Content pushing method and device, electronic equipment and storage medium
By using a multi-task probability model to predict users' switching rates and repurchase rates, this technology solves the problems of high cost and low accuracy in content delivery, achieving efficient and accurate content delivery and improving user experience.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHENZHEN DACHENG COMM TECH CO LTD
- Filing Date
- 2023-01-10
- Publication Date
- 2026-06-26
AI Technical Summary
Existing technologies suffer from high costs and low accuracy when pushing content to users via the Internet, especially when it is difficult to achieve precise targeting when identifying high-value user groups.
By employing a multi-task trained probability model, the system obtains the raw data of the target audience, predicts their switching rate and repurchase rate, and then determines the target audience for push notifications, thereby improving the accuracy of content delivery.
By using a probabilistic model with a multi-task learning architecture, the accuracy of content delivery is improved, costs are reduced, and the user experience is enhanced.
Smart Images
Figure CN116166884B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of data processing technology, and more specifically, to a content delivery method, apparatus, electronic device, and storage medium. Background Technology
[0002] With the development of mobile internet, big data, artificial intelligence, and other technologies, people are using the internet more and more widely. Currently, applications that push content to users via the internet are becoming increasingly common, such as sending advertisements and promotions. However, these technologies suffer from high costs and inconsistent accuracy in targeting users. Summary of the Invention
[0003] In view of the above problems, this application proposes a content push method, device, electronic device and storage medium, which can obtain the device replacement and repurchase rate of the push object through a multi-task trained probability model, and obtain the target push object of the content based on the device replacement and repurchase rate of the push object, thereby improving the accuracy of content push, reducing the cost of content push and improving the user experience.
[0004] In a first aspect, embodiments of this application provide a content push method, the method comprising: acquiring a push object dataset, wherein the push object dataset includes multiple push objects and original data corresponding to each push object; inputting the original data into a trained probability model to obtain a device replacement and repurchase rate output by the trained probability model corresponding to the original data, wherein the device replacement and repurchase rate is obtained based on the device replacement rate and the repurchase rate corresponding to the original data; determining a target push object from the multiple push objects based on the device replacement and repurchase rate; and pushing content to the target push object.
[0005] Secondly, embodiments of this application provide a content push device, comprising: a push object dataset acquisition module, a device replacement and repurchase rate acquisition module, a target push object determination module, and a content push module. The push object dataset acquisition module is used to acquire a push object dataset, wherein the push object dataset includes multiple push objects and original data corresponding to each push object; the device replacement and repurchase rate acquisition module is used to input the original data into a trained probability model to obtain the device replacement and repurchase rate output by the trained probability model corresponding to the original data, wherein the device replacement and repurchase rate is obtained based on the device replacement rate and the repurchase rate corresponding to the original data; the target push object determination module is used to determine a target push object from the multiple push objects based on the device replacement and repurchase rate; and the content push module is used to push content to the target push object.
[0006] Thirdly, embodiments of this application provide an electronic device, including a memory and a processor, wherein the memory is coupled to the processor, the memory stores instructions, and when the instructions are executed by the processor, the processor performs the above-described method.
[0007] Fourthly, embodiments of this application provide a computer-readable storage medium storing program code, which can be invoked by a processor to execute the above-described method.
[0008] The content push method, apparatus, electronic device, and storage medium provided in this application improve the accuracy of content push, reduce the cost of content push, and enhance the user experience by acquiring a push object dataset, which includes multiple push objects and their corresponding original data; inputting the original data into a trained probability model to obtain the device replacement and repurchase rate output by the trained probability model corresponding to the original data, wherein the device replacement and repurchase rate is obtained based on the device replacement rate and repurchase rate corresponding to the original data; determining the target push object from the multiple push objects based on the device replacement and repurchase rate; and pushing content to the target push object. This improves the accuracy of content push, reduces the cost of content push, and enhances the user experience by using a multi-task trained probability model to obtain the device replacement and repurchase rate corresponding to the push object, and obtaining the target push object based on the device replacement and repurchase rate of the push object. Attached Figure Description
[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0010] Figure 1 A flowchart illustrating a content push method according to an embodiment of this application is shown;
[0011] Figure 2 A schematic diagram of the structure of a trained probability model provided in an embodiment of this application is shown;
[0012] Figure 3 A flowchart illustrating a content push method according to an embodiment of this application is shown;
[0013] Figure 4 A flowchart illustrating a content push method according to an embodiment of this application is shown;
[0014] Figure 5 A flowchart illustrating a content push method according to an embodiment of this application is shown;
[0015] Figure 6 This paper shows a module block diagram of a content push device according to an embodiment of the present application;
[0016] Figure 7 A block diagram of an electronic device for performing a content push method according to an embodiment of this application is shown;
[0017] Figure 8 A storage unit for storing or carrying program code implementing the content push method according to an embodiment of this application is shown. Detailed Implementation
[0018] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.
[0019] With the development of mobile internet, big data, artificial intelligence, and other technologies, people are using the internet more and more widely. For example, as the internet becomes more widespread, people are increasingly relying on e-commerce platforms such as Tmall, Taobao, JD.com, and Youzan for shopping. These e-commerce platforms serve hundreds of millions of users every day, and each platform pushes different product content to users daily for marketing purposes.
[0020] Currently, in internet-based marketing scenarios, due to budget constraints and the need to avoid excessive disruption to users, it is usually necessary to select high-value users from the overall user base for marketing. How to accurately target high-value users for content delivery has become a significant challenge in marketing modeling.
[0021] For example, in mobile phone marketing scenarios, the lifecycle of an existing user is usually divided into the novice stage, growth stage, maturity stage, and replacement stage. In mobile phone marketing scenarios, it is important to mobile phone manufacturers to identify this high-repurchase-intention user group in advance and to take effective user marketing and outreach actions in advance to retain or convert more users.
[0022] In related technologies, methods for identifying high-potential repeat purchase users fall into two main categories: operational strategy-based and machine learning model-based. Operational strategy-based methods rely on data analysis and user operational experience, manually defining user segmentation criteria and manually identifying eligible users as high-potential repeat purchase users. Because operational strategy-based methods heavily depend on the experience and skill of operational personnel, they suffer from high costs. Machine learning model-based methods are mostly single-task machine learning models. They utilize historical information from repeat purchase users (e.g., device usage time, free memory usage percentage) to divide the user group into non-repurchase users and repeat purchase users, and then use this information as a label, along with corresponding user features (e.g., device usage time, free memory usage percentage), to train a model to identify high-potential repeat purchase users. Since all users who switch phones include not only repeat customers but also churned customers (those who purchased other brands), removing churned customers would bias the user sample used to train the model, making it difficult for the model to predict the repeat purchase intentions of these users. On the other hand, if these churned customers are retained, their user characteristics (such as holding time and free memory percentage) are highly similar to those of repeat customers, but their sample labels are completely opposite during model training, resulting in lower accuracy of the model trained to identify high-potential repeat customers.
[0023] To address the aforementioned problems, the inventors, through long-term research, discovered and proposed the content push method, apparatus, electronic device, and storage medium provided in the embodiments of this application. This method obtains the device replacement and repurchase rate corresponding to the push target through a multi-task trained probability model, and obtains the target push object for the content based on the device replacement and repurchase rate of the push target. This improves the accuracy of content push, reduces the cost of content push, and enhances the user experience. The specific content push method will be described in detail in subsequent embodiments.
[0024] Please see Figure 1 , Figure 1 This illustration shows a flowchart of a content push method according to an embodiment of this application. This content push method obtains the device switching and repurchase rate corresponding to the push target through a multi-task trained probability model, and obtains the target push target for content based on the device switching and repurchase rate of the push target. This improves the accuracy of content push, reduces the cost of content push, and also enhances the user experience. In specific embodiments, this content push method can be applied to, for example... Figure 6 The content push device 200 and the electronic device 100 equipped with the content push device 200 are shown. Figure 7The following will use an electronic device as an example to illustrate the specific process of this embodiment. Of course, it is understood that the electronic device used in this embodiment may include smartphones, tablets, wearable electronic devices, etc., and is not limited thereto. The following will focus on... Figure 1 The process shown will be explained in detail. The content push method may specifically include the following steps:
[0025] Step S110: Obtain the push object dataset, wherein the push object dataset includes multiple push objects and the original data corresponding to each push object.
[0026] In some implementations, the electronic device can obtain the push object dataset from the associated cloud or electronic device through wireless communication technologies (such as Bluetooth, WiFi, Zigbee, etc.); it can also obtain the push object dataset from the associated electronic device through a serial communication interface (such as a serial peripheral interface, etc.); the push object dataset can be pre-stored in the memory of the electronic device, and the processor of the electronic device can directly obtain the push object dataset from the memory of the electronic device.
[0027] The push target dataset can include multiple push targets and their corresponding raw data. A push target can be understood as the object to which content is pushed; it can be the electronic device receiving the content or the user receiving the content. Push targets can include regular mobile phone users, churned users, and repeat customers. Regular mobile phone users can be understood as users using the target brand of mobile phones; churned users can be understood as users who have switched phones and purchased a non-target brand phone; and repeat customers can be understood as users who have switched phones and then purchased a target brand phone again.
[0028] The original data corresponding to each push target may include the user's basic attributes (such as gender, age, ID, etc.) and the user's historical behavior data (such as the user's mobile phone usage behavior, such as the percentage of remaining mobile phone memory, the sequence of opening mobile applications, and mobile phone lag). These are not limited here.
[0029] In some implementations, the electronic device can support task submission to the Spark cluster. Furthermore, the electronic device can obtain user data through Spark tasks as the raw data corresponding to the push objects. The users obtained through Spark tasks can be understood as the push objects. The user data obtained through Spark tasks can include basic user attributes (e.g., user gender, age, ID, etc.) and historical user behavior data (e.g., user's phone usage behavior, including but not limited to the percentage of remaining phone memory, the sequence of opening phone applications, phone lag, etc.). It is understood that the electronic device can support task submission to the Spark cluster. Furthermore, the electronic device can obtain the user data corresponding to each Spark task as the raw data corresponding to each push object; that is, the electronic device can obtain the user dataset corresponding to the Spark cluster tasks as the push object dataset. The Spark cluster can continuously submit tasks to the electronic device, and the push object dataset obtained by the corresponding electronic device can be continuously updated.
[0030] For example, the push content is an OPPO brand smartphone. The electronic device can support Spark cluster task submission. Furthermore, the electronic device can obtain raw data of active users on the dashboard through Spark tasks. Here, active users on the dashboard can be understood as the push targets, or users currently using OPPO brand smartphones normally. In some implementations, the electronic device can be pre-configured with a preset neural network model, such as logistic regression or Classification and Regression Tree (CART) models. Furthermore, after obtaining the raw data of active users on the dashboard through Spark tasks, the electronic device can use this preset neural network model to obtain the user's basic attributes (e.g., gender, age), historical user behavior data (e.g., sequences of app openings, remaining phone memory percentage), etc., as the raw data corresponding to the push targets.
[0031] Step S120: Input the original data into the trained probability model to obtain the replacement and repurchase rate corresponding to the original data output by the trained probability model, wherein the replacement and repurchase rate is obtained based on the replacement rate and the repurchase rate corresponding to the original data.
[0032] In some implementations, the electronic device may have a pre-set trained probability model. The electronic device may also obtain the trained probability model from a connected cloud or other electronic device via wireless communication technology, or from a connected electronic device via a serial communication interface. The trained probability model may employ a multi-task learning architecture; the multi-tasks corresponding to the trained probability model may include a replacement rate learning task, a repurchase rate learning task, and a replacement and repurchase rate learning task, etc.
[0033] For example, please refer to Figure 2 , Figure 2 This application illustrates a trained probability model provided in one embodiment. This trained probability model is constructed for mobile marketing scenarios and is divided into two sub-modules, or two tasks. Tower 1 of the trained probability model predicts the replacement rate (pCTR) of the target device, i.e., the user's willingness to switch devices; Tower 2 predicts the repurchase rate (pCVR) of the target device, i.e., the user's repurchase willingness. It is understood that, considering the multi-task learning architecture simultaneously training multiple end-to-end targets, during the process of training the model using a dataset to obtain the trained probability model, the dataset may include randomly sampled users of the target brand's mobile phones who are using them normally, randomly sampled churned users, and randomly sampled repurchase users. During the training of Tower 1, churned users and repurchase users are treated as positive samples; during the training of Tower 2, churned users are treated as negative samples, and repurchase users are treated as positive samples. Furthermore, the trained probability model can obtain the device switching and repurchase rate (pCTVR) corresponding to the push target based on the device switching rate obtained from Tower 1 and the repurchase rate obtained from Tower 2, thereby avoiding sample conflicts (that is, the conflict between churned users and repurchase users can exist among device switching users), and reducing the inconsistency between the distribution of predicted samples and the distribution of training samples by making full use of training samples.
[0034] Understandably, the trained probability model targets a funnel-shaped conversion path, meaning that among users of the target brand's phones, some intend to upgrade, while others intend to repurchase the target brand's phones. By employing a multi-task learning architecture, the efficiency of the final conversion stage is improved by fully utilizing samples from the preceding stages of the model. This effectively avoids conflicts between positive and negative samples during training, thus fully leveraging information from different user groups, improving the consistency between the model's training and prediction sample distributions, and ultimately enhancing the model's accuracy.
[0035] In some implementations, after obtaining the push target dataset, the electronic device can input the raw data into a trained probability model to obtain the replacement and repurchase rate corresponding to the raw data, output by the trained probability model. The replacement and repurchase rate can be obtained based on the replacement rate and repurchase rate corresponding to the raw data.
[0036] In some implementations, before inputting raw data into the trained probability model, the electronic device can preprocess the raw data, such as through outlier handling or missing value imputation, to obtain preprocessed raw data. For example, the preprocessed raw data might include the user's ID and user characteristics (e.g., user's device usage time, user's phone type, etc.). For instance, if the age of a user included in the raw data is negative, it can be determined that the age of the user included in the raw data is an outlier. Furthermore, the electronic device can delete the raw data or replace the user's age included in the raw data with the mean, mode, median, etc., of the ages included in the push target dataset.
[0037] Step S130: Determine the target push object from the plurality of push objects based on the replacement and repurchase rate.
[0038] In some implementations, after obtaining the replacement and repurchase rates corresponding to each push object in the push dataset, the electronic device can determine the target push object from multiple push objects based on the replacement and repurchase rates corresponding to each push object. Optionally, a probability threshold can be preset in the electronic device. Further, the electronic device can determine the target push object from multiple push objects based on the replacement and repurchase rates by comparing the replacement and repurchase rates corresponding to each push object with the probability threshold. If the replacement and repurchase rate is greater than or equal to the probability threshold, the push object is determined to be the target push object; if the replacement and repurchase rate is greater than or equal to the probability threshold, the push object is determined not to be the target push object.
[0039] Optionally, the electronic device determines the target push object from multiple push objects based on the replacement and repurchase rate. This can be done by arranging the push objects in descending order of their corresponding replacement and repurchase rates, obtaining a push object list, and saving this push object list to a database in the electronic device. In some embodiments, the electronic device may have a preset ratio. Furthermore, the electronic device can obtain a preset ratio number of push objects from the head of the push object list as target push objects.
[0040] Step S140: Push the content to the target push object.
[0041] In some implementations, after obtaining the target object, the electronic device can push content to the target object. Optionally, the electronic device may pre-store the content to be pushed to the target object, or the electronic device may obtain the content to be pushed to the target object from a related cloud or electronic device via wireless communication technology, or the electronic device may obtain the content to be pushed to the target object from a related electronic device via a serial communication interface.
[0042] The content pushed to the target audience may include advertisements, promotional materials, etc.; and the content format may be audio / video, text-based interfaces, etc. For example, the content pushed to the target audience may include a video introducing the performance of the OPPO foldable phone.
[0043] In some implementations, the electronic device can push content to the target object via wireless communication technology or via a serial communication interface. In this embodiment, the method by which the electronic device delivers content to the target object is not limited.
[0044] One embodiment of this application provides a content push method that obtains a push object dataset, which includes multiple push objects and their corresponding original data; inputs the original data into a trained probability model to obtain the device replacement and repurchase rate output by the trained probability model corresponding to the original data, wherein the device replacement and repurchase rate is obtained based on the device replacement rate and repurchase rate corresponding to the original data; determines the target push object from the multiple push objects based on the device replacement and repurchase rate; pushes content to the target push object, and then obtains the device replacement and repurchase rate corresponding to the push object through a multi-task trained probability model, and obtains the target push object of the content based on the device replacement and repurchase rate of the push object, thereby improving the accuracy of content push, reducing the cost of content push, and improving the user experience.
[0045] Please see Figure 3 , Figure 3 A flowchart illustrating a content push method according to an embodiment of this application is shown. This method is applied to the aforementioned electronic device, and will be discussed below. Figure 3 The process shown will be explained in detail. The content push method may specifically include the following steps:
[0046] Step S210: Obtain the push object dataset, wherein the push object dataset includes multiple push objects and the original data corresponding to each push object.
[0047] For a detailed description of step S210, please refer to the previous description of step S110, which will not be repeated here.
[0048] Step S220: Preprocess the original data to obtain preprocessed original data, wherein the preprocessing includes at least one of outlier handling and missing value imputation.
[0049] In some implementations, after obtaining the raw data, the electronic device can preprocess the raw data to obtain preprocessed raw data. Preprocessing may include data processing methods such as outlier handling and missing value imputation.
[0050] For example, if the original data does not include the user's age, it can be determined that the original data has missing values; further, the electronic device can calculate the mean, mode, median, etc. of the ages included in the push target dataset to fill in the age of the user corresponding to the original data.
[0051] Step S230: Input the original data into the trained probability model to obtain the replacement and repurchase rate corresponding to the original data output by the trained probability model, wherein the replacement and repurchase rate is obtained based on the replacement rate and the repurchase rate corresponding to the original data.
[0052] In some implementations, the electronic device may have a pre-set trained probability model. Further, after obtaining the original data corresponding to the push target, the electronic device can input the original data into the trained probability model to obtain the replacement and repurchase rate output by the trained probability model corresponding to the original data. The replacement and repurchase rate is obtained based on the replacement rate and repurchase rate corresponding to the original data. Specifically, the replacement and repurchase rate corresponding to the original data is the same as the replacement and repurchase rate corresponding to the push target.
[0053] In some implementations, the trained probabilistic model may include a trained shared embedding layer, a trained device replacement probability model, and a trained repeat purchase probability model. See [link to relevant documentation]. Figure 4 Step S230 may include steps S231-S234.
[0054] Step S231: Input the raw data into the trained shared embedding layer to obtain the shared detection features output by the trained shared embedding layer.
[0055] In some implementations, considering the need to standardize neural network inputs to improve accuracy and speed, the trained probability model provided in this application embodiment may have preset standardization parameters. Furthermore, after obtaining the raw data, the electronic device can standardize the raw data based on the preset standardization parameters to obtain standardized raw data.
[0056] Furthermore, after obtaining standardized raw data, the electronic device can input the raw data into the trained shared embedding layer of the trained probability model to obtain the shared detection features output by the trained shared embedding layer.
[0057] The trained shared embedding layer can employ domain embedding methods, such as manifold learning, local linear embedding (LLE), Laplacian eigenmaps, and t-distributed random neighbor embedding, to transform the original data input to the trained shared embedding layer into shared detection features of the target dimension (e.g., 16×N).
[0058] Step S232: Input the shared detection features into the trained switching probability model to obtain the switching rate corresponding to the original data output by the trained switching probability model.
[0059] In some implementations, after the trained probability model obtains shared detection features, these shared detection features can be input into the trained phone-switching probability model to obtain the phone-switching rate corresponding to the raw data output by the trained phone-switching probability model. Here, the phone-switching rate corresponding to the raw data can be understood as the probability that a user normally using the target brand of mobile phone will become a phone-switching user.
[0060] Step S233: Input the shared detection features into the trained repurchase probability model to obtain the repurchase rate corresponding to the original data output by the trained repurchase probability model.
[0061] In some implementations, after the trained probability model obtains shared detection features, these shared detection features can be input into the trained repurchase probability model to obtain the repurchase rate corresponding to the raw data output by the trained repurchase probability model. Here, the repurchase rate corresponding to the raw data can be understood as the probability that a user who normally uses the target brand mobile phone will become a repeat customer.
[0062] Step S234: Based on the replacement rate and repurchase rate corresponding to the original data, obtain the replacement and repurchase rate corresponding to the original data.
[0063] In some implementations, after the trained probabilistic model obtains the replacement rate and repurchase rate corresponding to the original data, it can obtain the replacement and repurchase rate corresponding to the original data based on the replacement rate and repurchase rate corresponding to the original data. It is understood that the replacement and repurchase rate corresponding to the original data is a conditional probability relative to the replacement rate and repurchase rate corresponding to the original data.
[0064] The trained probability model obtains the replacement rate and repurchase rate (pCTVR) corresponding to the original data based on the original data's replacement rate (pCTR) and repurchase rate (pCVR). This can be achieved by using the joint probability distribution formula: p(y=1, z=1|x) = p(y=1|x) × p(z=1|y=1, x), thus revealing the relationship between pCTVR, pCTR, and pCVR. Here, p(y=1, z=1|x) represents the replacement rate and repurchase rate (pCTVR) corresponding to the original data, p(y=1|x) represents the replacement rate (pCTR), and p(z=1|y=1, x) represents the repurchase rate (pCVR). In this context, y represents the replacement tag (y=1 for replacement, y=0 for no replacement); x represents the normal user (i.e., the target audience); and z represents the repurchase tag (z=1 for repurchase, z=0 for no repurchase).
[0065] In some implementations, please refer to Figure 5 The content pushing method provided in this application embodiment may include steps S2310-S2320 before step S230.
[0066] Step S2310: Obtain the sample push object dataset, wherein the sample push object dataset includes multiple sample objects, sample data corresponding to each sample object, and sample tags corresponding to each sample object, wherein the sample tags include at least one of the following: device replacement tag, no device replacement tag, repurchase tag, and no repurchase tag.
[0067] In some implementations, the electronic device can obtain the sample push object dataset from an associated cloud or electronic device via wireless communication technology, or it can obtain the sample push object dataset from an associated electronic device via a serial communication interface. The sample push object dataset includes multiple sample objects, sample data corresponding to each sample object, and sample tags corresponding to each sample object; wherein the sample tags include at least one of replacement tags, non-replacement tags, repurchase tags, and non-repurchase tags.
[0068] For example, the electronic device can support task submission to a Spark cluster. It can obtain raw data (e.g., gender, age, and behavioral sequences of opening mobile applications) of a large number of users who are normally using the target brand's mobile phones through Spark tasks. These users are the sample objects. Furthermore, the electronic device can use the raw data of the large number of users corresponding to the Spark cluster task submissions as a dataset, and divide this dataset according to a preset ratio (e.g., training set: validation set: test set = 4:1:1). This dataset can be understood as the sample push object dataset. The training set can be used to train the model; the validation set can be used to determine the hyperparameters in the model; and the test set can be used to test and evaluate the model's accuracy, i.e., its generalization ability. Furthermore, the dataset obtained by the electronic device can also include sample labels corresponding to the users of each large number of mobile phones. For example, the sample label corresponding to a user may be "no phone replacement," or "phone replacement," or "phone replacement and no repeat purchase," or "phone replacement and repeat purchase."
[0069] In some implementations, after the electronic device obtains the sample push object dataset, it can preprocess the sample data, such as outlier handling and missing value imputation, to obtain preprocessed sample data.
[0070] Step S2320: Train the initial probability model based on the sample push object dataset to obtain the trained probability model.
[0071] In some implementations, the initial probability model provided in this application can be an ESMM model with a multi-task learning architecture. This initial probability model can be based on the funnel-shaped conversion link between users who regularly use the target brand mobile phone, users who upgrade their phones, and repeat purchase users. It constructs two auxiliary prediction tasks: a task to obtain the probability of upgrading and a task to obtain the rate of upgrading and repeat purchase. Based on these auxiliary tasks, repeat purchase users and churned users are treated as positive samples. On one hand, the sample data is passed through a shared embedding layer; on the other hand, the features output by the shared embedding layer based on the sample data are input into the two auxiliary prediction tasks. Finally, the initial probability model uses the rich positive sample information from the preceding tasks to learn the shared embedding layer, thereby mitigating the data label conflict when directly predicting the rate of upgrading and repeat purchase.
[0072] In some implementations, an initial probability model may be pre-set in the electronic device. Furthermore, after the electronic device obtains the sample push object dataset, it can train the initial probability model based on the sample push object dataset to obtain a trained probability model.
[0073] In some implementations, the initial probability model may include an initial shared embedding layer, an initial replacement probability model, and an initial repurchase probability model, and step S2320 may include steps S2321-S2325.
[0074] Step S2321: Input the sample data into the initial shared embedding layer to obtain the shared sample features output by the initial shared embedding layer.
[0075] In some implementations, the electronic device trains an initial probability model based on a sample push object dataset to obtain a trained probability model. This can be achieved by inputting sample data from the sample push object dataset into an initial shared embedding layer in the initial probability model to obtain shared sample features output by the initial shared embedding layer. The initial shared embedding layer can use a domain-specific approach to transform the sample data into feature outputs of different dimensions, thereby improving the efficiency of subsequent steps in the initial probability model.
[0076] In some implementations, step S2321 may include steps S3211-S3212.
[0077] Step S3211: Standardize the sample data based on the standardization parameters to obtain standard sample features.
[0078] In some implementations, the features input to the initial shared embedding layer correspond to standardized parameters; furthermore, the electronic device can standardize the sample data input to the shared embedding layer based on the standardized parameters to obtain standardized sample features.
[0079] During the training process, the standardized parameters can be continuously updated. After obtaining the trained probability model, the standardized parameters corresponding to the trained probability model can be saved as preset standardized parameters. In the process of using the trained probability model to predict the replacement and repurchase rate corresponding to the original data, the standardized parameters can be used to standardize the original data to obtain standardized original data, thereby improving the efficiency of the model.
[0080] In one training cycle of training the initial probability model to obtain the trained probability model, the electronic device can traverse the sample data in the sample push object dataset, standardize the sample data based on the standardization parameters, and obtain standard sample features.
[0081] Step S3212: Input the standard sample features into the initial shared embedding layer to obtain the shared sample features output by the initial shared embedding layer.
[0082] In some implementations, after the electronic device obtains the standard sample features corresponding to the sample data, it can input the standard sample features into the initial shared embedding layer to obtain the shared sample features output by the initial shared embedding layer.
[0083] The initial shared embedding layer can be based on domain embedding methods, such as manifold learning, local linear embedding, Laplacian feature mapping, t-distribution random neighbor embedding, etc., to convert the input data into feature outputs of different dimensions.
[0084] Step S2322: Input the shared sample features into the initial switching probability model to obtain the switching rate corresponding to the sample data output by the initial switching probability model.
[0085] In some implementations, after the initial probability model obtains shared sample features, these features can be input into the initial device replacement probability model to obtain the device replacement rate corresponding to the sample data output by the initial device replacement probability model. Here, the device replacement probability corresponding to the sample data can be understood as the probability that a user normally using the target brand will switch to a device replacement user.
[0086] Step S2323: Input the shared sample features into the initial repurchase probability model to obtain the repurchase rate corresponding to the sample data output by the initial repurchase probability model.
[0087] In some implementations, after the initial probability model obtains shared sample features, these features can be input into the initial repurchase probability model to obtain the repurchase rate corresponding to the sample data output by the initial repurchase probability model. Here, the repurchase probability corresponding to the sample data can be understood as the probability that a user who normally uses the target brand will become a repeat customer.
[0088] Steps S2322 and S2323 can be executed simultaneously or at different times, and this is not limited here.
[0089] Step S2324: Obtain the replacement and repurchase rate corresponding to the sample data based on the replacement rate and repurchase rate corresponding to the sample data.
[0090] In some implementations, after the initial probability model obtains the replacement rate and repurchase rate corresponding to the sample data, the replacement and repurchase rate corresponding to the sample data can be obtained based on the replacement rate and repurchase rate corresponding to the sample data.
[0091] The initial probability model can be based on the joint probability distribution formula: p(y=1, z=1|x) = p(y=1|x) × p(z=1|y=1, x), to obtain the replacement and repurchase rate. Here, p(y=1, z=1|x) represents the replacement rate corresponding to the sample data, p(y=1|x) represents the replacement rate corresponding to the sample data, and p(z=1|y=1, x) represents the repurchase rate corresponding to the sample data. Here, y represents the replacement label (y=1 for replacement, y=0 for no replacement); x represents the users who normally use the phone, i.e., the sample objects; z represents the repurchase label (z=1 for repurchase, z=0 for no repurchase). Furthermore, after obtaining the replacement rate and repurchase rate corresponding to the sample data, the initial probability model can calculate the product of the replacement rate and the repurchase rate, and use this product as the replacement and repurchase rate corresponding to the sample data.
[0092] Step S2325: Based on the device replacement rate, repurchase rate, and device replacement and repurchase rate corresponding to the sample data, the sample labels corresponding to each sample object, and the preset loss function, train the initial probability model to obtain the trained probability model.
[0093] In some implementations, after an electronic device obtains the replacement and repurchase rates corresponding to the sample data, it can train an initial probability model based on the replacement rate, repurchase rate, and sample label of each sample object, as well as a preset loss function, to obtain a trained probability model.
[0094] In some implementations, the electronic device may have a preset loss function, which may be a joint loss function designed based on the initial probability model of the multi-task learning architecture, a noise-robust loss function, or a label noise technique.
[0095] For example, the initial probability model may include two auxiliary prediction tasks that output the switching rate and repurchase rate respectively. These two auxiliary prediction tasks correspond to different models, and it can be understood that the output of each auxiliary prediction task has a loss value. Furthermore, in the process of obtaining the trained probability model, a joint loss function can be defined based on the switching and repurchase rate acquisition task and the switching rate acquisition task: Where, L(θ) cvr θ ctr ) represents the total loss value of the two auxiliary prediction tasks, l(y) i f(x) i ;θ ctr )) represents the loss value of obtaining the task based on the chance of switching, l(y) i &zi f(x) i ;;θ ctr )×f(x i ;θ cvr )) represents the loss value for the task of obtaining the replacement and repurchase rate; where N represents the number of sample objects, i represents a sample object, y represents the label of whether the sample data has been replaced, z represents the label of whether the sample data has been repurchased, x represents a sample data, and f(x) i ;θ ctr f(x) represents the switching rate corresponding to the sample data. i ;θ cvr The number ) represents the repurchase rate corresponding to the sample data.
[0096] It is understandable that the input sample data for the two losses corresponding to the two auxiliary prediction tasks are the same, but the sample labels corresponding to each sample data are different. That is, the same sample data needs to be input into two task modules, and the value of the sample label corresponding to the sample data in each task module may be 1 or 0.
[0097] Furthermore, the electronic device can use the joint loss function as the preset loss function, and perform backpropagation on the entire initial probability model based on the total loss value obtained from the preset loss function to update the parameters included in the initial probability model. It can also update the standardized parameters to obtain the trained probability model.
[0098] Optionally, the electronic device can use the Adam optimization function to backpropagate the entire network of the initial probability model based on the prediction results of the preset number of sample data after using the initial probability model to make predictions. Alternatively, the electronic device can train the initial probability model based on a preset learning rate and repeat steps S2321-S2325 in multiple training cycles until the joint loss function fully converges (e.g., the total loss value is less than or equal to the loss threshold, the training cycle reaches the cycle threshold, etc.) to obtain the trained probability model.
[0099] For example, the preset number is 1024, and the learning probability of the initial probability model is 0.02. The electronic device can use the Adam optimization function to backpropagate the initial probability model based on the prediction results of the batch size of 1024 and the initial learning rate of 0.02. The learning rate is halved after every two training cycles to shorten the gradient of the model, thereby improving the accuracy of the model and the efficiency of model training.
[0100] In some implementations, step S2325 may include steps S3251-S3252.
[0101] Step S3251: Input the switching rate, repurchase rate, and switching and repurchase rate corresponding to the sample data, as well as the sample label corresponding to each sample object, into the preset loss function to obtain the joint loss value.
[0102] In some implementations, during the process of training the initial probability model to obtain the trained probability model, the switching rate, repurchase rate, and switching and repurchase rate of the sample data, as well as the sample label of each sample object, can be input into a preset loss function, such as the logarithmic loss function or the cross-entropy loss function, to obtain the joint loss value.
[0103] Step S3252: If the joint loss value is greater than the loss threshold, update the parameters of the initial probability model and return to perform training on the initial probability model until the joint loss value is less than or equal to the loss threshold, and obtain the trained probability model.
[0104] In some implementations, a loss threshold is preset in the electronic device. After obtaining the joint loss value, the electronic device can compare the joint loss value with the loss threshold. If the joint loss value is greater than the loss threshold, the parameters of the initial probability model are updated, and the training of the initial probability model is returned until the joint loss value is less than or equal to the loss threshold, thus obtaining the trained probability model.
[0105] The process of returning to train the initial probability model can include: inputting sample data into the initial shared embedding layer to obtain the shared sample features output by the initial shared embedding layer; inputting the shared sample features into the initial switching probability model to obtain the switching rate corresponding to the sample data output by the initial switching probability model; inputting the shared sample features into the initial repurchase probability model to obtain the repurchase rate corresponding to the sample data output by the initial repurchase probability model; obtaining the switching and repurchase rate corresponding to the sample data based on the switching rate and repurchase rate corresponding to the sample data; and inputting the switching rate, repurchase rate, and repurchase rate corresponding to the sample data, along with the sample labels corresponding to each sample object, into a preset loss function to obtain a joint loss value until the joint loss value is less than or equal to a loss threshold, thus obtaining the trained probability model.
[0106] Step S240: Determine the target push object from the plurality of push objects based on the replacement and repurchase rate.
[0107] In some implementations, after obtaining the replacement and repurchase rates corresponding to the raw data, the electronic device can determine the target push object from multiple push objects based on the replacement and repurchase rates. Specifically, the electronic device can use the replacement and repurchase rates corresponding to each piece of raw data output by a trained probability model as a score for the push object, and arrange the push objects in descending order of score, selecting the top push object from high to low as the target push object for content delivery.
[0108] In some implementations, step S240 may include steps S2410-S2420.
[0109] Step S2410: Arrange the multiple push objects in descending order of their respective original data corresponding to device replacement and repurchase rates to obtain a push object list.
[0110] In some implementations, the electronic device can arrange multiple push targets in descending order of their respective original data corresponding to replacement and repurchase rates to obtain a push target list, and store the obtained push target list in the database of the electronic device.
[0111] Step S2420: Based on preset push conditions, determine the target push object from the list of push objects.
[0112] In some implementations, the electronic device may have preset push conditions. Furthermore, after obtaining the list of push objects, the electronic device may determine the target push object from the list of push objects based on the preset push conditions.
[0113] The preset push conditions may include content distribution budget, distribution scope ratio, etc. For example, the preset push condition is content distribution budget. Furthermore, the electronic device can determine the number of content items to be distributed based on the content distribution budget. Further, the electronic device can sort the push object list from high to low and determine the push objects with the same number of content items as the target push objects.
[0114] In some implementations, step S2420 may include steps S2421-S2422.
[0115] Step S2421: Determine the push range based on the preset push conditions.
[0116] In some implementations, the electronic device is pre-set with preset push conditions. If the preset push condition is content delivery budget, the electronic device can determine the number of content items to be delivered based on the preset push condition and use the number of content items to be delivered as the push range. If the preset push condition is a delivery range ratio, the electronic device can use the delivery range ratio as the push range. For example, if the delivery range ratio is 30%, the electronic device can determine the push range to be 30% of the push objects in the push object list.
[0117] Step S2422: Determine the push objects in the push object list that are within the push range as the target push objects.
[0118] In some implementations, after determining the push range, the electronic device can identify the push objects within the push range from the push object list as target push objects. In other implementations, considering operational strategies, the electronic device can select the top push objects within the push range from the top of the push object list as target push objects to improve the usability of content push and enhance the user experience.
[0119] Step S250: Push the content to the target push object.
[0120] In this embodiment, after the electronic device determines the target push object, it can push content to the target push object. In some implementations, the electronic device can be preset with a preset time length. After the electronic device pushes the content to the target push object for the preset time length, it can also return to execute steps S210-S250 to automatically generate a new push object list based on the continuously updated push object dataset, and determine the target push object for content delivery, thereby improving the accuracy and real-time performance of content push.
[0121] The content push method provided in one embodiment of this application is compared to... Figure 1 The content push method shown in this embodiment can further preprocess the raw data before inputting it into the trained probability model to obtain the device replacement and repurchase rate corresponding to the raw data output by the trained probability model. The preprocessing includes at least one of outlier handling and missing value imputation, thereby improving the efficiency of obtaining the device replacement and repurchase rate corresponding to the push object through the multi-task trained probability model. Based on the device replacement and repurchase rate of the push object, the target push object of the content can be obtained, thereby improving the accuracy of content push, reducing the cost of content push, and improving the user experience.
[0122] Please see Figure 6 , Figure 6This diagram illustrates a module block diagram of a content push device according to an embodiment of this application. The content push device 200 is applied to the aforementioned electronic device, and will be discussed below. Figure 6 The process is described in detail below. The content push device 200 includes: a push target dataset acquisition module 210, a device replacement and repurchase rate acquisition module 220, a target push target determination module 230, and a content push module 240, wherein:
[0123] The push object dataset acquisition module 210 is used to acquire the push object dataset, wherein the push object dataset includes multiple push objects and the original data corresponding to each push object.
[0124] The replacement and repurchase rate acquisition module 220 is used to input the original data into a trained probability model to obtain the replacement and repurchase rate corresponding to the original data output by the trained probability model, wherein the replacement and repurchase rate is obtained based on the replacement rate and the repurchase rate corresponding to the original data.
[0125] The target push object determination module 230 is used to determine the target push object from the plurality of push objects based on the replacement and repurchase rate.
[0126] The content push module 240 is used to push content to the target push object.
[0127] Furthermore, the target push object determination module 230 may include: a push object list acquisition unit and a target push object determination submodule, wherein:
[0128] The push object list acquisition module is used to arrange the multiple push objects in descending order of their respective original data corresponding to device replacement and repurchase rates, and obtain a push object list.
[0129] The target push object determination submodule is used to determine the target push object from the list of push objects based on preset push conditions.
[0130] Furthermore, the target push object determination submodule may include: a push range determination unit and a target push object determination subunit, wherein:
[0131] The push range determination unit is used to determine the push range based on the preset push conditions.
[0132] The target push object determination subunit is used to determine the push objects in the push object list that are within the push range as the target push objects.
[0133] Further, the trained probability model includes a trained shared embedding layer, a trained device replacement probability model, and a trained repurchase probability model. The device replacement and repurchase rate acquisition module 220 may include: a shared detection feature acquisition unit, a device replacement rate acquisition unit, a repurchase rate acquisition unit, and a device replacement and repurchase rate acquisition subunit, wherein:
[0134] A shared detection feature acquisition unit is used to input the original data into the trained shared embedding layer to obtain the shared detection features output by the trained shared embedding layer.
[0135] The device switching rate acquisition unit is used to input the shared detection features into the trained device switching probability model to obtain the device switching rate corresponding to the original data output by the trained device switching probability model.
[0136] The repurchase rate detection unit is used to input the shared detection features into the trained repurchase probability model to obtain the repurchase rate corresponding to the original data output by the trained repurchase probability model.
[0137] The device replacement and repurchase rate detection subunit is used to obtain the device replacement and repurchase rate corresponding to the original data based on the device replacement rate and the repurchase rate corresponding to the original data.
[0138] Furthermore, before inputting the raw data into the trained probability model to obtain the device switching and repurchase rates corresponding to the raw data output by the trained probability model, the content push device 200 may further include: a preprocessing module, wherein:
[0139] A preprocessing module is used to preprocess the original data to obtain preprocessed original data, wherein the preprocessing includes at least one of outlier handling and missing value imputation.
[0140] Furthermore, before inputting the raw data into the trained probability model to obtain the device replacement and repurchase rate corresponding to the raw data output by the trained probability model, the content push device 200 may further include: a sample push object dataset acquisition module and a model training module, wherein:
[0141] The sample push object dataset acquisition module is used to acquire the sample push object dataset, wherein the sample push object dataset includes multiple sample objects, sample data corresponding to each sample object, and sample tags corresponding to each sample object, wherein the sample tags include at least one of the following: device replacement tag, no device replacement tag, repeat purchase tag, and no repeat purchase tag.
[0142] The model training module is used to train the initial probability model based on the sample push object dataset to obtain the trained probability model.
[0143] Furthermore, the initial probability model includes an initial shared embedding layer, an initial device replacement probability model, and an initial repurchase probability model. The model training module may include: a shared sample feature acquisition unit, a sample device replacement rate acquisition unit, a sample repurchase rate acquisition unit, a sample device replacement and repurchase rate acquisition subunit, and a model training subunit, wherein:
[0144] A shared sample feature acquisition unit is used to input the sample data into the initial shared embedding layer and obtain the shared sample features output by the initial shared embedding layer.
[0145] The sample switching rate acquisition unit is used to input the shared sample features into the initial switching probability model to obtain the switching rate corresponding to the sample data output by the initial switching probability model.
[0146] The sample repurchase rate acquisition unit is used to input the shared sample features into the initial repurchase probability model to obtain the repurchase rate corresponding to the sample data output by the initial repurchase probability model.
[0147] The sample device replacement and repurchase rate acquisition subunit is used to obtain the device replacement and repurchase rate corresponding to the sample data based on the device replacement rate and the repurchase rate corresponding to the sample data.
[0148] The model training subunit is used to train the initial probability model based on the switching rate, repurchase rate, and switching and repurchase rate of the sample data, the sample labels of each sample object, and a preset loss function, so as to obtain the trained probability model.
[0149] Furthermore, the model training subunit may include: a joint loss value acquisition unit and a model parameter update unit, wherein:
[0150] The joint loss value acquisition unit is used to input the switching rate, repurchase rate, and switching and repurchase rate corresponding to the sample data, as well as the sample label corresponding to each sample object, into the preset loss function to obtain the joint loss value.
[0151] The model parameter update unit is used to update the parameters of the initial probability model if the joint loss value is greater than the loss threshold, and return to perform training on the initial probability model until the joint loss value is less than or equal to the loss threshold, thereby obtaining the trained probability model.
[0152] Furthermore, the shared sample feature acquisition unit may include: a standard sample feature unit and a shared sample feature acquisition subunit, wherein:
[0153] The standard sample feature unit is used to standardize the sample data based on standardization parameters to obtain standard sample features.
[0154] The shared sample feature acquisition subunit is used to input the standard sample features into the initial shared embedding layer to obtain the shared sample features output by the initial shared embedding layer.
[0155] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described device and module can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.
[0156] In the several embodiments provided in this application, the coupling between modules can be electrical, mechanical, or other forms of coupling.
[0157] Furthermore, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module. The integrated modules described above can be implemented in hardware or as software functional modules.
[0158] Please see Figure 7 This document illustrates a structural block diagram of an electronic device according to an embodiment of this application. The electronic device 100 can be a smartphone, tablet computer, e-reader, or other electronic device capable of running applications. The electronic device 100 in this application may include one or more of the following components: a processor 110, a memory 120, and one or more applications, wherein the one or more applications can be stored in the memory 120 and configured to be executed by one or more processors 110, and the one or more applications are configured to perform the methods described in the foregoing method embodiments.
[0159] The processor 110 may include one or more processing cores. The processor 110 connects to various parts within the electronic device 100 using various interfaces and lines, and performs various functions and processes data by running or executing instructions, programs, code sets, or instruction sets stored in the memory 120, and by calling data stored in the memory 120. Optionally, the processor 110 may be implemented using at least one hardware form of Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), or Programmable Logic Array (PLA). The processor 110 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the content to be displayed; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor 110 and may be implemented separately using a communication chip.
[0160] The memory 120 may include random access memory (RAM) or read-only memory (ROM). The memory 120 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 120 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for implementing at least one function (such as touch functionality, sound playback functionality, image playback functionality, etc.), and instructions for implementing the various method embodiments described below. The data storage area may also store data created by the electronic device 100 during use (such as phonebook data, audio and video data, chat log data, etc.).
[0161] Please see Figure 8 This diagram illustrates a structural block diagram of a computer-readable storage medium provided in an embodiment of this application. The computer-readable medium 300 stores program code that can be invoked by a processor to execute the methods described in the above method embodiments.
[0162] The computer-readable storage medium 300 may be an electronic memory such as flash memory, EEPROM (Electrically Erasable Programmable Read-Only Memory), EPROM, hard disk, or ROM. Optionally, the computer-readable storage medium 300 includes a non-transitory computer-readable storage medium. The computer-readable storage medium 300 has storage space for program code 310 that performs any of the method steps described above. This program code can be read from or written to one or more computer program products. The program code 310 may be compressed, for example, in a suitable form.
[0163] In summary, the content push method, apparatus, electronic device, and storage medium provided in this application improve the accuracy of content push and enhance the user experience by acquiring a push object dataset, which includes multiple push objects and their corresponding original data; inputting the original data into a trained probability model to obtain the device replacement and repurchase rate output by the trained probability model corresponding to the original data, wherein the device replacement and repurchase rate is obtained based on the device replacement rate and repurchase rate corresponding to the original data; determining the target push object from multiple push objects based on the device replacement and repurchase rate; and pushing content to the target push object. This improves the accuracy of content push and enhances the user experience by obtaining the device replacement and repurchase rate corresponding to the push object through a multi-task trained probability model and determining the target push object based on the device replacement and repurchase rate of the push object.
[0164] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.
Claims
1. A content push method, characterized in that, The method includes: Obtain a push object dataset, wherein the push object dataset includes multiple push objects and the original data corresponding to each push object; The original data is input into a trained probability model to obtain the replacement and repurchase rate corresponding to the original data, which is output by the trained probability model. The replacement and repurchase rate is obtained based on the replacement rate and the repurchase rate corresponding to the original data. The target push object is determined from the plurality of push objects based on the replacement and repurchase rates; Push the content to the target recipient; The trained probability model includes a trained shared embedding layer, a trained device replacement probability model, and a trained repurchase probability model. The step of inputting the raw data into the trained probability model to obtain the device replacement and repurchase rate corresponding to the raw data output by the trained probability model includes: The original data is input into the trained shared embedding layer to obtain the shared detection features output by the trained shared embedding layer; The shared detection features are input into the trained phone switching probability model to obtain the phone switching rate corresponding to the original data output by the trained phone switching probability model, wherein the phone switching rate is used to characterize the probability that a user who normally uses the target brand mobile phone will become a phone switching user. The shared detection features are input into the trained repurchase probability model to obtain the repurchase rate corresponding to the original data output by the trained repurchase probability model, wherein the repurchase rate is used to characterize the probability that a user who normally uses the target brand mobile phone will become a repurchase user. Based on the replacement rate and repurchase rate corresponding to the original data, the replacement and repurchase rate corresponding to the original data is obtained.
2. The method according to claim 1, characterized in that, The step of determining the target push object from the plurality of push objects based on the replacement and repurchase rate includes: Arrange the multiple push objects in descending order of their respective original data corresponding to device replacement and repurchase rates to obtain a push object list; Based on preset push conditions, the target push object is determined from the list of push objects.
3. The method according to claim 2, characterized in that, The step of determining the target push object from the list of push objects based on preset push conditions includes: The push range is determined based on the preset push conditions; The push objects that are within the push range in the push object list are determined as the target push objects.
4. The method according to claim 1, characterized in that, Before inputting the raw data into the trained probability model to obtain the replacement and repurchase rate corresponding to the raw data output by the trained probability model, the method further includes: The original data is preprocessed to obtain preprocessed original data, wherein the preprocessing includes at least one of outlier handling and missing value imputation.
5. The method according to claim 1, characterized in that, Before inputting the raw data into the trained probability model to obtain the replacement and repurchase rate corresponding to the raw data output by the trained probability model, the method further includes: Obtain a sample push object dataset, wherein the sample push object dataset includes multiple sample objects, sample data corresponding to each sample object, and sample tags corresponding to each sample object, wherein the sample tags include at least one of device replacement tag, no device replacement tag, repurchase tag, and no repurchase tag; The initial probability model is trained based on the sample push object dataset to obtain the trained probability model.
6. The method according to claim 5, characterized in that, The initial probability model includes an initial shared embedding layer, an initial device replacement probability model, and an initial repeat purchase probability model. Training the initial probability model based on the sample push object dataset to obtain the trained probability model includes: The sample data is input into the initial shared embedding layer to obtain the shared sample features output by the initial shared embedding layer; The shared sample features are input into the initial device replacement probability model to obtain the device replacement rate corresponding to the sample data output by the initial device replacement probability model; The shared sample features are input into the initial repurchase probability model to obtain the repurchase rate corresponding to the sample data output by the initial repurchase probability model; The replacement and repurchase rate corresponding to the sample data is obtained based on the replacement rate and repurchase rate corresponding to the sample data. Based on the device replacement rate, repurchase rate, and replacement and repurchase rate corresponding to the sample data, the sample labels corresponding to each sample object, and the preset loss function, the initial probability model is trained to obtain the trained probability model.
7. The method according to claim 6, characterized in that, The step of training the initial probability model based on the device replacement rate, repurchase rate, and both replacement and repurchase rate corresponding to the sample data, the sample labels corresponding to each sample object, and a preset loss function to obtain the trained probability model includes: Input the device replacement rate, repurchase rate, and both replacement and repurchase rate corresponding to the sample data, as well as the sample label corresponding to each sample object, into the preset loss function to obtain the joint loss value; If the joint loss value is greater than the loss threshold, the parameters of the initial probability model are updated, and training of the initial probability model is resumed until the joint loss value is less than or equal to the loss threshold, thus obtaining the trained probability model.
8. The method according to claim 6, characterized in that, The step of inputting the sample data into the initial shared embedding layer to obtain the shared sample features output by the initial shared embedding layer includes: The sample data is standardized based on the standardization parameters to obtain standard sample features; The standard sample features are input into the initial shared embedding layer to obtain the shared sample features output by the initial shared embedding layer.
9. A content push device, characterized in that, The device includes: The push object dataset acquisition module is used to acquire the push object dataset, wherein the push object dataset includes multiple push objects and the original data corresponding to each push object; The replacement and repurchase rate acquisition module is used to input the original data into a trained probability model to obtain the replacement and repurchase rate corresponding to the original data output by the trained probability model, wherein the replacement and repurchase rate is obtained based on the replacement rate and the repurchase rate corresponding to the original data. The target push object determination module is used to determine the target push object from the plurality of push objects based on the replacement and repurchase rate; The content push module is used to push content to the target push object; The trained probability model includes a trained shared embedding layer, a trained phone replacement probability model, and a trained repurchase probability model. The phone replacement and repurchase rate acquisition module is further configured to: input the original data into the trained shared embedding layer to obtain shared detection features output by the trained shared embedding layer; input the shared detection features into the trained phone replacement probability model to obtain the phone replacement rate corresponding to the original data output by the trained phone replacement probability model, wherein the phone replacement rate represents the probability that a user normally using the target brand phone will become a phone replacement user; and input the shared detection features into the trained repurchase probability model to obtain the repurchase rate corresponding to the original data output by the trained repurchase probability model, wherein the repurchase rate represents the probability that a user normally using the target brand phone will become a repurchase user. Based on the replacement rate and repurchase rate corresponding to the original data, the replacement and repurchase rate corresponding to the original data is obtained.
10. An electronic device, characterized in that, include: One or more processors; Memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications being configured to perform the method as described in any one of claims 1-8.
11. A computer-readable storage medium, characterized in that, The computer-readable storage medium contains program code that can be invoked by a processor to execute the method as described in any one of claims 1-8.