Data processing method and electronic device

Abstract

Embodiments of the present application disclose a data processing method and an electronic device. The method comprises: obtaining a target model; determining a target merchant user in a current prediction period that needs to be predicted, and obtaining features in multiple dimensions by counting historical behavior data of the target merchant user from a current prediction time; taking the feature values corresponding to the target merchant user as input data of the target model, and predicting a value contribution of the target merchant user in a future time period starting from the current prediction time, so as to make a decision or optimization on a target project / action according to the value contribution prediction result. Through the embodiments of the present application, an index that can be observed in a short term can be provided for a merchant cultivation project, and a desired merchant cultivation effect can be obtained by a system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a data processing method and an electronic device. BACKGROUND

[0002] In the cross-border e-commerce service system, mature merchants have a strong willingness to perceive overseas markets and establish self-brands. However, for novice cross-border e-commerce merchants and traditional foreign trade merchants who are transforming into cross-border e-commerce merchants, they may be in a state of confusion and uncertainty. Therefore, the cross-border e-commerce service system will also launch a merchant growth plan. After the merchant signs up as a system member, the merchant can be empowered from team building, product release, foreign trade basic training, and store operation to achieve rapid growth according to the development stage of the merchant, which is also called "incubation of merchants".

[0003] In the prior art, the renewal rate of merchants renewing as system members is usually used to measure the effect of the incubation of merchants. However, the renewal rate is an annual index, and it is difficult to prove the influence on the renewal rate in a short time (less than one year to the last day, it is impossible to confirm how much the final renewal rate is), which makes it difficult to prove the value of the incubation of merchants in a short time, and further makes the incubation path not clear enough. SUMMARY

[0004] The present application provides a data processing method and an electronic device, which can provide a short-term observable index for the incubation of merchants, and help the system to obtain the expected incubation effect.

[0005] The present application provides the following solutions:

[0006] A data processing method comprises:

[0007] Obtaining a target model, wherein the target model is obtained by: obtaining features in multiple dimensions from historical behavior data of a plurality of merchant users at at least one observation time point, and value contribution situations of the plurality of merchant users respectively corresponding to a future time period starting from the observation time point, taking feature values corresponding to the plurality of merchants respectively as training input data, and taking the value contribution situations as training targets, performing model training to obtain the target model;

[0008] Determining a target merchant user that needs to be predicted in a current prediction period, and obtaining features in multiple dimensions from historical behavior data of the target merchant user starting from a current prediction time;

[0009] The feature value corresponding to the target merchant user is taken as input data of the target model, and the value contribution of the target merchant user in a future time period starting from the current prediction time is predicted, so as to make a decision or optimization on the target item / action according to the value contribution prediction result.

[0010] The value contribution is divided into multiple value contribution components according to different service categories available to the merchant user, and different target models are respectively corresponded to the value contribution components, so as to select the feature dimensions and / or sort the importance of the feature dimensions for different target models, to complete the training of the models.

[0011] The service category includes a member service, and the corresponding value contribution is the value contribution made by the merchant user to the system by becoming a system member through renewal;

[0012] The target model corresponding to the member service includes a first classification model and a second classification model having a cascading relationship, the first classification model is used to predict the renewal probability of the merchant user in a future target time period, and is trained with the target of completing renewal of the merchant user in the future target time period; and the second classification model is used to predict the probability of whether the member level of the merchant user will change when renewing, and is trained with the target of whether the member level of the merchant user will change when renewing in the future target time period.

[0013] The first classification model and the second classification model having the cascading relationship are multiple groups, corresponding to multiple different member levels in which different target merchant users currently stay.

[0014] The service category includes an information promotion service, and the corresponding value contribution is the value contribution made by the merchant user to the system by using the information promotion service provided by the system;

[0015] The target model corresponding to the information promotion service includes multiple regression models corresponding to multiple different information promotion service types, and the regression models are trained with the target of resource consumption generated by the merchant user in the process of using the information promotion service of the corresponding type in a future time period.

[0016] The information promotion service type includes an information promotion service type in which the generated resource consumption is related to the clicking situation of a consumer user, and the corresponding regression model includes a deep structure model of multiple layers of random forests stacked in a cascading manner.

[0017] The service category includes a cross-border supply chain service, and the corresponding value contribution is the value contribution made by the merchant user to the system by using the cross-border supply chain service provided by the system;

[0018] The target model corresponding to the cross-border supply chain service includes: a plurality of regression models corresponding to a plurality of different cross-border supply chain service types, which are trained to target resource consumption generated by a merchant user in the future time period in the process of using the corresponding type of cross-border supply chain service.

[0019] Among them, the cross-border supply chain service type includes: a cross-border supply chain service type related to a transaction order, and a cross-border supply chain service type unrelated to a transaction order; wherein the cross-border supply chain service type related to the transaction order is further divided into a plurality of subtypes according to whether the generated resource consumption and the transaction order are one-to-one.

[0020] Among them, in the process of obtaining training samples for the target model corresponding to the cross-border supply chain service, or in the process of predicting the value of the target merchant user, the features of the target merchant user in multiple dimensions are obtained. For the dimension associated with the time window, determine the time distribution of the multiple resource consumptions related to the cross-border supply chain service in the corresponding time window, and the distance between the observation time point or the current prediction time, and determine different time decay coefficients for the multiple resource consumptions, and determine the features in the corresponding dimension.

[0021] Among them, based on the method of causal analysis, the selection of feature dimensions and / or the importance ranking of feature dimensions for the target model corresponding to the cross-border supply chain service.

[0022] Among them, in the process of obtaining training samples for the target model, a plurality of observation time points are determined for each merchant user, so as to obtain a plurality of training samples for the same merchant user at a plurality of different observation time points for model training; wherein the plurality of observation time points corresponding to the same merchant user are evenly distributed between different sub-time periods, and have randomness in the same sub-time period.

[0023] Among them, in the process of obtaining training samples for the target model corresponding to the information promotion service, for the plurality of training samples corresponding to the same merchant user at different observation time points, the features in the time dimension are constructed, and the features in the time dimension include: whether the sub-time period where the observation time point is located is associated with a target marketing activity, the distance between the observation time point and the start or end time of the target marketing activity, and / or the identifier of the sub-time period where the observation time point is located.

[0024] Among them, the target merchant user that needs to be predicted in the current prediction period includes:

[0025] In the process of AB testing on the target item / action, the target merchant user is determined from two groups of merchant users participating in the AB testing, wherein different target items / actions are online for different groups of merchant users, or the target target item / action is online for one group of merchant users, and the target target item / action is not online for another group of merchant users;

[0026] The method further comprises:

[0027] According to the prediction results of the value contribution of each target merchant user in the same user group in the future time period, the prediction results of the merchant nurturing effect that the same target item / action can obtain in the future time period are determined, so as to compare the prediction results between different target items / actions, and make a selection or decision on the target item / action.

[0028] Further comprising:

[0029] According to the prediction results of the value contribution of the target merchant user, the importance and influence direction of each feature in the target model in influencing the prediction results of the target merchant user are obtained;

[0030] According to the importance and influence direction, a first target feature that can be used as a potential factor for optimizing the prediction results and / or a second target feature that can be used as a risk factor for optimizing the prediction results are determined;

[0031] According to the first target feature and / or the second target feature, optimization suggestion information of the target item / action is provided to the target merchant user and / or a service personnel providing merchant nurturing service for the target merchant user.

[0032] A data processing apparatus comprises:

[0033] A model acquisition unit is configured to acquire a target model, wherein the target model is obtained by: obtaining features in multiple dimensions from historical behavior data of a plurality of merchant users at at least one observation time point, and value contribution of the plurality of merchant users corresponding to the features respectively in a future time period starting from the observation time point, taking feature values of the plurality of merchant users respectively as training input data, and taking the value contribution as a training target, and performing model training to obtain the target model;

[0034] A feature acquisition unit is configured to determine a target merchant user for which value prediction is needed in a current prediction period, and obtain features in multiple dimensions from historical behavior data of the target merchant user starting from a current prediction time;

[0035] A prediction unit is configured to take the feature values corresponding to the target merchant user as input data of the target model, to predict the value contribution of the target merchant user in a future time period starting from the current prediction time, and to make a decision or optimization on a target item / action based on the prediction result of the value contribution.

[0036] A computer readable storage medium having stored thereon a computer program which, when executed by a processor, implements the steps of any of the methods described above.

[0037] An electronic device comprising:

[0038] one or more processors; and

[0039] a memory associated with the one or more processors, the memory configured to store program instructions that, when executed by the one or more processors, implement the steps of any of the methods described above.

[0040] According to the embodiments provided in the present application, the following technical effects are disclosed:

[0041] According to the embodiments provided in the present application, the following technical effects are disclosed:

[0042] In the preferred embodiment, the merchant value contribution can be divided into multiple value contribution components, and each component corresponds to a different target model, so that the feature dimension selection and / or feature dimension importance ranking can be performed for different target models to complete the model training, thereby improving the accuracy of the model.

[0043] In terms of constructing training samples, by randomly selecting a day in each month as an observation time point for each merchant user, multiple training sample data can be generated for the same merchant user, thereby improving the richness of the sample and the generalization ability of the model.

[0044] For the value contribution component of the member service, a scheme of combining a first classification model and a second classification model with a cascade relationship is adopted, so that each model can be a binary classification model, which is beneficial to reduce the prediction difficulty of the model and improve the accuracy of the model.

[0045] For the value contribution component of the information promotion service, for the information promotion service type in which the resource consumption is related to the click of the consumer user, a deep structure model of multiple layers of random forests can be stacked in a cascading manner to make up for the accuracy problem caused by insufficient sample size and feature missing. In addition, for multiple training samples corresponding to different observation time points of the same merchant user, features in the time dimension can be constructed to smooth the differences between different observation time points.

[0046] For the value contribution component of the cross-border supply chain service, a time decay coefficient based on Newton's cooling law can be used to determine the time distribution of multiple resource consumptions related to the cross-border supply chain service within the corresponding time window and the distance between the observation time point or the current prediction time, and different time decay coefficients for the multiple resource consumptions, and then determine the features in the corresponding dimension. This can reflect the different effects of different times on the trend of future consumption of the merchant user in the corresponding aspect, thereby further improving the accuracy of the model.

[0047] In the application aspect, the value contribution in the future time period obtained by different merchant nurturing projects can be compared by AB testing, so as to select the merchant nurturing project or make a decision on whether a certain merchant nurturing project can be officially launched.

[0048] The importance and influence direction of each feature in the target model in influencing the prediction result of the target merchant user can be obtained according to the prediction result of the value contribution of the target merchant user, so as to identify the problems of the merchant user, help the merchant diagnose and operate, and improve the business effect of the merchant.

[0049] Of course, any product implementing the present application does not necessarily need to achieve all of the advantages described above at the same time. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0051] Figure 1 It is a schematic diagram of the system architecture provided by the embodiment of the present application;

[0052] Figure 2 is a flow chart of the method provided in an embodiment of the present application;

[0053] Figure 3 This is a schematic diagram of the model cascade relationship provided in the embodiment of the present application;

[0054] Figure 4 is a schematic diagram of a training sample construction method provided in an embodiment of the present application;

[0055] Figure 5 is a schematic diagram of a device provided in an embodiment of the present application;

[0056] Figure 6 Schematic diagram of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0057] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field are within the scope of protection of this application.

[0058] In the embodiments of the present application, first, an index that can be observed in a short term can be provided for a merchant growth plan (target project / action). Specifically, since the result of nurturing a merchant is finally reflected in the value improvement brought by the merchant to the system, the current nurturing operation action will affect the future nurturing result (including whether the member will renew after expiration, etc.). Therefore, in the embodiments of the present application, model training can be performed based on a large amount of historical behavior data of merchant users to predict the value contribution that a specific merchant user can bring to the system within a certain time period in the future (for example, one year in the future, etc.). The value contribution can be specifically referred to as the LTV (Life time value, customer lifetime value) of the merchant user. The LTV prediction results of a plurality of merchant users aggregated together can be used to measure the effect of a certain merchant growth plan. For example, in the process of executing a certain merchant growth plan, the LTV of a plurality of merchant users in a future time period is predicted, and after the LTV prediction results of the plurality of merchant users are aggregated, they can be used to represent the possible effect of the merchant growth plan. This prediction can be performed in a period of "day" or even shorter, therefore, the specific index can be a dynamic value, and the execution of each operation action can bring changes to the index value, so as to more sensitively reflect the effect of the specific operation action, thereby helping the operation personnel to determine whether the specific operation action is effective, whether it should be continued or interrupted, or optimized on the original basis, etc. In addition, the shap value of the model (used to represent the importance and influence direction of each feature in affecting the prediction result of a specific merchant user) can be used to identify the problems of the merchant user, help the merchant to diagnose and operate, and improve the merchant operation effect.

[0059] In the process of merchant LTV prediction, training samples can be obtained according to historical behavior data of a plurality of merchants. In the process of obtaining the training samples, a certain observation time point can be selected, and the historical behavior data of a plurality of merchant users can be counted to obtain features in a plurality of dimensions. In addition, the value contribution of the plurality of merchant users in a future time period (for example, one year in the future) starting from the observation time point (for example, the resource investment in using the services provided by the system in the future one year) can be obtained. Then, the feature values corresponding to the plurality of merchants can be used as training input data, and the value contribution can be used as a training target to train a target model. Then, in the process of predicting the merchant user, the features in a plurality of dimensions obtained by counting the historical behavior data of the target merchant user starting from the current prediction time point can be obtained. Then, the feature values corresponding to the target merchant user can be used as input data of the target model to predict the value contribution of the target merchant user in a future time period starting from the current prediction time point, for example, to predict the resource investment of the current merchant user in using various services provided by the system in the future one year.

[0060] However, in the process of model training, feature engineering is an important link. That is, there can be a large number of features counted from the user historical behavior data, but some of the features can be effective for model training, and different features can have different importance for model training. How to more accurately select and sort the features has an important influence on the prediction accuracy of the finally trained model.

[0061] In the embodiments of the present application, the resource cost invested by the merchant user in using the services provided by the system can be relatively complex, for example, there are investments in member services, investments in information promotion services, and investments in supply chain services. The influence factors corresponding to the investment of the merchant user in each aspect are very complex. In this case, it is almost impossible to use one model to complete the prediction of the merchant LTV.

[0062] Therefore, in the embodiments of the present application, the value contribution can be divided into a plurality of value contribution components according to different service categories available to the merchant user, and correspond to different target models. In this way, different types of models (for example, including classification models, regression models, etc.) can be selected according to the data characteristics of various components, and the selection of feature dimensions and / or the importance sorting of feature dimensions can be performed for different target models to complete the training of the model, thereby reducing the difficulty of model training and improving the accuracy of the model.

[0063] For example, for a certain system, it is assumed that the service categories provided by a merchant user specifically are mainly divided into: merchant member service (which is usually the basic service), information promotion service, and supply chain service. The main value contribution of the merchant to the platform is also reflected in these three aspects. And the purpose of the embodiments of the present application is to view the value of the merchant user in the system from a more global perspective, avoiding falling into local optimization. Therefore, the LTV value of the merchant user is composed of three parts: merchant member service usage fee (member service LTV), information promotion service usage fee (information promotion service LTV), and supply chain service usage fee (supply chain service LTV). In this way, different models can be trained for each LTV category, and feature selection and importance sorting can be performed separately for different models. That is, for the same merchant user, multiple component LTV values can be predicted by multiple different models, and then the components are added to obtain the comprehensive LTV value of the merchant user. That is:

[0064] Merchant user LTV = Member service LTV + Information promotion service LTV + Supply chain service TLV

[0065] In this way, from the perspective of system architecture, as shown in Figure 1 , member service LTV, information promotion service LTV, and supply chain service TLV can be predicted by multiple models, which can include classification models, regression models, etc. After obtaining each LTV component and the comprehensive LTV value, there can be multiple applications. For example, one application can be to compare the LTV values obtained by AB testing or other methods for user groups using a certain merchant project and not using a certain merchant project to obtain the effect that the merchant project can obtain. Or, the effects of two merchant projects can be compared by obtaining the LTV values of user groups using merchant project A and merchant project B, respectively, to select one of them for further execution, etc. This can help the operation to make decisions on the merchant project. In addition, LTV data dashboards can be provided to track changes in LTV, and when a decrease in LTV value is found, specific solutions can be optimized in a timely manner, etc. In addition, the LTV diagnosis module can output potential factors or risk factors that affect the final LTV value of the merchant user based on the prediction results and shap values of the LTV prediction model, and can provide the merchant workbench or operation workbench in the form of tasks, etc. to guide the merchant user or the operation personnel to improve or further strengthen specific factors in a timely manner, etc.

[0066] The specific implementation schemes provided by the embodiments of the present application will be described in detail below.

[0067] First, the embodiments of the present application provide a data processing method, as shown in Figure 2 , the method can comprise:

[0068] S201: Obtain a target model, the target model is obtained by the following way: a plurality of dimensions of features obtained by statistically processing historical behavior data of a plurality of merchant users from at least one observation time point, and value contribution situations of the plurality of merchant users respectively corresponding to a future time period starting from the observation time point, taking the feature values corresponding to the plurality of merchants as training input data, and taking the value contribution situations as training targets, performing model training to obtain the target model.

[0069] For specific implementation of the training of the model, see the description below.

[0070] S202: Determine target merchant users that need to be predicted in the current prediction period, and obtain features in a plurality of dimensions by statistically processing historical behavior data of the target merchant users starting from the current prediction time.

[0071] In the embodiments of the present application, after the model is trained, the LTV values of the merchant users can be predicted in a periodic manner, for example, one day can be a period, etc. In each prediction period, the target merchant users that need to be predicted in the current prediction period can be determined first, and features in a plurality of dimensions are obtained by statistically processing historical behavior data of the target merchant users starting from the current prediction time.

[0072] Among them, the target merchant users that need to be predicted in the current prediction period can be, in general, all merchant users who have registered in the current system or become members of the system, that is, as long as the specific merchant user has registered in the system or become a member of the system, and relevant behavior data is generated in the system, the LTV prediction can be performed. Of course, for the LTV prediction related to the member service, the main role of the model is to predict the renewal probability of the merchant user in the future time period, etc., but there are some two-year signed merchant users currently in the first service period, for these merchant users, the model may not be needed to predict their renewal in the next year. For this, there will be a detailed description below.

[0073] S203: Take the feature values corresponding to the target merchant users as input data of the target model, predict the value contribution situations of the target merchant users in the future time period starting from the current prediction time, so as to make decisions or optimization on target projects / actions according to the value contribution situation prediction results.

[0074] That is, in the embodiments of the present application, feature mining can be performed according to historical behavior data of multiple merchants, and model training can be performed, so that the value contribution of the merchants to the system in a future time period can be predicted through a specific trained model, and the prediction result can be used for decision-making, optimization, etc. of the merchant cultivation project. As described above, since the value contribution of the merchant is relatively complex, and the factors affecting the value contribution of the merchant are also very complex, in the preferred embodiments, the value contribution can be divided into multiple value contribution components according to different service categories available to the merchant user, and different target models are used for prediction, so that the selection of the model can be performed according to various components, and the selection of the feature dimension and / or the importance ranking of the feature dimension can be performed for different target models.

[0075] Specifically, in a specific implementation, the LTV value of the merchant user can have three parts of member service LTV, information promotion service LTV, and supply chain service LTV, which will be introduced below.

[0076] (1) Member service LTV

[0077] Regarding the member service LTV, that is, for the service category of the member service provided by the system for the merchant user, the value contribution of the merchant user to the system when using this service through the way of becoming a member of the system.

[0078] Among them, the merchant user can be divided into two types of new signing and renewal signing according to the signing mode when using the member service. New signing means that the merchant becomes a member of a certain system for the first time (without considering the case of re-purchasing after discontinuing the contract), and renewal signing means that the merchant is already a member of the system and continues to purchase the member product after the service period. In addition, according to the research feedback of the merchant user, the merchant user is more willing to pay for the value (inquiry / order, etc.) of the system. Therefore, the better the service effect of the merchant obtained from the system, the higher the probability of renewal signing (the service of the salesperson is important, but not the decisive factor). That is, renewal signing also means that the merchant user recognizes the current member identity, and the merchant user generally makes a renewal decision according to the platform effect. Therefore, the effect of the merchant cultivation project can be measured through the member renewal of the merchant user. In this way, the prediction of the member service LTV is the prediction of the renewal rate of the member service.

[0079] However, there can be multiple different member levels of services in the system, and the merchant user can choose one level when signing up, and can continue to maintain the previous level unchanged or can have a member level transition when renewing the subscription, which makes it difficult to predict the renewal rate of the merchant user.

[0080] For example, for a specific membership level, a system can provide membership levels including ordinary members, senior members, etc. Among them, the service effect of the merchant user in the system is different when purchasing different levels of member products. For example, for ordinary members, the system can provide one-stop shop decoration, product display, marketing promotion, business negotiation, and shop management, etc. Full series of online services and tools to help enterprises reduce costs and efficiently explore the foreign trade market. For senior members, the system can authenticate and verify the strength of the merchant's enterprise qualification, commodity qualification, enterprise capacity, etc. through online and offline combination, in addition to enjoying all the benefits and services of ordinary members, senior members can also enjoy exclusive marketing benefits, exclusive front-end scenarios, and rich tool benefits, etc. The difference between different membership levels and service content leads to differences in the renewal rate of the merchant group.

[0081] The membership level transition means that the merchant user can continue to retain the current membership level when renewing, or can also decide to change to another membership level according to the situation. For example, during the renewal process, ordinary members of the merchant can upgrade to senior members if they meet the system requirements. Similarly, senior members can also change to ordinary members according to their own operation situation, etc. The influencing factors of membership level transition are complex, and the merchant user will usually make a comprehensive decision on whether to upgrade or downgrade according to their own investment and system effect, etc. Among them, the influencing factors of ordinary members upgrading to senior members are different from those of senior members downgrading to ordinary members.

[0082] As can be seen from the above, for the prediction of membership service LTV, the following difficulties and challenges may be faced:

[0083] 1. The situation is complex, and there are many and complex factors affecting the merchant's renewal. In theory, the entire system full process each link has an impact on the merchant's renewal, so it is a big challenge for feature collection and feature engineering.

[0084] 2. The distinction between ordinary members and senior members is not high, and the difference is not big. Many features will perform better for ordinary members than for senior members, making it difficult for the model to accurately identify ordinary members and senior members.

[0085] 3. The time period of the merchant to the platform is very different, with some merchants serving for more than ten years, and a large number of merchants serving for less than a year, even less than three months. Therefore, there may be a problem of data sparseness, which may greatly affect the accuracy of the model.

[0086] To address the above difficulties and challenges, first, the effect of different merchant groups can be focused on respectively, and the renewal rate and the probability of membership level transition are focused on. Then, the LTV value of the merchant members as a whole is calculated by summing up the LTV values of different merchant groups.

[0087] For example, assuming that the membership levels in the system are divided into two levels, i.e., ordinary members and senior members, the membership fee of senior members is C as , and the membership fee of ordinary members is C gs .

[0088] If the current merchant is a senior member, the probability of the current merchant continuing to sign up as a senior member after the service expires is r as , and the probability of the current merchant downgrading to an ordinary member after the service expires is r gs . Then, the membership service LTV of this merchant is:

[0089] LTV 单个高级会员商家 =C as r as +C gs r gs (1)

[0090] If the current merchant is an ordinary member, the probability of the current merchant signing up as an ordinary member after the service expires is P gs , and the probability of the current merchant upgrading to a senior member after the service expires is P as . Then, the membership service LTV of this merchant is:

[0091] LTV 单个普通会员商家 =C as P as +C gs P gs (2)

[0092] It should be noted that if the merchant has completed the renewal before the service expires (assuming that the early renewal date is A and the service expiration date of the merchant is B), the time from A to B does not need to be predicted because the renewal has been completed, and therefore the corresponding renewal probability can be 1. After the date B, if the merchant has not renewed, the merchant can be automatically identified as a customer who will renew in the next fiscal year, and the LTV value will be adjusted accordingly.

[0093] From the above member level and transition analysis, it can be seen that the member level of the merchant is different, and the member level is changeable during the renewal. In view of the difference between different member levels, the possibility of member level transition during the renewal of the merchant, and the characteristics that users of different member levels are easy to predict and confuse, in the preferred embodiment of the present application, the target model corresponding to the member service can be divided into a first classification model and a second classification model with a cascade relationship. The first classification model is used to predict the renewal probability of the merchant user in a future target time period, and is trained with the target of completing the renewal of the merchant user in the future target time period. The second classification model is used to predict the probability of whether the member level of the merchant user will change during the renewal, and is trained with the target of whether the member level of the merchant user will change during the renewal in the future target time period. That is, through this division, the member service LTV becomes a classification problem, so the main task is to train multiple classification models, and the training of the classification model is relatively simple, and feature selection and parameter adjustment can be performed for each model separately, so as to obtain higher model accuracy.

[0094] Among them, the first classification model and the second classification model with a cascade relationship can also be multiple groups corresponding to multiple different member levels of different target merchant users. For example, in the case of dividing the member level into two levels of ordinary member and senior member, the merchant users to be predicted can be divided into two groups of ordinary member merchants and senior member merchants, and the cascade mode is used to predict the probability of the merchant renewing a certain member. For this purpose, four models can be designed as shown in the following table. Figure 3 Among them, for ordinary member merchants, two models correspond to them, which are both binary classification models. Among them, LTV model 1: mainly judges whether the ordinary member merchant will renew (at this time it cannot be determined which member level he will renew) in the future time period (for example, in the future one year). If the contract is terminated, it is the direct loss of the system. If it is renewed, the specific probability of renewing different member levels can be further predicted by the algorithm model. LTV model 2 is to further predict the specific probability of the ordinary member merchant renewing different member levels through this model on the premise that it is predicted that the merchant will not terminate the contract, including the probability of continuing to be an ordinary member and the probability of upgrading to a senior member.

[0095] After the prediction and calculation of the two models, three probability values can be obtained:

[0096] p1: the probability of the ordinary member merchant directly terminating the contract.

[0097] p2: the probability of the ordinary member merchant upgrading to a senior member.

[0098] p3: the probability that the general member merchant continues to renew the general member.

[0099] In any case, it is desirable that p1+p2+p3=1.

[0100] Specific details: the probability that model 1 predicts as a positive sample (renewal) is p11, the probability that model 2 predicts as a positive sample (renewal as a premium member) is p12, then p1=1-p11; p2=p11*p12; p3=p11*(1-p12).

[0101] For LTV model 3 and LTV model 4, similar to the case of LTV model 1 and LTV model 2 described above, it is not repeated here.

[0102] After completing the model selection, there are also problems related to feature engineering. As described above, in the embodiments of the present application, it is necessary to use the historical behavior data of the merchant to train the model, and such historical behavior data not only includes features in multiple dimensions, including the number of visits in the last 60 days, etc., but also needs to include label information, for example, our goal is to predict whether the merchant will renew before the contract expiration date in the future, then each piece of data in the training sample also needs to have its own label, which indicates whether the corresponding merchant has renewed before the contract expiration date in the future, whether the membership level has changed, etc. Therefore, when constructing the training sample, there will also be requirements on the merchant's time of entry. For example, it is necessary to select merchant users whose entry time exceeds one service period, so that the renewal status at the end of a service period is known, etc. However, since the number of merchants in the system is already very limited, and some newly entered merchants are removed, the problem of sample sparsity is more prominent.

[0103] Therefore, in the preferred embodiments of the present application, the following method can be used for sample construction: a plurality of observation time points are determined for each merchant user, so as to obtain a plurality of training samples for the same merchant user at a plurality of different observation time points for model training; wherein the plurality of observation time points corresponding to the same merchant user can be evenly distributed between different sub-time periods and can have randomness within the same sub-time period. For example, specifically, the sub-time period can be one month, so for each merchant user, a day is randomly selected as an observation time point every month, and the historical behavior data is counted from the observation time point, and the renewal status in the future time period is collected. In this way, a merchant user can correspond to multiple data, and each piece of data can be used as an independent training sample for model training.

[0104] The reason for constructing the sample in the manner of randomly selecting a day in each month as the observation time point is that, on the one hand, different behavior data of the same merchant at different time points can constitute independent samples, so that for the same merchant user, multiple data can be extracted at different observation time points to form multiple different training samples, thereby enriching the sample size. On the other hand, the prediction process in the embodiment of the present application is usually performed every day, so it is hoped that various samples cover every day of each month, so the manner of randomly selecting a day in each month is adopted to improve the generalization ability of the model. Furthermore, many indicators may be periodic in the "year" dimension, for example, the performance in June and November of each year is better than that in other months due to "big promotion month", so by randomly selecting a day in each month for the same merchant user, the problem of inaccuracy caused by statistics of only certain specific time points for some merchants can be avoided, and the like.

[0105] Regarding the division of positive and negative samples, the sample that is renewed before the contract expiration date is a positive sample, and the sample that is not renewed on the contract expiration date and is renewed after the contract expiration date is a negative sample. Because the model is fixed on the contract expiration date to determine whether to renew, the model should have a fixed judgment time node.

[0106] In the aspect of feature selection, because there are many and complex factors that affect the renewal of the merchant, all possible factors that affect the renewal can be collected in advance according to the service process of the right system. For example, it can be roughly classified as merchant attribute, website activity, product quality, customer service, transaction conversion, basic rights and interests, senior member rights and interests, information promotion investment and effect, performance and guarantee, and the like. Among them, each category can include multiple specific indicators, for example, regarding product quality, it can include the number of "good products", the number of "video products", the number of "exhibition products", and the like. Regarding customer service, it can include the number of visits in the last 60 days, and the like. In specific implementation, a special "indicator center" usually exists in the specific cross-border e-commerce system, which can output numerical values on multiple indicators according to various basic data, so in the embodiment of the present application, the feature values of the specific merchant user in multiple dimensions can also be obtained from the indicator center, and for the indicators specific to the embodiment of the present application, the specific indicator values can also be output by the "indicator center" by submitting a demand to the "indicator center", and the like.

[0107] In this way, for each merchant user, the historical behavior data of the plurality of merchant users can be counted according to each observation time point, so as to obtain the features in the above-mentioned multiple dimensions, for example, including the number of "good products", the number of "video products", the number of "exhibition products", the number of revisit times in the last 60 days, and the like. In this way, each training sample can include features in the above-mentioned multiple dimensions, and can also include label information such as whether the corresponding merchant user has signed a contract before the expiration date of the future contract, so as to train the model.

[0108] It should be noted that, although the LTV value of the merchant user in a certain fixed time period (for example, one year in the future) is predicted in the embodiments of the present application, and is used to guide the decision or optimization of the merchant cultivation project, it is still not desirable to look at the problem from a short-term or limited perspective. For example, if the renewal rate in the next year is improved through some means, but the renewal rate in the next year is affected, it is not the result expected. However, in actual application, the specific contract period of the specific merchant when signing the member service contract can also be different, for example, including a one-year contract (that is, after signing a contract, the contract expires after one year of service, and needs to be renewed every year), and a two-year contract (after signing a contract, the contract expires after two years of service, and can be renewed every two years). In the embodiments of the present application, the specific prediction time period can be one year, that is, the merchant service related to renewal in the current year (natural year or accounting year) is the main target of the embodiments of the present application. However, due to the existence of such a two-year contract, if more resources are given to the merchant to be renewed in the current year under the condition that the system resources (resources available for the merchant cultivation project) are limited, whether it will affect the merchant to be renewed in the next year needs to be considered.

[0109] Therefore, in the specific implementation of the embodiments of the present application, the following test is performed: after the model is trained in the above-mentioned manner, the model is used to predict whether each merchant in the test set will renew in the next year, the member level migration when renewing, and the like. The result shows that the accuracy of the prediction is still very high. That is, the difference between the prediction of the renewal in the next year, the next two years or even the next three years using the same model is not large. This shows that the use of the above-mentioned model provided by the embodiments of the present application to predict the LTV of the merchant in the member service, and the decision or optimization of the merchant cultivation project based on the prediction result to improve the renewal rate and the like in the next year will not have too much impact on the renewal rate in the next two years or the subsequent years.

[0110] In addition, through statistics, it is found that the proportion of service one-year contract merchants discontinuing the contract is very low among two-year contract merchants. Based on this fact, in the specific model, the probability of a two-year contract merchant renewing the contract after the first year of service is 1. In this way, the merchants that need to renew the contract in the current year can be divided into: ordinary merchants signed / renewed in the last year + two-year contract merchants signed in the last financial year (two-year contract merchants in the first service period, with a default renewal rate of 1) + two-year contract merchants signed in the last financial year (two-year contract merchants in the second service period). For example, assuming that the current year is FY22, the one-year contract merchants signed in the last year FY21 will renew the contract in the current year. Similarly, the two-year contract merchants signed in the last year FY20 will also renew the contract in the current year after one year of service, and so on. In addition, the merchants that need to renew the contract in the next year include: new merchants signed in the current year + merchants that have just completed the renewal in the current year.

[0111] Based on the above analysis, when determining the target merchant users that need to be predicted in the current prediction period, the merchants that sign a one-year contract and are in the service period, and the merchants that sign a two-year contract and are in the second service period can be included. For the merchants that sign a two-year contract and are in the first service period, the renewal probability is directly set to 1, and the member level will not change.

[0112] In the training process, multiple feature combinations and feature weights can be selected for multiple training, and then the prediction effect of the model is compared by constructing a test set to determine the better feature combination and feature weight.

[0113] When constructing the test set, 20% (or other proportions) of the merchants can be randomly selected, and the distribution of these merchants in the prediction time to the contract expiration time is relatively uniform. The distribution of first-year, second-year, and multi-year merchants is approximately the same as the distribution of the real prediction environment. The purpose of this is to enable the model to have good generalization performance on merchants in various service periods.

[0114] In summary, after the above model is trained, for a specific merchant user that needs to be tested, the probability of the merchant renewing the contract before the contract expires in the future can be obtained according to the current member level of the merchant through the above LTV model. If it is an ordinary member merchant, the predicted renewal rate value is (p2, p3). If it is a senior member merchant, the predicted renewal rate value is (t2, t3). Then, according to the member level of the merchant and the predicted renewal rate value, formula (1) or (2) is called to calculate the member service LTV value of the merchant.

[0115] (II) Information Promotion Service LTV

[0116] Information promotion service LTV, that is, the value contribution of the merchant user to the system through the use of the information promotion service provided by the system for the service category of information promotion service provided by the system for the merchant user.

[0117] Unlike the member service, the merchant user does not pay once a year or twice a year when using the information promotion service, but may involve continuous payment, so a regression model can be used for prediction. In addition, since there may be multiple types of specific information promotion services, the specific influencing factors will be different when the merchant user invests in different types of information promotion services. Therefore, in specific implementation, the target model corresponding to the information promotion service can include: multiple regression models corresponding to multiple different information promotion service types, and is trained to target the resource consumption generated by the merchant user in the use of the corresponding type of information promotion service in the future time period.

[0118] Specifically, when classifying information promotion services, information promotion services can be divided into multiple types according to different billing methods of specific services, etc. For example, it can specifically include: an information promotion service type related to the click of the consumer user, or a type unrelated to the click of the consumer user, etc.

[0119] For example, for the former, the merchant user can prepay in the relevant account and create a marketing plan, and the merchant with a high member level can also choose to target specific countries and populations. When the buyer user clicks on the specific promoted goods, consumption will be generated. After consumption is generated, the merchant user's account can be deducted. For the latter, the system can mainly open some specific resource positions in specific pages for information promotion, for example, the carousel position in the home page, or the resource position in the search result page. Since these resource positions have higher exposure, their value is also higher. For these resource positions, the merchant user can obtain the qualification for information promotion through specific resource positions through various competitive ways. For this case, whether the buyer user has a click operation on the resource position or not, the merchant user will generate corresponding consumption, etc. It can be seen that the above different types of information promotion services, due to the different ways of resource consumption generated by the merchant user, the specific influencing factors will be different, and the difficulty of prediction will also be different.

[0120] For the above type irrelevant to the click of the buyer user, although the merchant user can pay once every half year or year, whether the system can obtain the income in this aspect in the future, whether the merchant user will continue to purchase the corresponding keyword (usually related to the information promotion displayed in the relatively high position of the search result page, for example, if a merchant user purchases a keyword, the merchant user related information promotion content can be displayed in the relatively high position of the search result page when the buyer user searches the keyword), and since the merchant user can purchase more than one word, the expiration time of multiple schemes is also different, therefore, it can still be regarded as a regression problem, and a regression model is used to predict the investment of the merchant user in the information promotion service in the future time period.

[0121] For the information promotion service type related to the click behavior of the buyer user, the influencing factors of the consumption generated by the merchant user are relatively complex, and are related to the budget and operation of the merchant user, because the bid of the merchant and the commodity itself affect the ranking, and the commodity itself and the ranking affect the click of the buyer user, and finally the consumption is generated according to the click of the buyer and the bid of the merchant, so the consumption difference of the merchant and the merchant in this information promotion service will be large, and the difference will also be large in different time periods, so this scene is a standard regression problem scene.

[0122] However, because the billing logic of the information promotion service is different, the service mode is different, which will lead to great difference in feature selection, so the consumption of the merchant user in the above different types of information promotion service can be predicted by two models respectively, and finally the two parts are added to form the information promotion service LTV:

[0123] Total_ad_LTV = LTV1 + LTV2 (3)

[0124] However, in the model training and prediction process, some difficulties and challenges will still be faced, and the specific difficulties and challenges are as follows:

[0125] 1. Small sample size: for the information promotion service related to the click of the user, not every merchant has purchased, and for the information promotion service irrelevant to the click of the user, fewer merchants have purchased. The training of the regression model has a high requirement for the sample size, and in this case, the training difficulty is relatively large.

[0126] 2. Long prediction period: in the embodiment of the present application, the consumption of each merchant in the information promotion service in the future one year needs to be predicted, and because the time is long, the difference and fluctuation of each merchant are also large.

[0127] 3. Numerous factors influence the cost of information promotion services. For example, information promotion services that rely on clicks from buyers require merchants to develop a plan, set a budget, and initiate investment. Then, algorithms undergo a series of processing to ultimately determine the placement of the promotion. Whether this will result in cost depends on factors such as whether the buyer clicks, etc. Therefore, finding a suitable model and its key features is crucial in this process.

[0128] To address the above issues, firstly, since the sample size is relatively small, we can also adopt a solution of randomly selecting one day per month from each merchant as the observation time point when constructing the sample. In this way, for the same merchant user, we can construct multiple sample data. This can not only enrich the sample size and solve the problem of small sample size, but also use different time points as observation points to observe the future situation of the merchant and improve the generalization performance of the model. For example, Figure 4 As shown, we take the data of the past two years, starting from the first month, and divide the training set by month. For each merchant, we randomly select one day every month, use this time node as the dividing line, and count the consumption generated by information promotion services for one year as the target value. We count the indicators in different dimensions as features to realize model training.

[0129] Regarding feature selection, all factors that may influence merchants' investment in information promotion services can be collected in advance based on the system's service processes. For example, these can be broadly categorized into consumption-related factors (including daily account balance, average daily consumption over the past 30 days, etc.), product-related factors (number of active products, number of potential products, etc.), keyword and traffic-related factors (number of online product keywords, monthly online purchase keyword volume), sales-related factors (number of active customers, number of sales visits over the past 30 days), merchant attributes and strength, factors affecting previous information promotion effectiveness (such as inquiries over the past 30 days), and plan-related factors (average number of marketing plans over the past 30 days), etc. Relevant indicators can also be read or customized through the "Indicators" center to serve as model training data. During training, multiple training runs can be performed by selecting multiple feature combinations and feature weights. Then, by constructing a test set, the model's prediction performance can be compared to identify the most effective feature combinations and feature weights.

[0130] In addition, since the prediction time can span the whole year, and prediction can be made every day, the sampling time usually also needs to span the whole year, and in the case of randomly selecting a day in each month as the observation time point, the features collected in different time periods can be quite different. For example, the merchant's consumption of information promotion services in some "big promotion months" and "30-day average information promotion services" can be the highest in the whole year, etc. Therefore, in order to be able to smooth the differences between these features, some time-dimension features can also be added to the specifically constructed data, such as: whether it is a "big promotion month", the number of days from the specific observation time point to the "big promotion month", the last two digits of the natural month (01 represents January, 02 represents February, and so on), etc. Of course, in specific implementation, when specifically extracting the observation time point, the sub-time period can also be "week" or other time lengths, but similar methods can also be used to add time features to the collected data. Specifically including: whether the observation time point is associated with a target marketing activity, the distance between the observation time point and the start or end time of the target marketing activity, and / or the identifier of the sub-time period in which the observation time point is located, etc.

[0131] Regarding the selection of the model, for the aforementioned information promotion services unrelated to the buyer user click, a model such as random forest or LightGBM (Light Gradient Boosting Machine, light gradient boosting machine learning) can be used for training and prediction.

[0132] As for the information promotion services related to the buyer user click, since this type of situation is more complex and the prediction is more difficult, the demand for samples will be greater. However, due to the small amount of samples and the relatively serious missing of feature data, especially in the early stage of modeling, if traditional random forest and LightGBM are used, there can be problems such as insufficient learning. Therefore, for this type of LTV prediction, a Deep Forest model can be used. Based on the idea of ensemble learning, other learning models are stacked to obtain better representation performance. DeepForest is a deep structure based on this idea, which stacks multiple layers of random forests in a cascading manner to obtain better feature representation and learning performance, so that the model can perform better in small-scale training.

[0133] (Three) Cross-border supply chain service LTV

[0134] Due to the numerous links involved in cross-border trade, for example, including: customs, logistics, warehousing, foreign exchange settlement, tax refund, finance, etc., it is complicated and cumbersome. For large enterprises, these links may be self-absorbed and solved. For small and medium-sized enterprises, it is difficult to carry out cross-border trade. As a result, cross-border compliance supply chain services have emerged.

[0135] The cross-border supply chain service LTV, that is, the value contribution of the merchant user to the system through the use of the cross-border supply chain service provided by the system for this service category. For this cross-border supply chain service LTV, the value contribution of the supply chain service of multiple merchants in the future time period can be taken as the prediction target. Since the input of the supply chain service in the next year for each merchant is a continuous data, the regression model can also be used for the supply chain LTV model.

[0136] Among them, since the supply chain service involves many types of fees, the fees are very different, and the factors affecting each type of fee are different. The order distribution of the merchant determines the main revenue composition, and is also affected by the cost of the payment channel, the proportion of online payment, the proportion of withdrawal, and the level. Individual fee types are affected by the platform, the merchant, and the external environment, such as international exchange rates and international shipping prices. Therefore, it brings great challenges to the construction of the model.

[0137] Therefore, in the embodiments of the present application, the cross-border supply chain service can also be divided into multiple types, and different regression models are used for LTV prediction for various types of services. That is, the specific model can include multiple regression models corresponding to multiple different cross-border supply chain service types, and this model can be trained by targeting the resource consumption generated by the merchant user in the use of the corresponding type of cross-border supply chain service in the future time period.

[0138] Among them, the specific division method of cross-border supply chain service type can be various, for example, one way can be divided into the following types: cross-border supply chain service type related to transaction order and cross-border supply chain service type unrelated to transaction order; Among them, the cross-border supply chain service type related to transaction order is divided into several subtypes according to whether the generated resource consumption corresponds to the transaction order one by one. For example, the aforementioned transaction order related type can be divided into two parts: basic service fee part and logistics service fee part, wherein the basic service fee part is usually one-to-one corresponding to the transaction order, and each order generates a basic service fee. Although the logistics service fee is also related to the transaction order, it is not necessarily one-to-one. For example, the quantity of goods involved in the cross-border order may be large, and according to the urgency of the buyer's demand, the same transaction order may be split into multiple logistics orders, for example, part of the goods can be shipped by air to meet the buyer's urgent demand. The remaining part may be shipped by sea to save the logistics cost of both parties, etc. When different logistics channels are adopted, the logistics service fee generated is also different. In addition, the supply chain service unrelated to the transaction order can be financial and commission part, for example, the commission generated in the financing process of the merchant through the system, etc.

[0139] In this way, the supply chain service LTV can be expressed as:

[0140] LTV s =C order +C jr +C wl (4)

[0141] Among them, C order represents the value contribution of the order-related basic service aspect, C jr represents the value contribution of the financial and commission aspect, and C wl represents the value contribution of the logistics service aspect.

[0142] Regarding sample construction, the same can be used for each merchant user, randomly extracting one day as an observation time point every month to collect feature and consumption data. In this way, the richness of the sample can be met, and the generalization ability of the model can be improved.

[0143] Regarding feature engineering, the model features of the supply chain can be roughly divided into cost-related categories (buyer service fees in the last 30 days, foreign exchange income in the last 30 days, etc.), order and GMV (Gross Merchandise Volume, total merchandise transaction volume) related categories (historical order volume, etc.), merchant attributes (foreign trade strength, factory size, etc.), activity-related categories (number of active sub-accounts in the last 7 days, 7-day dwell time, etc.), supply chain service categories (online payment method proportion, each compliance channel proportion, etc.), and the like.

[0144] As for feature selection and feature importance ranking, for such a supply chain service LTV-related model, a causal analysis-based method can be used to select and / or rank the importance of feature dimensions for the target model corresponding to the cross-border supply chain service. The causal analysis-based method is based on causal verification and falsification of features and targets. The conclusion is whether there is a direct or indirect causal relationship between the features and the target to be predicted, etc. Through this causal analysis method, the importance of various feature dimensions to the target model can be more accurately determined, thereby selecting and ranking the feature dimensions.

[0145] In addition, in the supply chain service LTV-related prediction project, there will be some time window-related feature dimensions similar to the logistics service fee consumption in the last 7 days, the logistics service fee consumption in the last 30 days, and the logistics service fee consumption in the last 90 days. These features are all effective, and the importance of different time windows is different, which raises the question that the logistics income feature has a time dimension benefit, so it is necessary to use a method to reflect the difference in the time dimension of this feature.

[0146] For example, two merchant users A and B have both generated 10 pieces of logistics service consumption in the past 90 days. Among them, the 10 pieces of consumption of merchant A are all generated in the last 20 days, and the 10 pieces of consumption of the other merchant are generated 2 months ago. Obviously, the consumption of the former is more representative of the future consumption trend of the merchant A in this logistics aspect.

[0147] Therefore, in the preferred embodiments of the present application, the implementation scheme of processing features by using time decay coefficients based on Newton's cooling law can be adopted. Specifically, in the process of obtaining training samples for the target model corresponding to the cross-border supply chain service, or in the process of value prediction for the target merchant user, or in the process of obtaining the features of the target merchant user in multiple dimensions, for the dimensions associated with the time window, the time distribution of the multiple resource consumptions related to the cross-border supply chain service generated within the corresponding time window can be determined, as well as the distance between the time of specific consumption and the observation time point or the current prediction time, so as to determine different time decay coefficients for the multiple resource consumptions and then determine the features in the corresponding dimensions.

[0148] For example, in specific implementation, for the logistics service fee consumption in the last 7 days, the logistics service fee consumption in the last 30 days, the logistics service fee consumption in the last 90 days, and other features originally expressed by multiple feature dimensions, they can be combined into the same feature dimension, for example, unified as the feature dimension of “logistics service fee consumption in the last 90 days”. However, when obtaining the features in the dimension, the time distribution of the multiple resource consumptions can be obtained, that is, not only the total amount of “logistics service fee consumption in the last 90 days” needs to be obtained, but also how many pieces of logistics service fee consumption in the time window and the time of each consumption need to be obtained. In this way, the time difference between the time of each consumption and the current observation time point or the prediction time determines different time decay coefficients for each consumption, wherein the larger the time difference, the smaller the decay coefficient, that is, the lower the importance of the feature. In this way, the influence of the time of each consumption on model training and prediction can be better reflected, so as to improve the accuracy of model prediction.

[0149] The above describes the model training and prediction method of the three components of member service LTV, information promotion service LTV, and cross-border supply chain service LTV. In the specific prediction using a specific model, it can be performed according to a period, for example, once a day, etc. Among them, in each prediction, the target merchant user to be predicted can be determined first, and then the features of the target merchant user in multiple dimensions are obtained based on the current prediction time and input into each model (the feature dimensions required by each model may be different and can be obtained respectively), so as to predict multiple LTV components respectively, and then aggregate the multiple LTV components to obtain the comprehensive LTV value of the merchant user. Among them, each LTV component may also include multiple models respectively, for example, in the above example, the member service LTV may correspond to four classification models (of course, the same merchant user may only need to pass through two classification models for member service LTV prediction), the information promotion service LTV may correspond to two regression models, and the cross-border supply chain service LTV may correspond to three regression models. In this way, the values of multiple indicators can be predicted through the nine models, and the comprehensive LTV value of the same merchant user can also be aggregated.

[0150] Regarding the above merchant user LTV, there can be various applications. For example, in one way, after a merchant nurturing project A is developed, before it is officially put into service to provide services to merchants, the merchant nurturing effect that the merchant nurturing project A can bring in the future time period (for example, in the future one year) can be predicted, and then according to the prediction result, a final decision can be made on whether to officially put the merchant nurturing project A online.

[0151] Among them, for the above case, it can be realized through AB testing. Specifically, the merchant users can be randomly divided into groups M and N, among which the merchant nurturing project A is put online for the user group M, and the merchant nurturing project A is not put online for the user group N. Then, prediction can be performed every day, and in each prediction, each merchant user in the user groups M and N is taken as a target user to be tested, and the LTV of each target merchant user is predicted. Then, the LTV values of each merchant user in the user group M are aggregated to obtain the LTV total value of the user group M, and the LTV values of each merchant user in the user group N are aggregated to obtain the LTV total value of the user group N. In this way, by comparing the LTV values of different user groups in the same prediction period, the merchant nurturing effect that the merchant nurturing project A can obtain in the future can be determined, and then whether to officially put the merchant nurturing project A online to all merchant users can be determined according to the comparison of the merchant nurturing effect, etc.

[0152] Of course, when performing AB testing, it is also possible to put a certain target project / action online for one group of users of a merchant, and not put the target project / action online for another group of users of the merchant. In this way, by comparing the predicted results of LTV of different groups of users, a decision can be made on the target project / action.

[0153] In another application mode, since the prediction of LTV can be performed for each target merchant user respectively, the importance and direction of influence of each feature in the target model on the prediction result of the target merchant user can also be obtained according to the predicted result of LTV of the target merchant user. In this way, the first target feature that can be a potential factor for optimizing the prediction result and / or the second target feature that can be a risk factor for optimizing the prediction result can be determined according to the importance and direction of influence. Furthermore, growth help plan optimization suggestion information can also be provided to the target merchant user and / or a service personnel who provides growth help for the target merchant user according to the first target feature and / or the second target feature.

[0154] Among them, the importance and direction of influence of each feature on the prediction result of the target merchant user can be determined by Shap value. Shap is the abbreviation of Shapley Additive explanations, that is, Shapley Additive explanations. Through Shap value calculation, a prediction value is generated for each sample model, and the Shap value is the value allocated to each feature in the sample. That is, through the Shap value, the prediction result can be decomposed to the feature dimension, which is used to analyze the influence degree and direction of the feature on the prediction result. After obtaining the importance and direction of the feature through the Shap value, the top 3 (or other number) factors in the positive and negative directions can be taken according to the importance of the feature, and the positive direction is taken as a potential factor for optimizing LTV, and the negative direction is taken as a risk factor for optimizing LTV. In this way, differential improvement strategies can be specified for these indicators.

[0155] In summary, by the embodiments of the present application, the features in multiple dimensions obtained by statistics on the historical behavior data of the plurality of merchant users at at least one observation time point, and the value contribution situations of the plurality of merchant users respectively corresponding to a future time period starting from the observation time point can be obtained, then the feature values of the plurality of merchants respectively corresponding are taken as training input data, and the value contribution situations are taken as training targets, model training is performed to obtain a target model. Subsequently, the target merchant user that needs to be predicted in the current prediction period can be determined, and the features in multiple dimensions obtained by statistics on the historical behavior data of the target merchant user starting from the current prediction time point can be obtained, then the feature values of the target merchant user can be taken as input data of the target model, the value contribution situation of the target merchant user in a future time period starting from the current prediction time point is predicted, so as to make a decision or optimization on the target project / action according to the prediction result of the value contribution situation. Since the execution of each merchant nurturing action can affect the value contribution situation of the merchant user in the future, the influence is quantitatively expressed by the embodiments of the present application, thereby providing a sensitive and short-term observable index for the merchant nurturing project, thereby helping the operation service personnel to obtain data support when making a selection decision on the merchant nurturing project, and the existing merchant nurturing project can be optimized or adjusted in time when a negative influence on the future value contribution situation is found, thereby helping the system to obtain the expected merchant nurturing effect.

[0156] In the preferred embodiments, the merchant value contribution situation can be divided into a plurality of value contribution components, and each corresponds to a different target model, so that the selection of feature dimensions and / or the importance ranking of feature dimensions can be performed for different target models to complete the training of the model, thereby improving the accuracy of the model.

[0157] In terms of constructing training samples, by randomly selecting a day in each month as an observation time point for each merchant user, a plurality of training sample data of the same merchant user can be generated, thereby improving the richness of the samples and the generalization ability of the model.

[0158] For the value contribution component of the member service, a scheme of combining the first classification model and the second classification model with a cascade relationship is adopted, so that each model can be a binary classification model, which is beneficial to reduce the prediction difficulty of the model and improve the accuracy of the model.

[0159] For the value contribution component of the information promotion service, for the information promotion service type related to the resource consumption generated and the click situation of the consumer user, a deep structure model of a multi-layer random forest can be stacked in a cascading manner to make up for the accuracy problem caused by insufficient sample size, feature missing, and the like. In addition, for multiple training samples corresponding to the same merchant user at different observation time points, features in the time dimension can also be constructed to smooth the differences between different observation time points.

[0160] For the value contribution component of the cross-border supply chain service, a time decay coefficient based on Newton's cooling law can be used to determine the time distribution of multiple resource consumptions related to the cross-border supply chain service in the corresponding time window and the distance between the observation time point or the current prediction time, and different time decay coefficients for the multiple resource consumptions, and determine the features in the corresponding dimension. This can reflect the different effects of different times of specific consumption on the trend of future consumption of the merchant user in the corresponding aspect, thereby further improving the accuracy of the model.

[0161] In the application aspect, the value contribution of different merchant nurturing projects in the future time period can be compared through AB testing and the like, so as to make a selection of the merchant nurturing project or make a decision on whether a certain merchant nurturing project can be officially launched and released.

[0162] The importance and influence direction of each feature in the target model in influencing the prediction result of the target merchant user can also be obtained according to the prediction result of the value contribution of the target merchant user, so as to identify the problems of the merchant user, help the merchant diagnose and operate, and improve the operating effect of the merchant.

[0163] It should be noted that the present application embodiments can involve the use of user data. In actual application, user-specific personal data can be used in the schemes described herein within the scope allowed by applicable laws and regulations, for example, with the explicit consent of the user, the actual notification to the user, and the like.

[0164] Corresponding to the foregoing method embodiments, the present application embodiments also provide a data processing apparatus, which is described with reference to Figure 5 The apparatus can include:

[0165] The model obtaining unit 501 is configured to obtain a target model, wherein the target model is obtained by the following manner: a plurality of dimensions of features obtained by statistically processing historical behavior data of a plurality of merchant users at at least one observation time point, and a value contribution situation of the plurality of merchant users respectively corresponding to a future time period starting from the observation time point, taking the feature values of the plurality of merchants respectively corresponding as training input data, and taking the value contribution situation as a training target, and performing model training to obtain the target model;

[0166] The feature obtaining unit 502 is configured to determine a target merchant user in a current prediction period for which value prediction needs to be performed, and obtain a plurality of dimensions of features by statistically processing historical behavior data of the target merchant user starting from a current prediction time point;

[0167] The prediction unit 503 is configured to take the feature values corresponding to the target merchant user as input data of the target model, to predict a value contribution situation of the target merchant user in a future time period starting from the current prediction time point, so as to make a decision or optimization on a target item / action according to the value contribution situation prediction result.

[0168] In this embodiment of the present application, the value contribution situation can be divided into a plurality of value contribution components according to different service categories available to the merchant user, and the different target models correspond to the different value contribution components respectively, so as to select the feature dimensions and / or sort the importance of the feature dimensions for the different target models respectively, to complete the training of the models.

[0169] In this embodiment of the present application, the service category includes a member service, and the corresponding value contribution situation is a value contribution situation made by the merchant user to the system by becoming a system member through renewal;

[0170] The target model corresponding to the member service includes a first classification model and a second classification model having a cascading relationship, the first classification model is configured to predict a renewal probability of the merchant user in a future target time period, and is trained with the target of completing renewal of the merchant user in the future target time period; and the second classification model is configured to predict a probability of whether a member level of the merchant user will change when the merchant user renews, and is trained with the target of whether the member level of the merchant user changes when the merchant user renews in the future target time period.

[0171] In this embodiment of the present application, the first classification model and the second classification model having the cascading relationship can be a plurality of groups, corresponding to a plurality of different member levels in which different target merchant users currently stay.

[0172] In addition, the service category includes an information promotion service, and the corresponding value contribution situation is a value contribution situation made by the merchant user to the system by using the information promotion service provided by the system.

[0173] The target model corresponding to the information promotion service includes: a plurality of regression models corresponding to a plurality of different information promotion service types, which are trained to target resource consumption generated by the merchant user in the use of the corresponding type of information promotion service in a future time period.

[0174] Among them, the information promotion service type includes: an information promotion service type related to the generated resource consumption and the click situation of the consumer user, and the corresponding regression model includes: a deep structure model of stacking multiple layers of random forests in a cascading manner.

[0175] In addition, the service category includes cross-border supply chain services, and the corresponding value contribution situation is: the value contribution situation made by the merchant user to the system by using the cross-border supply chain services provided by the system.

[0176] The target model corresponding to the cross-border supply chain service includes: a plurality of regression models corresponding to a plurality of different cross-border supply chain service types, which are trained to target resource consumption generated by the merchant user in the use of the corresponding type of cross-border supply chain service in a future time period.

[0177] Among them, the cross-border supply chain service type includes: a cross-border supply chain service type related to a transaction order, and a cross-border supply chain service type unrelated to a transaction order; wherein, the cross-border supply chain service type related to the transaction order is further divided into a plurality of subtypes according to whether the generated resource consumption is one-to-one corresponding to the transaction order.

[0178] Specifically, in the process of obtaining training samples for the target model corresponding to the cross-border supply chain service, or in the process of value prediction for the target merchant user, in the process of obtaining the features of the target merchant user in multiple dimensions, for the dimension associated with the time window, the time distribution of the multiple resource consumptions related to the cross-border supply chain service generated in the corresponding time window and the distance between the observation time point or the current prediction time are determined, and different time decay coefficients are determined for the multiple resource consumptions, and the features in the corresponding dimension are determined.

[0179] In addition, the selection of feature dimensions and / or the importance ranking of feature dimensions for the target model corresponding to the cross-border supply chain service can be based on the method of causal analysis.

[0180] Further, in the process of obtaining the training samples for the target model, a plurality of observation time points can be determined for each merchant user, so as to obtain a plurality of training samples for the same merchant user at a plurality of different observation time points, for model training; wherein the plurality of observation time points corresponding to the same merchant user are evenly distributed between different sub-time periods and have randomness within the same sub-time period.

[0181] In the process of obtaining the training samples for the target model corresponding to the information promotion service, a time dimension feature can be constructed for the plurality of training samples corresponding to the same merchant user at different observation time points, and the time dimension feature includes: whether the sub-time period in which the observation time point is located is associated with a target marketing activity, the distance between the observation time point and the start or end time of the target marketing activity, and / or the identifier of the sub-time period in which the observation time point is located.

[0182] Specifically, the feature acquisition unit can be specifically configured to:

[0183] In the process of AB testing of the target item / action, the target merchant user is determined from two merchant user groups participating in the AB testing, wherein different target items / actions are online for different merchant user groups, or a target target item / action is online for one of the merchant user groups, and the target target item / action is not online for the other merchant user group;

[0184] At this time, the device can further include:

[0185] The aggregation comparison unit is configured to aggregate the prediction results of the value contribution of each target merchant user in the same user group in the future time period to determine the prediction result of the merchant nurturing effect that the same target item / action can obtain in the future time period, so as to compare the prediction results between different target items / actions to select or decide the target item / action.

[0186] In another application mode, the device can further include:

[0187] The feature importance analysis unit is configured to obtain the importance and influence direction of each feature in the target model in influencing the prediction result of the target merchant user according to the prediction result of the value contribution of the target merchant user;

[0188] The target feature determination unit is configured to determine a first target feature that can be a potential factor for optimizing the prediction result and / or a second target feature that can be a risk factor for optimizing the prediction result according to the importance and influence direction;

[0189] The suggestion information providing unit is configured to provide optimization suggestion information of a target item / action to the target merchant user and / or a service personnel providing a service for the target merchant user according to the first target feature and / or the second target feature.

[0190] In addition, the embodiments of the present application further provide a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the steps of the method in any one of the foregoing method embodiments.

[0191] An electronic device includes:

[0192] one or more processors; and

[0193] a memory associated with the one or more processors, the memory configured to store program instructions that, when executed by the one or more processors, perform the steps of the method in any one of the foregoing method embodiments.

[0194] wherein, Figure 6 An exemplary architecture of an electronic device is shown, which can specifically include a processor 610, a video display adapter 611, a disk drive 612, an input / output interface 613, a network interface 614, and a memory 620. The processor 610, the video display adapter 611, the disk drive 612, the input / output interface 613, the network interface 614, and the memory 620 can be communicatively connected through a communication bus 630.

[0195] The processor 610 can be implemented in a general-purpose CPU (Central Processing Unit), a microprocessor, an ASIC (Application Specific Integrated Circuit), or one or more integrated circuits, and is configured to execute related programs to implement the technical solutions provided by the present application.

[0196] The memory 620 can be implemented in the form of a ROM (Read Only Memory), a RAM (Random Access Memory), a static storage device, a dynamic storage device, etc. The memory 620 can store an operating system 621 for controlling the operation of the electronic device 600, a basic input / output system (BIOS) for controlling the low-level operation of the electronic device 600. In addition, a web browser 623, a data storage management system 624, and a data processing system 625, etc. can also be stored. The data processing system 625 can be an application program for implementing the above-mentioned steps in the embodiments of the present application. In summary, when the technical solutions provided in the present application are implemented by software or firmware, the relevant program codes are stored in the memory 620 and executed by the processor 610.

[0197] The input / output interface 613 is configured to connect an input / output module to realize information input and output. The input / output module can be configured in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. The input device can include a keyboard, a mouse, a touch screen, a microphone, various sensors, etc., and the output device can include a display, a speaker, a vibrator, an indicator light, etc.

[0198] The network interface 614 is configured to connect a communication module (not shown in the figure) to realize the communication interaction between the device and other devices. The communication module can realize communication through a wired manner (such as USB, network cable, etc.) or a wireless manner (such as mobile network, WIFI, Bluetooth, etc.).

[0199] The bus 630 includes a path for transmitting information between various components (such as the processor 610, the video display adapter 611, the disk drive 612, the input / output interface 613, the network interface 614, and the memory 620) of the device.

[0200] It should be noted that although the above device only shows the processor 610, the video display adapter 611, the disk drive 612, the input / output interface 613, the network interface 614, the memory 620, and the bus 630, etc., in the specific implementation process, the device can also include other components necessary for normal operation. In addition, those skilled in the art can understand that the above device can also only contain the components necessary for implementing the solutions of the present application, and does not necessarily contain all the components shown in the figure.

[0201] Those skilled in the art can clearly understand the application by the description of the above embodiments that the application can be implemented by means of software and the necessary universal hardware platforms. Based on such an understanding, the technical solutions of the application can be embodied in the form of a software product, and the computer software product can be stored in a storage medium, such as a ROM / RAM, a magnetic disk, an optical disk, and the like, and includes a plurality of instructions to make a computer device (which can be a personal computer, a server, or a network device, and the like) execute the methods described in each embodiment or some parts of the embodiments of the application.

[0202] Each of the embodiments in the specification is described in a progressive manner, and the same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the difference from other embodiments. In particular, for the system or system embodiments, since it is basically similar to the method embodiments, the description is relatively simple, and the related parts can be referred to the part of the method embodiments. The above-described system and system embodiments are merely illustrative, and the units described as separate components can be or can not be physically separated, and the components displayed as units can be or can not be physical units, that is, they can be located in one place, or can be distributed on multiple network units. According to the actual needs, some or all of the modules can be selected to achieve the purpose of the embodiment. Those skilled in the art can understand and implement it without creative labor.

[0203] The data processing method and the electronic device provided by the application are described in detail above, and the principle and implementation manner of the application are described by applying specific examples. The above embodiment is only used to help understand the method and the core idea of the application; meanwhile, for those skilled in the art, according to the idea of the application, the specific implementation manner and application range can be changed. In conclusion, the content of the specification should not be understood as a limitation of the application.

Claims

1. A data processing method, characterized by, The method comprises the following steps: obtaining a target model by the following method: obtaining features in multiple dimensions from historical behavior data of a plurality of merchant users at at least one observation time point, and the value contribution of the plurality of merchant users respectively corresponding to a future time period starting from the observation time point, taking the feature values of the plurality of merchants respectively corresponding as training input data, and taking the value contribution as the training target, and training the model to obtain the target model; wherein, according to the different service categories available to the merchant users, the value contribution is divided into a plurality of value contribution components, and different target models are respectively corresponded, so as to select the feature dimensions and / or importance of the feature dimensions for different target models respectively, so as to complete the training of the model; determining a target merchant user who needs to be predicted in a current prediction period, and obtaining features in multiple dimensions from historical behavior data of the target merchant user at a current prediction time; wherein, in the process of AB testing of the target item / action, the target merchant user is determined from two merchant user groups participating in the AB testing, wherein different target items / actions are online for different merchant user groups, or the target target item / action is online for one merchant user group, and the target item / action is not online for the other merchant user group; taking the feature values corresponding to the target merchant user as the input data of the target model, predicting the value contribution of the target merchant user in a future time period starting from the current prediction time, so as to make a decision or optimization on the target item / action according to the value contribution prediction result; aggregating the value contribution prediction results of each target merchant user in the same user group in the future time period to determine the prediction result of the merchant nurturing effect of the same target item / action in the future time period, so as to select or decide the target item / action by comparing the prediction results of different target items / actions.

2. The method of claim 1, wherein: the service category includes member service, and the corresponding value contribution situation is the value contribution made by the merchant user to the system by becoming a system member through re-signing; the target model corresponding to the member service includes a first classification model and a second classification model having a cascade relationship, the first classification model is used to predict the re-signing probability of the merchant user in a future target time period, and is trained with the completion of re-signing of the merchant user in the future target time period as the target; the second classification model is used to predict the probability of whether the member level of the merchant user will change when re-signing, and is trained with whether the member level of the merchant user changes when re-signing in the future target time period as the target.

3. The method of claim 1, wherein: the service category includes information promotion service, and the corresponding value contribution situation is the value contribution made by the merchant user to the system by using the information promotion service provided by the system; The target model corresponding to the information promotion service includes: a plurality of regression models corresponding to a plurality of different information promotion service types, and the plurality of regression models are trained to target resource consumption generated by a merchant user in using a corresponding type of information promotion service in a future time period.

4. The method of claim 3, wherein, The information promotion service type includes an information promotion service type in which generated resource consumption is related to click conditions of a consumer user, and the corresponding regression model includes a deep structure model of a plurality of stacked random forests.

5. The method of claim 1, wherein, The service category includes cross-border supply chain services, and the corresponding value contribution condition is a value contribution condition made by a merchant user to the system by using cross-border supply chain services provided by the system. The target model corresponding to the cross-border supply chain service includes: a plurality of regression models corresponding to a plurality of different cross-border supply chain service types, and the plurality of regression models are trained to target resource consumption generated by a merchant user in using a corresponding type of cross-border supply chain service in a future time period.

6. The method of claim 5, wherein, The cross-border supply chain service type includes a cross-border supply chain service type related to a transaction order and a cross-border supply chain service type unrelated to a transaction order, and the cross-border supply chain service type related to a transaction order is further divided into a plurality of subtypes according to whether the generated resource consumption is one-to-one corresponding to a transaction order.

7. The method of claim 5, wherein, In the process of obtaining training samples for the target model corresponding to the cross-border supply chain service, or in the process of predicting the value for a target merchant user, the features of the target merchant user in multiple dimensions are obtained, for the dimensions associated with the time window, the time distribution of a plurality of resource consumptions related to the cross-border supply chain service generated in the corresponding time window and the distance between the observation time point or the current prediction time are determined, and different time decay coefficients are determined for the plurality of resource consumptions, and the features in the corresponding dimension are determined.

8. The method of any one of claims 2, 3 or 5, wherein, In the process of obtaining training samples for the target model, a plurality of observation time points are determined for each merchant user, so as to obtain a plurality of training samples for the same merchant user at a plurality of different observation time points for model training, and the plurality of observation time points corresponding to the same merchant user are evenly distributed between different sub-time periods and have randomness in the same sub-time period.

9. The method of claim 8, wherein, In the process of obtaining training samples for the target model corresponding to the information promotion service, for multiple training samples corresponding to the same merchant user at different observation time points, a feature in the time dimension is constructed, the feature in the time dimension including whether a sub-time period in which the observation time point is located is associated with a target marketing activity, a distance between the observation time point and a start or end time of the target marketing activity, and / or an identifier of the sub-time period in which the observation time point is located.

10. The method according to any one of claims 1 to 7, 9, characterized in that, Further comprising: obtaining, according to a prediction result of a value contribution of the target merchant user, an importance degree and an influence direction of each feature in the target model in influencing the prediction result of the target merchant user; determining, according to the importance degree and the influence direction, a first target feature that can be a potential factor for optimizing the prediction result and / or a second target feature that can be a risk factor for optimizing the prediction result; providing, according to the first target feature and / or the second target feature, optimization suggestion information of a target item / action to the target merchant user and / or a service personnel providing a merchant service for the target merchant user.

11. A computer readable storage medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the steps of the method of any one of claims 1 to 10.

12. An electronic device, comprising: Comprise: one or more processors; and a memory associated with the one or more processors, the memory being configured to store program instructions, the program instructions being configured to be read and executed by the one or more processors to perform the steps of the method of any one of claims 1 to 10.

Images