An electronic resource consumption information prediction method, device, equipment and storage medium

Abstract

The application provides an electronic resource consumption information prediction method and device, equipment and a storage medium, which can be applied to the field of cloud computing or artificial intelligence, and is used for solving the problem of low prediction accuracy of electronic resource consumption information. The method comprises the following steps: obtaining a target feature matrix of a target account, wherein the target feature matrix is obtained based on electronic resource historical consumption information of the target account; using a trained electronic resource prediction model to predict the use probability of the target account for each target service node based on the target feature matrix, wherein each target service node corresponds to one kind of electronic resource reference consumption information; and obtaining the electronic resource prediction consumption information of the target account based on the obtained use probability and the electronic resource reference consumption information corresponding to each target service node.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, apparatus, device and storage medium for predicting electronic resource consumption information. Background Technology

[0002] With the continuous development of technology, different online information can be displayed to different types of accounts on the same traffic platform. Account types can be categorized based on the predicted total amount of electronic resources consumed by each account on the platform. Accounts with higher electronic resource consumption are classified as high-consumption accounts, and those with lower consumption are classified as low-consumption accounts. Media platforms can then use this traffic platform to deliver high-consumption-type online information to high-consumption-type accounts and low-consumption-type online information to low-consumption-type accounts.

[0003] However, accounts consume electronic resources differently. If the total electronic resource consumption of an account on a traffic platform is directly predicted, low-consumption accounts may frequently perform low-consumption operations, resulting in a higher total electronic resource consumption for low-consumption accounts. This could lead to a situation where the predicted total electronic resource consumption of low-consumption accounts is higher, thus incorrectly identifying low-consumption accounts as high-consumption accounts. Consequently, media platforms may be unable to accurately deliver online information to various accounts. Summary of the Invention

[0004] This application provides a method, apparatus, computer equipment, and storage medium for predicting electronic resource consumption information, which addresses the problem of low accuracy in predicting electronic resource consumption information.

[0005] Firstly, a method for predicting electronic resource consumption information is provided, including:

[0006] Obtain the target feature matrix of the target account, wherein the target feature matrix is ​​obtained based on the historical electronic resource consumption information of the target account;

[0007] Using a trained electronic resource prediction model, based on the target feature matrix, the probability of the target account's use for each target service node is predicted, wherein each target service node corresponds to a type of electronic resource reference consumption information;

[0008] Based on the obtained usage probabilities and the electronic resource reference consumption information corresponding to each target service node, the predicted electronic resource consumption information of the target account is obtained.

[0009] Secondly, an electronic resource consumption information prediction device is provided, comprising:

[0010] Acquisition module: used to obtain the target feature matrix of the target account, wherein the target feature matrix is ​​obtained based on the historical electronic resource consumption information of the target account;

[0011] Processing module: Used to employ a trained e-resource prediction model, based on the target feature matrix, to predict the usage probability of the target account for each target service node, wherein each target service node corresponds to an e-resource reference consumption information; and, based on the obtained usage probabilities and the e-resource reference consumption information corresponding to each target service node, to obtain the e-resource prediction consumption information of the target account.

[0012] Optionally, the acquisition module is further configured to: before using the trained electronic resource prediction model to predict the usage probability of the target account for each target service node based on the target feature matrix, obtain an initial sample set, wherein the initial sample set includes the sample feature matrix corresponding to each sample account and the electronic resource sample consumption information corresponding to each sample account.

[0013] The processing module is also used to: determine the target node combination corresponding to each sample account based on the obtained electronic resource sample consumption information and the electronic resource reference consumption information corresponding to each candidate service node.

[0014] The processing module is further configured to: train the electronic resource prediction model to be trained using a training sample set composed of each sample feature matrix and each target node combination, until the training loss of the electronic resource prediction model to be trained meets the preset convergence condition, thereby obtaining the trained electronic resource prediction model.

[0015] Optionally, the processing module is specifically used for:

[0016] The candidate service nodes are randomly combined to obtain various candidate node combinations, wherein each candidate node combination includes at least one candidate service node.

[0017] Based on the electronic resource reference consumption information corresponding to each candidate service node, the combination result of the electronic resource reference consumption information corresponding to each candidate node is determined respectively.

[0018] For each of the sample accounts, the following operations are performed: From each of the candidate node combinations, a candidate node combination whose combination result matches the electronic resource sample consumption information corresponding to one of the sample accounts is selected as the target node combination corresponding to that sample account.

[0019] Optionally, the processing module is specifically used for:

[0020] From the various candidate node combinations, at least one candidate node combination is selected, wherein the error between the combination result of each candidate node combination in the at least one candidate node combination and the electronic resource sample consumption information corresponding to the sample account is within a preset error range.

[0021] If the at least one candidate node combination includes only one candidate node combination, then the one candidate node combination is determined as the target node combination corresponding to the sample account;

[0022] If the at least one candidate node combination includes multiple candidate node combinations, then the number of candidate service nodes included in each of the multiple candidate node combinations is determined, and the candidate node combinations whose number of nodes is within a preset range are selected as the target node combination corresponding to the sample account.

[0023] Optionally, the processing module is further configured to:

[0024] After determining the target node combination corresponding to each sample account, based on the obtained target node combination, the total number of times each candidate service node appears in each target node combination is counted to obtain the statistical result of each candidate service node.

[0025] From the candidate service nodes, the candidate service nodes whose statistical results meet the preset statistical conditions are selected as the target service nodes.

[0026] Optionally, the acquisition module is further configured to: obtain the historical usage records of each sample account for each candidate service node before using the trained electronic resource prediction model to predict the usage probability of the target account for each target service node based on the target feature matrix;

[0027] The processing module is also used to: based on the obtained historical usage records, count the number of times each sample account uses electronic resources at each candidate service node;

[0028] The processing module is also used to: filter out candidate service nodes whose usage frequency falls within a preset range, and use them as target service nodes.

[0029] Optionally, the processing module is specifically used for:

[0030] The probability of the target account being used for each target service node is determined as the weight coefficient corresponding to each target service node.

[0031] Based on the obtained weight coefficients, the electronic resource reference consumption information corresponding to each target service node is weighted and fused to obtain the electronic resource prediction consumption information of the target account.

[0032] Thirdly, a computer device is provided, comprising:

[0033] Memory, used to store program instructions;

[0034] A processor is configured to invoke program instructions stored in the memory and execute the method described in the first aspect according to the obtained program instructions.

[0035] Fourthly, a storage medium is provided that stores computer-executable instructions for causing a computer to perform the method as described in the first aspect.

[0036] In this embodiment, a trained e-resource prediction model is used to predict the usage probability of a target account for each target service node based on the target account's target feature matrix. This allows for the determination of whether the target account consumes e-resources at each target service node. Since the e-resource consumption information of the target account varies under different circumstances, the correlation between whether the target account consumes e-resources at each target service node and the characteristics of the target account's e-resource consumption is stronger than the target account's consumption information. This makes the predicted usage probability of the target account for each target service node based on the target account's target feature matrix more accurate. Therefore, based on the obtained usage probabilities and the corresponding e-resource reference consumption information for each target service node, the predicted e-resource consumption information of the target account is more accurate than the predicted e-resource consumption information directly predicted based on the target account's target feature matrix, thus solving the problem of low prediction accuracy of e-resource consumption information. Attached Figure Description

[0037] Figure 1 This is one application scenario of the electronic resource consumption information prediction method provided in the embodiments of this application;

[0038] Figure 2a A schematic diagram of the principle of the electronic resource consumption information prediction method provided in the embodiments of this application. Figure 1 ;

[0039] Figure 2b A flowchart illustrating the electronic resource consumption information prediction method provided in this application embodiment. Figure 1 ;

[0040] Figure 3aA flowchart illustrating the electronic resource consumption information prediction method provided in this application embodiment is shown in Figure 2.

[0041] Figure 3b A schematic diagram of the principle of the electronic resource consumption information prediction method provided in the embodiments of this application (II);

[0042] Figure 3c A schematic diagram three illustrating the principle of the electronic resource consumption information prediction method provided in this application embodiment;

[0043] Figure 3d A schematic diagram four illustrating the principle of the electronic resource consumption information prediction method provided in this application embodiment;

[0044] Figure 3e A schematic diagram of the principle of the electronic resource consumption information prediction method provided in the embodiments of this application. Figure 5 ;

[0045] Figure 4a A schematic diagram of the principle of the electronic resource consumption information prediction method provided in the embodiments of this application. Figure 6 ;

[0046] Figure 4b A flowchart illustrating the electronic resource consumption information prediction method provided in this application is shown in Figure 3.

[0047] Figure 4c A schematic diagram seven illustrating the principle of the electronic resource consumption information prediction method provided in this application embodiment;

[0048] Figure 5 Schematic diagram of the electronic resource consumption information prediction device provided in the embodiments of this application Figure 1 ;

[0049] Figure 6 Schematic diagram 2 of the electronic resource consumption information prediction device provided in the embodiments of this application. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings.

[0051] The following explanations of some terms used in the embodiments of this application are provided to facilitate understanding by those skilled in the art.

[0052] (1) Online multimedia information:

[0053] In the internet age, media platforms influence the actions of various accounts by distributing media information to them. For example, online advertising allows advertisers to quickly reach consumers and influence their behavior, leading to short-term conversions such as registrations and purchases. Online advertising no longer uses offline contracted advertising space but leverages the traffic and influence of media outlets for promotion.

[0054] (2) Directional multimedia information:

[0055] Targeted multimedia messaging refers to media platforms delivering multimedia information to specific types of accounts. For example, in the advertising industry, pay-per-click advertising allows advertisers to precisely target specific user groups through search, thereby purchasing targeted traffic from that audience. Pay-per-click advertising does not guarantee ad delivery; advertisers need to adjust the balance between performance and traffic themselves. Pay-per-click advertising also includes real-time bidding, which typically uses a pay-per-impression model. In this model, media platforms have greater flexibility in selecting specific account types. For instance, advertisers can more flexibly segment and select their target audience, leading to the rapid development of broader data usage and transactions.

[0056] (3) Gini coefficient:

[0057] The Gini coefficient is the ratio of the area enclosed by the line of equality and the Lorenz curve to the area below the line of equality.

[0058] The Gini coefficient can be used to assess the accuracy of the ordering relationship in numerical predictions, with a greater focus on the prediction accuracy of larger values. Here, the vertical axis represents the cumulative percentage of positive examples out of the total number of positive examples, and the horizontal axis represents the cumulative number of samples. The larger the number of positive examples, the closer the Lorenz curve will be to the lower right corner, resulting in a larger area under the curve (A) and a higher Gini coefficient. The more accurate the ordering relationship of the predicted values, the closer it will be to the curve generated by the true values, thus achieving a higher Gini coefficient.

[0059] (4) User Lifetime Value:

[0060] User lifetime value is typically used to measure the revenue a user can generate over their lifetime (from registration to churn). Due to the limitation of the observation period, the calculation of LTV usually sets a number of days, such as: LTV1 and LTV7 represent the user's spending within one day / one week from registration, respectively.

[0061] (5) Zero-inflation normal distribution and zero-inflation log-normal distribution:

[0062] Zero-inflated normal distribution: In real-world data, there are too many zero values, causing the normal distribution to lose its predictive power in the case of large numbers and failing to adequately characterize the data distribution. The first proposed zero-inflated model is the zero-inflated Poisson model, which corresponds to two components: the first process is controlled by a binary distribution that generates structural zeros, and the second process is controlled by a Poisson distribution.

[0063] Zero-inflated log-normal distribution: The log-normal distribution is often used to fit asymmetrical data, such as age, which is naturally a value greater than 0. The log-normal distribution, which describes the normal distribution of the logarithm of a random variable, is very good at characterizing this distribution. The structure of the zero-inflated log-normal distribution is similar to that of the zero-inflated normal distribution, except that the second process is controlled by the log-normal distribution.

[0064] This application's embodiments relate to cloud technology and artificial intelligence (AI). They are designed based on cloud computing and cloud storage within cloud technology, and on machine learning (ML) within artificial intelligence.

[0065] Cloud technology refers to a hosting technology that unifies hardware, software, and network resources within a wide area network (WAN) or local area network (LAN) to achieve data computation, storage, processing, and sharing. Cloud technology is a collective term for network technology, information technology, integration technology, management platform technology, and application technology applied to the cloud computing business model. It can form resource pools, providing flexible and convenient on-demand access. Cloud computing technology will become a crucial support. Backend services of technical network systems require substantial computing and storage resources, such as video websites, image websites, and many portal websites. With the rapid development and application of the internet industry, every item may have its own identification mark in the future, requiring transmission to backend systems for logical processing. Data at different levels will be processed separately, and various industry data will require robust system support, which can only be achieved through cloud computing.

[0066] Cloud computing is a computing model that distributes computing tasks across a resource pool composed of a large number of computers, enabling various application systems to access computing power, storage space, and information services as needed. The network providing these resources is called the "cloud." From the user's perspective, the resources in the "cloud" are infinitely scalable, readily available, on-demand, expandable, and pay-as-you-go.

[0067] As a provider of fundamental cloud computing capabilities, a cloud resource pool (referred to as a cloud platform, generally called an Infrastructure as a Service (IaaS) platform) is established. Various types of virtual resources are deployed in the resource pool for external customers to choose from. The cloud resource pool mainly includes: computing devices (virtualized machines containing operating systems), storage devices, and network devices.

[0068] Based on logical function, a Platform as a Service (PaaS) layer can be deployed on top of the IaaS layer, followed by a Software as a Service (SaaS) layer. Alternatively, SaaS can be deployed directly on top of IaaS. PaaS is the platform for running software, such as databases and web containers. SaaS refers to various types of business software, such as web portals and bulk SMS senders. Generally, SaaS and PaaS are upper layers compared to IaaS.

[0069] Cloud storage is a new concept that has been extended and developed from the concept of cloud computing. A distributed cloud storage system (hereinafter referred to as a storage system) refers to a storage system that uses cluster applications, grid technology and distributed storage file systems to bring together a large number of storage devices of various types in the network (storage devices are also called storage nodes) to work together through application software or application interfaces to provide data storage and business access functions to the outside world.

[0070] Currently, the storage method of storage systems is as follows: Logical volumes are created. During the creation of a logical volume, physical storage space is allocated to each logical volume. This physical storage space may consist of a single storage device or the disks of several storage devices. Clients store data on a logical volume, which means storing the data on the file system. The file system divides the data into many parts, each part being an object. Each object contains not only the data but also additional information such as a data identifier (ID entity, ID). The file system writes each object to the physical storage space of that logical volume and records the storage location information of each object. Therefore, when a client requests access to data, the file system can allow the client to access the data based on the storage location information of each object.

[0071] The process by which a storage system allocates physical storage space to a logical volume is as follows: the physical storage space is pre-divided into strips according to the capacity estimate of the objects stored in the logical volume (this estimate often has a large margin relative to the actual capacity of the objects to be stored) and the grouping of Redundant Array of Independent Disks (RAID). A logical volume can be understood as a strip, thus allocating physical storage space to the logical volume.

[0072] Artificial intelligence (AI) is the theory, methods, technology, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. AI studies the design principles and implementation methods of various intelligent machines, enabling them to possess perception, reasoning, and decision-making capabilities. AI technology mainly includes several major areas such as computer vision, natural language processing, machine learning, and deep learning.

[0073] With the research and advancement of artificial intelligence technology, AI is being studied and applied in various fields, such as smart homes, intelligent recommendation systems, virtual assistants, smart speakers, intelligent marketing, intelligent translation, autonomous driving, robotics, and smart healthcare. It is believed that with the development of technology, AI will be applied in more fields and play an increasingly important role.

[0074] Machine learning is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, and algorithm complexity theory, among others. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to endow computers with intelligence; its applications span all areas of artificial intelligence. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and instructional learning.

[0075] The application fields of the electronic resource consumption information prediction method provided in the embodiments of this application will be briefly introduced below.

[0076] With the continuous development of technology, the way information is delivered to various accounts is no longer limited to offline public places. Instead, online information can be delivered on online traffic platforms and displayed on the screens of each account. This not only increases the number of accounts that can access online information but also improves the timeliness of information delivery. At the same time, delivering information through a virtual network environment reduces the cost of information delivery.

[0077] Within the same traffic platform, different online information can be displayed to different types of accounts. Account types can be categorized based on the predicted total amount of digital resources consumed by each account on the platform; accounts with higher digital resource consumption are classified as high-consumption accounts, and those with lower consumption are classified as low-consumption accounts. Media platforms can then use this traffic platform to deliver high-consumption-type online information to high-consumption-type accounts and low-consumption-type online information to low-consumption-type accounts, and so on.

[0078] For example, in the advertising industry, advertisers can target ads to various accounts, influencing their behavior such as app usage or electronic resource consumption through online advertising. To further target ads to different accounts, advertisers can use bidding-based advertising on the same traffic platform to deliver different ads to different accounts.

[0079] Taking the gaming industry as an example, advertisers can target high-spending accounts with ads for limited-edition virtual character skins, and low-spending accounts with ads for the latest discount promotions. This reduces the number of advertising messages received by accounts, and allows advertisers to achieve better advertising results with minimal cost.

[0080] Continuing with the gaming industry as an example, the typical method for implementing bidding-based advertising is to predict the total amount of electronic resources an account will consume in the game, and then categorize accounts based on this total consumption. Account categories could include, for example, high-consumption and low-consumption types. Advertisers can then target high-consumption accounts with ads for limited-edition virtual character skins, and low-consumption accounts with ads for the latest discount promotions, etc.

[0081] However, accounts consume electronic resources differently. If the total electronic resource consumption of an account on a traffic platform is directly predicted, low-consumption accounts may frequently perform low-consumption operations, resulting in a higher total electronic resource consumption for low-consumption accounts. This could lead to a situation where the predicted total electronic resource consumption of low-consumption accounts is higher, thus incorrectly identifying low-consumption accounts as high-consumption accounts. Consequently, media platforms may be unable to accurately deliver online information to various accounts.

[0082] Continuing with the gaming industry as an example, some accounts prefer to spend digital resources to obtain limited-edition virtual character skins, while others prefer to spend them on discounted items. In reality, the release dates of limited-edition virtual character skins and the duration of discount events are uncertain. This can lead to situations where, for example, during the end of a discount event, accounts that prefer spending digital resources on limited-edition virtual character skins are classified as low-spending accounts, while accounts that prefer spending them on discounted items are classified as high-spending accounts. Consequently, advertising targeting to these accounts can be inaccurate.

[0083] For example, existing user lifecycle prediction methods are mainly divided into two categories: regression-based and classification-based models.

[0084] Regression-based methods model the LTV prediction problem as an attempt to directly predict LTV values. Three representative methods are naive regression, two-stage method, and log regression method.

[0085] Naive regression employs the most basic least squares-based regression model, regressing user lifetime value based on user characteristics. However, according to the least squares principle, the response variable needs to follow a normal distribution. But in real-world gaming scenarios, the vast majority of users (80-90%) spend zero during their game lifetime, meaning they are non-paying players. Therefore, in reality, the distribution of user spending has a very high peak at zero, which significantly deviates from a normal distribution.

[0086] The two-stage approach, based on the assumption that user payments follow a zero-inflated normal distribution, attempts to address this problem by introducing a two-stage prediction method. The first stage uses a binary classification model to determine whether a user will pay. If a user pays, the second stage uses a naive regression model to estimate the amount paid. While the two-stage approach, by introducing a binary classification model in the first stage, models the distribution of non-paying users, thus better reflecting reality, the zero-inflated normal distribution assumption still has problems. Specifically, the normal distribution is symmetrical, while in reality, the higher the payment amount, the fewer the number of paying users, exhibiting a long-tailed distribution.

[0087] The Ziln method introduces the assumption of a zero-inflated log-normal distribution in an attempt to model the length distribution. This method employs a novel loss function calculation approach, estimating user lifetime value by predicting the mean and variance of the log-normal distribution.

[0088] Regression-based methods make assumptions about user value distribution that don't align with reality, leading to unsatisfactory results. Although regression-based methods have continuously revised their distribution assumptions—from normal to zero-inflated normal, and then to zero-inflated log-normal—in an attempt to better match reality, the actual user spending is not continuous but discretely distributed across various payment points. Therefore, regression methods based on continuous distribution assumptions are inherently unsuitable for discrete payment scenarios.

[0089] While multi-class classification methods seemingly transform continuous regression problems into discrete classification problems, they are essentially quantile regressions, thus still exhibiting inconsistencies with reality and resulting in unsatisfactory performance. Furthermore, severe class imbalance exists between different categories, with a large number of users belonging to the low-value category and a small number belonging to the high-value category. This causes the classification model to heavily favor predicting the low-value category, further impacting the model's performance. In addition, determining how to reasonably categorize users by value is also a challenging problem.

[0090] It is evident that prediction models trained solely on the total electronic resource consumption of individual sample accounts in the game are ill-equipped to accurately predict the total electronic resource consumption of a target account, thus failing to accurately target ads to that account. Similar issues exist in other fields as well.

[0091] To address the issue of low accuracy in predicting electronic resource consumption information, this application provides a method for predicting electronic resource consumption information. After obtaining the target feature matrix of a target account, this method employs a trained electronic resource prediction model to predict the usage probability of the target account for each target service node based on the target feature matrix. After obtaining each usage probability, and based on the obtained usage probabilities and the corresponding electronic resource reference consumption information for each target service node, the predicted electronic resource consumption information of the target account is obtained.

[0092] In this embodiment, a trained e-resource prediction model is used to predict the usage probability of a target account for each target service node based on the target account's target feature matrix. This allows for the determination of whether the target account consumes e-resources at each target service node. Since the e-resource consumption information of the target account varies under different circumstances, the correlation between whether the target account consumes e-resources at each target service node and the characteristics of the target account's e-resource consumption is stronger than the target account's consumption information. This makes the predicted usage probability of the target account for each target service node based on the target account's target feature matrix more accurate. Therefore, based on the obtained usage probabilities and the corresponding e-resource reference consumption information for each target service node, the predicted e-resource consumption information of the target account is more accurate than the predicted e-resource consumption information directly predicted based on the target account's target feature matrix, thus solving the problem of low prediction accuracy of e-resource consumption information.

[0093] The application scenarios of the electronic resource consumption information prediction method provided in this application are described below.

[0094] Please refer to Figure 1 This application scenario describes one application of the electronic resource consumption information prediction method provided in this application. The application scenario includes a client 101 and a server 102. The client 101 and the server 102 can communicate with each other. The communication method can be wired, such as through a network cable or serial cable; or wireless, such as through Bluetooth or Wireless Fidelity (WIFI). No specific limitation is imposed.

[0095] Client 101 generally refers to a device that can provide the target account to server 102, such as a terminal device, a third-party application accessible by the terminal device, or a webpage accessible by the terminal device. Terminal devices include, for example, mobile phones, tablets, or personal computers. Server 102 generally refers to a device that can obtain predicted electronic resource consumption information for the target account, such as a terminal device or a server. Servers include, for example, cloud servers or local servers. Both client 101 and server 102 can utilize cloud computing to reduce the consumption of local computing resources; similarly, they can utilize cloud storage to reduce the consumption of local storage resources.

[0096] As one embodiment, the client 101 and the server 102 can be the same device, and there is no specific limitation. In this embodiment, the client 101 and the server 102 are described as different devices.

[0097] Before introducing the electronic resource consumption information prediction method provided in the embodiments of this application, the method for training the electronic resource prediction model to be trained will be introduced first. Please refer to... Figure 2a This is a schematic diagram illustrating the principle of a method for training an electronic resource prediction model. The server 102 obtains combinations of target nodes based on an initial sample set, and then obtains various target service nodes and a training sample set based on these combinations. Finally, it trains the electronic resource prediction model based on the training sample set.

[0098] Please refer to Figure 2b This is a flowchart illustrating a method for training an electronic resource prediction model.

[0099] Server 102 can train different electronic resource prediction models for different electronic resource consumption scenarios. If each electronic resource consumption scenario can have a type label, server 102 can also train different electronic resource prediction models for electronic resource consumption scenarios with different type labels, etc., without any specific limitations. There are several methods to determine the type label of an electronic resource consumption scenario. The first method can be based on the scenario type, such as game type; the second method can be based on the account operation method set in the electronic resource consumption scenario, such as two-handed operation, one-handed operation, swipe operation, or click operation; the third method can be based on the function options set in the electronic resource consumption scenario, such as dress-up function, stamina increase function, or level completion item selection function. In this embodiment, the training of different electronic resource prediction models for different electronic resource consumption scenarios by server 102 is used as an example for description.

[0100] S201, Obtain the initial sample set.

[0101] The initial sample set includes the sample feature matrix corresponding to each sample account, as well as the electronic resource sample consumption information corresponding to each sample account. The server 102 can obtain the initial sample set in several ways. For example, the first method is for the server 102 to read a pre-stored initial sample set; the second method is for the server 102 to receive an initial sample set sent by other devices; and the third method is for the server 102 to download the initial sample set from network resources, etc. No specific method is limited.

[0102] The sample feature matrix of a sample account is used to characterize the electronic resource consumption characteristics of the sample account. The server 102 can determine the sample feature matrix of the sample account through the user profile of the sample account. The server 102 can also extract features from the stored historical electronic resource consumption information of the sample account to obtain the sample feature matrix of the sample account. For example, the historical electronic resource consumption information of the sample account may be the payment records of the sample account in each game in the past month; the user profile of the sample account may be the account information, login time or recharge records of the sample account.

[0103] The electronic resource consumption information of a sample account can be the electronic resource consumption information of the sample account for the current electronic resource consumption scenario; it can also be the electronic resource consumption information of the sample account for multiple electronic resource consumption scenarios, without any specific restrictions. Multiple electronic resource consumption scenarios can include the current electronic resource consumption scenario, electronic resource consumption scenarios related to the current electronic resource consumption scenario, or electronic resource consumption scenarios associated with the sample account, etc.

[0104] S202, based on the obtained electronic resource consumption information of each sample and the electronic resource reference consumption information of each candidate service node, determine the target node combination corresponding to each sample account.

[0105] Since the initial sample set includes the feature matrix of each sample account and the electronic resource consumption information of each sample account, the prediction model trained on the initial sample set can only directly predict the total electronic resource consumption of the target account based on the target feature matrix. Taking a game scenario as an example, the time period for pushing limited-edition virtual character skin ads may differ from the time period for pushing discounted product ads. Therefore, during the discounted product ad period, it's easy to predict target accounts that like to consume electronic resources for discounted products as high-consumption accounts, and target accounts that like to consume electronic resources for limited-edition virtual character skins as low-consumption accounts. However, in reality, since the unit price of limited-edition virtual character skins is much higher than discounted products, target accounts that like to consume electronic resources for limited-edition virtual character skins should actually be high-consumption accounts. Therefore, directly predicting the total electronic resource consumption of the target account has low predictive accuracy.

[0106] Therefore, after obtaining the initial sample set, server 102 can determine the target node combination corresponding to each sample account based on the obtained initial sample set. By combining the target node combination corresponding to each sample account, the electronic resource prediction model can be trained. This allows the trained electronic resource prediction model to predict whether a target account consumes electronic resources at each target service node. Thus, even during the period when advertising information for discounted products is pushed, target accounts that like to consume electronic information for discounted products will not be predicted as high-consumption accounts, thereby improving prediction accuracy.

[0107] Please refer to Figure 3a This is a flowchart illustrating the process of determining the target node combination corresponding to each sample account.

[0108] S301, randomly combine the set candidate service nodes to obtain the candidate node combinations.

[0109] The current electronic resource consumption scenario can include multiple designated candidate service nodes. The server 102 can obtain each candidate service node in the current electronic resource consumption scenario. Each candidate service node can correspond to a type of electronic resource reference consumption information. For example, in the current electronic resource consumption scenario, different electronic resource consumption nodes can serve as different candidate service nodes; for another example, nodes with the same electronic resource consumption can serve as one candidate service node; for yet another example, nodes with electronic resource consumption belonging to the same electronic resource consumption range can serve as one candidate service node, and so on.

[0110] For example, in a game, a 10-yuan payment point can be used as a candidate service node, and a 50-yuan payment point can be used as a candidate service node; or, for example, a payment point with an amount in the range [0, 10] can be used as a candidate service node, and a payment point with an amount in the range [10, 50] can be used as a candidate service node, and so on.

[0111] After obtaining the candidate service nodes, server 102 can randomly combine the candidate service nodes to obtain candidate node combinations. For example, each candidate service node can be combined as a candidate node combination; every two candidate service nodes can be combined as a candidate node combination, and so on, to obtain all possible candidate node combinations.

[0112] The server 102 can also pre-set the number of candidate service nodes included in the candidate node combination, and randomly combine each candidate service node according to the preset number of nodes to obtain all candidate node combinations that meet the node number requirement. The preset number of nodes can be user-defined, determined by the server 102 according to the total number of candidate service nodes, or a model parameter learned when training the electronic resource prediction model, etc. There are no specific restrictions.

[0113] The server 102 can also pre-set specified candidate service nodes, randomly combine each candidate service node with the specified candidate service node, and obtain all candidate node combinations that include the specified candidate service node. The pre-set specified candidate service nodes can be user-defined, randomly selected by the server 102 according to the total number of candidate service nodes, or learned during the training of the electronic resource prediction model, etc. There are no specific restrictions.

[0114] For example, if the candidate service nodes include a first candidate service node, a second candidate service node, and a third candidate service node, then a random combination of the first, second, and third candidate service nodes can be performed. Please refer to [reference needed]. Figure 3b Seven candidate node combinations can be obtained, represented in vector form as (1, 0, 0), (0, 1, 0), (0, 0, 1), (1, 1, 0), (1, 0, 1), (0, 1, 1) and (1, 1, 1).

[0115] S302, based on the electronic resource reference consumption information corresponding to each candidate service node, determine the combination result of the electronic resource reference consumption information corresponding to each candidate node combination.

[0116] After obtaining the various candidate node combinations, server 102 can determine the combination result of the electronic resource reference consumption information corresponding to each candidate service node combination based on the electronic resource reference consumption information corresponding to each candidate service node. When a candidate node combination includes only one candidate service node, the electronic resource reference consumption information of that single candidate service node can be used as the combination result of that candidate node combination. When a candidate node combination includes multiple candidate service nodes, the electronic resource reference consumption information corresponding to each of the multiple candidate service nodes can be fused to obtain the combination result of the candidate node combination.

[0117] The information fusion processing method varies depending on the content included in the electronic resource reference consumption information. For example, when the electronic resource reference consumption information includes the electronic resource reference consumption amount, the information fusion processing method is to sum the electronic resource reference consumption amounts corresponding to each of the multiple candidate service nodes and use the summed result as the combined result. When the electronic resource reference consumption information includes an electronic resource reference consumption range, the information fusion processing method is to randomly select a consumption value from the electronic resource reference consumption range corresponding to each candidate service node, sum the consumption values ​​selected from each candidate service node, obtain all summed results, and use the range including all summed results as the combined result, and so on.

[0118] Continuing with the previous example, for instance, the reference electronic resource consumption information for the first candidate service node is 10, the reference electronic resource consumption information for the second candidate service node is 50, and the reference electronic resource consumption information for the third candidate service node is 60. Please refer to... Figure 3c The combination result of the first candidate node combination (1, 0, 0) is 10, the combination result of the second candidate node combination (0, 1, 0) is 50, the combination result of the third candidate node combination (0, 0, 1) is 60, the combination result of the fourth candidate node combination (1, 1, 0) is 60, the combination result of the fifth candidate node combination (1, 0, 1) is 70, the combination result of the sixth candidate node combination (0, 1, 1) is 110, and the combination result of the seventh candidate node combination (1, 1, 1) is 120.

[0119] S303, determine the target node combination corresponding to each sample account.

[0120] After obtaining the combination results of the electronic resource reference consumption information corresponding to each candidate node combination, server 102 can determine the target node combination corresponding to each sample account. The following describes the method for determining the target node combination corresponding to one sample account, using one account from each sample account as an example. The process for determining the target node combination corresponding to other sample accounts is the same.

[0121] Server 102 can filter out candidate node combinations from various candidate node combinations that match the electronic resource consumption information corresponding to the sample account, and use these as the target node combinations for the sample account. Matching the electronic resource consumption information corresponding to the sample account can mean that the combined result and the electronic resource consumption information are identical; it can also mean that the error between the combined result and the electronic resource consumption information is within a preset error range, etc.

[0122] Server 102 can determine the error between the combination result corresponding to each candidate node combination and the electronic resource sample consumption information of the sample account. The method for determining the error differs depending on the content included in the electronic resource reference consumption information. For example, when the electronic resource reference consumption information includes the electronic resource reference consumption amount, the combination result corresponding to the candidate node combination is the sum of the electronic resource reference consumption amounts, and the electronic resource sample consumption information is the electronic resource sample consumption amount. The method for determining the error is to determine the difference between the sum of the electronic resource reference consumption amounts and the electronic resource sample consumption amount, and use this difference as the error between the combination result and the electronic resource sample consumption information of the sample account. When the electronic resource reference consumption information includes an electronic resource reference consumption range, the combination result corresponding to the candidate node combination is also a numerical range, and the electronic resource sample consumption information is an electronic resource sample consumption range. The method for determining the error is to determine the union between the numerical range corresponding to the combination result and the electronic resource sample consumption range, and the intersection between the numerical range corresponding to the combination result and the electronic resource sample consumption range, and use the union after removing the intersection as the error between the combination result and the electronic resource sample consumption information of the sample account, etc.

[0123] Server 102 can filter at least one candidate node combination whose error is within a preset error range based on the error between the combination result corresponding to each candidate node combination and the electronic resource sample consumption information of the sample account. For example, it can filter at least one candidate node combination whose error is less than a preset error threshold.

[0124] The preset error range can be user-defined, determined by the server 102 based on the average value of the electronic resource reference consumption information of the candidate service nodes, or it can be a model parameter learned when training the electronic resource prediction model, etc. There are no specific restrictions.

[0125] The preset error range varies depending on the content of the determined error. When the determined error is numerical, the preset error range can be a numerical interval corresponding to the difference between the order of magnitude and the total electronic resource reference consumption and the electronic resource sample consumption. When the determined error is the union after removing the intersection, the preset error range can be a numerical interval corresponding to the order of magnitude and the ratio between the intersection and the union after removing the intersection. For example, if the difference between the total electronic resource reference consumption and the electronic resource sample consumption is 10, then the preset error range can be a numerical interval of [0, 12]. Another example: if the intersection is [5, 12], the union is [0, 24], and the union after removing the intersection is [0, 5] and [12, 24], then the ratio between the intersection and the union after removing the intersection is 7 / 17, and the preset error range can be a numerical interval of [0, 1].

[0126] If only one candidate node combination is selected, then that candidate node combination is used as the target node combination corresponding to the sample account. If multiple candidate node combinations are selected, then the server 102 can determine the number of candidate service nodes included in each of the multiple candidate node combinations. The server 102 can use candidate node combinations with a number of nodes within a preset range as the target node combination corresponding to the sample account. For example, the candidate node combination with the fewest nodes can be used as the target node combination corresponding to the sample account; another example is using the candidate node combination with the most nodes; yet another example is using the candidate node combination with a relatively even number of nodes, etc. If multiple candidate node combinations include the same number of candidate service nodes, then the total number of occurrences of all candidate service nodes included in each candidate node combination can be determined based on the total number of occurrences of the candidate service node in all candidate node combinations, and the candidate service node with the largest sum of occurrences can be used as the target service node.

[0127] Continuing with the previous example, please refer to... Figure 3d For example, if the electronic resource consumption information of a sample account is 10, then among the seven candidate node combinations, the first candidate node combination with the smallest error between the combination result and the electronic resource consumption information of 10 is determined as the target node combination. If the electronic resource consumption information of a sample account is 60, then among the seven candidate node combinations, the third and fourth candidate node combinations with the smallest error between the combination result and the electronic resource consumption information of 60 are determined. The number of nodes included in the third and fourth candidate node combinations is determined respectively. The third candidate node combination includes one candidate service node, and the fourth candidate node combination includes two candidate service nodes. The third node combination with the fewest nodes is selected as the target node combination.

[0128] S203 uses a training sample set composed of feature matrices of each sample and combinations of target nodes to train the electronic resource prediction model until the training loss of the electronic resource prediction model meets a preset convergence condition, thus obtaining a trained electronic resource prediction model. The preset convergence condition can be user-defined or determined by the server 102 based on the amount of data in the training sample set, etc., with no specific restrictions. The preset convergence condition is used to stabilize the training loss of the electronic resource prediction model within a certain range, thereby making the output accuracy of the electronic resource prediction model more stable.

[0129] During the training of the electronic resource prediction model to be trained, the candidate service nodes included in each target node combination can be used as target service nodes to train the electronic resource prediction model to be trained; as an example, some candidate service nodes included in each target node combination can also be used as target service nodes to train the electronic resource prediction model to be trained.

[0130] One way to select a subset of candidate service nodes from the candidate service nodes included in each target node combination as target service nodes is to calculate the total number of times each candidate service node appears in each target node combination, obtaining the statistical result for each candidate service node. From all candidate service nodes, those whose statistical results meet preset statistical conditions are selected as target service nodes. The preset statistical conditions can be user-defined, or determined by the server 102 based on the number of users served by the candidate service node, etc., with no specific restrictions.

[0131] One way to select candidate service nodes whose statistical results meet preset statistical conditions is to select candidate service nodes whose total occurrences exceed a preset total occurrence threshold, and use these as target service nodes. The preset total occurrence threshold can be user-defined, or it can be determined by the server 102 based on the number of users served by the candidate service nodes, etc., without any specific restrictions.

[0132] Another way to select candidate service nodes whose statistical results meet the preset statistical conditions is to first sort the total number of occurrences in descending order, and then select the candidate service nodes that rank before the preset number as the target service nodes, etc., without any specific restrictions. The preset number can be user-defined, or it can be determined by the server 102 based on the number of users served by the candidate service nodes, etc., without any specific restrictions.

[0133] For example, please refer to Figure 3e The target node combination includes the first target node combination (1, 0, 0) and the second target node combination (1, 1, 0). The total number of occurrences of the first candidate service node is 2, the total number of occurrences of the second candidate service node is 1, and the total number of occurrences of the third candidate service node is 0. The candidate service nodes with a total occurrence count greater than 0 are taken as the target service nodes, that is, the first candidate service node and the second candidate service node are determined as the target service nodes.

[0134] As one embodiment, server 102 can train the electronic resource prediction model to be trained using a number of target service nodes, and score each trained electronic resource prediction model. The trained electronic resource prediction model with the highest score is selected as the final trained electronic resource prediction model used to predict electronic resource consumption information. The prediction score can use evaluation criteria such as the Gini coefficient to evaluate the prediction accuracy of the electronic resource prediction model, and there are no specific limitations.

[0135] As one embodiment, if the server 102 can obtain the historical usage records of each sample account for each candidate service node, it can count the number of times each sample account used electronic resources on each candidate service node based on the obtained historical usage records. Candidate service nodes whose counts fall within a preset range are then selected as target service nodes. While the historical usage records of each sample account for each candidate service node allow for more accurate determination of target service nodes, in some cases, due to the need to protect account information and prevent its leakage, a method for obtaining more accurate target service nodes with less data is required. Therefore, the method described above for determining target service nodes based on a combination of target nodes can improve the security of account information. In practical use, the method for determining target service nodes can be selected according to the actual scenario.

[0136] There are several methods for training the electronic resource prediction model to be trained. Two of them will be introduced below as examples.

[0137] Training Method 1:

[0138] After obtaining the sample feature matrix and the target node combination corresponding to each sample account, server 102 obtains a training sample set based on the obtained sample feature matrix and target node combination. Each training sample includes the sample feature matrix corresponding to a sample account and the target node combination corresponding to that sample account.

[0139] After obtaining the target node combinations corresponding to each sample account, server 102 can determine each target service node from among the candidate service nodes based on the obtained target node combinations. Server 102 can count the total number of occurrences of each candidate service node in each obtained target node combination to obtain the statistical result of each candidate service node. Server 102 can filter out candidate service nodes whose statistical results meet preset statistical conditions from each candidate service node and use them as target service nodes. The method for filtering out candidate service nodes whose statistical results meet the preset statistical conditions is, for example, to filter out candidate service nodes whose total occurrences are greater than a preset occurrence threshold; or, for example, to sort the statistical results of each candidate service node in descending order and use the candidate service nodes whose sorting order is before the preset sequence number as target service nodes, etc.

[0140] As one embodiment, if server 102 can obtain the historical usage records of each sample account for each candidate service node, then server 102 can determine the target service node based on the historical usage records. After obtaining the historical usage records of each sample account for each candidate service node, server 102 can count the number of times each sample account consumes electronic resources on each candidate service node based on the obtained historical usage records, and obtain the statistical results of each candidate service node. Server 102 can filter out candidate service nodes whose statistical usage count is within a preset range as target service nodes. Filtering out candidate service nodes whose statistical usage count is within the preset range can be, for example, filtering out candidate service nodes whose statistical usage count is greater than a preset usage count threshold; or, for example, first sorting the statistical usage counts of each candidate service node in descending order, and selecting the candidate service node whose sorting order is before a preset sequence number as the target service node, etc.

[0141] Server 102 can use each training sample in the training sample set to train the electronic resource prediction model to be trained. Based on the error between the usage probability of each target service node output by the electronic resource prediction model to be trained and the candidate service nodes included in the target node combination, the training loss of the electronic resource prediction model to be trained is determined. If the training loss of the electronic resource prediction model to be trained does not converge, the model parameters of the electronic resource prediction model to be trained are adjusted, and the electronic resource prediction model to be trained continues to be trained until the training loss of the electronic resource prediction model to be trained converges, thus obtaining the trained electronic resource prediction model.

[0142] Training Method Two:

[0143] After obtaining the sample feature matrix corresponding to each sample account and the target node combination corresponding to each sample account, server 102 obtains a training sample set based on the obtained sample feature matrix, target node combination, and electronic resource sample consumption information. Each training sample includes the sample feature matrix corresponding to a sample account, the target node combination corresponding to that sample account, and the electronic resource sample consumption information corresponding to that sample account.

[0144] Server 102 can use each training sample in the training sample set to train the e-resource prediction model to be trained. The following describes the process of training with a single training sample as an example. The sample feature matrix of the sample account is input into the e-resource prediction model to be trained, obtaining the usage probability for each target service node output by the model. The usage probability for each target service node is determined as the weight coefficient corresponding to each target service node. Based on the obtained weight coefficients, the e-resource reference consumption information corresponding to each target service node is weighted and fused to obtain the e-resource training consumption information of the sample account.

[0145] After obtaining the electronic resource training consumption information of the sample accounts, server 102 determines the error between the electronic resource training consumption information and the electronic resource sample consumption information of the sample accounts, and determines the training loss of the electronic resource prediction model to be trained. If the training loss of the electronic resource prediction model to be trained does not converge, the model parameters of the electronic resource prediction model are adjusted, and the electronic resource prediction model to be trained continues to be trained until the training loss of the electronic resource prediction model to be trained converges, thus obtaining the trained electronic resource prediction model.

[0146] The method for weighted fusion of the electronic resource reference consumption information corresponding to each target service node is similar to the process of information fusion of the electronic resource reference consumption information corresponding to multiple candidate service nodes in S302, and will not be repeated here. The method for determining the target service node is similar to the method for determining the target service node introduced in training method one, and will not be repeated here.

[0147] Please refer to Figure 4a This is a schematic diagram illustrating the principle of a method for predicting electronic resource consumption information. Server 102 obtains the target feature matrix of the target account, and using a trained electronic resource prediction model, based on the target feature matrix and the electronic resource reference consumption information corresponding to each target service node, obtains the predicted electronic resource consumption information of the target account.

[0148] Please refer to Figure 4b This is a flowchart illustrating a method for predicting electronic resource consumption information.

[0149] S401, obtain the target feature matrix of the target account.

[0150] There are multiple opportunities for the server 102 to obtain the target feature matrix of the target account. For example, when the target account enters the current electronic resource consumption scenario through the client 101, or after the target account successfully registers through the client 101, or after the target account consumes electronic resources through the client 101. There are no specific restrictions on when the server 102 obtains the target feature matrix of the target account.

[0151] There are several methods for server 102 to obtain the target feature matrix of the target account. For example, server 102 can obtain the user profile of the target account and extract the target feature matrix based on the user profile; alternatively, it can obtain the historical e-resource consumption information of the target account from a database or other devices, extract features from the historical e-resource messages of the target account, and obtain the target feature matrix of the target account; or it can directly obtain the target feature matrix of the target account from a database or other devices. The user profile can be determined based on the historical e-resource consumption information of the target account, or it can be determined based on the account information of the target account, etc. The user profile can be used to characterize the behavioral characteristics of the target account. The process of server 102 obtaining the target feature matrix of the target account is similar to the process of server 102 obtaining the sample feature matrix of the sample account described in S201, and will not be repeated here.

[0152] S402 uses a trained electronic resource prediction model to predict the probability of a target account using each target service node based on the target feature matrix.

[0153] After obtaining the target feature matrix of the target account, server 102 can use a trained e-resource prediction model to predict the probability of the target account's use on each target service node based on the target feature matrix. The probability of the target account's use on each target service node can be used to characterize the likelihood of the target account consuming e-resources on each target service node. For example, if the probability of the target account's use on target service node A is 0.8, it means that the likelihood of the target account consuming e-resources on target service node A is 0.8.

[0154] Each target service node corresponds to a type of electronic resource reference consumption information. This can be that each target service node corresponds to a type of electronic resource range. Taking a game scenario as an example, the first target service node corresponds to the electronic resource range [0, 20], and the second target service node corresponds to the electronic resource range [50, 80]. In the game, the positions that require 0 to 20 electronic resources all belong to the first target service node, and the positions that require 50 to 80 electronic resources all belong to the second target service node, and so on.

[0155] Each target service node can correspond to a specific electronic resource reference consumption information, or it can correspond to a specific electronic resource consumption value. For example, in a game scenario, the first target service node corresponds to an electronic resource consumption value of 10, and the second target service node corresponds to an electronic resource consumption value of 50. Therefore, in the game, locations requiring 10 electronic resources belong to the first target service node, locations requiring 50 electronic resources belong to the second target service node, and so on. Electronic resource reference consumption information can also include the number of vouchers consumed, the number of virtual tasks completed, etc., which will not be detailed here.

[0156] S403. Based on the obtained usage probabilities and the electronic resource reference consumption information corresponding to each target service node, obtain the electronic resource prediction consumption information of the target account.

[0157] After obtaining the various usage probabilities, server 102 can obtain the predicted electronic resource consumption information of the target account based on the obtained usage probabilities and the electronic resource reference consumption information corresponding to each target service node. Server 102 can use the obtained usage probabilities as weight coefficients corresponding to each target service node. Based on the weight coefficients, server 102 performs weighted fusion processing on the electronic resource reference consumption information corresponding to each target service node to obtain the predicted electronic resource consumption information of the target account. The process of weighted fusion processing on the electronic resource reference consumption information corresponding to each target service node is similar to the information fusion processing process described in S303, and will not be repeated here.

[0158] For example, the probability of the target account using the first, second, and third target service nodes is 0.1, 0.2, and 0.5, respectively, and the reference consumption information of electronic resources for each of the first, second, and third target service nodes is 10, 50, and 100, respectively. Therefore, please refer to... Figure 4c The predicted consumption of electronic resources for the target account is 0.1×10 + 0.2×50 + 0.5×100 = 61.

[0159] The following example, using a game scenario, illustrates the electronic resource consumption information prediction method provided in this application. Each consumption information item is represented by its consumption amount.

[0160] Server 102 obtains the initial sample set D = {(x i y i )}, where i represents the i-th sample account, x i Let y represent the sample feature matrix of the i-th sample account. iThis represents the electronic resource consumption information for the i-th sample account.

[0161] Based on the electronic resource sample consumption information of the sample account, the server 102 sorts the initial samples in the initial sample set D in order of size, and determines the target node combination corresponding to each sample account in turn.

[0162] Taking candidate service nodes Z = {z1 = 1, z2 = 2, z3 = 3, z3 = 4} as an example, this paper provides an example of determining the target node combination corresponding to the sample account. The candidate node combination is represented in vector form.

[0163] If the electronic resource consumption information of the sample account is 10, then from the various candidate node combinations, the candidate node combination (1, 1, 1, 1) with a combination result of 10 is determined as the target node combination, and the combination result of the candidate node combination is 1+2+3+4=10.

[0164] If the electronic resource consumption information of the sample account is 11, and there is no candidate node combination with a node combination of 10 among the candidate node combinations, then the candidate node combination with the smallest error between the combination result and 11 is determined from the candidate node combinations and taken as the target node combination, that is, the candidate node combination with a combination result of 10 (1, 1, 1, 1).

[0165] If the electronic resource consumption information of the sample account is 3, and each candidate node combination includes two candidate node combinations with a node combination of 3, namely (1, 1, 0, 0) and (0, 0, 1, 0), then determine the number of candidate service nodes contained in (1, 1, 0, 0) and (0, 0, 1, 0) respectively, and take the candidate node combination with the fewest nodes, namely (0, 0, 1, 0), as the target node combination.

[0166] If the electronic resource consumption information of the sample account is 5, and each candidate node combination includes two candidate node combinations with node combination 5, namely (1, 0, 0, 1) and (0, 1, 1, 0), then determine the sum of the total occurrences of z1 and z4 in all candidate node combinations, and the sum of the total occurrences of z2 and z3 in all candidate node combinations. If the sum of the total occurrences of z1 and z4 in all candidate node combinations is greater, then (1, 0, 0, 1) will be taken as the target node combination.

[0167] After determining the target node combination corresponding to each sample account, server 102 can determine each target service node based on each target node combination. For example, if the number of target service nodes is k, then server 102 counts the total number of occurrences of each candidate service node in all target node combinations, sorts the candidate service nodes from largest to smallest according to the total number of occurrences, and selects the top k as target service nodes, thus obtaining the target service node Z = (z1, z2, ..., zk). k The value of k can generally be between 5 and 15.

[0168] After obtaining the target node combination, server 102 can train the electronic resource prediction model to be trained based on the sample feature matrix corresponding to each sample account and the training sample set composed of the target node combinations corresponding to each sample account, thus obtaining the trained electronic resource prediction model. During the training process, the electronic resource prediction model can obtain a classification [0, 1] for each target service node regarding whether it consumes electronic resources. k The corresponding usage probabilities are 0, which means that the account will not consume electronic resources at the k-th target service node, and 1, which means that the account will consume electronic resources at the k-th target service node.

[0169] Server 102 can obtain classification results for all target service nodes. These results can be recorded as 0-1 combinations. Based on the electronic resource reference consumption information corresponding to each target service node, server 102 obtains electronic resource training consumption information for each 0-1 combination. Server 102 can determine the training loss of the electronic resource prediction model to be trained based on the error between the electronic resource training consumption information and the electronic resource sample consumption information. It then adjusts the model parameters of the electronic resource prediction model to be trained according to the training loss until the training loss converges, thus obtaining the trained electronic resource prediction model.

[0170] After obtaining the trained e-resource prediction model, server 102 can predict the e-resource usage of the target account. It inputs the target account's target feature matrix into the trained e-resource prediction model to obtain the probability of whether the target account consumes e-resources for each target service node. For example, the probability of use for each target service node can be represented by a vector image, i.e., (0.1, 0.2, 0.1, 0.3). Based on these usage probabilities and the e-resource reference consumption information corresponding to each target service node, server 102 can determine the predicted e-resource consumption information of the target account, i.e., y^=1×0.1+0.2×2+3×0.1+4×0.3=2.

[0171] In a real-world game scenario, the electronic resource consumption information prediction method provided in this application is used for prediction. Based on the initial sample data of 7 consecutive days in the game, the electronic resource prediction model to be trained is trained. The next day in the game is used as training data, and about 50 structured numerical features are used, including the number of games played and the number of paid games.

[0172] For example, if the amounts (reference consumption information) corresponding to the target service nodes are 6, 12, 28, 30, 68, 88, 98, and 108 respectively, and if server 102, after obtaining the target feature matrix of the target account, uses a trained e-resource prediction model to determine the usage probability of the target account for each target service node based on the target feature matrix, which is (0.1, 0.2, 0.1, 0.3, 0.1, 0.1, 0.8, 0.2), then server 102, based on the obtained usage probabilities, determines the predicted e-resource consumption information of the target account as 6×0.1+12×0.2+28×0.1+30×0.3+68×0.1+88×0.1+98×0.8+108×0.2=124.4.

[0173] Using Bi-DNN and Ziln as reference models, the prediction accuracy of the electronic resource prediction model Mbi-DNN in this application embodiment is compared. Bi-DNN is a binary classification model based on a three-layer fully connected deep neural network, which essentially categorizes account consumption information into two levels: consuming electronic resources or not consuming electronic resources. Ziln is a Ziln loss regression model based on a three-layer fully connected deep neural network. The electronic resource prediction model Mbi-DNN is based on a three-layer fully connected deep neural network.

[0174] Finally, the Gini coefficient is used to evaluate the e-resource prediction model. A higher Gini coefficient means that the e-resource prediction model can identify accounts that consume more e-resources each time as high-consumption accounts and accounts that consume less e-resources each time as low-consumption accounts. This can optimize the bidding for online ads and thus improve the ROI of ad placement.

[0175] The evaluation results based on the Gini coefficient are shown in Table 1.

[0176] Table 1

[0177]

[0178]

[0179] It is evident that the prediction accuracy of the electronic resource prediction model Mbi-dnn in this embodiment is far superior to the other two models, with an improvement of approximately 5 to 10%.

[0180] Based on the same inventive concept, this application provides an electronic resource consumption information prediction device, which is equivalent to the server 102 described above and can realize the functions corresponding to the aforementioned application launch method. Please refer to... Figure 5 The device includes an acquisition module 501 and a processing module 502, wherein:

[0181] Acquisition module 501: used to obtain the target feature matrix of the target account, wherein the target feature matrix is ​​obtained based on the historical consumption information of electronic resources of the target account;

[0182] Processing module 502: Used to employ a trained e-resource prediction model, based on the target feature matrix, to predict the usage probability of the target account for each target service node, wherein each target service node corresponds to an e-resource reference consumption information; and, based on the obtained usage probabilities and the e-resource reference consumption information corresponding to each target service node, to obtain the e-resource prediction consumption information of the target account.

[0183] In one possible embodiment, the acquisition module 501 is further configured to: before using a trained electronic resource prediction model to predict the usage probability of a target account for each target service node based on the target feature matrix, obtain an initial sample set, wherein the initial sample set includes the sample feature matrix corresponding to each sample account and the electronic resource sample consumption information corresponding to each sample account.

[0184] The processing module 502 is also used to: determine the target node combination corresponding to each sample account based on the obtained electronic resource sample consumption information and the electronic resource reference consumption information corresponding to each candidate service node.

[0185] The processing module 502 is also used to: train the electronic resource prediction model to be trained using a training sample set composed of each sample feature matrix and each target node combination until the training loss of the electronic resource prediction model to be trained meets the preset convergence condition, thereby obtaining the trained electronic resource prediction model.

[0186] In one possible embodiment, the processing module 502 is specifically used for:

[0187] The candidate service nodes are randomly combined to obtain various candidate node combinations, wherein each candidate node combination includes at least one candidate service node.

[0188] Based on the electronic resource reference consumption information corresponding to each candidate service node, the combination result of the electronic resource reference consumption information corresponding to each candidate node is determined respectively.

[0189] For each sample account, perform the following operations: From each candidate node combination, select the candidate node combination whose combination result matches the electronic resource sample consumption information of one of the sample accounts, and use it as the target node combination for one sample account.

[0190] In one possible embodiment, the processing module 502 is specifically used for:

[0191] From the various candidate node combinations, at least one candidate node combination is selected, wherein the error between the combination result of each candidate node combination in the at least one candidate node combination and the electronic resource sample consumption information corresponding to a sample account is within a preset error range.

[0192] If at least one candidate node combination includes only one candidate node combination, then that candidate node combination is determined as the target node combination corresponding to a sample account.

[0193] If at least one candidate node combination includes multiple candidate node combinations, then determine the number of candidate service nodes included in each candidate node combination, and select the candidate node combinations whose number of nodes is within the preset range as the target node combination corresponding to a sample account.

[0194] In one possible embodiment, the processing module 502 is further configured to:

[0195] After determining the target node combination corresponding to each sample account, based on the obtained target node combinations, the total number of times each candidate service node appears in each target node combination is counted to obtain the statistical results of each candidate service node.

[0196] From all candidate service nodes, select those whose statistical results meet the preset statistical conditions and use them as target service nodes.

[0197] In one possible embodiment, the acquisition module 501 is further configured to: obtain the historical usage records of each sample account for each candidate service node before using the trained electronic resource prediction model to predict the usage probability of the target account for each target service node based on the target feature matrix.

[0198] The processing module 502 is also used to: count the number of times each sample account uses electronic resources at each candidate service node based on the obtained historical usage records;

[0199] The processing module 502 is also used to: filter out candidate service nodes whose usage counts are within a preset range, and use them as target service nodes.

[0200] In one possible embodiment, the processing module 502 is specifically used for:

[0201] The probability of a target account being used on each target service node is determined as the weighting coefficient for each target service node.

[0202] Based on the obtained weight coefficients, the electronic resource reference consumption information corresponding to each target service node is weighted and fused to obtain the electronic resource prediction consumption information of the target account.

[0203] Based on the same inventive concept, this application provides a computer device, which is described below.

[0204] Please refer to Figure 6 The aforementioned electronic resource consumption information prediction device can run on computer device 600. The current and historical versions of the electronic resource consumption information prediction program, as well as the application software corresponding to the electronic resource consumption information prediction program, can be installed on computer device 600. Computer device 600 includes display unit 640, processor 680, and memory 620. The display unit 640 includes display panel 641, which is used to display user interactive operation interface, etc.

[0205] In one possible embodiment, the display panel 641 may be configured in the form of a liquid crystal display (LCD) or an organic light-emitting diode (OLED).

[0206] The processor 680 is used to read a computer program and then execute the methods defined by the computer program. For example, the processor 680 reads an electronic resource consumption information prediction program or file, thereby running the electronic resource consumption information prediction program on the computer device 600 and displaying the corresponding interface on the display unit 640. The processor 680 may include one or more general-purpose processors, and may also include one or more DSPs (Digital Signal Processors) for performing related operations to implement the technical solutions provided in the embodiments of this application.

[0207] The memory 620 generally includes main memory and secondary storage. The main memory can be random access memory (RAM), read-only memory (ROM), and cache, etc. The secondary storage can be a hard disk, optical disk, USB flash drive, floppy disk, or magnetic tape drive, etc. The memory 620 is used to store computer programs and other data. The computer programs include applications corresponding to each client, and the other data may include data generated after the operating system or applications are run, including system data (e.g., operating system configuration parameters) and user data. In this embodiment, program instructions are stored in the memory 620, and the processor 680 executes the program instructions stored in 620 to implement any of the electronic resource consumption information prediction methods discussed in the preceding figures.

[0208] The aforementioned display unit 640 is used to receive input digital information, character information, or contact touch operations / non-contact gestures, and to generate signal inputs related to user settings and function control of the computer device 600. Specifically, in this embodiment, the display unit 640 may include a display panel 641. The display panel 641, for example, is a touch screen, which can collect touch operations performed by the user on or near it (such as operations performed by the user using a finger, stylus, or any suitable object or accessory on or on the display panel 641), and drive corresponding connection devices according to a pre-set program.

[0209] In one possible embodiment, the display panel 641 may include two parts: a touch detection device and a touch controller. The touch detection device detects the player's touch position and the signal generated by the touch operation, and transmits the signal to the touch controller. The touch controller receives touch information from the touch detection device, converts it into touch point coordinates, and sends it to the processor 680. It can also receive and execute commands sent by the processor 680.

[0210] The display panel 641 can be implemented using various types such as resistive, capacitive, infrared, and surface acoustic wave. In addition to the display unit 640, the computer device 600 may also include an input unit 630, which may include a graphical input device 631 and other input devices 632. These other input devices may include, but are not limited to, one or more of the following: a physical keyboard, function keys (such as volume control buttons, power buttons, etc.), a trackball, a mouse, and a joystick.

[0211] In addition to the above, the computer device 600 may also include a power supply 690 for powering other modules, an audio circuit 660, a near-field communication module 670, and an RF circuit 610. The computer device 600 may also include one or more sensors 650, such as an accelerometer, a light sensor, a pressure sensor, etc. The audio circuit 660 specifically includes a speaker 661 and a microphone 662, for example, the computer device 600 can use the microphone 662 to collect the user's voice and perform corresponding operations.

[0212] As one embodiment, the number of processors 680 can be one or more, and the processors 680 and memory 620 can be coupled together or relatively independent.

[0213] As one example, Figure 6 The processor 680 in the middle can be used to implement, for example Figure 5 The functions of the acquisition module 501 and the processing module 502 in the process.

[0214] As one example, Figure 6 The processor 680 can be used to implement the functions corresponding to the test device 103 discussed above.

[0215] Those skilled in the art will understand that all or part of the steps of the above method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When the program is executed, it performs the steps of the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0216] Alternatively, if the integrated units of this invention are implemented as software functional modules and sold or used as independent products, they can also be stored in a computer-readable storage medium. Based on this understanding, the technical solutions of the embodiments of this invention, or the parts that contribute to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as mobile storage devices, ROM, RAM, magnetic disks, or optical disks.

[0217] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for predicting electronic resource consumption information, characterized in that, include: Obtain the target feature matrix of the target account, wherein the target feature matrix is ​​obtained based on the historical electronic resource consumption information of the target account; Using a trained electronic resource prediction model, based on the target feature matrix, the probability of the target account's use for each target service node is predicted, wherein each target service node corresponds to a type of electronic resource reference consumption information; Based on the obtained usage probabilities and the electronic resource reference consumption information corresponding to each target service node, the electronic resource predicted consumption information of the target account is obtained. The trained electronic resource prediction model was obtained in advance using the following method: An initial sample set is obtained, wherein the initial sample set includes the sample feature matrix corresponding to each sample account and the electronic resource sample consumption information corresponding to each sample account; Based on the obtained electronic resource consumption information of each sample and the electronic resource reference consumption information of each candidate service node, the target node combination corresponding to each sample account is determined respectively. The electronic resource prediction model to be trained is trained using a training sample set composed of feature matrices of each sample and combinations of target nodes until the training loss of the electronic resource prediction model to be trained meets the preset convergence condition, thus obtaining the trained electronic resource prediction model.

2. The method according to claim 1, characterized in that, Based on the obtained electronic resource consumption information of each sample, and the electronic resource reference consumption information corresponding to each candidate service node, the target node combination corresponding to each sample account is determined, including: The candidate service nodes are randomly combined to obtain various candidate node combinations, wherein each candidate node combination includes at least one candidate service node. Based on the electronic resource reference consumption information corresponding to each candidate service node, the combination result of the electronic resource reference consumption information corresponding to each candidate node is determined respectively. For each of the sample accounts, the following operations are performed: From each of the candidate node combinations, a candidate node combination whose combination result matches the electronic resource sample consumption information corresponding to one of the sample accounts is selected as the target node combination corresponding to that sample account.

3. The method according to claim 2, characterized in that, From the candidate node combinations, candidate node combinations whose combination results match the electronic resource sample consumption information corresponding to one of the sample accounts are selected as the target node combination corresponding to the sample account, including: From the various candidate node combinations, at least one candidate node combination is selected, wherein the error between the combination result of each candidate node combination in the at least one candidate node combination and the electronic resource sample consumption information corresponding to the sample account is within a preset error range. If the at least one candidate node combination includes only one candidate node combination, then the one candidate node combination is determined as the target node combination corresponding to the sample account; If the at least one candidate node combination includes multiple candidate node combinations, then the number of candidate service nodes included in each of the multiple candidate node combinations is determined, and the candidate node combinations whose number of nodes is within a preset range are selected as the target node combination corresponding to the sample account.

4. The method according to any one of claims 1 to 3, characterized in that, After determining the target node combination corresponding to each of the sample accounts, the process also includes: Based on the obtained combinations of target nodes, the total number of times each candidate service node appears in each combination of target nodes is counted to obtain the statistical results of each candidate service node; From the candidate service nodes, the candidate service nodes whose statistical results meet the preset statistical conditions are selected as the target service nodes.

5. The method according to any one of claims 1 to 3, characterized in that, Before using the trained electronic resource prediction model to predict the probability of the target account's use for each target service node based on the target feature matrix, the method further includes: Obtain the historical usage records of each sample account for each candidate service node; Based on the obtained historical usage records, the number of times each sample account used electronic resources at each candidate service node is counted. Candidate service nodes whose usage frequency falls within a preset range are selected as target service nodes.

6. The method according to any one of claims 1 to 3, characterized in that, Based on the obtained usage probabilities and the electronic resource reference consumption information corresponding to each target service node, the predicted electronic resource consumption information of the target account is obtained, including: The probability of the target account being used for each target service node is determined as the weight coefficient corresponding to each target service node. Based on the obtained weight coefficients, the electronic resource reference consumption information corresponding to each target service node is weighted and fused to obtain the electronic resource prediction consumption information of the target account.

7. An electronic resource consumption information prediction device, characterized in that, include: Acquisition module: used to obtain the target feature matrix of the target account, wherein the target feature matrix is ​​obtained based on the historical electronic resource consumption information of the target account; Processing module: Used to employ a trained e-resource prediction model, based on the target feature matrix, to predict the usage probability of the target account for each target service node, wherein each target service node corresponds to an e-resource reference consumption information; and, based on the obtained usage probabilities and the e-resource reference consumption information corresponding to each target service node, to obtain the e-resource prediction consumption information of the target account. The trained electronic resource prediction model is pre-trained using the following method, and the acquisition module is further used for: An initial sample set is obtained, wherein the initial sample set includes the sample feature matrix corresponding to each sample account and the electronic resource sample consumption information corresponding to each sample account; Based on the obtained electronic resource consumption information of each sample and the electronic resource reference consumption information of each candidate service node, the target node combination corresponding to each sample account is determined respectively. The electronic resource prediction model to be trained is trained using a training sample set composed of feature matrices of each sample and combinations of target nodes until the training loss of the electronic resource prediction model to be trained meets the preset convergence condition, thus obtaining the trained electronic resource prediction model.

8. A computer device, characterized in that, include: Memory, used to store program instructions; A processor is configured to invoke program instructions stored in the memory and execute the method as described in any one of claims 1 to 6 according to the obtained program instructions.

9. A storage medium, characterized in that, The storage medium stores computer-executable instructions for causing a computer to perform the method as described in any one of claims 1 to 6.

Images