Data prediction method, electronic device, storage medium, and computer program product

Abstract

Embodiments of the present disclosure provide a data prediction method, an electronic device, a storage medium and a computer program product. The method can include: obtaining portrait data and life cycle data of a target user account, and generating portrait features corresponding to the portrait data; the portrait features include a paid feature and a non-paid feature, and the life cycle data is used to represent a payment state of the target user account; inputting the portrait features and the life cycle data into an LTV prediction model to obtain an LTV prediction result corresponding to the target user account; wherein the LTV prediction model includes: a gating network configured to determine a first weight corresponding to the paid feature and a second weight corresponding to the non-paid feature according to the payment state represented by the life cycle data; a payment amount prediction network configured to perform weighted processing on the paid feature according to the first weight, perform weighted processing on the non-paid feature according to the second weight, and generate the LTV prediction result based on the paid feature and the non-paid feature obtained by the weighted processing.

Description

TECHNICAL FIELD

[0001] Embodiments of the present disclosure relate to the field of artificial intelligence, and more particularly, to a data prediction method, an electronic device, a storage medium and a computer program product. BACKGROUND

[0002] This section is intended to provide background information to facilitate a better understanding of embodiments of the disclosure recited in the claims. The description herein does not constitute admission that the information provided herein is prior art merely because it is included herein.

[0003] Life Time Value (LTV) is an index for measuring the total value of a user's contribution to a platform or enterprise during the entire product cycle (or a time period). Predicting the LTV of a user helps the enterprise or platform to more accurately assess the value of an individual user, so as to take different marketing strategies for users of different values, to maximize the long-term value of users and improve user retention rate.

[0004] In related technologies, LTV prediction is usually performed according to user portrait data and historical payment records. However, this approach fails to fully consider the differences in payment behavior among different users, resulting in inaccurate LTV prediction results. SUMMARY

[0005] In a first aspect of the embodiments of the present disclosure, a data prediction method is provided, comprising:

[0006] Obtaining portrait data and life cycle data of a target user account, and generating portrait features corresponding to the portrait data; the portrait features include payment features and non-payment features, and the life cycle data is used to represent the payment state of the target user account;

[0007] Inputting the portrait features and the life cycle data into an LTV prediction model to obtain an LTV prediction result corresponding to the target user account; wherein the LTV prediction model comprises:

[0008] a gating network for determining a first weight corresponding to the payment features and a second weight corresponding to the non-payment features according to the payment state represented by the life cycle data;

[0009] a payment amount prediction network for weighting the payment features according to the first weight, weighting the non-payment features according to the second weight, and generating the LTV prediction result based on the weighted payment features and non-payment features.

[0010] Optionally, the portrait data comprises static attribute data, historical active behavior data and payment data of the target user account, and the generating the portrait feature corresponding to the portrait data comprises: performing feature extraction on the payment data to obtain payment features; performing feature extraction on the static attribute data and the historical active behavior data respectively to obtain non-payment features, the non-payment features comprising static attribute features corresponding to the static attribute data and historical active behavior features corresponding to the historical active behavior data; performing normalization processing on the payment features and the non-payment features respectively, and splicing the processed payment features and the processed non-payment features to obtain the portrait feature.

[0011] Optionally, the LTV prediction model further comprises at least one payment behavior prediction sub-network, each payment behavior prediction sub-network corresponding to a preset time window; wherein any payment behavior prediction sub-network is configured to generate a payment behavior prediction result of the target user account in a corresponding preset time window according to the portrait feature and the first weight and the second weight output by the gating network; and the payment amount prediction network is further configured to generate the LTV prediction result according to the payment features and the non-payment features obtained through weighting processing and the payment behavior prediction results generated by each payment behavior prediction sub-network.

[0012] Optionally, the LTV prediction model further comprises at least one payment amount prediction sub-network, each payment amount prediction sub-network corresponding to a preset time window, and all preset time windows being different from the time window corresponding to the payment amount prediction network; wherein any payment amount prediction sub-network is configured to generate an LTV prediction result of the target user account in a corresponding preset time window according to the payment features and the non-payment features obtained through weighting processing.

[0013] Optionally, the payment amount prediction network comprises a prediction layer and at least one feature processing layer, each feature processing layer corresponding to a different feature dimension and each feature processing layer corresponding to a gating network; wherein any feature processing layer is configured to perform feature analysis based on the output of a previous feature processing layer and the first weight and the second weight output by a corresponding gating network, and output features of a corresponding feature dimension; the input of the first feature processing layer is the payment features, the non-payment features and the first weight and the second weight output by the corresponding gating network; and the prediction layer is configured to perform prediction based on the output of the last feature processing layer to generate the LTV prediction result.

[0014] In a second aspect of the embodiments of the present disclosure, a training method of a data prediction model is provided, comprising:

[0015] obtaining sample portrait data, sample life cycle data and sample LTV of a sample user account, and generating sample portrait features corresponding to the sample portrait data; wherein the sample portrait features comprise sample payment features and sample non-payment features, and the sample life cycle data is used to represent a sample payment state of the sample user account;

[0016] inputting the sample portrait features and the sample life cycle data into a to-be-trained LTV prediction model to obtain a sample LTV prediction result corresponding to the sample user account;

[0017] calculating an LTV loss value based on the sample LTV and the sample LTV prediction result, and iteratively training the to-be-trained LTV prediction model according to the LTV loss value;

[0018] The to-be-trained LTV prediction model comprises:

[0019] a gating network configured to determine a first sample weight corresponding to the sample payment features and a second sample weight corresponding to the sample non-payment features according to a sample payment state represented by the sample life cycle data;

[0020] a payment amount prediction network configured to perform weighted processing on the sample payment features according to the first sample weight, perform weighted processing on the sample non-payment features according to the second sample weight, and generate the sample LTV prediction result based on the sample payment features and the sample non-payment features obtained through the weighted processing.

[0021] Optionally, the to-be-trained LTV prediction model further comprises at least one payment behavior prediction sub-network, and each payment behavior prediction sub-network corresponds to a sample preset time window; wherein any payment behavior prediction sub-network is configured to generate a sample payment behavior prediction result of the sample user account in a corresponding sample preset time window according to the sample portrait features, the first sample weight and the second sample weight output by the gating network; and the payment amount prediction network is further configured to generate the sample LTV prediction result according to the sample payment features and the sample non-payment features obtained through the weighted processing, and the sample payment behavior prediction results generated by each payment behavior prediction sub-network.

[0022] Optionally, the method further comprises: obtaining sample payment behaviors in each sample preset time window, and calculating a payment behavior loss value corresponding to each sample preset time window according to the sample payment behaviors and the sample payment behavior prediction results in each sample preset time window; and the iteratively training the to-be-trained LTV prediction model according to the LTV loss value comprises: iteratively training the to-be-trained LTV prediction model according to the LTV loss value and the payment behavior loss values corresponding to each sample preset time window.

[0023] Optionally, the LTV prediction model to be trained further comprises at least one payment amount prediction sub-network, each payment amount prediction sub-network corresponding to a sample preset time window, and all sample preset time windows being different from the sample time window corresponding to the payment amount prediction network; wherein any payment amount prediction sub-network is configured to generate a sample LTV prediction result of the sample user account within the corresponding sample preset time window according to the sample payment features and the sample non-payment features obtained through the weighted processing.

[0024] Optionally, the method further comprises: obtaining the sample LTV within each sample preset time window, and calculating the LTV loss value corresponding to each sample preset time window according to the sample LTV and the sample LTV prediction result corresponding to each sample preset time window; and the iterative training of the LTV prediction model to be trained according to the LTV loss value comprises: iterative training of the LTV prediction model to be trained according to the LTV loss value and the LTV loss value corresponding to each sample preset time window.

[0025] In a third aspect of the embodiments of the present disclosure, an electronic device is provided, comprising:

[0026] a processor;

[0027] a memory for storing executable instructions of the processor;

[0028] The processor implements the method of any one of the first aspect or the second aspect by running the executable instructions.

[0029] In a fourth aspect of the embodiments of the present disclosure, a computer readable storage medium is provided, which stores computer instructions, and the instructions are executed by a processor to implement the method of any one of the first aspect or the second aspect.

[0030] In a fifth aspect of the embodiments of the present disclosure, a computer program product is provided, which comprises computer program / instructions, and the computer program / instructions are executed by a processor to implement the steps of the method of any one of the first aspect or the second aspect.

[0031] The above embodiments of the present disclosure have at least the following beneficial effects:

[0032] The portrait data and the life cycle data of the target user account are input into an LTV prediction model to generate an LTV prediction result corresponding to the target user account by using the LTV prediction model. The LTV prediction model includes a gating network and a payment amount prediction network. In the prediction process, the gating network can determine the weights of the payment features and the non-payment features included in the portrait features corresponding to the portrait data according to the payment state of the target user account. Then, the payment amount prediction network can generate the LTV prediction result according to the weighted payment features and the weighted non-payment features.

[0033] Since the payment features of different users reflect the differences in payment behaviors between different users, the weights of the payment features and the non-payment features of the users are controlled by the gating network, so as to control the influence degree of the payment features and the non-payment features on the LTV prediction result based on the corresponding weights, which can make the generated LTV prediction result more consistent with the payment characteristics of the users. This prediction method fully considers the influence of the differences in payment behaviors of different users on the LTV prediction result, which can significantly improve the accuracy of the LTV prediction result. It can be seen that the technical solution of the present disclosure can generate accurate LTV prediction results for users with different payment behaviors. BRIEF DESCRIPTION OF DRAWINGS

[0034] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will be more apparent from the following detailed description taken in conjunction with the accompanying drawings, in which:

[0035] Figure 1 FIG. 1 is a schematic diagram of a system architecture of a data prediction method according to an example embodiment;

[0036] Figure 2 FIG. 2 is a flowchart of a data prediction method according to an example embodiment;

[0037] Figure 3 FIG. 3 is a schematic diagram of a structure of an LTV prediction model according to an example embodiment;

[0038] Figure 4 FIG. 4 is a schematic diagram of another structure of an LTV prediction model according to an example embodiment;

[0039] Figure 5 FIG. 5 is a schematic diagram of a structure of a gating network according to an example embodiment;

[0040] Figure 6a FIG. 6 is a schematic diagram of a structure of an LTV prediction model including a payment behavior prediction sub-network according to an example embodiment;

[0041] Figure 6bFIG. 6 is a structural diagram of an LTV prediction model including a payment behavior prediction sub-network according to an example embodiment;

[0042] Figure 6c FIG. 7 is a structural diagram of a payment behavior prediction sub-network according to an example embodiment;

[0043] Figure 7a FIG. 8 is a structural diagram of an LTV prediction model including a payment amount prediction sub-network according to an example embodiment;

[0044] Figure 7b FIG. 9 is a structural diagram of an LTV prediction model including a payment amount prediction sub-network according to another example embodiment;

[0045] Figure 8a FIG. 10 is a structural diagram of an LTV prediction model including both a payment amount prediction sub-network and a payment behavior prediction sub-network according to an example embodiment;

[0046] Figure 8b FIG. 11 is a structural diagram of an LTV prediction model according to another example embodiment;

[0047] Figure 9 FIG. 12 is a flowchart of a training method of a data prediction model according to an example embodiment;

[0048] Figure 10 FIG. 13 is a block diagram of a data prediction apparatus according to an example embodiment;

[0049] Figure 11 FIG. 14 is a block diagram of a training apparatus of a data prediction model according to an example embodiment;

[0050] Figure 12 FIG. 15 is a schematic diagram of a readable storage medium corresponding to the method according to an example embodiment;

[0051] Figure 13 FIG. 16 is a schematic diagram of an electronic device capable of implementing the method according to an example embodiment.

[0052] In the drawings, identical or corresponding reference signs indicate identical or corresponding parts. DETAILED DESCRIPTION

[0053] The principles and spirits of the present disclosure will be described below with reference to several example embodiments. It should be understood that these embodiments are given only to enable those skilled in the art to better understand and implement the present disclosure, and do not limit the scope of the present disclosure in any way. On the contrary, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0054] Those skilled in the art understand that the embodiments of the present disclosure can be implemented as a system, a device, an apparatus, a method or a computer readable storage medium. Therefore, the present disclosure can be embodied in the form of entirely hardware, entirely software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0055] According to embodiments of the present disclosure, a data prediction method, an electronic device, a storage medium and a computer program product are provided.

[0056] In this document, it should be understood that the number of any elements in the drawings is merely for example and not a limitation, and any naming is merely for differentiation and does not have any limiting meaning. Moreover, the data involved in the present disclosure can be data authorized by a user or data sufficiently authorized by all parties.

[0057] The principles and spirits of the present disclosure will be explained in detail below with reference to several representative embodiments of the present disclosure.

[0058] Overview of application scenarios

[0059] In the field of Internet, LTV is usually used to measure the economic benefits brought by users of an application program to the operator of the application program. For example, for a music application program, the LTV of a user can be understood as the total amount of money that the user charges for the music application program within a time interval, which can include the amount of money that the user spends to purchase a membership, the amount of money that the user spends to purchase songs, and the amount of money that the user spends for other charging activities. Charging activities refer to charging activities participated in the music application program. Predicting the LTV of a user helps to improve the retention rate or conversion rate of the user in the application program.

[0060] In the related art, LTV prediction is usually performed according to user portrait data and historical payment records. However, this method fails to fully consider the differences in payment behaviors among different users, resulting in inaccurate LTV prediction results. For example, for a user who has never paid, the user does not have historical payment records, and therefore the absence of this part of data in the historical payment records during the prediction process leads to inaccurate LTV prediction results.

[0061] As can be seen, in the scenario of predicting the LTV of a user, there can be a problem that accurate LTV prediction results cannot be generated for some users.

[0062] It should be noted that the above application scenarios are only shown for the purpose of facilitating the understanding of the spirits and principles of the present disclosure, and the embodiments of the present disclosure are not limited in this respect. On the contrary, the embodiments of the present disclosure can be applied to any applicable scenario. SUMMARY

[0064] Therefore, the present specification provides a technical solution of data prediction.

[0065] The core idea of the present specification is that a gating network is introduced into an LTV prediction model to control the weights of the payment features and non-payment features of users on the LTV prediction result through the gating network, and then generate the LTV prediction result based on the weighted payment features and weighted non-payment features. In this way, since the payment features of different users reflect the differences in payment behaviors / records between different users, the influence degree of the payment features and non-payment features on the LTV prediction result can be controlled through the corresponding weights, so that the generated LTV prediction result can be more consistent with the payment characteristics of the users, significantly improve the accuracy of the LTV prediction result, and realize accurate LTV prediction results for users with different payment behaviors.

[0066] Exemplary method

[0067] The technical idea of the present specification will be described in detail below (in combination with the system architecture diagram as shown in Figure 1

[0068]

System Architecture

[0069] In an embodiment of the present disclosure, the terminal can be an electronic device used by a user. Specifically, the terminal can include but is not limited to a smartphone, a desktop computer, a tablet computer, a notebook computer, an e-book reader, a smart watch, a smart bracelet, and other electronic devices with certain computing capabilities, and the electronic device can run software or websites of instant messaging, social networking, gaming, education, and other categories, and can be applied to the scene of multi-person audio / video call.

[0070] The server can include a server, a server cluster formed by multiple servers, or a cloud server. Specifically, the server can be a standalone physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN (Content Delivery Network), and basic cloud computing services such as big data and artificial intelligence platforms.

[0071] The terminal and the server can be directly or indirectly connected through wired or wireless communication, and the present disclosure does not make special limitations.

[0072] Based on Figure 1 ​As shown in the system architecture diagram, the data prediction system includes a server 11 and terminals 12-14, and the LTV prediction model is deployed on the server 11. Any terminal can obtain the portrait data and the life cycle data of a target user account, and upload the obtained data to the server. The server can generate corresponding portrait features for the received portrait data, and input the portrait features and the received life cycle data into the LTV prediction model. Then, the server returns the LTV prediction result output by the LTV prediction model to the terminal that submits the portrait data and the life cycle data.

[0073] In an embodiment of the present disclosure, the data prediction system can also include a terminal, and the LTV prediction model is deployed on the terminal. The terminal can obtain the portrait data and the life cycle data of a target user account, and generate corresponding portrait features for the obtained portrait data. Then, the terminal inputs the portrait features and the life cycle data into the locally deployed LTV prediction model to obtain the LTV prediction result output by the LTV prediction model.

[0074] The following will take a terminal as an example to execute the data prediction method provided by the embodiments of the present disclosure for example-based description. Please refer to Figure 2 , Figure 2 is a flowchart of a data prediction method provided by an example embodiment. The method can include the following steps:

[0075] Step 201: Obtain the portrait data and the life cycle data of a target user account, and generate the portrait features corresponding to the portrait data; the portrait features include the paid features and the non-paid features, and the life cycle data is used to represent the payment status of the target user account.

[0076] Generally, a user performs various activities on a system or platform through the corresponding account of the user. The user account is an identity created by the user for accessing online services, applications or websites, which usually includes login credentials such as username and password. The activity records (such as purchase history, browsing behavior, comments, etc.) of the user on the platform or application are associated with the account created by the user on the corresponding platform or application. If allowed, a user can have multiple accounts on the same platform / application.

[0077] In this embodiment, the account of the target user is referred to as a "target user account", which can be any account for which the LTV needs to be predicted. The portrait data and the life cycle data of the target user account refer to the portrait data and the life cycle data associated with the target user account. The portrait data refers to a collection of data of various attributes and behavioral characteristics of the target user, aiming to depict a specific user image, which can include but is not limited to age, gender, geographic location, interests, consumption habits, active duration, payment record. The life cycle of a user on a platform / application can be understood as the entire process of subscribing to a service in the platform / application: from the first registration of the user, through the occurrence of important events (such as completing a purchase, upgrading a membership, etc.), to the user's cancellation of subscription or termination of service. In this embodiment, the life cycle data of the target user account can include the payment status of the target user account. For example, the payment status of a user account can be divided into a new member (i.e. a user who has never paid), a valid member (i.e. a current paid user), and a member loss (i.e. a paid user who has become a non-paid user). Alternatively, the payment status of a user account can be divided into a non-paid user, an active member (i.e. a non-paid user who has become a paid user), a high-value member (i.e. a long-term stable paid user), and a lost member (i.e. a paid user who has become a non-paid user). It should be noted that the above division is only an example, and those skilled in the art can divide the payment status of a user account according to their own needs, and the present disclosure does not limit this.

[0078] If the LTV of the target user on any platform / system is to be predicted, the portrait data and the life cycle data associated with the account of the target user on the platform / system can be obtained. Then, the terminal can perform feature extraction on the portrait data to generate corresponding portrait features. There are many ways to perform feature extraction, such as performing feature extraction through a machine learning model, using Word2Vec, GloVe, etc. The present disclosure does not particularly limit the specific implementation of feature extraction. The generated portrait features include payment features and non-payment features, wherein the payment features are generated based on data related to payment behavior in the portrait data, and the non-payment features are generated based on data unrelated to payment behavior in the portrait data.

[0079] In some possible implementations, the portrait data of the target user account can include static attribute data, historical active behavior data, and payment data. The payment data refers to data related to payment behavior, including but not limited to: historical purchase times, historical purchase amounts, historical purchase unit prices, payment status remaining days, and payment status expiration days. Feature extraction on the payment data can obtain payment features.

[0080] The static attribute data and the historical active behavior data belong to data irrelevant to the payment behavior. The static attribute data can include age, gender, city level, account level, price interval of the device used by the user, and other long-term unchanged data. The historical active behavior data refers to data generated by other behaviors of the user in the corresponding platform / application except the payment behavior in the past period of time, including but not limited to login days, playing times, and content and times of collection / comment / like. Feature extraction is performed on the static attribute data and the historical active behavior data respectively, and static attribute features and historical active behavior features can be obtained. The two parts of features constitute the non-payment features.

[0081] Further, the payment features and the non-payment features can be normalized, and the normalized payment features and the non-payment features can be spliced to obtain the portrait features of the target user account. In this way, the scales of the payment features and the non-payment features can be unified.

[0082] In the embodiment, by obtaining the static attribute data and the historical active data of the target user account, the comprehensiveness of the portrait data is improved, which is helpful to predict more accurate LTV prediction results according to more comprehensive portrait data. Moreover, after the payment features and the non-payment features are normalized and spliced to obtain the portrait features, and the portrait features are input into the LTV prediction model, it can be avoided that the LTV prediction model ignores the features with smaller scales in the prediction process due to the non-uniform scales of the payment features and the non-payment features, thereby further improving the accuracy of the LTV prediction results.

[0083] Step 202: inputting the portrait features and the life cycle data into the LTV prediction model to obtain the LTV prediction result corresponding to the target user account; wherein the LTV prediction model comprises: a gating network, configured to determine a first weight corresponding to the payment features and a second weight corresponding to the non-payment features according to the payment state represented by the life cycle data; a payment amount prediction network, configured to perform weighted processing on the payment features according to the first weight, and perform weighted processing on the non-payment features according to the second weight, and generate the LTV prediction result based on the payment features and the non-payment features obtained by the weighted processing.

[0084] Figure 3 is a structural schematic diagram of an LTV prediction model provided by an example embodiment. Please refer to Figure 3The LTV prediction model includes an embedding layer, a gate network, and a payment amount prediction network. The life cycle data and the portrait features are input into the LTV prediction model, and the embedding layer can perform vector conversion on the life cycle data to convert the life cycle data into life cycle features. For example, the embedding layer can include a randomly initialized matrix V, and the size of the matrix V is s*d, where s is the number of payment states. For example, if the payment states include new member, valid member, and member loss, s is equal to 3. The three payment states can be numbered as payment state 1, payment state 2, and payment state 3. The life cycle data of the target user account can include the number of the represented payment state (denoted as stage_id). Thus, the number of the payment state is input into the embedding layer, and the content of the (stage_id-1) row in the matrix V is found as the life cycle feature output. For example, in the case of stage_id=2, the first row of the matrix V is output as the life cycle feature of the target user account.

[0085] The input of the gate network includes the portrait features and the life cycle features output by the embedding layer. The gate network can determine the first weight corresponding to the payment feature and the second weight corresponding to the non-payment feature in the portrait features according to the payment state represented by the life cycle features, and output the first weight and the second weight. For example, for a user with a payment state of never paid, the first weight will be much smaller than the second weight, thereby increasing the influence of the non-payment feature on the LTV prediction result. For a user with a payment state of currently paid, the gate network will appropriately increase the proportion of the first weight and the second weight, thereby increasing the influence of the payment feature on the LTV prediction result.

[0086] The input of the payment amount prediction network includes the portrait features, the first weight, and the second weight output by the gate network. The payment amount prediction network can perform weighted processing on the payment feature according to the first weight, and perform weighted processing on the non-payment feature according to the second weight, and then generate and output an LTV prediction result based on the payment feature and the non-payment feature obtained by the weighted processing. The LTV prediction result is the predicted future total payment amount for the target user account.

[0087] In the above embodiments, since the payment features of different users reflect the differences in payment behaviors between different users, by controlling the weights of the payment features and the non-payment features of the users through the gating network, and controlling the influence degree of the payment features and the non-payment features on the LTV prediction result based on the corresponding weights, the generated LTV prediction result can be more consistent with the payment characteristics of the users. This prediction method fully considers the influence of the differences in payment behaviors of different users on the LTV prediction result, and can significantly improve the accuracy of the LTV prediction result, so as to generate accurate LTV prediction results for users with different payment behaviors. By predicting accurate LTV prediction results for users with different payment behaviors, it is helpful to finely divide the users in the platform / system according to the LTV prediction results, and to perform targeted marketing strategy delivery, thereby improving the marketing delivery effect.

[0088] In some possible embodiments, the payment amount prediction network can include a prediction layer and at least one feature processing layer. Each feature processing layer corresponds to a different feature dimension, and each feature processing layer corresponds to one gating network. Figure 4 FIG. 2 is a structural schematic diagram of another LTV prediction model provided by an example embodiment. Figure 4 In FIG. 2, the payment amount prediction network includes a prediction layer (logit layer) and three feature processing layers (neural layer 1-3), each of which corresponds to one gating network.

[0089] After inputting the life cycle data and the portrait features of the target user account into the LTV prediction model, first, the life cycle data is converted into life cycle features through the vector layer (embedding layer), and at the same time, the portrait features are nonlinearly processed through the feature input layer. For example, the portrait features are nonlinearly processed by using the ReLu activation function, and the principle of this process can be referred to formula (1):

[0090]

[0091] In formula (1), represents the portrait features, represents the weight of the feature input layer, represents the bias of the feature input layer, represents the portrait features after nonlinear processing.

[0092] Then, the image features after the nonlinear processing are spliced with the life cycle features as the input of the gate network. The input of the three gate networks is consistent in structure but different in parameters, so that different gate networks can learn features in different feature dimensions. For example, the gate network can adopt a two-layer neural network, the first layer uses a relu activation function, and the second layer uses a sigmiod activation function, as shown in Figure 5 It should be noted that Figure 5 This is only an example, and the specific structure of the gate network is not particularly limited in the present disclosure. The working principle of any gate network can refer to formula (2):

[0093]

[0094] In formula (2), characterizes the image features after the nonlinear processing, characterizes the life cycle features, characterizes the splicing operation, characterizes the first weight and the second weight of the gate network output.

[0095] For any feature processing layer, the input includes the output of the previous feature processing layer and the output of the corresponding gate network. The input of the first feature processing layer is the image features after the nonlinear processing (including the paid features and the non-paid features) and the first weight and the second weight of the gate network output. In order to facilitate the splicing of the output of the feature processing layer and the output of the corresponding gate network, the feature dimension corresponding to any feature processing layer and the feature dimension corresponding to the corresponding gate network remain consistent, such as the feature dimension corresponding to the feature processing layer neural layer1 and the gate network gate network1 is 20 dimensions, the feature dimension corresponding to the feature processing layer neural layer2 and the gate network gate network2 is 50 dimensions, and the feature dimension corresponding to the feature processing layer neural layer3 and the gate network gate network3 is 100 dimensions. The principle of the above splicing process can refer to formula (3):

[0096]

[0097] In formula (3), t represents the tth feature processing layer, represents the output of the (t-1)th feature processing layer or the image features after the nonlinear processing, represents the first weight and the second weight of the tth gate network output, represents the spliced features corresponding to the tth feature processing layer.

[0098] Since each feature processing layer corresponds to different feature dimensions, any feature processing layer is used to perform feature analysis on the corresponding spliced feature and output the feature in the corresponding feature dimension, and the principle of the feature analysis process can be referred to formula (4):

[0099]

[0100] In formula (4), representing the weight of the t-th feature processing layer, representing the bias of the t-th feature processing layer, representing the spliced feature corresponding to the t-th feature processing layer, representing the output of the t-th feature processing layer.

[0101] The output of the last feature processing layer is input into the prediction layer to obtain the LTV prediction result output by the prediction layer, and the principle of the prediction process can be referred to formula (5):

[0102]

[0103] In formula (5), representing the weight of the prediction layer, representing the bias of the prediction layer, representing the input of the prediction layer, representing the LTV prediction result output by the prediction layer.

[0104] In this embodiment, by setting feature processing layers and gate networks of different dimensions, the depth and width of the LTV prediction model are increased, which can significantly improve the expression ability of the LTV prediction model, so that it can learn more complex features. And, the portrait features and life cycle data input into the LTV prediction model can be gradually extracted from low-level features to high-level features, which helps to capture more comprehensive and accurate feature representations, so as to generate more accurate LTV prediction results based on more comprehensive and accurate feature representations.

[0105] In some possible implementations, the LTV prediction model may further include at least one payment behavior prediction subnetwork, each corresponding to a preset time window. A preset time window can be understood as a preset duration. As the name suggests, the payment behavior prediction subnetwork is used to predict whether the target user account will make a payment (i.e., whether a payment behavior exists) within the corresponding preset time window. For example, the input to each payment behavior prediction subnetwork may include profile features and a first weight and a second weight output from the gating network. The payment behavior prediction subnetwork can generate a payment behavior prediction result for the target user account within the corresponding preset time window based on the profile features and the first and second weights. This prediction result includes the probability of the target user account making a payment within the corresponding preset time window. Furthermore, the payment behavior prediction results generated by each payment behavior prediction subnetwork are input to the payment amount prediction network, so that the payment amount prediction network generates an LTV prediction result based on these inputs and the weighted payment and non-payment features.

[0106] Figure 6a This is a schematic diagram illustrating the structure of an LTV prediction model that includes a paid behavior prediction subnetwork, as provided in an exemplary embodiment. Figure 6a In this system, each payment behavior prediction subnetwork is directly connected to the payment amount prediction network. During the prediction process, the input to each payment behavior prediction subnetwork consists of the user profile features and the first and second weights output by the gating network. The payment amount prediction network can generate the LTV prediction result for the target user account based on the payment behavior prediction results (i.e., payment probability) output by each payment behavior prediction subnetwork, the weighted payment features, and the non-payment features. Assuming the time window corresponding to the payment amount prediction network is 180 days, it can predict the total payment amount of the target user account within the next 180 days. The six payment behavior prediction subnetworks (Buy1, Buy7, Buy30, Buy60, Buy90, ​​Buy180) correspond to preset time windows of 1 day, 7 days, 30 days, 60 days, 90 days, and 180 days, respectively. In other words, these six payment behavior prediction subnetworks can predict the payment probability of the target user account within the next 1 day, 7 days, 30 days, 60 days, 90 days, and 180 days, respectively.

[0107] Figure 6b This is a schematic diagram of the structure of another LTV prediction model including a paid behavior prediction subnetwork, provided in an exemplary embodiment. Figure 6bIn the specific implementation, the six payment behavior prediction sub-networks (Buy1, Buy7, Buy30, Buy60, Buy90, Buy180) are connected in stages according to the length of the corresponding preset time window, and the payment behavior prediction sub-network with the largest preset time window (i.e., Buy180) is connected with the payment amount prediction network. In the prediction process, the input of the payment behavior prediction sub-network farthest from the payment amount prediction network (i.e., Buy1) is the portrait feature, the first weight and the second weight output by the gating network. Then, the payment behavior prediction sub-network Buy1 inputs the generated payment behavior prediction result into the payment behavior prediction sub-network Buy7. That is, the input of the payment behavior prediction sub-network Buy7 includes the portrait feature, the first weight and the second weight output by the gating network, and the payment behavior prediction result generated by the payment behavior prediction sub-network Buy1, and then the payment behavior prediction sub-network Buy7 inputs the payment behavior prediction result generated by itself into the payment behavior prediction sub-network Buy30. In this way, the input of the payment behavior prediction sub-networks other than Buy1 includes the portrait feature, the first weight and the second weight output by the gating network, and the payment behavior prediction result generated by the previous payment behavior prediction sub-network. In this case, the input of the payment amount prediction network includes the payment behavior prediction result output by the payment behavior prediction sub-network Buy180, the payment feature and the non-payment feature after the weighting processing. The architecture of the connection in stages can enable each payment behavior prediction sub-network to learn the feature representation of the portrait feature and the life cycle feature in the previous payment behavior prediction sub-network, thereby improving the accuracy of the payment probability output by each payment behavior prediction sub-network. At the same time, on the basis of ensuring that the payment amount prediction network learns the payment probability features in multiple preset time windows, the data amount that needs to be processed by the payment amount prediction network is further reduced, and the prediction efficiency is improved. It should be noted that, Figure 6a With Figure 6b All the time windows involved in the specific implementation are examples only, and the disclosure does not particularly limit the time windows.

[0108] In the specific implementation, at least one payment behavior prediction sub-network is set in the LTV prediction model, so as to split the target of predicting the LTV of the user into the target of predicting the total payment amount and the targets of predicting multiple payment probabilities. Compared with directly predicting the total payment amount, the specific implementation is helpful to further improve the accuracy of the predicted total payment amount by predicting multiple payment probabilities to assist the prediction of the total payment amount under the architecture of multi-target / multi-task joint learning.

[0109] Figure 6c FIG. 1 is a structural diagram of a payment behavior prediction sub-network according to an example embodiment. Figure 6cAs shown, taking the architecture of the step-by-step connection of each payment behavior prediction sub-network as an example, each payment behavior prediction sub-network can include a probability generation layer (sigmoid layer), a prediction layer (logit layer), and multiple feature processing layers (neural layer), each feature processing layer corresponds to different feature dimensions, and each feature processing layer corresponds to a gating network respectively. Taking the Buy1 payment behavior prediction sub-network as an example for illustrative description: in the payment behavior prediction sub-network Buy1, the feature dimension corresponding to each feature processing layer is consistent with the feature dimension corresponding to the corresponding gating network. The input of the first feature processing layer neural layer 11 includes the portrait feature and the first weight and the second weight output by the first gating network. The portrait feature and the output of the first gating network are spliced, and then the spliced feature is input into the first feature processing layer neural layer 11 to obtain the feature representation in the feature dimension corresponding to the first feature processing layer 11. In addition to the first feature processing layer neural layer 11, the input of other feature processing layers includes the feature representation output by the previous feature processing layer and the first weight and the second weight output by the corresponding gating network. The prediction layer logit layer can predict according to the feature representation output by the last feature processing layer. The principle of the above process can be referred to formula (6):

[0110] (6)

[0111] In formula (6), characterizes the output processed in the three-layer feature processing layer in the payment behavior prediction sub-network Buy1, characterizes the weight of the prediction layer in the payment behavior prediction sub-network Buy1, characterizes the bias of the prediction layer in the payment behavior prediction sub-network Buy1, characterizes the output of the prediction layer in the payment behavior prediction sub-network Buy1.

[0112] Then, the output of the prediction layer is input into the probability generation layer sigmoid layer to obtain the payment probability of the target user account in the future 1 day, and the principle can be referred to formula (7):

[0113] (7)

[0114] In formula (7), characterizes the output of the prediction layer in the payment behavior prediction sub-network Buy1, The payment probability output by the payment behavior prediction subnetwork Buy1 is represented. The structure of such a payment behavior prediction subnetwork can help the payment behavior subnetwork learn the feature representation of the portrait feature and the life cycle feature in different feature dimensions, thereby improving the diversity and accuracy of feature learning, and helping to improve the accuracy of the payment probability output by the payment behavior prediction subnetwork.

[0115] The input and output of each feature processing layer and the probability generation layer in other payment behavior prediction subnetworks (Buy7, Buy30, Buy60, Buy90, and Buy180) can refer to the payment behavior prediction subnetwork Buy1, which will not be described here. The input of the prediction layer in other payment behavior prediction subnetworks includes the feature representation output by the previous feature processing layer in the network and the output of the prediction layer of the previous payment behavior prediction subnetwork. That is, the input of the prediction layer in the payment behavior prediction subnetwork Buy7 includes the feature representation output by the previous feature processing layer in the network and the output of the prediction layer in the payment behavior prediction subnetwork Buy1. The input of the prediction layer in the payment behavior prediction subnetwork Buy30 includes the feature representation output by the previous feature processing layer in the network and the output of the prediction layer in the payment behavior prediction subnetwork Buy7, and so on.

[0116] In combination with Figure 6c , the working principle of the prediction layer of the payment behavior prediction subnetwork Buy7 can refer to formula (8-1), and the working principle of the probability generation layer of the payment behavior prediction subnetwork Buy7 can refer to formula (8-2):

[0117]

[0118] (8-2)

[0119] In formula (8-1) and (8-2), represents the output of the prediction layer in the payment behavior prediction subnetwork Buy1, represents the output after being processed by the three-layer feature processing layer in the payment behavior prediction subnetwork Buy7, represents a concatenation operation, represents the weight of the prediction layer in the payment behavior prediction subnetwork Buy7, represents the bias of the prediction layer in the payment behavior prediction subnetwork Buy7, represents the output of the prediction layer in the payment behavior prediction subnetwork Buy7, represents the payment probability output by the payment behavior prediction subnetwork Buy7.

[0120] The working principle of the prediction layer of the payment behavior prediction subnetwork Buy30 can refer to formula (9-1), and the working principle of the probability generation layer of the payment behavior prediction subnetwork Buy30 can refer to formula (9-2):

[0121]

[0122] (9-2)

[0123] In formula (9-1), (9-2), denotes the output of the prediction layer in the paid behavior prediction sub-network Buy7, denotes the output after being processed by the three-layer feature processing layer in the paid behavior prediction sub-network Buy30, denotes the concatenation operation, denotes the weight of the prediction layer in the paid behavior prediction sub-network Buy30, denotes the bias of the prediction layer in the paid behavior prediction sub-network Buy30, denotes the output of the prediction layer in the paid behavior prediction sub-network Buy30, denotes the paid probability output by the paid behavior prediction sub-network Buy30.

[0124] For the paid behavior prediction sub-network Buy60, the working principle of the prediction layer thereof can refer to formula (10-1), and the working principle of the probability generation layer thereof can refer to formula (10-2):

[0125]

[0126] (10-2)

[0127] In formula (10-1), (10-2), denotes the output of the prediction layer in the paid behavior prediction sub-network Buy30, denotes the output after being processed by the three-layer feature processing layer in the paid behavior prediction sub-network Buy60, denotes the concatenation operation, denotes the weight of the prediction layer in the paid behavior prediction sub-network Buy60, denotes the bias of the prediction layer in the paid behavior prediction sub-network Buy60, denotes the output of the prediction layer in the paid behavior prediction sub-network Buy60, denotes the paid probability output by the paid behavior prediction sub-network Buy60.

[0128] For the paid behavior prediction sub-network Buy90, the working principle of the prediction layer thereof can refer to formula (11-1), and the working principle of the probability generation layer thereof can refer to formula (11-2):

[0129]

[0130] (11-2)

[0131] In formula (11-1), (11-2), denotes the output of the prediction layer in the paid behavior prediction sub-network Buy60, denotes the output after the three-layer feature processing layer in the paid behavior prediction sub-network Buy90, denotes the concatenation operation, denotes the weight of the prediction layer in the paid behavior prediction sub-network Buy90, denotes the bias of the prediction layer in the paid behavior prediction sub-network Buy90, denotes the output of the prediction layer in the paid behavior prediction sub-network Buy90, denotes the paid probability output by the paid behavior prediction sub-network Buy90.

[0132] For the paid behavior prediction sub-network Buy180, the working principle of the prediction layer thereof can refer to formula (12-1), and the working principle of the probability generation layer thereof can refer to formula (12-2):

[0133]

[0134] (12-2)

[0135] In formula (12-1), (12-2), denotes the output of the prediction layer in the paid behavior prediction sub-network Buy90, denotes the output after the three-layer feature processing layer in the paid behavior prediction sub-network Buy180, denotes the concatenation operation, denotes the weight of the prediction layer in the paid behavior prediction sub-network Buy180, denotes the bias of the prediction layer in the paid behavior prediction sub-network Buy180, denotes the output of the prediction layer in the paid behavior prediction sub-network Buy180, denotes the paid probability output by the paid behavior prediction sub-network Buy180.

[0136] In Figure 6c , the working principle of the paid amount prediction network can refer to formula (13):

[0137] (13)

[0138] In formula (13), denotes the output of the prediction layer in the paid behavior prediction sub-network Buy180, denotes the weight of the prediction layer in the paid amount prediction network, Characterizes the bias of the prediction layer in the payment amount prediction network. The output of the feature processing layer in the payment amount prediction network is represented. The LTV prediction result output by the characterization prediction layer is the total amount paid by the target user account in the next 180 days.

[0139] In some possible implementations, the LTV prediction model may also include at least one payment amount prediction sub-network, each payment amount prediction sub-network corresponding to a preset time window, and all preset time windows are different from the time window corresponding to the payment amount prediction network. Figure 7a This is a schematic diagram of the structure of an LTV prediction model including a payment amount prediction subnetwork, provided as an exemplary embodiment. Figure 7a As shown, the payment amount prediction network is LTV, and the payment amount prediction sub-networks include LTV1, LTV7, LTV30, LTV60, and LTV90. According to... Figure 7a The illustrated model structure shows that each payment amount prediction subnetwork is independent, but their inputs are identical, including profile features and the first and second weights output by the gating network. The payment features are weighted using the first weight to obtain the weighted payment features; similarly, the non-paying features are weighted using the second weight to obtain the weighted non-paying features. Each payment amount prediction subnetwork can generate an LTV (total payment amount) prediction result for the target user account within a corresponding preset time window based on its own input. Assuming the time window corresponding to LTV is 180 days, the preset time windows for these payment amount prediction subnetworks (LTV1, LTV7, LTV30, LTV60, LTV90) are 1 day, 7 days, 30 days, 60 days, and 90 days, respectively. Therefore, LTV can predict the total payment amount for the target user account within the next 180 days, while these five payment amount prediction subnetworks can predict the total payment amount for the target user account within the next 1 day, 7 days, 30 days, 60 days, and 90 days, respectively. In this way, a single LTV prediction model can be used to predict the total payment amount for the same target user account within different time windows, significantly improving the prediction range and expanding the versatility of the LTV prediction model. Furthermore, users can control the amount of total payment amount output by the LTV prediction model according to their own needs, such as controlling the LTV prediction model to only output the total payment amount for the next 30 days and 60 days, thus helping to meet users' personalized needs for LTV prediction.

[0140] Figure 7b This is a schematic diagram of the structure of another LTV prediction model including a payment amount prediction subnetwork, provided in an exemplary embodiment. For example... Figure 7bAs shown, the five payment amount prediction sub-networks (LTV1, LTV7, LTV30, LTV60, LTV90) can be connected in stages according to the length of the corresponding preset time window, and the payment behavior prediction sub-network with the largest preset time window (i.e. LTV90) is connected with the payment amount prediction network (i.e. LTV). In the prediction process, the input of the payment amount prediction sub-network farthest from the payment amount prediction network (i.e. LTV1) is the portrait feature, the first weight and the second weight output by the gating network. Then, the payment amount prediction sub-network LTV1 inputs the generated LTV prediction result into the payment amount prediction sub-network LTV7. That is, the input of the payment amount prediction sub-network LTV7 includes the portrait feature, the first weight and the second weight output by the gating network, and the LTV prediction result generated by the payment amount prediction sub-network LTV1. Then, the payment amount prediction sub-network LTV7 inputs the LTV prediction result generated by itself into the payment amount prediction sub-network LTV30. In this way, the input of the other payment amount prediction sub-networks, except for the payment amount prediction sub-network LTV1, includes the portrait feature, the first weight and the second weight output by the gating network, and the LTV prediction result generated by the previous payment amount prediction sub-network. In this case, the input of the payment amount prediction network LTV includes the LTV prediction result output by the payment amount prediction sub-network LTV90, and the weighted payment features and non-payment features. This kind of hierarchical connection architecture can make each payment amount prediction sub-network learn the feature representation of the portrait feature and the life cycle feature in the previous payment amount prediction sub-network, thereby improving the accuracy of the total payment amount output by each payment amount prediction sub-network. At the same time, it can ensure that the payment amount prediction network learns the payment amount features in multiple preset time windows, further improving the accuracy of the LTV prediction result output by the payment amount prediction network.

[0141] In some possible embodiments, the LTV prediction model can include a plurality of payment behavior prediction sub-networks and payment amount prediction sub-networks corresponding to a plurality of preset time windows. Please refer to Figure 8a , the preset time window corresponding to the payment amount prediction network LTV is 180 days, and the preset time windows corresponding to the two payment amount prediction sub-networks (LTV60, LTV90) are 60 days and 90 days, respectively. The preset time windows corresponding to the three payment behavior prediction sub-networks (Buy 60, Buy 90, Buy 180) are 60 days, 90 days and 180 days, respectively. The payment behavior prediction sub-networks and the payment amount prediction sub-networks / payment amount prediction network corresponding to the same preset time window are connected with each other, and each payment behavior prediction sub-network corresponding to a preset time window is also connected with a payment amount prediction sub-network corresponding to another preset time window, except for the network Buy 60.

[0142] In this case, for the payment amount prediction sub-network LTV60, the input includes: portrait features, the first weight and the second weight of the gating network output, the payment behavior prediction result output by the payment behavior prediction sub-network (i.e. Buy60) corresponding to the same preset time window; and the output is the total payment amount of the target user account within the future 60 days. For any other payment amount prediction sub-network / payment amount prediction network, the input includes: portrait features, the first weight and the second weight of the gating network output, the LTV prediction result output by the previous payment amount prediction sub-network, and the payment behavior prediction result output by the payment behavior prediction sub-network corresponding to the same preset time window. The previous payment amount prediction sub-network refers to the payment amount prediction sub-network in the preset time window adjacent to the preset time window corresponding to the payment amount prediction sub-network itself. For example, for the payment amount prediction network LTV, the corresponding preset time window is 180 days, and the previous payment amount prediction sub-network thereof is the payment amount prediction sub-network LTV90 in the preset time window of 90 days. The payment behavior prediction sub-network corresponding to the same preset time window refers to the payment behavior prediction sub-network Buy180 corresponding to the preset time window of 180 days.

[0143] For the payment behavior prediction sub-network Buy60, the input includes portrait features, the first weight and the second weight of the gating network output, and the output is the payment probability of the target user account within the future 60 days. For the payment behavior prediction sub-networks Buy90 and Buy180, the input includes: portrait features, the first weight and the second weight of the gating network output, and the payment behavior prediction result output by the previous payment behavior prediction sub-network; and the output is the payment probability of the target user account within the corresponding preset time window. Buy60 is the previous payment behavior prediction sub-network of Buy90, and Buy90 is the previous payment behavior prediction sub-network of Buy180. It should be noted that, in order to facilitate the display of the model structure, Figure 8a The total payment amount output by each payment amount prediction sub-network and the payment probability output by each payment behavior prediction sub-network are not shown in the model structure.

[0144] Based on Figure 8a the model structure shown, the payment probability and the total payment amount within multiple preset time windows can be output by using one LTV prediction model. Through this multi-task / target learning architecture, the prediction difficulty of each task can be further reduced, and the prediction accuracy and efficiency of each task can be improved.

[0145] On the basis of Figure 8a the portrait features and the life cycle features can be learned from multiple feature dimensions to further improve the accuracy of the prediction result. Please refer to Figure 8b , Figure 8bThe structure of the two networks is shown in the case of the payment amount prediction sub-network LTV90 and the payment behavior prediction sub-network Buy90. Figure 8b The input and output of each feature processing layer neural, and the input and output of the probability generation layer in the payment behavior prediction sub-network Buy90 can refer to the embodiment shown in Figure 6c The input of the prediction layer (logit layer) in the payment behavior prediction sub-network Buy90 includes the output of the last feature processing layer contained in the network itself, and the payment behavior prediction result output by the previous payment behavior prediction sub-network (i.e. Buy60). The input of the prediction layer (logit layer) in the payment amount prediction sub-network LTV90 includes the output of the last feature processing layer contained in the network itself, the payment behavior prediction result output by the payment behavior prediction sub-network (i.e. Buy90) under the same preset time window, and the LTV prediction result output by the previous payment amount prediction sub-network (i.e. LTV60).

[0146] Corresponding to the above data prediction method, in another embodiment of the present disclosure, a training method of a data prediction model is also provided, by which the LTV prediction model is trained to implement the above data prediction method based on the trained LTV prediction model.

[0147] Please refer to Figure 9 , Figure 9 is a flowchart of a training method of a data prediction model provided by an exemplary embodiment. The method can include the following steps:

[0148] Step 901: obtaining sample portrait data, sample life cycle data and sample LTV of a sample user account, and generating sample portrait features corresponding to the sample portrait data; wherein the sample portrait features include sample payment features and sample non-payment features, and the sample life cycle data is used to represent the sample payment state of the sample user account.

[0149] The sample user account can include a user account in a paid state, a user account that has never paid, and a user account whose payment state has expired. In this embodiment, a sample user account at date t can be selected, and then the sample portrait data, sample life cycle data and sample LTV of the sample user account in a preset time period in the corresponding platform / system are obtained. The sample life cycle data is used to represent the sample payment state of the sample user account at date t. The sample LTV in the preset time period refers to the sample LTV (i.e. sample total payment amount) of the sample user account after a preset time period from date t, such as the sample LTV in [t, t+180 days].

[0150] Different payment states can be numbered, such as a member new customer numbered 1, a valid member numbered 2, and a member loss numbered 3. The number corresponding to the sample user account is taken as the sample life cycle data, denoted as stage_id. Feature extraction can be performed on the sample portrait data of each sample user account to generate the corresponding sample portrait feature, denoted as vector X. In this way, the sample composition corresponding to each sample user account is {t, user_id, stage_id, X, LTV}, where user_id is the sample user account.

[0151] Step 902: input the sample portrait feature and the sample life cycle data into the to-be-trained LTV prediction model to obtain a sample LTV prediction result corresponding to the sample user account. The to-be-trained LTV prediction model includes: a gating network configured to determine a first sample weight corresponding to the sample payment feature and a second sample weight corresponding to the sample non-payment feature according to a sample payment state represented by the sample life cycle data; and a payment amount prediction network configured to perform weighted processing on the sample payment feature according to the first sample weight, perform weighted processing on the sample non-payment feature according to the second sample weight, and generate the sample LTV prediction result based on the sample payment feature and the sample non-payment feature obtained by the weighted processing.

[0152] The specific implementation of this step can refer to the related content in the embodiments shown in Figure 2 The specific implementation of this step can refer to the related content in the embodiments shown in

[0153] Step 903: calculate an LTV loss value based on the sample LTV and the sample LTV prediction result, and iteratively train the to-be-trained LTV prediction model according to the LTV loss value.

[0154] In this embodiment, in the case that the sample LTV is the total sample payment amount of the sample user account after 180 days from date t, the sample LTV prediction result is used to represent the predicted total payment amount of the sample user account after 180 days from date t. The LTV loss value can be calculated using the MSE loss function, and the principle can refer to formula (14):

[0155]

[0156] In formula (14), i represents the i-th sample user account, N represents the total number of sample user accounts, represents the sample LTV corresponding to the i-th sample user account, represents the sample LTV prediction result corresponding to the i-th sample user account, represents the LTV loss value.

[0157] After the LTV prediction model is trained through multiple iterations, the training of the LTV prediction model is completed when the obtained LTV prediction model meets a predefined training target (e.g., minimization of LTV loss value) or reaches a predefined number of iterations, and can be used to implement the data prediction scheme as shown in Figure 2 It should be noted that, in this embodiment, when the gradient is back propagated, the gradient weight from the gating network will not be back propagated to the sample portrait feature X in order not to affect the weight training of the sample portrait feature.

[0158] In some possible implementation manners, the LTV prediction model to be trained can further include at least one payment behavior prediction sub-network, and each payment behavior prediction sub-network corresponds to a sample preset time window. In this way, the LTV prediction model to be trained can learn the payment probability in multiple different time windows. Therefore, in this case, the sample portrait data of the sample user account can further include sample payment behaviors in multiple sample preset time windows. Still taking date t as an example, the sample portrait data can further include payment behaviors of the sample user account in [t, t+k], and k can be 1, 7, 30, 60, 90, or 180. If the sample user account has payment in the corresponding sample preset time window, the sample payment behavior in the sample preset time window is recorded as 1; if the sample user account has no payment in the corresponding sample preset time window, the sample payment behavior in the sample preset time window is recorded as 0. The sample composition corresponding to each sample user account can be {t, user_id, stage_id, X, LTV, Buy1, Buy7, Buy30, Buy60, Buy90, Buy180}, where Buy1-Buy180 are respectively used to represent the payment behaviors of the sample user account in multiple sample preset time windows.

[0159] In combination with Figures 6a-6c In the training process, any payment behavior prediction sub-network is configured to generate a sample payment behavior prediction result (i.e., a sample payment probability) of the sample user account in the corresponding sample preset time window according to the sample portrait feature and the first sample weight and the second sample weight output by the gating network. The payment amount prediction network can be configured to generate a sample LTV prediction result according to the sample payment feature and the sample non-payment feature obtained through the weighting processing and the sample payment behavior prediction result generated by each payment behavior prediction sub-network. For details, refer to the related content in the embodiment shown in Figure 2

[0160] ​In this embodiment, by setting at least one paid behavior prediction sub-network in the LTV prediction model to be trained, the training target of predicting the LTV of a user is split into the training target of predicting the total amount of payment and the training target of predicting the plurality of payment probabilities, which greatly simplifies the training difficulty of the LTV prediction model and improves the model training efficiency. Moreover, in the related art, there is only one training target of predicting the LTV of a user, and the model training effect may be poor due to the sample LTV of a part of sample user accounts being too high. In this embodiment, even if the sample LTV of a part of sample user accounts is too high, this part of sample data will only affect the payment amount prediction network in the LTV prediction model and will not affect the learning of the paid behavior prediction sub-network, thereby further improving the prediction accuracy of the LTV prediction model after training.

[0161] In some possible implementations, in the case where the LTV prediction model to be trained includes at least one paid behavior prediction sub-network, the paid behavior loss value of each paid behavior prediction sub-network also needs to be calculated. For example, the sample payment behaviors in each sample preset time window are obtained, and the paid behavior loss value corresponding to each sample preset time window is calculated according to the sample payment behavior corresponding to each sample preset time window and the sample payment behavior prediction result. For example, the ordinary cross-entropy loss function can be used to calculate the paid behavior loss value, and the principle can be referred to formula (15):

[0162]

[0163] In formula (15), t represents the tth sample preset time window, represents the sample payment behavior corresponding to the tth sample preset time window, represents the sample payment behavior prediction result corresponding to the tth sample preset time window, represents the ordinary cross-entropy function, represents the sum of the paid behavior loss values corresponding to each sample preset time window.

[0164] Further, the total loss value of the model is calculated according to the LTV loss value and the paid behavior loss value corresponding to each sample preset time window, and the LTV prediction model to be trained is iteratively trained based on the total loss value. The principle of the calculation process can be referred to formula (16):

[0165] (16)

[0166] In formula (16), represents the LTV loss value, the meanings of and are the same as formula (15), represents the preset loss weight, characterize the total loss value.

[0167] In this way, the payment amount prediction network and each payment behavior prediction sub-network can be optimized simultaneously, the feature representation that the LTV prediction model can learn is enriched, the generalization ability of the model is enhanced, the stability of the model training is improved, and the risk of gradient vanishing or explosion is reduced.

[0168] In some possible implementation manners, the LTV prediction model to be trained can further include at least one payment amount prediction sub-network, each payment amount prediction sub-network corresponds to a sample preset time window, and all sample preset time windows are different from the sample time window corresponding to the payment amount prediction network. In this way, the LTV prediction model to be trained can learn the payment amount in multiple different time windows. Therefore, in this case, the sample portrait data of the sample user account can further include sample LTVs (i.e., sample payment amounts) in multiple sample preset time windows. Still taking date t as an example, the sample portrait data can further include the payment amount of the sample user account in [t, t+k], and k can be 1, 7, 30, 60, 90, or 180. The sample composition corresponding to each sample user account can be {t, user_id, stage_id, X, LTV, LTV1, LTV7, LTV 30, LTV 60, LTV 90}, where LTV1-LTV90 are respectively used to characterize the sample LTV of the sample user account in multiple sample preset time windows.

[0169] In combination Figures 7a-7b During the training process, any payment amount prediction sub-network is configured to generate a sample LTV prediction result (i.e., a sample predicted payment amount) of the sample user account in the corresponding sample preset time window according to the sample portrait feature and the first sample weight and the second sample weight output by the gating network. For details, refer to the related content in the embodiment shown in Figure 2

[0170] In this embodiment, by setting at least one payment amount prediction sub-network in the LTV prediction model to be trained, the training target of predicting the user LTV is split into training targets of predicting the payment amount in multiple different time windows, thereby greatly simplifying the training difficulty of the LTV prediction model and improving the model training efficiency.

[0171] ​In some possible implementation manners, in the case that the LTV prediction model to be trained comprises at least one payment amount prediction subnetwork, the LTV loss value of each payment amount prediction subnetwork also needs to be calculated. For example, the sample LTVs in each sample preset time window are obtained respectively, and the LTV loss value corresponding to each sample preset time window is calculated according to the sample LTV corresponding to each sample preset time window and the sample LTV prediction result. The calculation principle can refer to formula (14).

[0172] Further, the total loss value of the LTV prediction model to be trained is calculated according to the LTV loss value of the payment amount prediction network and the LTV loss value of each payment amount prediction subnetwork, so as to iteratively train the LTV prediction model to be trained based on the total loss value. In this way, the payment amount prediction network and each payment amount prediction subnetwork can be optimized at the same time, the feature representation that can be learned by the LTV prediction model is enriched, the generalization ability of the model is enhanced, the stability of the model training is improved, and the risk of gradient vanishing or explosion is reduced.

[0173] Exemplary apparatus

[0174] In an exemplary embodiment of the present disclosure, a data prediction apparatus is also provided.

[0175] Please refer to Figure 10 , Figure 10 is a block diagram of a data prediction apparatus provided by an exemplary embodiment.

[0176] As Figure 10 shown, the data prediction apparatus 1000 can include an acquisition module 1001 and an input module 1002. Wherein:

[0177] The acquisition module 1001 is configured to acquire portrait data and life cycle data of a target user account, and generate portrait features corresponding to the portrait data; the portrait features include payment features and non-payment features, and the life cycle data is used to represent a payment state of the target user account.

[0178] The input module 1002 is configured to input the portrait features and the life cycle data into an LTV prediction model to obtain an LTV prediction result corresponding to the target user account; wherein the LTV prediction model comprises: a gating network configured to determine a first weight corresponding to the payment features and a second weight corresponding to the non-payment features according to the payment state represented by the life cycle data; and a payment amount prediction network configured to perform weighted processing on the payment features according to the first weight, perform weighted processing on the non-payment features according to the second weight, and generate the LTV prediction result based on the payment features and the non-payment features obtained by the weighted processing.

[0179] In an embodiment, the portrait data comprises static attribute data, historical active behavior data and payment data of the target user account, and the obtaining module 1001 is specifically configured to: perform feature extraction on the payment data to obtain payment features; perform feature extraction on the static attribute data and the historical active behavior data respectively to obtain non-payment features, the non-payment features comprising static attribute features corresponding to the static attribute data and historical active behavior features corresponding to the historical active behavior data; perform normalization processing on the payment features and the non-payment features respectively, and splice the processed payment features and the processed non-payment features to obtain the portrait features.

[0180] In an embodiment, the LTV prediction model further comprises at least one payment behavior prediction sub-network, each payment behavior prediction sub-network corresponding to a preset time window; wherein any payment behavior prediction sub-network is configured to generate a payment behavior prediction result of the target user account in a corresponding preset time window according to the portrait features and the first weight and the second weight output by the gating network; and the payment amount prediction network is further configured to generate the LTV prediction result according to the payment features and the non-payment features obtained through weighting processing and the payment behavior prediction results generated by each payment behavior prediction sub-network.

[0181] In an embodiment, the LTV prediction model further comprises at least one payment amount prediction sub-network, each payment amount prediction sub-network corresponding to a preset time window and all preset time windows being different from the time window corresponding to the payment amount prediction network; wherein any payment amount prediction sub-network is configured to generate an LTV prediction result of the target user account in a corresponding preset time window according to the payment features and the non-payment features obtained through weighting processing.

[0182] In an embodiment, the payment amount prediction network comprises a prediction layer and at least one feature processing layer, each feature processing layer corresponding to a different feature dimension and each feature processing layer corresponding to a gating network; wherein any feature processing layer is configured to perform feature analysis based on the output of a previous feature processing layer and the first weight and the second weight output by a corresponding gating network, and output features of a corresponding feature dimension; the input of the first feature processing layer is the payment features, the non-payment features and the first weight and the second weight output by the corresponding gating network; and the prediction layer is configured to perform prediction based on the output of the last feature processing layer to generate the LTV prediction result.

[0183] In an exemplary embodiment of the present disclosure, a training device of a data prediction model is also provided.

[0184] See Figure 11 , Figure 11A block diagram of a training device of a data prediction model provided by an example embodiment.

[0185] As shown in Figure 11 The training device 1100 of the data prediction model can include a sample acquisition module 1101, an input module 1102, and a training module 1103. Among them:

[0186] The sample acquisition module 1101 is configured to acquire sample portrait data, sample life cycle data, and sample LTV of a sample user account, and generate sample portrait features corresponding to the sample portrait data; wherein the sample portrait features include sample payment features and sample non-payment features, and the sample life cycle data is used to represent the sample payment state of the sample user account.

[0187] The input module 1102 is configured to input the sample portrait features and the sample life cycle data into a to-be-trained LTV prediction model to obtain a sample LTV prediction result corresponding to the sample user account; wherein the to-be-trained LTV prediction model includes: a gating network configured to determine a first sample weight corresponding to the sample payment features and a second sample weight corresponding to the sample non-payment features according to the sample payment state represented by the sample life cycle data; a payment amount prediction network configured to perform weighted processing on the sample payment features according to the first sample weight, and perform weighted processing on the sample non-payment features according to the second sample weight, and generate the sample LTV prediction result based on the sample payment features and the sample non-payment features obtained by the weighted processing.

[0188] The training module 1103 is configured to calculate an LTV loss value based on the sample LTV and the sample LTV prediction result, and iteratively train the to-be-trained LTV prediction model according to the LTV loss value.

[0189] In an embodiment, the to-be-trained LTV prediction model further includes at least one payment behavior prediction sub-network, and each payment behavior prediction sub-network corresponds to a sample preset time window; wherein any payment behavior prediction sub-network is configured to generate a sample payment behavior prediction result of the sample user account in a corresponding sample preset time window according to the sample portrait features and the first sample weight and the second sample weight output by the gating network; and the payment amount prediction network is further configured to generate the sample LTV prediction result according to the sample payment features and the sample non-payment features obtained by the weighted processing, and the sample payment behavior prediction result generated by each payment behavior prediction sub-network.

[0190] In an embodiment, the device 1100 further includes:

[0191] The payment behavior loss value calculation module 1104 is configured to obtain sample payment behaviors in each sample preset time window respectively, and calculate a payment behavior loss value corresponding to each sample preset time window according to the sample payment behavior corresponding to each sample preset time window and the sample payment behavior prediction result.

[0192] The training module 1103 is specifically configured to iteratively train the LTV prediction model to be trained according to the LTV loss value and the payment behavior loss value corresponding to each sample preset time window.

[0193] In an embodiment, the LTV prediction model to be trained further includes at least one payment amount prediction sub-network, each payment amount prediction sub-network corresponding to one sample preset time window, and all sample preset time windows being different from the sample time window corresponding to the payment amount prediction network; wherein any payment amount prediction sub-network is configured to generate a sample LTV prediction result of the sample user account in the corresponding sample preset time window according to the sample payment features and the sample non-payment features obtained through the weighting processing.

[0194] In an embodiment, the device 1100 further includes:

[0195] The LTV loss value calculation module 1105 is configured to obtain sample LTVs in each sample preset time window respectively, and calculate an LTV loss value corresponding to each sample preset time window according to the sample LTV corresponding to each sample preset time window and the sample LTV prediction result.

[0196] The training module 1103 is specifically configured to iteratively train the LTV prediction model to be trained according to the LTV loss value and the LTV loss value corresponding to each sample preset time window.

[0197] The specific details of each module of the data prediction device 1000 and the training device 1100 of the data prediction model have been described in detail in the description of the data prediction method flow, and thus will not be described here.

[0198] It should be noted that, although several modules or units of the data prediction device 1000 and the training device 1100 of the data prediction model are mentioned in the foregoing detailed description, such a division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. Conversely, the features and functions of one module or unit described above can be further divided into a plurality of modules or units.

[0199] Exemplary medium

[0200] In exemplary embodiments of the present disclosure, a computer-readable storage medium having stored thereon a program product capable of implementing the above-described method of the present specification is also provided. In some possible embodiments, various aspects of the present disclosure can also be implemented in the form of a program product including program code for causing an end device to perform the steps described in the above "Exemplary Method" section according to various exemplary embodiments of the present disclosure when the program product is run on the end device.

[0201] Referring to Figure 12 , Figure 12 is a schematic diagram of a readable storage medium provided by an exemplary embodiment of a data prediction method or a training method of a data prediction model.

[0202] Referring to Figure 12 , a readable storage medium 1200 for implementing the above-described method according to embodiments of the present disclosure is described, which can adopt a portable compact disc read-only memory (CD-ROM) and include program code, and can be run on an end device such as a personal computer. However, the readable storage medium of the present disclosure is not limited thereto, and in the present disclosure, the readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, device or apparatus.

[0203] The readable storage medium can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can be, for example but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or apparatus, or any combination thereof. More specific examples (non-exhaustive list) of the readable storage medium include an electrical connection having one or more wires, a portable disc, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0204] The computer-readable signal medium can include a data signal propagated in a baseband or as a part of a carrier wave, in which readable program code is borne. Such a propagated data signal can adopt various forms, including but not limited to an electromagnetic signal, an optical signal or any suitable combination thereof. The readable signal medium can also be any readable medium other than the readable storage medium, which can send, propagate or transmit a program for use by or in conjunction with an instruction execution system, device or apparatus.

[0205] The program code embodied on the computer readable medium can be transmitted using any appropriate medium, including but not limited to wireless, wired, optical fiber cable, RF, and the like, or any suitable combination of the foregoing.

[0206] The program code for carrying out operations of the present disclosure can be written in any combination of one or more programming languages, including an object oriented programming language such as Java, C++, and the like, and conventional procedural programming languages, such as the "C" programming language or similar programming languages. The program code can execute entirely on the user's computing device, partly on the user's device, as a stand-alone software package, partly on the user's device and partly on a remote computing device or entirely on the remote computing device or server. In the latter scenario, the remote computing device can be connected to the user's device through any type of network, including a local area network (LAN) or a wide area network (WAN), or the connection can be made to an external computing device, such as through the Internet using an Internet Service Provider.

[0207] Exemplary electronic device

[0208] In an exemplary embodiment of the present disclosure, an electronic device capable of implementing the training method of the data prediction method or the data prediction model is also provided.

[0209] Referring to Figure 13 , Figure 13 is a schematic diagram of an electronic device capable of implementing the method according to an exemplary embodiment.

[0210] The electronic device 1300 according to this embodiment of the present disclosure will be described below with reference to Figure 13 Figure 13 The electronic device 1300 shown is merely an example, and should not impose any limitation on the function and scope of use of the embodiments of the present disclosure.

[0211] As shown in Figure 13 , the electronic device 1300 is in the form of a general computing device. The components of the electronic device 1300 can include, but are not limited to, the at least one processing unit 1301 described above, the at least one storage unit 1302 described above, and a bus 1303 connecting different system components, including the storage unit 1302 and the processing unit 1301.

[0212] The storage unit stores program code that can be executed by the processing unit 1301, so that the processing unit 1301 performs the steps of the various embodiments described in the specification.

[0213] ​Storage 1302 can include a readable medium that can be a volatile memory, such as random access memory (RAM) 13021 and / or cache memory 13022, and can further include a non-volatile memory, such as read-only memory (ROM) 13023.

[0214] Storage 1302 can also include a program / utility 13024 having a set of program modules 13025 such as an operating system, one or more application programs, other program modules, and program data, each of which can operate in conjunction with the processor 13001 and / or other components of the electronic device 1300. These examples are intended to be illustrative. For example, the electronic device 1300 can be a network appliance. A person of ordinary skill in the art can recognize other examples.

[0215] Bus 1303 can represent one or more of several types of bus structures, including a storage bus or bus controller, peripheral bus, graphics bus, processor or local bus using any of a variety of bus architectures.

[0216] Electronic device 1300 can also communicate with one or more external devices 1304 such as a keyboard or pointing device, using one or more communication interfaces 1305. Communication interfaces 1305 can include a network interface 1306 that can be used to provide access to one or more networks. Access can be provided by a wireless interface, a wired interface, or a combination of both. Examples of networks include a local area network (LAN), a wide area network (WAN), and / or the Internet. Examples of communication interfaces 1305 include a modem, a network interface card (e.g. Ethernet), a personal computer card Figure 13 Other hardware and / or software modules that can be used in conjunction with the electronic device 1300 can also include, but are not limited to, application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), and / or other customized or non-standard hardware, software, or firmware.

[0217] Exemplary computer program product

[0218] In an example embodiment of the present disclosure, a computer program product is also provided, comprising computer program / instructions which, when executed by a processor, implement the steps of the method according to any of the above embodiments.

[0219] Those skilled in the art can easily understand that the example embodiments described herein can be implemented by software and also can be implemented by software in combination with necessary hardware, through the above description of the embodiments. Therefore, the technical solutions according to the embodiments of the disclosure can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (which can be a CD-ROM, a U disk, a mobile hard disk, or the like) or a network, and includes a number of instructions to make a computing device (which can be a personal computer, a server, a terminal device, or a network device, etc.) execute the method according to the embodiments of the disclosure.

[0220] Other embodiments of the disclosure will be apparent to those skilled in the art from consideration of the specification and practice of the disclosure disclosed herein. This application is intended to cover any variations, uses or adaptations of the disclosure including its general principles and specific embodiments disclosed herein. The specification and examples given herein are illustrative only and not restrictive of the true scope and spirit of the disclosure. The true scope and spirit of the disclosure are indicated by the appended claims.

[0221] It should be noted that although several units / modules or sub-units / modules of the apparatus are mentioned in the above detailed description, such division is merely exemplary and not mandatory. Indeed, according to an embodiment of the disclosure, features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, features and functions of one unit / module described above can be further divided into units / modules.

[0222] Furthermore, although the operations of the method of the disclosure are described in a particular order in the drawings, this does not require or imply that the operations must be performed in that particular order, or that all of the illustrated operations must be performed to achieve desirable results. Additionally or alternatively, certain steps can be omitted, combined, performed simultaneously, or performed in a different order.

[0223] While the principles and spirit of the disclosure have been described with reference to several specific embodiments, it is to be understood that the disclosure is not limited to the specific embodiments disclosed and that the division into aspects does not imply that features from the aspects cannot be combined to benefit, but is merely for ease of presentation. The disclosure is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

Claims

1. A data prediction method characterized by, The method comprises: obtaining portrait data and life cycle data of a target user account, and generating portrait features corresponding to the portrait data; the portrait features include a paid feature and a non-paid feature, and the life cycle data is used to represent a payment state of the target user account; inputting the portrait features and the life cycle data into an LTV prediction model to obtain an LTV prediction result corresponding to the target user account; wherein the LTV prediction model comprises: a gating network configured to determine a first weight corresponding to the paid feature and a second weight corresponding to the non-paid feature according to the payment state represented by the life cycle data; a payment amount prediction network configured to perform weighted processing on the paid feature according to the first weight, perform weighted processing on the non-paid feature according to the second weight, and generate the LTV prediction result based on the paid feature and the non-paid feature obtained through the weighted processing.

2. The method of claim 1, wherein, The portrait data includes static attribute data, historical active behavior data and payment data of the target user account, and the generation of the portrait features corresponding to the portrait data comprises: performing feature extraction on the payment data to obtain the paid feature; performing feature extraction on the static attribute data and the historical active behavior data respectively to obtain the non-paid feature, which includes static attribute features corresponding to the static attribute data and historical active behavior features corresponding to the historical active behavior data; performing normalization processing on the paid feature and the non-paid feature respectively, and splicing the processed paid feature and the processed non-paid feature to obtain the portrait features.

3. The method of claim 1, wherein, The LTV prediction model further comprises at least one payment behavior prediction sub-network, and each payment behavior prediction sub-network corresponds to a preset time window; wherein any payment behavior prediction sub-network is configured to generate a payment behavior prediction result of the target user account in a corresponding preset time window according to the portrait features, the first weight and the second weight output by the gating network; The payment amount prediction network is further configured to generate the LTV prediction result according to the paid feature and the non-paid feature obtained through the weighted processing and the payment behavior prediction results generated by each payment behavior prediction sub-network.

4. The method of claim 1, wherein, The LTV prediction model further comprises at least one payment amount prediction sub-network, and each payment amount prediction sub-network corresponds to a preset time window, and all preset time windows are different from the time window corresponding to the payment amount prediction network; wherein any payment amount prediction sub-network is configured to generate an LTV prediction result of the target user account in a corresponding preset time window according to the paid feature and the non-paid feature obtained through the weighted processing.

5. The method of claim 1, wherein, The payment amount prediction network comprises a prediction layer and at least one feature processing layer, each feature processing layer corresponds to different feature dimensions, and each feature processing layer corresponds to a gating network respectively; Any feature processing layer is configured to perform feature analysis on the output of the previous feature processing layer and the first weight and the second weight of the corresponding gating network output, and output the features of the corresponding feature dimension; the input of the first feature processing layer is the paid feature, the non-paid feature, and the first weight and the second weight of the corresponding gating network output; The prediction layer is configured to perform prediction based on the output of the last feature processing layer to generate the LTV prediction result. 6.A method for training a data prediction model, the method comprising: The method comprises: obtaining sample portrait data, sample life cycle data and sample LTV of a sample user account, and generating sample portrait features corresponding to the sample portrait data; wherein the sample portrait features include sample paid features and sample non-paid features, and the sample life cycle data is used to represent the sample payment state of the sample user account; inputting the sample portrait features and the sample life cycle data into a to-be-trained LTV prediction model to obtain a sample LTV prediction result corresponding to the sample user account; calculating an LTV loss value based on the sample LTV and the sample LTV prediction result, and iteratively training the to-be-trained LTV prediction model according to the LTV loss value; The to-be-trained LTV prediction model comprises: a gating network configured to determine a first sample weight corresponding to the sample paid feature and a second sample weight corresponding to the sample non-paid feature according to the sample payment state represented by the sample life cycle data; a payment amount prediction network configured to perform weighted processing on the sample paid feature according to the first sample weight, perform weighted processing on the sample non-paid feature according to the second sample weight, and generate the sample LTV prediction result based on the weighted processed sample paid feature and sample non-paid feature.

7. The method of claim 6, wherein, The to-be-trained LTV prediction model further comprises at least one payment behavior prediction sub-network, and each payment behavior prediction sub-network corresponds to a sample preset time window; wherein any payment behavior prediction sub-network is configured to generate a sample payment behavior prediction result of the sample user account in a corresponding sample preset time window based on the sample portrait features and the first sample weight and the second sample weight output by the gating network; The payment amount prediction network is further configured to generate the sample LTV prediction result based on the weighted processed sample paid feature and sample non-paid feature, and the sample payment behavior prediction results generated by each payment behavior prediction sub-network.

8. The method of claim 7, further comprising: obtaining sample payment behaviors in each sample preset time window, and calculating a payment behavior loss value corresponding to each sample preset time window based on the sample payment behavior in the sample preset time window and the sample payment behavior prediction result corresponding to the sample preset time window; iteratively training the to-be-trained LTV prediction model according to the LTV loss value and the payment behavior loss value corresponding to each sample preset time window. ​ 9. The method of claim 6, wherein, The to-be-trained LTV prediction model further comprises at least one payment amount prediction sub-network, each payment amount prediction sub-network corresponds to a sample preset time window, and all sample preset time windows are different from the sample time window corresponding to the payment amount prediction network; wherein, Any payment amount prediction sub-network is configured to generate a sample LTV prediction result of the sample user account in the corresponding sample preset time window according to the sample payment features and sample non-payment features obtained through the weighted processing.

10. The method of claim 9, further comprising: respectively obtaining sample LTVs in each sample preset time window, and respectively calculating LTV loss values corresponding to each sample preset time window according to the sample LTV corresponding to each sample preset time window and the sample LTV prediction result; iteratively training the to-be-trained LTV prediction model according to the LTV loss value and the LTV loss value corresponding to each sample preset time window. The device comprises:

11. A data prediction apparatus, characterized by comprising: an acquisition module configured to acquire portrait data and life cycle data of a target user account, and generate portrait features corresponding to the portrait data; the portrait features comprise payment features and non-payment features, and the life cycle data is used to represent a payment state of the target user account; an input module configured to input the portrait features and the life cycle data into an LTV prediction model to obtain an LTV prediction result corresponding to the target user account; wherein the LTV prediction model comprises: a gating network configured to determine a first weight corresponding to the payment features and a second weight corresponding to the non-payment features according to the payment state represented by the life cycle data; a payment amount prediction network configured to perform weighted processing on the payment features according to the first weight, perform weighted processing on the non-payment features according to the second weight, and generate the LTV prediction result based on the payment features and the non-payment features obtained through the weighted processing. The portrait data comprises static attribute data, historical active behavior data and payment data of the target user account, 12. The apparatus of claim 11, wherein, The acquisition module is further configured to: extract features from the payment data to obtain the payment features; extract features from the static attribute data and the historical active behavior data respectively to obtain the non-payment features, the non-payment features comprising static attribute features corresponding to the static attribute data and historical active behavior features corresponding to the historical active behavior data; The payment features and the non-payment features are normalized respectively, and the processed payment features and the processed non-payment features are spliced to obtain the portrait features. The LTV prediction model further comprises at least one payment behavior prediction sub-network, each payment behavior prediction sub-network corresponding to a preset time window; wherein, 13. The apparatus of claim 11, wherein, ​ Any payment behavior prediction sub-network is configured to generate a payment behavior prediction result of the target user account within a corresponding preset time window according to the portrait feature and first and second weights output by the gating network. The payment amount prediction network is further configured to generate the LTV prediction result according to the payment feature and the non-payment feature obtained through the weighting processing and the payment behavior prediction result generated by each payment behavior prediction sub-network.

14. The apparatus of claim 11, wherein, The LTV prediction model further comprises at least one payment amount prediction sub-network, each payment amount prediction sub-network corresponding to a preset time window, and all preset time windows being different from the time window corresponding to the payment amount prediction network; wherein any payment amount prediction sub-network is configured to generate an LTV prediction result of the target user account within a corresponding preset time window according to the payment feature and the non-payment feature obtained through the weighting processing.

15. The apparatus of claim 11, wherein, The payment amount prediction network comprises a prediction layer and at least one feature processing layer, each feature processing layer corresponding to a different feature dimension and each feature processing layer corresponding to a gating network; Any feature processing layer is configured to perform feature analysis based on the output of the previous feature processing layer and first and second weights output by the corresponding gating network, and output features of the corresponding feature dimension; the input of the first feature processing layer is the payment feature, the non-payment feature and the first and second weights output by the corresponding gating network. The prediction layer is configured to perform prediction based on the output of the last feature processing layer to generate the LTV prediction result.

16. A training device for a data prediction model, characterized in that, The device comprises: A sample acquisition module configured to acquire sample portrait data, sample life cycle data and a sample LTV of a sample user account, and generate sample portrait features corresponding to the sample portrait data; wherein the sample portrait features comprise sample payment features and sample non-payment features, and the sample life cycle data is configured to represent a sample payment state of the sample user account; An input module configured to input the sample portrait features and the sample life cycle data into a to-be-trained LTV prediction model to obtain a sample LTV prediction result corresponding to the sample user account; A training module configured to calculate an LTV loss value based on the sample LTV and the sample LTV prediction result, and iteratively train the to-be-trained LTV prediction model according to the LTV loss value; The to-be-trained LTV prediction model comprises: A gating network configured to determine first sample weights corresponding to the sample payment features and second sample weights corresponding to the sample non-payment features according to a sample payment state represented by the sample life cycle data; A payment amount prediction network configured to perform weighting processing on the sample payment features according to the first sample weights, perform weighting processing on the sample non-payment features according to the second sample weights, and generate the sample LTV prediction result based on the sample payment features and the sample non-payment features obtained through the weighting processing.

17. The apparatus of claim 16, wherein, The to-be-trained LTV prediction model further comprises at least one payment behavior prediction sub-network, each payment behavior prediction sub-network corresponding to a sample preset time window; wherein, Any payment behavior prediction sub-network is configured to generate a sample payment behavior prediction result of the sample user account in a corresponding sample preset time window according to the sample portrait feature and the first sample weight and the second sample weight output by the gating network. The payment amount prediction network is further configured to generate the sample LTV prediction result according to the sample payment feature and the sample non-payment feature obtained through the weighting processing and the sample payment behavior prediction result generated by each payment behavior prediction sub-network.

18. The apparatus of claim 17, wherein, The device further includes: A payment behavior loss value calculation module configured to obtain the sample payment behavior in each sample preset time window and calculate a payment behavior loss value corresponding to each sample preset time window according to the sample payment behavior and the sample payment behavior prediction result corresponding to each sample preset time window. The training module is specifically configured to iteratively train the to-be-trained LTV prediction model according to the LTV loss value and the payment behavior loss value corresponding to each sample preset time window.

19. The apparatus of claim 16, wherein, The to-be-trained LTV prediction model further includes at least one payment amount prediction sub-network, each payment amount prediction sub-network corresponding to a sample preset time window, and all sample preset time windows being different from the sample time window corresponding to the payment amount prediction network; wherein Any payment amount prediction sub-network is configured to generate a sample LTV prediction result of the sample user account in a corresponding sample preset time window according to the sample payment feature and the sample non-payment feature obtained through the weighting processing.

20. The apparatus of claim 19, wherein, The device further includes: An LTV loss value calculation module configured to obtain the sample LTV in each sample preset time window and calculate an LTV loss value corresponding to each sample preset time window according to the sample LTV and the sample LTV prediction result corresponding to each sample preset time window. The training module is specifically configured to iteratively train the to-be-trained LTV prediction model according to the LTV loss value and the LTV loss value corresponding to each sample preset time window.

21. An electronic device, comprising: It includes: A processor; A memory for storing executable instructions of the processor; The processor implements the method of any one of claims 1-10 by running the executable instructions.

22. A computer readable storage medium having stored thereon computer instructions, wherein, The instructions are executed by the processor to implement the method of any one of claims 1-10.

23. A computer program product, characterised in that, It includes a computer program / instruction, which is executed by the processor to implement the steps of the method of any one of claims 1-10.

Images