[0160] S202. Input the user service information and the intervention action set into a pre-trained intervention gain model; the intervention gain model is trained by the intervention gain model training method described in any embodiment of the present disclosure.
[0161] S203: Determine, by means of the intervention gain model, a predicted response result corresponding to each element in the intervention action set according to the user service information and the intervention action set.
[0162] The intervention gain model can process user business information and intervention action sets, and can obtain the predicted response results of each element. Or the intervention gain model only outputs the best element in the predicted response results, and the predicted response result of the element. Exemplarily, the intervention action set includes: not pushing information, pushing information of interest, and pushing information with high popularity. After processing by the intervention gain model, it can be determined that the number of resource feedbacks corresponding to not pushing information is 1, the number of resource feedbacks corresponding to pushing information of interest is 2, and the number of resource feedbacks corresponding to pushing information with high popularity is 3. Among them, the best predicted response result at this time may refer to the largest number of resource feedbacks, that is, the number of resource feedbacks is 3. At this time, the intervention gain model can output all resource feedback numbers, or only output the identifier of the information with high push popularity, and the number of resource feedbacks is 3.
[0163] Optionally, the intervention gain prediction method also includes: determining the user's tendency to act probability based on the user's business information and the intervention action set through the intervention gain model; and determining the intervention gain result of the target user based on the user's tendency to act probability and the predicted response results corresponding to each element in the intervention action set.
[0164] The intervention gain model can also determine the probability of the user's preferred action based on the user's business information and the intervention action set. The user's preferred action probability can include the probability corresponding to each element. The intervention gain result is used to describe the comprehensive effect of the intervention action set on the response result of the target user.
[0165] In one example, according to the tendency probability of the target user for each element in the intervention action set described in the user tendency action probability, the gain value corresponding to each element is weighted and summed to obtain the intervention gain result of the target user. The intervention gain result includes the gain result of at least one other intervention action. The gain result of each intervention action is the sum of the first result and the second result. Among them, the first result is the product of the first gain value of the intervention action and the user tendency action probability of the benchmark intervention action, and the second result is the sum of the products of the second gain value of each intervention action and the user tendency action probability of the intervention action.
[0166] The intervention gain result of the target user can be calculated based on the following formula:
[0167]
[0168] Among them, score k represents the gain result of the kth intervention action. p0 represents the probability that the sample user prefers the baseline intervention action. It represents the first gain value of the kth intervention action, that is, the gain value of the response result generated from the baseline intervention action to the application of the intervention action treatment=k. Represents the second gain value of the j-th element, that is, the gain value of the response result generated from applying the intervention action treatment=j to the baseline intervention action.
[0169] By calculating the probability and gain value of each element through the intervention gain model, and calculating the intervention gain result based on the probability and gain value of each element, it is possible to calculate the comprehensive effect of the intervention action on the target user, so as to comprehensively evaluate the intervention effect of the intervention action and increase the content of the predicted results of the intervention gain to assist users in better utilizing the predicted response results.
[0170] Optionally, the intervention action set includes: a distribution method for at least one service of the target traffic; and the predicted response result includes an interaction result between the user and each service.
[0171] Different intervention actions are different allocation methods, and the specific different allocation methods refer to different allocated services. The base intervention action can be any allocation method.
[0172] In one example, the services to be allocated include services A, B, and C. The allocation method may include allocating the target traffic to service A, allocating the target traffic to service B, and allocating the target traffic to service C. The baseline intervention action is allocating the target traffic to service A, and the different intervention actions are allocating the target traffic to service B and allocating the target traffic to service C. The interaction result is LTV (Lifetime Value, i.e., user lifetime value).
[0173] By setting different traffic allocation methods as intervention actions and setting the predicted response results as the interaction results between users and various services, we determine the intervention scenarios that can adapt to the traffic allocation service, apply the intervention gain model, obtain the predicted response results of the intervention of traffic allocation to different services, analyze the impact of different business allocation methods on users, and achieve reasonable allocation of traffic resources.
[0174] Optionally, the intervention action set includes: at least one resource provision operation; and the predicted response result includes an interaction result between the user and a resource-associated service.
[0175] Resources are used to provide to users. Different resource provision operations may include at least one of: different resource amounts, different resource types, and different operations. For example, a baseline intervention action may refer to providing a resource amount of 0, or a resource provision operation being empty (i.e., no resource provision operation is performed). Different intervention actions may have at least one of different resource amounts, different resource types, and different resource provision operations. The interaction result may be the result of a user's behavior when using the services provided by the resource-related business. The interaction result may be the user's behavior itself, or the result caused by the behavior.
[0176] In one example, in the marketing field, the resource provision operation is to issue red envelopes. Different face values can be defined as different treatments. By calculating the treatment effects brought by different treatments and designing corresponding red envelope face values, marketing costs can be reduced. The resource-related business can be the food delivery business, and the interaction result, i.e. the treatment effect, can be the number of times the food delivery business is used or the red envelope conversion rate.
[0177] In one example, in the area of user growth: different incentive amounts can be defined as different treatments, and the interaction result is the retention improvement rate.
[0178] In one example, in the medical field: different drugs are defined as different treatments, and the interaction result is the cure rate.
[0179] By setting the intervention action set to include at least one resource provision operation and limiting the predicted response result to the interaction result between the user and the resource-related business, it is possible to adapt to multiple resource provision intervention scenarios, apply the intervention gain model, obtain the predicted response results of the intervention of different resource provision operations, analyze the impact of different resource provision operations on users, reduce resource waste and reduce interference to users.
[0180] According to the technical solution disclosed in the present invention, by obtaining the user business information of the target user and the intervention action set and inputting them into a pre-trained intervention gain model, the predicted response result output by the intervention gain model is obtained, and end-to-end intervention gain prediction is achieved. This can reduce the accumulation of model errors and improve the accuracy of intervention prediction. In addition, the intervention action set can set multiple intervention actions to adapt to multi-intervention scenarios. It can achieve simultaneous intervention gain prediction of multiple interventions through only one model and one input, thereby reducing the complexity of multi-intervention prediction.
[0181] In an optional embodiment, if Figure 3 As shown, the embodiment of the present disclosure proposes an end-to-end uplift model algorithm based on a neural network.
[0182] The intervention gain prediction model of the disclosed embodiment includes: a feature extraction network, a response network ResponseNetwork, a pseudo treatment effect network PTE Network, and a propensity score matching network PSM Network. Among them, the feature extraction network may include a multilayer perceptron (Multilayer Perceptron, MLP). Among them, the Response Network may include an MLP, and the PTE Network includes two MLPs, one MLP is used to calculate the first gain value, and one MLP is used to calculate the second gain value. The PSM Network includes an MLP and a SoftMax Layer (multi-classification layer). Among them, each MLP is independent of each other.
[0183] like Figure 4 As shown in the figure, three networks are designed in the overall model architecture to realize their respective functions;
[0184] Response Network: used to learn the real Label;
[0185] PTE Network: used to learn treatment effect.
[0186] PSM Network: used to estimate an individual's tendency towards treatment.
[0187] Treatment is input as an embedding into the Response Network, and the treatment embedding layer (interference action feature layer) is moved up, that is, Figure 4 In the figure, the treatment embedding layer and the shared network are placed in the same layer. The output of the treatment embedding layer and the output of the shared network are both used as the input of the response network, thereby increasing the weight of the treatment embedding in the input of the response network. This can solve the problem that the S-learner may not be able to fully utilize the treatment information due to the large feature dimension. The design here can have many operations, such as using gate to weight each layer, and the core purpose is the same.
[0188] In the design of PSM Network, a special gate is used to expand the processing of X-Learner to be applicable to multitreatment business scenarios.
[0189] Since the uplift model is more prone to overfitting than other supervised models, the overall model structure adopts the underlying shared embedding method. A shared network (used to extract features) is used to extract features from the input data to obtain shared features, which are then input into the Response Network, PTE Network, and PSM Network to achieve the sharing of underlying features among the Response Network, PTE Network, and PSM Network.
[0190] Loss will include two types of loss, two counterfactual losses and one PSM loss:
[0191] Loss treatment Used to indicate that treatment effect loss occurs when the control group is affected by treatment.
[0192] Loss control Used to represent the treatment effect loss generated when the treatment group is not affected by the treatment.
[0193] Loss psm Used to indicate the tendency score Loss.
[0194] Specifically, the predicted response results of each element in the intervention action set output by the Response Network are as follows:
[0195] μ0(x)=model0(Y 0 ~X 0 )
[0196] μ1(x)=model1(Y 1 ~X 1 )
[0197] …
[0198] μ t (x) = model m (Y m ~X m )
[0199] The expected expression of the first gain value corresponding to each other intervention action output by PTE Network is as follows:
[0200]
[0201]
[0202] …
[0203]
[0204] The expected expression of the second gain value corresponding to each other intervention action output by PTE Network is as follows:
[0205]
[0206]
[0207] …
[0208]
[0209] Based on the following formula, for the i-th sample user, calculate the first gain loss Loss control :
[0210]
[0211] Based on the following formula, for the i-th sample user, the second gain loss Loss is calculated: treatment :
[0212]
[0213] PSM Network output j .
[0214] Based on the following formula, for the i-th sample user, calculate the propensity score loss Loss psm :
[0215] Loss psm =-∑ j∈m t j log (p j )
[0216] When applied, the intervention gain model can calculate the intervention gain results in addition to outputting the predicted response results of each element in the intervention action set. The intervention gain results are calculated based on the following formula:
[0217]
[0218] The intervention gain model in the embodiment of the present disclosure is generally applicable. In theory, it can be used in any application scenario where the uplift model can be used to solve problems. It mainly solves the following two application scenarios in business scenarios, so it has a good effect on similar business problems:
[0219] It is impossible to collect RCT samples in business scenarios, and only observational data can be used for modeling.
[0220] Flexible expansion or additional treatment is required in business scenarios.
[0221] Specific areas of application include:
[0222] Marketing: For example, when a takeaway business distributes red envelopes, different face values can be defined as different treatments. By calculating the treatment effects brought about by different treatments and designing corresponding red envelope face values, marketing costs can be reduced. Figure 5 As shown, you can receive red envelopes on the food ordering page.
[0223] User growth area: The commonly used user growth methods in the industry are basically to increase user retention by issuing incentives. The modeling between the amount of incentives issued and the retention improvement rate can also be estimated using the intervention gain model. Figure 6 As shown, by watching the video, you can enter the page for issuing incentives from the incentive entrance.
[0224] like Figure 7 As shown in the figure, in the field of efficiency evaluation: in order to support some new businesses or explore new business forms, a part of "free traffic" is often allocated to explore traffic for some businesses. The question that follows is how to evaluate the efficiency of the use of this free traffic in various businesses. Figure 7 As shown:
[0225] S701. The service obtains a certain flow.
[0226] The service competes for a certain traffic flow through a virtual budget.
[0227] S702: Collect the interaction results between the user and the service.
[0228] After the user interacts with the business, a piece of sample data is generated, and the sample is used to train the intervention gain model.
[0229] S703, training intervention gain model.
[0230] S704: Predict response results.
[0231] After the intervention gain model is trained, the intervention gain model is used to estimate the processing effect of the traffic if it is competed for by other services, that is, to obtain the predicted response result.
[0232] S705: Business virtual budget allocation strategy.
[0233] Then, according to the treatment effect, the virtual budget value of a certain business in the next quarter is calculated through the designed virtual budget allocation strategy.
[0234] S706. Refresh the business virtual budget.
[0235] After refreshing the virtual budget of the service, the service can continue to participate in the traffic auction in the next quarter with the obtained virtual budget. After executing S706, the service can continue to be used or other services can be selected to obtain the traffic.
[0236] In the existing intervention gain model:
[0237] T-Learner:
[0238] The two models are trained using the experimental group data and the control group data respectively, and the uplift score is obtained by subtracting the estimated values of the two models.
[0239]
[0240]
[0241]
[0242] The disadvantage is that the data is not fully utilized, the difference between the two groups (i.e., the lift signal) is not well fitted, and the model error is easily amplified. Assuming that the accuracy of a single model is 90%, the final lift prediction accuracy is only 81%.
[0243] S-Learner:
[0244] The single model method uses the treatment variable t as the input feature for training. When predicting, the same sample feature is input multiple times, and each time only a different t value is changed. Generally speaking, t = 0 represents the control group, …represents each experimental group.
[0245]
[0246]
[0247] The disadvantage is that when the sample features are too large, the features of a single t may make the prediction ability of Y insufficient and fail to fully utilize the treatment information, especially when t is a high-dimensional feature, which often leads to an uplift score of 0.
[0248] X-Learner:
[0249] X-Learner is an optimization based on T-Learner, which solves the problem of insufficient data utilization of T-Learner, and also has better effects when the data of the Treatment group and the Control group are unbalanced. X-Learner estimates the feedback signals of the experimental group and the control group in the first step, and then uses the experimental group model to calculate the counterfactual elasticity of the control group in the second step, and uses the control group model to calculate the counterfactual elasticity of the experimental group.
[0250] The disadvantage is that the process is relatively complicated. When processing a single treatment, four models need to be maintained. It is not suitable for multi-treatment application scenarios (the number of models will increase linearly with the increase in treatment), and the problem of the T-Learner model is not solved, and there is still cumulative error.
[0251] The embodiment of the present disclosure uses an end-to-end model, and in the end-to-end model, the counterfactual treatment effect is learned through the PTE Network, so as to avoid learning the counterfactual treatment effect by adding models, reduce the problem of cumulative errors of multiple models of T-Learner, thereby avoiding the cumulative errors between multiple models and reducing model errors; and by increasing the weight proportion of the intervention action set in the features, the problem that S-Learner cannot fully utilize the information of the intervention action is solved, and the information of the treatment is fully utilized, thereby increasing the ability to estimate the intervention gain; by setting multiple intervention actions and performing learning optimization for the response of each intervention action, the problem that X-Learner is not suitable for multi-treatment application scenarios is solved, and one or more treatments can be added quickly, conveniently and flexibly. Only minor adjustments need to be made under the current model to conveniently expand and support new treatments, and the cumulative errors can be effectively reduced. More importantly, the model effect on the observed data is more stable; and the end-to-end training will not increase the number of models due to the increase in treatments, thereby maintaining the cost, and when the data of the Treatment group and the Control group are unbalanced or when only the observed data is used for modeling, the intervention gain model obtained by training performs better.
[0252] According to an embodiment of the present disclosure, Figure 8 1 is a structural diagram of a training device for an intervention gain model in an embodiment of the present disclosure, and the present disclosure embodiment is applicable to the case of training an intervention gain model. The device is implemented by software and / or hardware, and is specifically configured in an electronic device with certain data computing capabilities.
[0253] like Figure 8The training device 800 of an intervention gain model shown in the figure comprises: a sample acquisition module 801, an intervention gain model 802 and a model parameter adjustment module 803.
[0254] The sample acquisition module 801 is used to acquire user service information of a sample user, service interaction observation data of the sample user, and an intervention action set; the intervention action set includes a baseline intervention action and at least one other intervention action;
[0255] An intervention gain model 802, used to determine, based on the user service information and the intervention action set, a predicted response result generated by the sample user for each element in the intervention action set;
[0256] The intervention gain model 802 is used to determine the gain value corresponding to each of the other intervention actions according to the user service information and the intervention action set;
[0257] The intervention gain model 802 is used to calculate the gain response result corresponding to each of the other intervention actions according to the gain value corresponding to each of the other intervention actions and the predicted response result generated by each element in the intervention action set;
[0258] The intervention gain model 802 is used to calculate the difference between the gain response result corresponding to each of the other intervention actions and the corresponding observation response result in the business interaction observation data to obtain the observation gain loss;
[0259] The model parameter adjustment module 803 is used to adjust the parameters of the intervention gain model according to the observed gain loss.
[0260] According to the technical solution disclosed in the present invention, by predicting the predicted response results generated by sample users for each intervention element in the intervention action set and the gain value of each intervention action, and fusing the gain value of each intervention action with the predicted response result of each element, a gain response result based on the intervention action gain can be obtained, and the observed response result corresponding to the gain response result is queried in the observation data as the true value, and the difference between the gain response result and the true value is calculated to obtain the observed gain loss. The predicted gain response result can be associated with the observation data, and the true value corresponding to the predicted gain response result can be directly obtained from the observation data, thereby reducing the sample acquisition cost. In addition, by implementing intervention gain prediction through an end-to-end model, the accumulation of model errors can be reduced, thereby improving the accuracy of intervention gain prediction. At the same time, multiple intervention actions can be set in the intervention action set to realize an intervention gain prediction scenario that is adapted to multiple interventions.
[0261] Furthermore, the intervention gain model is used to:
[0262] Determine, according to the user service information and the intervention action set, a first gain value corresponding to each of the other intervention actions; the first gain value corresponding to the other intervention action is a gain value of a response result generated from a reference intervention action to applying the other intervention action;
[0263] Determine, according to the user service information and the intervention action set, a second gain value corresponding to each of the other intervention actions; the second gain value corresponding to the other intervention action is a gain value of a response result generated from applying the other intervention action to the benchmark intervention action;
[0264] The gain value corresponding to each of the other intervention actions is determined according to the first gain value corresponding to each of the other intervention actions and the second gain value corresponding to each of the other intervention actions.
[0265] Furthermore, the intervention gain model is used to:
[0266] For each of the other intervention actions, calculating a difference between a first gain value corresponding to the other intervention action and a predicted response result of the element corresponding to the other intervention action, and determining the difference as a first response result corresponding to the other intervention action;
[0267] For each of the other intervention actions, a sum of the second gain value corresponding to the other intervention action and the predicted response result of the element corresponding to the other intervention action is calculated, and determined as the second response result corresponding to the other intervention action.
[0268] The first response result corresponding to the other intervention action and the second response result corresponding to the other intervention action are determined as the gain response result corresponding to the other intervention action.
[0269] Furthermore, the model parameter adjustment module includes:
[0270] A first gain loss calculation unit, configured to calculate a first gain loss according to a difference between a first response result corresponding to each of the other intervention actions and a corresponding observed response result in the business interaction observation data;
[0271] A second gain loss calculation unit, configured to calculate a second gain loss according to a difference between a second response result corresponding to each of the other intervention actions and a corresponding observation response result in the service interaction observation data;
[0272] The gain loss fusion unit is used to determine an observed gain loss according to the first gain loss and the second gain loss.
[0273] Furthermore, the first gain loss calculation unit includes:
[0274] A first observation result acquisition subunit is used to query the observation response result corresponding to each other intervention action in the business interaction observation data for the first response result corresponding to the other intervention action;
[0275] A first observation loss calculation subunit, configured to calculate, for each other intervention action, a difference between a first response result corresponding to the other intervention action and an observed response result corresponding to the other intervention action;
[0276] The first intervention loss fusion subunit is used to sum and average the differences of the first response results corresponding to the other intervention actions to calculate the first gain loss.
[0277] Furthermore, the second gain loss calculation unit includes:
[0278] A second observation result acquisition subunit is used to query the observation response result corresponding to each other intervention action in the business interaction observation data for the second response result corresponding to the other intervention action;
[0279] A second observation loss calculation subunit is used to calculate, for each other intervention action, a difference between a second response result corresponding to the other intervention action and an observed response result corresponding to the other intervention action;
[0280] The second intervention loss fusion subunit is used to sum and average the differences of the second response results corresponding to the other intervention actions to calculate the second gain loss.
[0281] Furthermore, the intervention gain model is also used to:
[0282] Determining the user's preferred action probability according to the user service information and the intervention action set;
[0283] Calculating the tendency score loss according to the user's tendency action probability and each element in the intervention action set;
[0284] The model parameter adjustment module comprises:
[0285] The tendency loss adjustment unit is used to adjust the intervention gain model parameters according to the observed gain loss and the tendency score loss.
[0286] Furthermore, the intervention gain model is used to:
[0287] generating an initial probability according to the user service information and the intervention action set;
[0288] The first gain value and the second gain value are combined with the initial probability to obtain the user's preferred action probability.
[0289] Furthermore, the intervention gain model is used to:
[0290] Determining a shared feature according to the user service information and the intervention action set;
[0291] Determining, according to the shared features, a predicted response result generated by the sample user for each element in the intervention action set;
[0292] The gain value corresponding to each of the other intervention actions is determined according to the shared feature.
[0293] Furthermore, the intervention gain model is used to:
[0294] Data enhancement is performed on the intervention action set, and shared features are determined according to the user business information.
[0295] Furthermore, the intervention gain model is used to:
[0296] generating initial features according to the user service information and the intervention action set;
[0297] The intervention action set and the initial features are fused to generate shared features.
[0298] The training device for the intervention gain model mentioned above can execute the training method for the intervention gain model provided by any embodiment of the present disclosure, and has the corresponding functional modules and beneficial effects for executing the training method for the intervention gain model.
[0299] According to an embodiment of the present disclosure, Fig. 9 is a structural diagram of an intervention gain prediction device in an embodiment of the present disclosure, and the present disclosure embodiment is applicable to the case where a pre-trained intervention gain model is used to predict intervention gain. The device is implemented using software and / or hardware, and is specifically configured in an electronic device with certain data computing capabilities.
[0300] like Fig. 9 The intervention gain prediction device 900 shown in the figure comprises: a module for obtaining data to be predicted 901, a module for inputting data to be predicted 902, and the intervention gain model 903.
[0301] The module 901 for obtaining data to be predicted is used to obtain user service information and an intervention action set of a target user; the intervention action set includes a baseline intervention action and at least one other intervention action;
[0302] The to-be-predicted data input module 902 is used to input the user service information and the intervention action set into a pre-trained intervention gain model; the intervention gain model is trained by the intervention gain model training method as described in any embodiment of the present disclosure;
[0303] The intervention gain model 903 is used to determine the predicted response result corresponding to each element in the intervention action set according to the user service information and the intervention action set.
[0304] According to the technical solution disclosed in the present invention, by obtaining the user business information of the target user and the intervention action set and inputting them into a pre-trained intervention gain model, the predicted response result output by the intervention gain model is obtained, and end-to-end intervention gain prediction is achieved. This can reduce the accumulation of model errors and improve the accuracy of intervention prediction. In addition, the intervention action set can set multiple intervention actions to adapt to multi-intervention scenarios. It can achieve simultaneous intervention gain prediction of multiple interventions through only one model and one input, thereby reducing the complexity of multi-intervention prediction.
[0305] Furthermore, the intervention gain model is also used to:
[0306] Determining the user's preferred action probability according to the user service information and the intervention action set;
[0307] The intervention gain result of the target user is determined according to the user's tendency action probability and the predicted response result corresponding to each element in the intervention action set.
[0308] Furthermore, the intervention action set includes: a distribution method for at least one service of the target traffic; and the predicted response result includes an interaction result between the user and each service.
[0309] Furthermore, the intervention action set includes: at least one resource provision operation; and the predicted response result includes an interaction result between the user and the resource-associated service.
[0310] The above-mentioned intervention gain prediction device can execute the intervention gain prediction method provided by any embodiment of the present disclosure, and has corresponding functional modules and beneficial effects for executing the intervention gain prediction method.
[0311] In the technical solution of the present disclosure, the collection, storage, use, processing, transmission, provision and disclosure of user personal information involved are in compliance with the provisions of relevant laws and regulations and do not violate public order and good morals.
[0312] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device, a readable storage medium and a computer program product.
[0313] Fig.10A schematic area diagram of an example electronic device 1000 that can be used to implement an embodiment of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present disclosure described and / or claimed herein.
[0314] like Fig.10 As shown, the device 1000 includes a computing unit 1001, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 1002 or a computer program loaded from a storage unit 1008 into a random access memory (RAM) 1003. In the RAM 1003, various programs and data required for the device 1000 instructions can also be stored. The computing unit 1001, the ROM 1002, and the RAM 1003 are connected to each other via a bus 1004. An input / output (I / O) interface 1005 is also connected to the bus 1004.
[0315] A number of components in the device 1000 are connected to the I / O interface 1005, including: an input unit 1006, such as a keyboard, a mouse, etc.; an output unit 1007, such as various types of displays, speakers, etc.; a storage unit 1008, such as a disk, an optical disk, etc.; and a communication unit 1009, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 1009 allows the device 1000 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.
[0316] The computing unit 1001 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the computing unit 1001 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The computing unit 1001 performs the various methods and processes described above, such as the training method of the intervention gain model or the intervention gain prediction method. For example, in some embodiments, the training method of the intervention gain model or the intervention gain prediction method may be implemented as a computer software program, which is tangibly included in a machine-readable medium, such as a storage unit 1008. In some embodiments, part or all of the computer program may be loaded and / or installed on the device 1000 via the ROM 1002 and / or the communication unit 1009. When the computer program is loaded into the RAM 1003 and executed by the computing unit 1001, one or more steps of the training method of the intervention gain model or the intervention gain prediction method described above may be performed. Alternatively, in other embodiments, the computing unit 1001 may be configured to execute the intervention gain model training method or the intervention gain prediction method in any other appropriate manner (for example, by means of firmware).
[0317] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard objects (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.
[0318] The program code for implementing the method of the present disclosure may be written in any combination of one or more programming languages. These program codes may be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that the program code, when executed by the processor or controller, implements the functions / instructions specified in the flow chart and / or the area diagram. The program code may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.
[0319] In the context of the present disclosure, a machine-readable medium may be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, device, or equipment. A machine-readable medium may be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.
[0320] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).
[0321] The systems and techniques described herein may be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.
[0322] A computer system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The relationship between the client and the server is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or a cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services. The server may also be a server of a distributed system, or a server combined with a blockchain.
[0323] Artificial intelligence is a discipline that studies how to use computers to simulate certain human thought processes and intelligent behaviors (such as learning, reasoning, thinking, planning, etc.). It includes both hardware-level and software-level technologies. Artificial intelligence hardware technologies generally include technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, and big data processing; artificial intelligence software technologies mainly include computer vision technology, speech recognition technology, natural language processing technology, machine learning / deep learning technology, big data processing technology, knowledge graph technology, and other major directions.
[0324] Cloud computing refers to a technology system that uses network access to elastically scalable shared physical or virtual resource pools. Resources can include servers, instruction systems, networks, software, applications, and storage devices, and can be deployed and managed on demand and in a self-service manner. Cloud computing technology can provide efficient and powerful data processing capabilities for the application of technologies such as artificial intelligence and blockchain, as well as model training.
[0325] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps recorded in this disclosure can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution provided by this disclosure can be achieved, and this document does not limit this.
[0326] The above specific implementations do not constitute a limitation on the protection scope of the present disclosure. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present disclosure shall be included in the protection scope of the present disclosure.
Claims
1. A training method for an intervention gain model, comprising: Obtaining user service information of sample users, service interaction observation data of the sample users, and a set of intervention actions; The intervention action set includes a baseline intervention action and at least one other intervention action; Determining, by means of an intervention gain model, a predicted response result generated by the sample user for each element in the intervention action set according to the user service information and the intervention action set; Determining, by means of the intervention gain model, a gain value corresponding to each of the other intervention actions according to the user service information and the intervention action set; Calculating the gain response result corresponding to each of the other intervention actions through the intervention gain model according to the gain value corresponding to each of the other intervention actions and the predicted response result generated by each element in the intervention action set; By using the intervention gain model, the difference between the gain response result corresponding to each of the other intervention actions and the corresponding observation response result in the business interaction observation data is calculated to obtain the observation gain loss; Parameters of the intervention gain model are adjusted according to the observed gain loss.
2. The method according to claim 1, wherein: The determining, according to the user service information and the intervention action set, a gain value corresponding to each of the other intervention actions includes: Determine, according to the user service information and the intervention action set, a first gain value corresponding to each of the other intervention actions; the first gain value corresponding to the other intervention action is a gain value of a response result generated from the reference intervention action to the application of the other intervention action; Determine, according to the user service information and the intervention action set, a second gain value corresponding to each of the other intervention actions; the second gain value corresponding to the other intervention action is a gain value of a response result generated from applying the other intervention action to the benchmark intervention action; The gain value corresponding to each of the other intervention actions is determined according to the first gain value corresponding to each of the other intervention actions and the second gain value corresponding to each of the other intervention actions.
3. The method according to claim 2, wherein: The calculating, according to the gain value corresponding to each of the other intervention actions and the predicted response result generated by each element in the intervention action set, the gain response result corresponding to each of the other intervention actions comprises: For each of the other intervention actions, calculating a difference between a first gain value corresponding to the other intervention action and a predicted response result of the element corresponding to the other intervention action, and determining the difference as a first response result corresponding to the other intervention action; For each of the other intervention actions, calculating a sum of a second gain value corresponding to the other intervention action and a predicted response result of an element corresponding to the other intervention action, and determining the sum as a second response result corresponding to the other intervention action; The first response result corresponding to the other intervention action and the second response result corresponding to the other intervention action are determined as the gain response result corresponding to the other intervention action.
4. The method according to claim 3, wherein: The calculating the difference between the gain response result corresponding to each of the other intervention actions and the corresponding observation response result in the business interaction observation data to obtain the observed gain loss includes: Calculating a first gain loss according to a difference between a first response result corresponding to each of the other intervention actions and a corresponding observed response result in the business interaction observation data by using the intervention gain model; Calculating a second gain loss according to a difference between a second response result corresponding to each of the other intervention actions and a corresponding observed response result in the business interaction observation data by using the intervention gain model; An observed gain loss is determined based on the first gain loss and the second gain loss.
5. The method according to claim 4, wherein: The calculating the first gain loss according to the difference between the first response result corresponding to each of the other intervention actions and the corresponding observation response result in the business interaction observation data includes: For each first response result corresponding to the other intervention actions, querying the business interaction observation data for the observation response result corresponding to the other intervention actions; For each of the other intervention actions, calculating a difference between a first response result corresponding to the other intervention action and an observed response result corresponding to the other intervention action; The differences calculated for the first response results corresponding to the other intervention actions are summed and averaged to calculate the first gain loss.
6. The method according to claim 4, wherein: The calculating the second gain loss according to the difference between the second response result and the corresponding observation response result in the service interaction observation data includes: For each second response result corresponding to the other intervention actions, query the business interaction observation data for the observation response result corresponding to the other intervention actions; For each of the other intervention actions, calculating a difference between a second response result corresponding to the other intervention action and an observed response result corresponding to the other intervention action; The differences calculated for the second response results corresponding to the other intervention actions are summed and averaged to calculate the second gain loss.
7. The method according to claim 1, further comprising: Determine the user's preferred action probability according to the user service information and the intervention action set through the intervention gain model; Calculating the tendency score loss according to the user's tendency action probability and each element in the intervention action set through the intervention gain model; The adjusting the parameters of the intervention gain model according to the observed gain loss includes: The intervention gain model parameters are adjusted according to the observed gain loss and the propensity score loss.
8. The method according to claim 7, wherein: The determining, according to the user service information and the intervention action set, the probability of the user's preferred action comprises: Generate an initial probability according to the user service information and the intervention action set; The first gain value and the second gain value are combined with the initial probability to obtain the user's preferred action probability.
9. The method according to claim 1, wherein: The determining, according to the user service information, a predicted response result generated by the sample user for each element in the intervention action set includes: Determining a shared feature according to the user service information and the intervention action set; Determining, according to the shared features, a predicted response result generated by the sample user for each element in the intervention action set; The determining, according to the user service information and the intervention action set, a gain value corresponding to each of the other intervention actions includes: The gain value corresponding to each of the other intervention actions is determined according to the shared feature.
10. The method according to claim 9, wherein: The determining of the shared features according to the user service information and the intervention action set includes: Data enhancement is performed on the intervention action set, and shared features are determined according to the user business information.
11. The method according to claim 10, wherein: The performing data enhancement on the intervention action set and determining the shared features according to the user service information includes: generating initial features according to the user service information and the intervention action set; The intervention action set and the initial features are fused to generate shared features.
12. A method for predicting intervention gain, comprising: Obtain user business information and intervention action set of target users; The intervention action set includes a baseline intervention action and at least one other intervention action; Inputting the user service information and the intervention action set into a pre-trained intervention gain model; the intervention gain model is trained by the intervention gain model training method according to any one of claims 1 to 11; The intervention gain model is used to determine the predicted response result corresponding to each element in the intervention action set according to the user service information and the intervention action set.
13. The method according to claim 12, further comprising: Determine the user's preferred action probability and the gain value corresponding to each element in the intervention action set through the intervention gain model according to the user service information and the intervention action set; The intervention gain result of the target user is determined according to the user's tendency action probability and the gain value corresponding to each element in the intervention action set.
14. The method according to claim 12, wherein: The intervention action set includes: a distribution method for at least one service of the target traffic; and the predicted response result includes an interaction result between the user and each service.
15. The method according to claim 12, wherein: The intervention action set includes: at least one resource provision operation; the predicted response result includes an interaction result between the user and the resource-related service.
16. A training device for an intervention gain model, comprising: A sample acquisition module, used to acquire user service information of sample users, service interaction observation data of the sample users, and a set of intervention actions; The intervention action set includes a baseline intervention action and at least one other intervention action; An intervention gain model, used to determine, based on the user service information and the intervention action set, a predicted response result generated by the sample user for each element in the intervention action set; The intervention gain model is used to determine the gain value corresponding to each of the other intervention actions according to the user service information and the intervention action set; The intervention gain model is used to calculate the gain response result corresponding to each of the other intervention actions according to the gain value corresponding to each of the other intervention actions and the predicted response result generated by each element in the intervention action set; The intervention gain model is used to calculate the difference between the gain response result corresponding to each of the other intervention actions and the corresponding observation response result in the business interaction observation data to obtain the observation gain loss; A model parameter adjustment module is used to adjust the parameters of the intervention gain model according to the observed gain loss.
17. The device according to claim 16, wherein: The intervention gain model is used to: Determine, according to the user service information and the intervention action set, a first gain value corresponding to each of the other intervention actions; the first gain value corresponding to the other intervention action is a gain value of a response result generated from the reference intervention action to the application of the other intervention action; Determine, according to the user service information and the intervention action set, a second gain value corresponding to each of the other intervention actions; the second gain value corresponding to the other intervention action is a gain value of a response result generated from applying the other intervention action to the benchmark intervention action; The gain value corresponding to each of the other intervention actions is determined according to the first gain value corresponding to each of the other intervention actions and the second gain value corresponding to each of the other intervention actions.
18. The device according to claim 17, wherein: The intervention gain model is used to: For each of the other intervention actions, calculating a difference between a first gain value corresponding to the other intervention action and a predicted response result of the element corresponding to the other intervention action, and determining the difference as a first response result corresponding to the other intervention action; For each of the other intervention actions, calculating a sum of a second gain value corresponding to the other intervention action and a predicted response result of an element corresponding to the other intervention action, and determining the sum as a second response result corresponding to the other intervention action; The first response result corresponding to the other intervention action and the second response result corresponding to the other intervention action are determined as the gain response result corresponding to the other intervention action.
19. The device according to claim 18, wherein: The model parameter adjustment module comprises: A first gain loss calculation unit, configured to calculate a first gain loss according to a difference between a first response result corresponding to each of the other intervention actions and a corresponding observed response result in the business interaction observation data; A second gain loss calculation unit, configured to calculate a second gain loss according to a difference between a second response result corresponding to each of the other intervention actions and a corresponding observation response result in the service interaction observation data; The gain loss fusion unit is used to determine an observed gain loss according to the first gain loss and the second gain loss.
20. The device according to claim 19, wherein The first gain loss calculation unit includes: A first observation result acquisition subunit is used to query the observation response result corresponding to the other intervention action in the business interaction observation data for the first response result corresponding to each of the other intervention actions; A first observation loss calculation subunit, configured to calculate, for each of the other intervention actions, a difference between a first response result corresponding to the other intervention action and an observed response result corresponding to the other intervention action; The first intervention loss fusion subunit is used to sum and average the differences of the first response results corresponding to the other intervention actions to calculate the first gain loss.
21. The device according to claim 19, wherein The second gain loss calculation unit includes: A second observation result acquisition subunit is used to query the observation response result corresponding to the other intervention actions in the business interaction observation data for the second response result corresponding to each of the other intervention actions; A second observation loss calculation subunit, configured to calculate, for each of the other intervention actions, a difference between a second response result corresponding to the other intervention action and an observed response result corresponding to the other intervention action; The second intervention loss fusion subunit is used to sum and average the differences of the second response results corresponding to the other intervention actions to calculate the second gain loss.
22. The device according to claim 16, wherein: The intervention gain model is also used to: Determining the user's preferred action probability according to the user service information and the intervention action set; Calculating the tendency score loss according to the user's tendency action probability and each element in the intervention action set; The model parameter adjustment module comprises: The tendency loss adjustment unit is used to adjust the intervention gain model parameters according to the observed gain loss and the tendency score loss.
23. The device according to claim 22, wherein: The intervention gain model is used to: generating an initial probability according to the user service information and the intervention action set; The first gain value and the second gain value are combined with the initial probability to obtain the user's preferred action probability.
24. The device according to claim 16, wherein: The intervention gain model is used to: Determining a shared feature according to the user service information and the intervention action set; Determining, according to the shared features, a predicted response result generated by the sample user for each element in the intervention action set; The gain value corresponding to each of the other intervention actions is determined according to the shared feature.
25. The device according to claim 24, wherein: The intervention gain model is used to: Data enhancement is performed on the intervention action set, and shared features are determined according to the user business information.
26. The device according to claim 25, wherein The intervention gain model is used to: generating initial features according to the user service information and the intervention action set; The intervention action set and the initial features are fused to generate shared features.
27. An intervention gain prediction device, comprising: The module for obtaining the data to be predicted is used to obtain the user business information and intervention action set of the target user; The intervention action set includes a baseline intervention action and at least one other intervention action; A data input module to be predicted, used for inputting the user service information and the intervention action set into a pre-trained intervention gain model; the intervention gain model is trained by the intervention gain model training method according to any one of claims 1 to 11; The intervention gain model is used to determine the predicted response result corresponding to each element in the intervention action set according to the user service information and the intervention action set.
28. The device according to claim 27, wherein The intervention gain model is also used to: Determining the user's preferred action probability according to the user service information and the intervention action set; The intervention gain result of the target user is determined according to the user's tendency action probability and the predicted response result corresponding to each element in the intervention action set.
29. The device according to claim 27, wherein: The intervention action set includes: a distribution method for at least one service of the target traffic; and the predicted response result includes an interaction result between the user and each service.
30. The device according to claim 27, wherein: The intervention action set includes: at least one resource provision operation; the predicted response result includes an interaction result between the user and the resource-related service.
31. An electronic device comprising: at least one processor; as well as a memory communicatively connected to the at least one processor; wherein, The memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the training method of the intervention gain model described in any one of claims 1-11, or the intervention gain prediction method described in any one of claims 12-15.
32. A non-transitory computer-readable storage medium storing computer instructions, wherein: The computer instructions are used to enable the computer to execute the training method of the intervention gain model according to any one of claims 1-11, or the intervention gain prediction method according to any one of claims 12-15.
33. A computer program product, comprising a computer program, which, when executed by a processor, implements the training method of the intervention gain model according to any one of claims 1-11, or the intervention gain prediction method according to any one of claims 12-15.
Citation Information
Cited By
Road network state prediction method and system combined with ETC portal frame data
CN119339550A
Inspection intelligent analysis method and system for electricity marketing
CN120338553A
