Click rate estimation model generation method and device, electronic equipment and storage medium

Abstract

The invention relates to a click rate prediction model generation method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining a model training set, the model training set comprises a plurality of samples, each sample comprises user information, scene identifiers and click information, and the scene identifiers are used for representing different information release scenes; generating a first loss function according to the user information, and generating a second loss function and a third loss function according to the scene identifier and the click information; and determining a target loss function based on the first loss function, the second loss function and the third loss function, and training the click rate estimation model by using the target loss function. According to the method and the device, the target loss function is obtained by utilizing the user information, the scene identifier and the click information, and the finally obtained click rate prediction model can be more accurate based on the target loss function.

Description

Technical Field

[0001] This disclosure relates to the field of Internet technology, and in particular to a method, apparatus, electronic device, and storage medium for generating a click-through rate prediction model. Background Technology

[0002] With the rapid development of internet technology, click-through rate (CTR) has become one of the core criteria for traffic allocation on internet platforms. Taking advertising on internet platforms as an example, based on users' historical click information, CTR prediction models analyze user interests and behaviors to predict future click behavior, thereby promoting accurate ad placement. It is evident that CTR prediction models play a crucial role in the CTR prediction process. Therefore, how to obtain better CTR prediction models is a pressing technical problem that needs to be solved. Summary of the Invention

[0003] To overcome the problems existing in related technologies, this disclosure provides a method, apparatus, electronic device, and storage medium for generating a click-through rate prediction model.

[0004] According to a first aspect of the present disclosure, a method for generating a click-through rate (CTR) prediction model is provided, comprising:

[0005] Obtain a model training set, which includes multiple samples. Each sample includes user information, scene identifier, and click information. The scene identifier is used to represent different information delivery scenarios.

[0006] A first loss function is generated based on the user information, and a second and a third loss function are generated based on the scene identifier and the click information;

[0007] The target loss function is determined based on the first loss function, the second loss function, and the third loss function, and the click-through rate prediction model is trained using the target loss function.

[0008] Optionally, the user information includes user profile information, user behavior characteristic information, and user purchase characteristic information, and the click-through rate prediction model includes a main model;

[0009] The step of generating the first loss function based on the user information includes:

[0010] The user profile information, user behavior feature information, and user purchase feature information are input into the main model to obtain the estimated click-through rate of multiple samples;

[0011] The first loss function is determined based on the estimated click-through rates of the multiple samples.

[0012] Optionally, the click-through rate prediction model further includes a first contrastive learning module and a second contrastive learning module;

[0013] The step of generating the second loss function and the third loss function based on the scene identifier and the click information includes:

[0014] The scene identifier and the estimated click-through rate are input into the first contrastive learning module to obtain the second loss function. The first contrastive learning module is used to obtain the differences between different scenes.

[0015] The click information and the estimated click-through rate are input into the second contrastive learning module to obtain the third loss function. The click information of each sample is represented by a label, and the second contrastive learning module is used to obtain the differences between different labels.

[0016] Optionally, the step of inputting the scene identifier and the estimated click-through rate into the first contrastive learning module to obtain the second loss function includes:

[0017] A first anchor sample is determined, and the scene identifier of the first anchor sample is compared with the scene identifier of each other sample to filter out multiple first target samples, wherein the scene identifier of the first target sample is the same as the scene identifier of the first anchor sample.

[0018] The second loss function is determined based on multiple first target samples.

[0019] Optionally, the second loss function is determined based on the first mask function, the similarity score between every two samples in the plurality of samples, the temperature coefficient, and the hyperparameters, wherein the first mask function is used to determine whether the scenes between every two samples in the plurality of samples are consistent.

[0020] Optionally, inputting the click information and the estimated click-through rate into the second contrastive learning module to obtain the third loss function includes:

[0021] A second anchor sample is determined, and the click information of the second anchor sample is compared with the scene information of each other sample to filter out multiple second target samples. The click information of the second target samples is the same as the click information of the second anchor sample.

[0022] The third loss function is determined based on multiple second target samples.

[0023] Optionally, the third loss function is determined based on the second mask function, the similarity score between every two samples in the plurality of samples, the temperature coefficient, and the hyperparameters, wherein the second mask function is used to determine whether the labels between every two samples in the plurality of samples are consistent.

[0024] Optionally, the method further includes:

[0025] Get the data to be selected;

[0026] The data to be selected is input into the click-through rate prediction model to obtain the target audience. The target audience includes a first group and a second group, and the scene identifiers of the first group and the second group are different.

[0027] Send promotional text messages to the target audience.

[0028] According to a second aspect of the present disclosure, an apparatus for generating a click-through rate (CTR) prediction model is provided, comprising:

[0029] The acquisition module is configured to acquire a model training set, which includes multiple samples. Each sample includes user information, scene identifier, and click information. The scene identifier is used to represent different information delivery scenarios.

[0030] The generation module is configured to generate a first loss function based on the user information, and to generate a second loss function and a third loss function based on the scene identifier and the click information;

[0031] The determination module is configured to determine a target loss function based on the first loss function, the second loss function, and the third loss function, and to train the click-through rate prediction model using the target loss function.

[0032] According to a third aspect of the present disclosure, an electronic device is provided, comprising:

[0033] processor;

[0034] Memory used to store processor-executable instructions;

[0035] The processor is configured as follows:

[0036] Obtain a model training set, which includes multiple samples. Each sample includes user information, scene identifier, and click information. The scene identifier is used to represent different information delivery scenarios.

[0037] A first loss function is generated based on the user information, and a second and a third loss function are generated based on the scene identifier and the click information;

[0038] The target loss function is determined based on the first loss function, the second loss function, and the third loss function, and the click-through rate prediction model is trained using the target loss function.

[0039] According to a fourth aspect of the present disclosure, a computer-readable storage medium is provided that stores computer program instructions thereon, which, when executed by a processor, implement the steps of the click-through rate prediction model generation method provided in the first aspect of the present disclosure.

[0040] According to a fifth aspect of the present disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the click-through rate prediction model generation method provided in the first aspect of the present disclosure.

[0041] This disclosure utilizes scene identifiers and click-through rate (CTR) information to obtain a second loss function and a third loss function. The second and third loss functions are then combined with the first loss function to determine the target loss function. Since the target loss function fully considers different scene identifiers, the CTR prediction model can not only learn the differences between scenes but also does not require separate modeling for multiple scenes. This reduces the computational and maintenance costs associated with separate modeling to some extent. Furthermore, by introducing CTR information, this disclosure can achieve CTR prediction.

[0042] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit this disclosure. Attached Figure Description

[0043] The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the present disclosure and, together with the description, serve to explain the principles of the present disclosure.

[0044] Figure 1 This is a flowchart illustrating a method for generating a click-through rate prediction model according to an exemplary embodiment.

[0045] Figure 2 This is a flowchart illustrating another method for generating a click-through rate prediction model according to an exemplary embodiment.

[0046] Figure 3 This is an example diagram illustrating the structure of a click-through rate (CTR) prediction model in another method for generating a CTR prediction model according to an exemplary embodiment.

[0047] Figure 4 This is an example diagram illustrating scene-aware contrastive learning in another method for generating a click-through rate prediction model, according to an exemplary embodiment.

[0048] Figure 5 This is an example diagram illustrating multi-label contrastive learning in another method for generating a click-through rate prediction model, according to an exemplary embodiment.

[0049] Figure 6This is an example diagram illustrating the prediction of a selected audience using a click-through rate (CTR) prediction model in another method for generating a CTR prediction model according to an exemplary embodiment.

[0050] Figure 7 This is a block diagram illustrating a click-through rate prediction model generation apparatus according to an exemplary embodiment.

[0051] Figure 8 This is a block diagram illustrating an electronic device according to an exemplary embodiment.

[0052] Figure 9 This is a block diagram illustrating a chip system according to an exemplary embodiment. Detailed Implementation

[0053] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0054] The embodiments described in the following examples of this disclosure are not representative of all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0055] It should be noted that all actions involving the acquisition of signals, information, or data in this disclosure are carried out in compliance with the relevant data protection laws and policies of the country where the location is situated, and with authorization from the owner of the relevant device.

[0056] With the development of artificial intelligence technology, multi-scenario click-through rate (CTR) prediction in the recommendation field has received increasing attention, especially in the advertising field. This prediction operation, known as multi-scenario CTR prediction, refers to estimating the click-through rate (CTR) under different scenarios / businesses to obtain the predicted click-through rate (PCTR). Previously, multi-scenario modeling was required, which involved uniformly modeling when predicting the same task under different scenarios / businesses. However, modeling each scenario separately leads to problems such as wasted computational resources, increased maintenance costs, and inaccurate predictions in long-tail scenarios.

[0057] In the recommendation field, there is a significant need to predict CTR (Click-Through Rate) or CVR (Conversion Rate) in similar scenarios. For example, in SMS marketing, CTR prediction is required for different audiences during major promotions and new product launches. The main approach is to build a separate model for each scenario, but this results in too many online models, increasing computational and maintenance costs. Furthermore, in scenarios with lower traffic volumes, there is limited user behavior data, making it difficult for independent models to converge to a satisfactory solution during training.

[0058] To address these issues, multi-scenario modeling has emerged, where a single model is used across multiple scenarios to reduce computational and maintenance costs. However, multi-scenario modeling suffers from significant differences between scenarios, including variations in sample distribution, bias in training scenarios with larger sample sizes, and challenges in learning the commonalities and differences between scenarios. Furthermore, multi-scenario click-through rate prediction also suffers from sample selection bias and highly sparse data.

[0059] Figure 1 This is a flowchart illustrating a method for generating a click-through rate prediction model according to an exemplary embodiment, such as... Figure 1 As shown, the method may include the following steps.

[0060] In step S11, the model training set is obtained.

[0061] In some implementations, the model training set can be used to train the click-through rate prediction model. This set may include multiple samples, each of which may include user information, scene identifiers, and click information. The user information may include user profile information, user behavior characteristic information, and user purchase characteristic information.

[0062] User profiles are descriptions of a user's characteristics built upon data such as their online behavior, interests, preferences, and social relationships. User profiles can be used to understand a user's personalized needs and behavioral habits, thereby enabling personalized recommendations and targeted marketing. Specifically, user profiles may include information such as the user's basic information, interests and preferences, behavioral habits, spending power, social relationships, emotional attitudes, and health status.

[0063] For example, a user's basic information may include basic attributes such as gender, age, geographical location, occupation, and education level, which form the basis for building a user profile; a user's interests and preferences describe the degree of liking for specific topics, fields, or themes, and this embodiment of the disclosure can infer a user's interests and preferences based on data such as browsing history, search records, and click behavior; a user's behavioral habits describe a user's behavior and habits on the Internet, including information such as access frequency, active time periods, and devices used; a user's spending power describes a user's purchasing power and consumption preferences, which can be inferred from data such as purchase history and consumption behavior; a user's social relationships describe a user's relationship network and social behavior in social networks, including information such as friend relationships, following lists, and follower lists; a user's emotional attitude describes a user's emotional tendency towards specific content or events, including emotional expressions such as liking, disliking, joy, anger, sorrow, and happiness; and a user's health status describes a user's physical health status and health management needs, including information such as physical examination records, health problems, and health habits.

[0064] Optionally, user profile information may also include other characteristics, such as family situation, vehicle ownership, and hobbies, depending on different business needs.

[0065] User behavior characteristics can be records and patterns of a user's behavior over a period of time, and these characteristics can be historical user behavior characteristics. Historical user behavior characteristics can include click history, add-to-cart history, browsing history, purchase history, favorites history, comment history, rating history, and sharing history, etc.

[0066] User purchase characteristic information can be a record and pattern of a user's purchase behavior over a period of time, and this user purchase characteristic information can be a user's historical purchase characteristics. Here, user purchase characteristic information can include purchase frequency, purchase amount, purchase time, purchase type preference, purchase brand preference, and purchase channel preference, etc.

[0067] In other implementations, scenario identifiers can be used to represent different information delivery scenarios; that is, different information delivery scenarios correspond to different scenario identifiers. Information delivery scenarios can be those where relevant information or content is pushed to users at appropriate times, locations, and in appropriate ways based on factors such as user behavior, preferences, and environment. These scenarios can include user lifecycle stages, specific time points (such as holidays or promotional days), geographical location, social influence, user behavior, and so on.

[0068] For example, information delivery scenarios may include promotional days, new product launch days, user behavior triggers, seasonal events, geolocation information, social influence, user lifecycle stages, special events, and inventory management.

[0069] Among these features, Promotion Day can push relevant promotional information or special offers to users based on specific promotional activities or holidays; New Product Launch Day can push relevant product information or recommend new products to users when they are launched or released; User Behavior Triggers can recommend relevant content to users based on user behavior triggers, such as browsing history, purchase records, and search behavior; Seasonal Activities can adjust recommended content according to different seasons or holidays, such as Christmas and Chinese New Year; Geographical Location Information can push relevant content to users based on the characteristics of their location, such as weather and local popular events; Social Influence can recommend content related to users' social circles based on their social network relationships or influence, such as friends' purchasing behavior and social sharing; User Lifecycle Stage can recommend different types of content to users based on their current lifecycle stage, such as new user onboarding, active user maintenance, and churned user reactivation; Special Events can push relevant content based on specific themes or activities, such as limited-time discounts and holiday specials; Inventory Management can recommend promotional clearance items or best-selling products to users based on product inventory status.

[0070] For example, since attribution focuses on the group of people who make a purchase after clicking on the SMS message, this embodiment of the disclosure can model the SMS click-through rate group. The modeling objective is to determine the probability of clicking on the SMS message in scenario 1 and scenario 2, and to maximize the estimated score of the conversion group among the SMS click-through group.

[0071] In other implementations, click information can be a user's click on content in the recommendation system within a specific context, or a click after sending a text message. This click information can be used to measure the user's interest in and the attractiveness of the recommended content. Here, click information can be represented by labels; that is, the click information can be the labels of the model training set, and the labels of the model training set can be multi-labeled, with each sample having a corresponding label.

[0072] Here, the tags can include whether a user clicked, whether they purchased after clicking the SMS message, or whether they purchased without clicking the SMS message. "Whether a user clicked" refers to whether the user clicked on a recommended item in the recommendation system; "whether a user purchased after clicking the SMS message" refers to whether the user clicked on a link and made a purchase after receiving a recommendation message sent by the recommendation system via SMS; "whether a user purchased after not clicking the SMS message" refers to whether the user did not click on a link to make a purchase after receiving a recommendation message sent by the recommendation system via SMS, meaning the user purchased the item through other channels instead of using the link in the SMS message.

[0073] It should be noted that in this embodiment of the disclosure, whether or not a click is made can be the main label for estimating the click-through rate, while whether or not a purchase is made after clicking the SMS message and whether or not a purchase is made without clicking the SMS message can be auxiliary labels for estimating the click-through rate.

[0074] It should also be noted that, in addition to the above, click information processing may also include information such as click content, click time, click location, number of clicks, and user behavior after clicking the content.

[0075] Among these, the clicked content can be the specific recommended content clicked by the user, such as products, articles, videos, etc.; the click time can be the timestamp of the user's click on the content, which can be used to analyze the time patterns and trends of user activity; the click location can be the location information of the content clicked by the user on the recommendation system interface, which can be used to understand the user's browsing habits and preferences; the number of clicks can be the number of times the user clicks on a certain piece of content, which can be used to assess the popularity of the content and user interest; and the follow-up behavior after clicking on the content can be the user's behavior after clicking on the content, such as whether to continue browsing or whether to share it, which is used to analyze the user's conversion path and behavioral trajectory.

[0076] In step S12, a first loss function is generated based on user information, and a second loss function and a third loss function are generated based on scene identifiers and click information.

[0077] As an optional approach, after obtaining the model training set, embodiments of this disclosure can generate a first loss function based on the user information in the model training set, wherein the first loss function may be a cross-entropy loss function.

[0078] In this embodiment of the disclosure, the click-through rate (CTR) prediction model may include a main model. During the process of obtaining the first loss function based on user information, the user information may be input into the main model to obtain multiple predicted CTRs (PCTRs). Based on these PCTRs, the first loss function is determined.

[0079] Alternatively, embodiments of this disclosure can generate a second loss function and a third loss function based on scene identifiers and click information from the model training set. Specifically, embodiments of this disclosure can generate a second loss function based on scene identifiers and a third loss function based on click information. Furthermore, embodiments of this disclosure can combine the estimated click-through rate output by the main model and the scene identifiers to obtain the second loss function. Similarly, embodiments of this disclosure can also combine the estimated click-through rate output by the main model and the click information to obtain the third loss function.

[0080] In step S13, a target loss function is determined based on the first loss function, the second loss function, and the third loss function, and the click-through rate prediction model is trained using the target loss function.

[0081] As an alternative approach, after obtaining the first loss function, the second loss function, and the third loss function, embodiments of this disclosure can determine a target loss function based on the first loss function, the second loss function, and the third loss function, and use the target loss function to train a click-through rate prediction model.

[0082] For example, in embodiments of this disclosure, the first loss function, the second loss function, and the third loss function can be summed to obtain the target loss function, or the three loss functions can be weighted and summed to obtain the target loss function.

[0083] Optionally, embodiments of this disclosure can evaluate the performance of the click-through rate (CTR) prediction model based on a target loss function. That is, the target loss function can be used to evaluate the predictive performance of the CTR prediction model and measure how well the CTR prediction model fits the click behavior in the training data. By observing the magnitude of the target loss function, embodiments of this disclosure can intuitively understand the prediction accuracy and generalization ability of the CTR prediction model.

[0084] Optionally, embodiments of this disclosure can guide the optimization of the click-through rate (CTR) prediction model based on a target loss function. During the optimization process, embodiments of this disclosure can adjust the CTR prediction model parameters by minimizing the loss function, making the prediction results of the CTR prediction model as close as possible to the true labels. For example, through optimization algorithms such as gradient descent, embodiments of this disclosure can update the model parameters of the CTR prediction model, gradually reducing the target loss function, thereby improving the prediction accuracy of the CTR prediction model.

[0085] Furthermore, in click-through rate (CTR) prediction tasks, clicked samples are typically in the minority class, while non-clicked samples are in the majority class, leading to an imbalanced dataset. Embodiments of this disclosure can address this imbalance problem by utilizing a target loss function through weighting or other methods, thus enabling a fairer evaluation of the CTR prediction model's performance across different categories.

[0086] Optionally, according to the target loss function, embodiments of this disclosure can prevent the click-through rate prediction model from overfitting the training data, such as by penalizing the complexity of the click-through rate prediction model through regularization terms or other methods, making the click-through rate prediction model more generalizable, so that the click-through rate prediction model can also have good prediction performance for unseen data.

[0087] It should be noted that the embodiments disclosed herein can not only solve the problem of unified click-through rate modeling in multiple scenarios, but can also be extended to the field of multi-scenario click-through rate prediction in recommendation, search and other fields.

[0088] This disclosure embodiment can obtain a second loss function and a third loss function by utilizing scene identifiers and click-through rate information, and comprehensively determine the target loss function by combining the second loss function and the third loss function with the first loss function. Since the target loss function fully considers different scene identifiers, the click-through rate prediction model can not only learn the differences between scenes, but also does not need to model multiple scenes separately. To a certain extent, it can reduce the computational and maintenance costs caused by separate modeling. Furthermore, by introducing click-through rate information, this disclosure can achieve click-through rate prediction.

[0089] Figure 2 This is a flowchart illustrating another method for generating a click-through rate prediction model according to an exemplary embodiment, such as... Figure 2 As shown, the method may include the following steps.

[0090] In step S21, the model training set is obtained.

[0091] The specific implementation method of step S21 has been described in detail in the above embodiments, and will not be repeated here.

[0092] In step S22, a first loss function is generated based on user information, and a second loss function and a third loss function are generated based on scene identifiers and click information.

[0093] As described above, user information can include user profile information, user behavior feature information, and user purchase feature information. Furthermore, the click-through rate (CTR) prediction model can include a main model. In the process of generating the first loss function based on the user information, this embodiment can input the user profile information, user behavior feature information, and user purchase feature information into the main model to obtain the predicted CTR of multiple samples. Based on this, the first loss function is determined according to the predicted CTR of multiple samples.

[0094] In this embodiment of the disclosure, the structure of the main model can be as follows: Figure 3 As shown in 301, through Figure 3 The main model can be divided into an input layer, an encoding layer, an encoding processing layer, a model layer, and an output layer. The input layer can include user profile information, user behavior features, and user purchase features. The encoding layer maps each feature of the input layer into a multi-dimensional embedding vector, such as a 16-dimensional embedding vector. The encoding processing layer performs attention processing on the user behavior features, and then concatenates the attention-processed tensor with the embedding tensors of other features as input to the entire model.

[0095] The model layer can include multiple fully connected layers (FC). Multiple fully connected layers can be used to further learn the complex interaction relationships between features and extract higher-level feature representations, thereby better capturing the relationship between users and items and improving the performance of the recommendation system.

[0096] In this embodiment, a three-layer fully connected network can be used as the main model network structure. Alternatively, various other model structures can be used, such as XDeepFM (eXtreme Deep Factorization Machine) or DCNMix (Deep & Cross Network with Feature Mixing). The final fully connected layer (FC) can be the output of the model layer.

[0097] The output layer of the main model passes through a fully connected layer of size 1 to obtain the PCTR prediction score for that sample. In other words, the output layer in this embodiment can output the predicted click-through rate (PCTR) for multiple samples.

[0098] It should be noted that the embodiments of this disclosure can train the click-through rate prediction model in batches, where each batch can include multiple samples. The output layer of the main model can output the predicted click-through rate (PCTR) of multiple samples within a batch. Based on this, the embodiments of this disclosure can determine a first loss function based on multiple predicted click-through rates. The calculation formula for this first loss function can be:

[0099]

[0100] Among them, y i This represents the actual click data for the i-th sample. Let k be the predicted click-through rate of the i-th sample, and k be the number of samples. This refers to the logarithmic loss between the actual clicks of the i-th sample and the predicted click rate. The first loss function can be obtained by summing the logarithmic losses of all samples in a batch.

[0101] In some implementations, the click-through rate prediction model may further include a first contrastive learning module and a second contrastive learning module, wherein the first contrastive learning module may be... Figure 3 The scene-aware contrastive learning module shown can be used to acquire differences between different scenes. After obtaining the estimated click-through rate for each sample using the main model, this embodiment of the disclosure can input the scene identifier and the estimated click-through rate into the first contrastive learning module to obtain the second loss function.

[0102] Specifically, in this embodiment, a first anchor sample can be determined, and the scene identifier of the first anchor sample can be compared with the scene identifiers of each other sample to filter out multiple first target samples. The scene identifier of the first target sample is the same as that of the first anchor sample. Based on this, a second loss function is determined according to the multiple first target samples. It should be noted that the first anchor sample can be any sample within a batch; that is, each sample within a batch can be used as a first anchor sample. By comparing its scene identifier with the scene identifiers of other samples, multiple first target samples can be obtained. The process of obtaining the first target samples can be the process of obtaining the similarity scores of the scenes corresponding to all samples.

[0103] The primary function of the first contrastive learning module is to model the differences between scenes. Here, the InfoNceLoss (Noise Contrastive Estimation Loss) contrastive learning loss can be used. In other words, the first contrastive learning module can be used to minimize the distance between samples of the same scene and maximize the distance between samples of different scenes, thereby achieving the modeling of differences between scenes.

[0104] As described above, the input data for the first contrastive learning module can be the output of the main model and scene identifiers, where the output of the main model can be multiple estimated click-through rates. Furthermore, the first contrastive learning module can include a first MASK module, which can be used to find the output score of the model layer of the main model for each sample in each batch based on the scene identifiers. In other words, through the first MASK module in this embodiment, the scene identifier of each sample can be compared with the scene identifiers of every other sample, thereby filtering out multiple first target samples.

[0105] To better understand the process of acquiring the first target sample, embodiments of this disclosure provide, as follows: Figure 4 The example diagram shown is as follows. Figure 4 Two scene identifiers are given, scene A and scene B. For example, a first sample is identified as the first anchor sample, and its scene identifier is scene A. By comparing the first sample with the scenes of other samples (including the first sample), multiple first target samples can be obtained. Figure 4As shown, the first sample is identical to the first sample and can be represented by 401; the first sample and the fifth sample both have scene A as their scene identifier and can be represented by 402. Additionally, the sixth sample and the third sample both have scene B as their scene identifier and can be represented by 403. Here, 402 and 403 are samples with different scene identifiers, and the first target sample obtained through comparison can be a sample related to 402 and 403.

[0106] In this embodiment of the disclosure, the second loss function can be determined based on the first masking function, the similarity score between every two samples in the multiple samples, the temperature coefficient, and hyperparameters. The first masking function can be used to determine whether the scenes between every two samples in the multiple samples are consistent.

[0107] For example, the formula for calculating the second loss function can be:

[0108]

[0109] Where Loss1 is the second loss function, α is a hyperparameter, and this embodiment takes each batch as an example, N is the number of samples in a training batch, i is the first anchor sample, and Z i and Z j For the outputs of the i-th and j-th samples, sim(Z) i Z j Let be the similarity score between sample i and sample j. In this embodiment of the disclosure, a pre-selected similarity score can be used to calculate the similarity score between samples. τ is the temperature coefficient, and MASK1 is the first mask function, which can be used to determine whether the scenes of sample i and sample j are consistent. The calculation formula for MASK1 is:

[0110]

[0111] Based on the above formula, we know that the total loss of each batch of samples can be the sum of the losses of all samples multiplied by a hyperparameter α. The purpose of the hyperparameter α is to adjust the degree to which the model learns the differential information of the scene.

[0112] In summary, this embodiment of the present disclosure uses the i-th sample as the anchor point and compares it with all samples in the same batch to calculate a similarity score. The calculation includes the i-th sample itself, meaning that sample i is most similar to itself. Furthermore, in this embodiment, the similarity scores of all samples are processed and summed to obtain the denominator. The difference between calculating the numerator and the denominator is that the numerator can be determined using the MASK function.

[0113] As can be seen, in the process of calculating similarity, this embodiment of the disclosure can only calculate the similarity of samples that are in the same scene as sample i. That is, the first Mask function can be defined as follows: if sample i and sample j are in the same scene, then the similarity between sample j and sample i is calculated; if the scenes of sample i and sample j are not the same, then the similarity between sample j and sample i is not calculated.

[0114] This embodiment of the disclosure can minimize the distance between samples of the same scene and maximize the distance between samples of different scenes by utilizing the first contrastive learning module. Since the similarity information between scenes has been fully learned at the model layer, but the differences between scenes have not been learned, this embodiment of the disclosure can achieve differentiated modeling between scenes through the first contrastive learning module.

[0115] In this embodiment of the disclosure, the second comparison learning module may be: Figure 3 The multi-label contrastive learning module shown can be used to obtain the differences between different labels, wherein the click information of each sample can be represented by a label. After obtaining the estimated click-through rate corresponding to each sample using the main model, this embodiment of the disclosure can input the click information and the estimated click-through rate into the second contrastive learning module to obtain the third loss function.

[0116] Specifically, in this embodiment, a second anchor sample can be determined, and the click information of the second anchor sample can be compared with the click information of each other sample to filter out multiple second target samples. The click information of the second target samples is the same as that of the second anchor sample. Based on this, a third loss function is determined according to the multiple second target samples. It should be noted that the second anchor sample can be any sample within a batch; that is, each sample within a batch can be used as a second anchor sample. By comparing its click information (label) with the click information (label) of other samples, multiple second target samples can be obtained. The process of obtaining the second target samples can be the process of obtaining the similarity scores of the labels corresponding to all samples.

[0117] To improve the estimated score of the audience after click-through rate conversion while simultaneously predicting click-through rate, this embodiment introduces multi-label contrastive learning, specifically a second contrastive learning module. The main function of this second contrastive learning module is to model the differences between multiple labels. Similarly, this embodiment can use the InfoNceLoss contrastive learning loss. In other words, the second contrastive learning module can be used to minimize the distance between samples with the same label and maximize the distance between samples with different labels, thereby learning to model the differences between multiple labels. This can, to some extent, improve the estimated score of the audience after click-through rate conversion.

[0118] As described above, the input data for the second contrastive learning module can be the output of the main model and click information (labels). Here, the output of the main model can be multiple estimated click-through rates; the click information can include whether a click was made, whether a purchase was made after clicking the SMS message, and whether a purchase was made without clicking the SMS message. Whether a click was made can be the primary label for the estimated click-through rate, while whether a purchase was made after clicking the SMS message and whether a purchase was made without clicking the SMS message can be secondary labels. To better understand the click information of the samples, the embodiments of this disclosure provide Table 1 below.

[0119] Table 1

[0120]

[0121] As shown in Table 1, a single sample can correspond to multiple tags. That is, click information can include multiple tags. For example, a sample can include three tags: whether the user clicked the SMS message, whether they clicked the SMS message and purchased an item, and whether they did not click the SMS message but purchased an item. For instance, Sample 1 corresponds to the user clicking the SMS message and purchasing an item through that click; Sample 2 corresponds to the user clicking the SMS message but not purchasing an item; and Sample 3 corresponds to the user not clicking the SMS message but purchasing an item.

[0122] Additionally, the second contrastive learning module may include a second MASK module, which can be used to find the output score of the model layer of the main model for each sample in each batch based on click information (label). In other words, through the second MASK module, this embodiment of the disclosure can compare the click information of each sample with the click information of every other sample, thereby filtering out multiple second target samples.

[0123] For example, in embodiments of this disclosure, each sample can be non-conflictingly labeled into four categories, which can be click and purchase, click but not purchase, no click and no purchase, and no click and purchase, respectively.

[0124] To better understand the process of obtaining the second target sample, embodiments of this disclosure provide, as follows: Figure 5 The example diagram shown is as follows. Figure 5 Two labels are provided: "Click and buy" and "Don't click and don't buy". This image is for illustrative purposes only, and other labels may also be included, but are not intended as actual limitations.

[0125] For example, a first sample is identified as the second anchor sample, and the first sample is labeled "click and purchase." By comparing the labels of the first sample with those of other samples (including the first sample), multiple second target samples can be obtained. Figure 5As shown, the first sample is identical to the second sample and can be represented by 501; both the first and second samples are labeled "clicked and purchased" and can be represented by 502. Furthermore, the fourth and fifth samples are labeled "no click, no purchase" and can be represented by 503. Here, 502 and 503 are samples with different labels, and the final target sample obtained through comparison can be a sample related to 502 and 503.

[0126] In this embodiment of the disclosure, the third loss function can be determined based on the second masking function, the similarity score between every two samples in the multiple samples, the temperature coefficient, and hyperparameters. The second masking function can be used to determine whether the labels of every two samples in the multiple samples are consistent.

[0127] For example, the formula for calculating the third loss function can be:

[0128]

[0129] Where Loss2 is the third loss function, α is a hyperparameter, and this embodiment takes each batch as an example, N is the number of samples in a training batch, i is the second anchor sample, and Z i and Z j For the outputs of the i-th and j-th samples, sim(Z) i Z j Let be the similarity score between sample i and sample j. In this embodiment of the disclosure, a pre-selected similarity score can be used to calculate the similarity score between samples. τ is a temperature coefficient, and MASK2 is a second mask function, which can be used to determine whether the labels of sample i and sample j are consistent. The formula for calculating MASK2 is:

[0130]

[0131] Based on the above formula, we know that the total loss of each batch of samples can be the sum of the losses of all samples multiplied by a hyperparameter α. The purpose of the hyperparameter α is to adjust the degree to which the model learns multi-label differential information.

[0132] In summary, this embodiment of the present disclosure uses the i-th sample as the anchor point and compares it with all samples in the same batch to calculate a similarity score. The calculation includes the i-th sample itself, meaning that sample i is most similar to itself. Furthermore, in this embodiment, the similarity scores of all samples are processed and summed to obtain the denominator. The difference between calculating the numerator and the denominator is that the numerator can be determined using the MASK function.

[0133] As can be seen, in the process of calculating similarity, this embodiment of the disclosure can only calculate the similarity of samples with the same label as sample i. That is, the second Mask function can be defined as follows: if the labels of sample i and sample j are the same, then the similarity between sample j and sample i is calculated; if the labels of sample i and sample j are different, then the similarity between sample j and sample i is not calculated.

[0134] This embodiment of the disclosure utilizes a second contrastive learning module to minimize the distance between samples of various tags and maximize the distance between samples of different tags. Therefore, the click-through rate prediction model using the second contrastive learning module can learn the similarity relationships between clicked and converted samples, thereby improving the identification accuracy of such samples.

[0135] In summary, by introducing the first and second contrastive learning modules, this embodiment of the present disclosure can not only improve the accuracy of CTR evaluation but also increase the post-click purchase rate. Specifically, the first contrastive learning module can learn the differences between scenarios, solving the problems of high computational and maintenance costs caused by separate modeling; the second contrastive learning module learns the differences between different categories of auxiliary tags, solving the problem that multi-scenario models cannot improve the predicted score of the post-click conversion audience while predicting the click-through rate.

[0136] In step S23, a target loss function is determined based on the first loss function, the second loss function, and the third loss function, and the click-through rate prediction model is trained using the target loss function.

[0137] The specific implementation method of step S23 has been described in detail in the above embodiments, and will not be repeated here.

[0138] In step S24, the data to be selected is obtained.

[0139] As an alternative approach, embodiments of this disclosure can be modeled using user historical purchase, browsing, and add-to-cart data, as well as modeled based on scenarios, such as promotional days (scenario 1) or new product launch days (scenario 2).

[0140] The data to be selected may include multiple user information items and scene identifiers. User information may include, for example, Figure 6 The data displayed includes historical purchase, browsing, and add-to-cart information. Based on this, this embodiment of the disclosure can input the data to be selected into the click-through rate prediction model, i.e., proceed to step S25.

[0141] In step S25, the data to be selected is input into the click-through rate prediction model to obtain the target audience.

[0142] Based on the click-through rate prediction model, this embodiment of the disclosure can select high-intent audiences who are likely to click on the SMS message and convert, and send promotional or new product SMS messages to the audience. After receiving the SMS message, users can enter the activity page through the SMS link to browse and place an order, thereby improving the conversion rate and GMV (Gross Merchandise Volume) on promotional days or new product launches.

[0143] As explained above, different scene identifiers correspond to different selected groups. That is, the target selected group can include the first group and the second group, and the scene identifiers of the first group and the second group are different.

[0144] As an example, such as Figure 6 The embodiments of this disclosure shown can use a click-through rate (CTR) prediction model to select people who may click on the SMS message and make a purchase on promotional days (Scenario 1), and can use the CTR prediction model to select people who may click on the SMS message and make a purchase on new product launch days (Scenario 2).

[0145] In step S26, an activity SMS message is sent to the target selected group of people.

[0146] As an optional approach, after obtaining the target audience, this embodiment of the disclosure can send promotional SMS messages to the target audience. It is evident that the click-through rate prediction model in this embodiment of the disclosure can adapt to multiple scenarios, and to a certain extent can reduce computational costs, lower maintenance costs, and enhance prediction accuracy in long-tail scenarios.

[0147] This disclosure embodiment can obtain a second loss function and a third loss function by utilizing scene identifiers and click-through rate information, and comprehensively determine the target loss function by combining the second loss function and the third loss function with the first loss function. Since the target loss function fully considers different scene identifiers, the click-through rate prediction model can not only learn the differences between scenes, but also does not need to model multiple scenes separately. To a certain extent, it can reduce the computational and maintenance costs caused by separate modeling. Furthermore, by introducing click-through rate information, this disclosure can achieve click-through rate prediction.

[0148] Figure 7 This is a block diagram illustrating a click-through rate prediction model generation apparatus according to an exemplary embodiment. (Refer to...) Figure 7 The click-through rate prediction model generation device 700 includes an acquisition module 710, a generation module 720, and a determination module 730.

[0149] The acquisition module 710 is configured to acquire a model training set, which includes multiple samples. Each sample includes user information, scene identifier, and click information. The scene identifier is used to represent different information delivery scenarios.

[0150] The generation module 720 is configured to generate a first loss function based on the user information, and to generate a second loss function and a third loss function based on the scene identifier and the click information;

[0151] The determining module 730 is configured to determine a target loss function based on the first loss function, the second loss function, and the third loss function, and to train the click-through rate prediction model using the target loss function.

[0152] In some implementations, the user information includes user profile information, user behavior characteristic information, and user purchase characteristic information; the click-through rate prediction model includes a master model; the generation module 720 may include:

[0153] The input submodule is configured to input the user profile information, the user behavior feature information, and the user purchase feature information into the main model to obtain the estimated click-through rate of multiple samples;

[0154] The first loss determination submodule is configured to determine the first loss function based on the estimated click-through rates of the plurality of samples.

[0155] In some implementations, the click-through rate prediction model further includes a first contrastive learning module and a second contrastive learning module; the generation module 720 may further include:

[0156] The second loss determination submodule is configured to input the scene identifier and the estimated click rate into the first contrast learning module to obtain the second loss function. The first contrast learning module is used to obtain the differences between different scenes.

[0157] The third loss determination submodule is configured to input the click information and the estimated click rate into the second contrastive learning module to obtain the third loss function. The click information of each sample is represented by a label, and the second contrastive learning module is used to obtain the differences between different labels.

[0158] In some implementations, the second loss determination submodule is further configured to determine a first anchored sample and compare the scene identifier of the first anchored sample with the scene identifier of each other sample to filter out a plurality of first target samples, wherein the scene identifier of the first target samples is the same as the scene identifier of the first anchored sample; and determine the second loss function based on the plurality of first target samples.

[0159] In some implementations, the second loss function is determined based on a first mask function, the similarity score between every two samples in the plurality of samples, a temperature coefficient, and hyperparameters, wherein the first mask function is used to determine whether the scenes between every two samples in the plurality of samples are consistent.

[0160] In some implementations, the third loss determination submodule is further configured to determine a second anchor sample and compare the click information of the second anchor sample with the scene information of each other sample to filter out a plurality of second target samples, wherein the click information of the second target samples is the same as the click information of the second anchor sample; and determine the third loss function based on the plurality of second target samples.

[0161] In some implementations, the third loss function is determined based on the second mask function, the similarity score between every two samples in the plurality of samples, the temperature coefficient, and hyperparameters, wherein the second mask function is used to determine whether the labels between every two samples in the plurality of samples are consistent.

[0162] In some embodiments, the click-through rate prediction model generation apparatus 700 may further include:

[0163] The data acquisition module is configured to acquire the data to be selected.

[0164] The data input module is configured to input the data to be selected into the click-through rate prediction model to obtain the target audience. The target audience includes a first type of audience and a second type of audience, and the scene identifiers of the first type of audience and the second type of audience are different.

[0165] The sending module is configured to send promotional text messages to the target selected group of people.

[0166] This disclosure embodiment can obtain a second loss function and a third loss function by utilizing scene identifiers and click-through rate information, and comprehensively determine the target loss function by combining the second loss function and the third loss function with the first loss function. Since the target loss function fully considers different scene identifiers, the click-through rate prediction model can not only learn the differences between scenes, but also does not need to model multiple scenes separately. To a certain extent, it can reduce the computational and maintenance costs caused by separate modeling. Furthermore, by introducing click-through rate information, this disclosure can achieve click-through rate prediction.

[0167] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated upon here.

[0168] This disclosure also provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the steps of the click-through rate prediction model generation method provided in this disclosure.

[0169] Figure 8This is a block diagram of an electronic device 800 illustrating a method for generating a click-through rate prediction model according to an exemplary embodiment. For example, the electronic device 800 may be a mobile phone, computer, digital broadcasting terminal, messaging device, game console, tablet device, medical device, fitness equipment, personal digital assistant, etc.

[0170] Reference Figure 8 The electronic device 800 may include one or more of the following components: processing component 802, memory 804, power supply component 806, multimedia component 808, audio component 810, input / output interface 812, sensor component 814, and communication component 816.

[0171] Processing component 802 typically controls the overall operation of electronic device 800, such as operations associated with display, telephone calls, data communication, camera operation, and recording. Processing component 802 may include one or more processors 820 to execute instructions to complete all or part of the steps in the above-described click-through rate prediction model generation method. Furthermore, processing component 802 may include one or more modules to facilitate interaction between processing component 802 and other components. For example, processing component 802 may include a multimedia module to facilitate interaction between multimedia component 808 and processing component 802.

[0172] Memory 804 is configured to store various types of data to support the operation of electronic device 800. Examples of this data include instructions for any application or method operating on electronic device 800, contact data, phonebook data, messages, pictures, videos, etc. Memory 804 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0173] Power supply component 806 provides power to various components of electronic device 800. Power supply component 806 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to electronic device 800.

[0174] Multimedia component 808 includes a screen that provides an output interface between the electronic device 800 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touchscreen to receive input signals from the user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors may sense not only the boundaries of the touch or swipe action but also the duration and pressure associated with the touch or swipe operation. In some embodiments, multimedia component 808 includes a front-facing camera and / or a rear-facing camera. When the electronic device 800 is in an operating mode, such as a shooting mode or a video mode, the front-facing camera and / or the rear-facing camera may receive external multimedia data. Each front-facing camera and rear-facing camera may be a fixed optical lens system or have focal length and optical zoom capabilities.

[0175] Audio component 810 is configured to output and / or input audio signals. For example, audio component 810 includes a microphone (MIC) configured to receive external audio signals when electronic device 800 is in an operating mode, such as call mode, recording mode, and voice recognition mode. The received audio signals may be further stored in memory 804 or transmitted via communication component 816. In some embodiments, audio component 810 also includes a speaker for outputting audio signals.

[0176] Input / output interface 812 provides an interface between processing component 802 and peripheral interface modules, such as keyboards, click wheels, buttons, etc. These buttons may include, but are not limited to, home buttons, volume buttons, power buttons, and lock buttons.

[0177] Sensor assembly 814 includes one or more sensors for providing state assessments of various aspects of electronic device 800. For example, sensor assembly 814 can detect the on / off state of electronic device 800, the relative positioning of components such as the display and keypad of electronic device 800, changes in position of electronic device 800 or a component of electronic device 800, the presence or absence of user contact with electronic device 800, orientation or acceleration / deceleration of electronic device 800, and temperature changes of electronic device 800. Sensor assembly 814 may include a proximity sensor configured to detect the presence of nearby objects without any physical contact. Sensor assembly 814 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, sensor assembly 814 may also include an accelerometer, gyroscope, magnetometer, pressure sensor, or temperature sensor.

[0178] Communication component 816 is configured to facilitate wired or wireless communication between electronic device 800 and other devices. Electronic device 800 can access wireless networks based on communication standards, such as WiFi, 2G, or 3G, or combinations thereof. In one exemplary embodiment, communication component 816 receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, communication component 816 also includes a near-field communication (NFC) module to facilitate short-range communication. For example, the NFC module may be implemented based on radio frequency identification (RFID) technology, Infrared Data Association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0179] In an exemplary embodiment, the electronic device 800 may be implemented by one or more application-specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field-programmable gate arrays (FPGAs), controllers, microcontrollers, microprocessors, or other electronic components to perform the above-described click-through rate prediction model generation method.

[0180] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 804 including instructions, which can be executed by a processor 820 of an electronic device 800 to complete the above-described method for generating a click-through rate prediction model. For example, the non-transitory computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0181] In another exemplary embodiment, a computer program product is also provided, the computer program product comprising a computer program executable by a programmable device, the computer program having a code portion for performing the above-described click-through rate prediction model generation method when executed by the programmable device.

[0182] Some embodiments of this disclosure also provide a chip system, such as Figure 9As shown, the chip system includes at least one processor 1301 and at least one interface circuit 1302. The processor 1301 and the interface circuit 1302 are interconnected via lines. For example, the interface circuit 1302 can be used to receive signals from other devices (e.g., the memory of an electronic device). As another example, the interface circuit 1302 can be used to send signals to other devices (e.g., the processor 1301). Exemplarily, the interface circuit 1302 can read instructions stored in memory and send those instructions to the processor 1301. When the instructions are executed by the processor 1301, the click-through rate prediction model generation apparatus can perform the steps in the above embodiments. Of course, the chip system may also include other discrete components, and some embodiments of this disclosure do not specifically limit this.

[0183] In some embodiments of this disclosure, the interface circuit 1302 can acquire data, program instructions, and / or information from the internal storage area of ​​the chip system; it can also acquire data, program instructions, and / or information from outside the chip system.

[0184] Optionally, the chip system also includes a memory 1303 for storing necessary computer programs and data.

[0185] Those skilled in the art will also understand that the various illustrative logical blocks and steps listed in the embodiments of this application can be implemented by electronic hardware, computer software, or a combination of both. Whether such functionality is implemented through hardware or software depends on the specific application and the overall system design requirements. Those skilled in the art can implement the described functionality using various methods for each specific application, but such implementation should not be construed as exceeding the scope of protection of the embodiments of this application.

[0186] In the above detailed description, reference has been made to the accompanying drawings, which illustrate specific aspects of this disclosure by way of illustration. In this regard, terms indicating direction or positional relationship, such as “center,” “longitudinal,” “lateral,” “length,” “width,” “thickness,” “upper,” “lower,” “front,” “rear,” “left,” “right,” “vertical,” “horizontal,” “top,” “bottom,” “inner,” “outer,” “clockwise,” “counterclockwise,” “axial,” “radial,” and “circumferential,” are used with reference to the orientation of the described figures. Since components of the described device can be positioned in multiple different orientations, directional terms are used for illustrative purposes and not for limitation. It should be understood that other aspects can be utilized and structural or logical changes can be made without departing from the concept of this disclosure. Therefore, the following detailed description should not be considered limiting.

[0187] It should be understood that, unless otherwise specifically indicated, features of various embodiments of this disclosure described herein can be combined with each other. As used herein, the term “and / or” includes any one of the relevant listed items and any combination of any two or more; similarly, “at least one of…” includes any one of the relevant listed items and any combination of any two or more.

[0188] It should be understood that, unless otherwise expressly specified and limited, the terms "joining," "attaching," "installing," "connecting," "linking," "fixing," etc., used in the embodiments of this disclosure should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, an electrical connection, or a connection that allows communication between them; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components, unless otherwise expressly limited. Those skilled in the art can understand the specific meaning of the above terms herein based on the specific circumstances.

[0189] Furthermore, the term "above" as used herein with respect to components, elements, or material layers formed or located "above" a surface may be used to indicate that the component, element, or material layer is "indirectly" positioned (e.g., placed, formed, deposited, etc.) on the surface such that one or more additional components, elements, or layers are arranged between the surface and the component, element, or material layer. However, the term "above" as used with respect to components, elements, or material layers formed or located "above" a surface may also optionally have a specific meaning: that the component, element, or material layer is "directly" positioned (e.g., placed, formed, deposited, etc.) on the surface, for example, in direct contact with the surface.

[0190] Although terms such as “first,” “second,” and “third” may be used herein to describe various components, parts, regions, layers, or sections, these components, parts, regions, layers, or sections are not limited to these terms. Rather, these terms are used only to distinguish one component, part, region, layer, or section from another. Therefore, without departing from the teachings of the examples described herein, the first component, part, region, layer, or section mentioned in the examples may also be referred to as the second component, part, region, layer, or section. Furthermore, the terms “first” and “second” are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as “first” or “second” may explicitly or implicitly include at least one of that feature. In the description herein, “a plurality” means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0191] It should be understood that spatial relative terms, such as “above,” “upper,” “below,” and “lower,” are used herein to describe the relationship between one element and another shown in the figures. In addition to the orientation depicted in the figures, these spatial relative terms are also intended to encompass different orientations of the device in use or operation. For example, if the device in the figures is flipped, an element described as “above” or “upper” relative to another element would be “below” or “lower” relative to that other element. Thus, depending on the spatial orientation of the device, the term “above” encompasses both above and below orientations. Devices may have other orientations (e.g., rotated 90 degrees or in other orientations), and the spatial relative terms used herein should be interpreted accordingly.

[0192] Furthermore, the term “exemplary” is used herein to mean serving as an example, instance, or illustration. Any aspect or design described herein as “exemplary” is not necessarily to be construed as advantageous compared to other aspects or designs. Rather, the use of the term “exemplary” is intended to present the concept in a concrete manner. As used herein, the term “or” is intended to mean an inclusive “or” rather than an exclusive “or.” That is, unless otherwise specified or clear from the context, “X applies A or B” is intended to mean any of the natural inclusive arrangements. That is, “X applies A or B” satisfies any of the foregoing instances if X applies A; X applies B; or both X applies A and B. Additionally, unless otherwise specified or clear from the context to refer to the singular form, the articles “a” and “an” as used in this application and the appended claims are generally understood to mean “one or more.”

[0193] Similarly, although this disclosure has been shown and described with respect to one or more implementations, equivalent variations and modifications will occur to those skilled in the art upon reading and understanding this specification and the accompanying drawings. This disclosure includes all such modifications and variations and is limited only by the scope of the claims. In particular, with respect to the various functions performed by the components described above (e.g., elements, resources, etc.), unless otherwise indicated, the terminology used to describe such components is intended to correspond to any component (functionally equivalent) that performs the specific function of the described component, even if structurally not equivalent to the disclosed structure. Furthermore, although specific features of this disclosure may have been disclosed with respect to only one of several implementations, such features may be combined with one or more other features of other implementations, as may be desired and advantageous for any given or particular application. Moreover, with regard to the terms “comprising,” “owning,” “having,” “having,” or variations thereof as used in the detailed description or claims, such terms are intended to be inclusive in a manner similar to the term “including.”

[0194] Those skilled in the art will readily appreciate other embodiments of the present disclosure after considering the specification and practicing the invention disclosed herein. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or customary techniques in the art not disclosed herein. The description and examples are to be considered as exemplary only, with the true scope and spirit of the present disclosure being indicated by the appended claims.

[0195] It should be understood that this disclosure is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this disclosure is limited only by the appended claims.

Claims

1. A method for generating a click-through rate (CTR) prediction model, characterized in that, include: Obtain a model training set, which includes multiple samples. Each sample includes user information, scene identifier, and click information. The scene identifier is used to represent different information delivery scenarios. A first loss function is generated based on the user information, and a second and a third loss function are generated based on the scene identifier and the click information; The target loss function is determined based on the first loss function, the second loss function, and the third loss function, and the click-through rate prediction model is trained using the target loss function.

2. The method according to claim 1, characterized in that, The user information includes user profile information, user behavior characteristic information, and user purchase characteristic information; the click-through rate prediction model includes a main model. The step of generating the first loss function based on the user information includes: The user profile information, user behavior feature information, and user purchase feature information are input into the main model to obtain the estimated click-through rate of multiple samples; The first loss function is determined based on the estimated click-through rates of the multiple samples.

3. The method according to claim 2, characterized in that, The click-through rate prediction model also includes a first contrastive learning module and a second contrastive learning module; The step of generating the second loss function and the third loss function based on the scene identifier and the click information includes: The scene identifier and the estimated click-through rate are input into the first contrastive learning module to obtain the second loss function. The first contrastive learning module is used to obtain the differences between different scenes. The click information and the estimated click-through rate are input into the second contrastive learning module to obtain the third loss function. The click information of each sample is represented by a label, and the second contrastive learning module is used to obtain the differences between different labels.

4. The method according to claim 3, characterized in that, The step of inputting the scene identifier and the estimated click-through rate into the first contrastive learning module to obtain the second loss function includes: A first anchor sample is determined, and the scene identifier of the first anchor sample is compared with the scene identifier of each other sample to filter out multiple first target samples, wherein the scene identifier of the first target sample is the same as the scene identifier of the first anchor sample. The second loss function is determined based on multiple first target samples.

5. The method according to claim 4, characterized in that, The second loss function is determined based on the first mask function, the similarity score between every two samples in the plurality of samples, the temperature coefficient, and the hyperparameters. The first mask function is used to determine whether the scenes between every two samples in the plurality of samples are consistent.

6. The method according to claim 3, characterized in that, The step of inputting the click information and the estimated click-through rate into the second contrastive learning module to obtain the third loss function includes: A second anchor sample is determined, and the click information of the second anchor sample is compared with the scene information of each other sample to filter out multiple second target samples. The click information of the second target samples is the same as the click information of the second anchor sample. The third loss function is determined based on multiple second target samples.

7. The method according to claim 6, characterized in that, The third loss function is determined based on the second mask function, the similarity score between every two samples in the plurality of samples, the temperature coefficient, and the hyperparameters. The second mask function is used to determine whether the labels between every two samples in the plurality of samples are consistent.

8. The method according to any one of claims 1 to 7, characterized in that, The method further comprises: Get the data to be selected; The data to be selected is input into the click-through rate prediction model to obtain the target audience. The target audience includes a first group and a second group, and the scene identifiers of the first group and the second group are different. Send promotional text messages to the target audience.

9. A device for generating a click-through rate prediction model, characterized in that, include: The acquisition module is configured to acquire a model training set, which includes multiple samples. Each sample includes user information, scene identifier, and click information. The scene identifier is used to represent different information delivery scenarios. The generation module is configured to generate a first loss function based on the user information, and to generate a second loss function and a third loss function based on the scene identifier and the click information; The determination module is configured to determine a target loss function based on the first loss function, the second loss function, and the third loss function, and to train the click-through rate prediction model using the target loss function.

10. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured as follows: Obtain a model training set, which includes multiple samples. Each sample includes user information, scene identifier, and click information. The scene identifier is used to represent different information delivery scenarios. A first loss function is generated based on the user information, and a second and a third loss function are generated based on the scene identifier and the click information; The target loss function is determined based on the first loss function, the second loss function, and the third loss function, and the click-through rate prediction model is trained using the target loss function.

11. A computer-readable storage medium storing computer program instructions thereon, characterized in that, When executed by a processor, the program instructions implement the steps of the method described in any one of claims 1 to 8.

12. A computer program product, characterized in that, It includes a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 8.

Images