Click rate estimation model training method and device, click rate estimation method and device

Abstract

The disclosure provides a click rate estimation model training method and device, and relates to the technical field of artificial intelligence. A specific embodiment of the method comprises: obtaining a preset sample set, the sample set comprising at least one sample; obtaining a click rate estimation network comprising a feature prior module and a model prior module, which is constructed in advance; performing the following training steps: selecting a sample from the sample set; inputting the sample into the click rate estimation network to obtain a click rate estimation value output by the click rate estimation network; calculating a feature loss value of a feature prior layer of the feature prior module and a model loss value common to a feature processing layer of the feature prior module and the model prior module; and in response to the click rate estimation network satisfying a training completion condition, taking the click rate estimation network as a click rate estimation model, which improves the generalization ability of the model.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, specifically to the field of artificial intelligence technology, and in particular to click-through rate prediction model training methods and apparatus, click-through rate prediction methods and apparatus, electronic devices, and computer-readable storage media. Background Technology

[0002] In existing recommendation systems, click-through rate prediction typically uses incremental updates, meaning the model is retrained based on incremental data from the previous service. This update method can cause the model to overfit to the incremental data, excessively fitting the distribution of that data, resulting in a lack of generalization ability and negatively impacting training performance. Summary of the Invention

[0003] Embodiments of this disclosure provide a click-through rate (CTR) prediction model training method and apparatus, a click-through rate prediction method and apparatus, an electronic device, and a computer-readable storage medium.

[0004] In a first aspect, embodiments of this disclosure provide a method for training a click-through rate (CTR) prediction model. The method includes: acquiring a pre-set sample set, the sample set including at least one sample, the sample including user features and product features; acquiring a pre-constructed CTR prediction network, the CTR prediction network including a feature prior module and a model prior module, wherein the feature prior module includes a feature prior layer and a feature processing layer, the feature prior layer being used to acquire prior CTR information for each feature in the sample, the feature processing layer obtaining a feature vector of the sample based on the prior CTR information; the model prior module being used to predict the CTR value from the feature vector; and performing the following training steps: selecting a sample from the sample set; inputting the sample into the CTR prediction network to obtain the CTR prediction value output by the CTR prediction network; calculating the feature loss value of the feature prior layer, the model loss value of the feature processing layer, and the model prior module, respectively; and, in response to the CTR prediction network meeting the training completion condition, using the CTR prediction network as a CTR prediction model.

[0005] In some embodiments, the feature processing layer includes a discretization layer and a high-dimensional embedding layer; the discretization layer is used to discretize the prior click-through rate information; the high-dimensional embedding layer is used to extract high-dimensional features of the discretized prior click-through rate information to obtain a first high-dimensional feature vector, and the first high-dimensional feature vector is used as a feature vector.

[0006] In some embodiments, the feature processing layer further includes: an initial embedding layer, used to extract high-dimensional features of each feature in the sample to obtain a second high-dimensional feature vector, and to connect the first high-dimensional feature vector and the second high-dimensional feature vector to obtain a feature vector.

[0007] In some embodiments, the above calculation of the feature loss value of the feature prior layer includes: constructing the loss function of the feature prior layer using the maximum likelihood estimation method; and obtaining the feature loss value of the feature prior layer based on the loss function of the feature prior layer.

[0008] In some embodiments, the above-mentioned calculation of the model loss value of the feature processing layer and the model prior module includes: constructing the cross-entropy loss function of the feature processing layer and the model prior module; and obtaining the model loss value of the feature processing layer and the model prior module based on the cross-entropy loss function.

[0009] In some embodiments, the above-mentioned calculation of the model loss value of the feature processing layer and the model prior module includes: constructing the cross-entropy loss function of the feature processing layer and the model prior module; constructing the distance function between the current parameter distribution and the historical parameter distribution of the feature processing layer and the model prior module; and obtaining the model loss value of the feature processing layer and the model prior module based on the cross-entropy loss function and the distance function.

[0010] Secondly, embodiments of this disclosure provide a click-through rate (CTR) prediction method, which includes: acquiring user data and product data to be tested; obtaining user features and product features based on the user data and product data; inputting the user features and product features into a CTR prediction model generated by the CTR prediction model training method of any embodiment of the first aspect, and obtaining a CTR prediction value output by the CTR prediction model.

[0011] Thirdly, embodiments of this disclosure provide a click-through rate (CTR) prediction model training apparatus, comprising: a sample acquisition unit configured to acquire a preset sample set, the sample set including at least one sample, the sample including: user features and product features; and a network acquisition unit configured to acquire a pre-constructed CTR prediction network, the CTR prediction network including: a feature prior module and a model prior module, wherein the feature prior module includes: a feature prior layer and a feature processing layer, the feature prior layer being used to acquire prior CTR information for each feature in the sample, and the feature processing layer based on prior points. The system extracts click-through rate (CTR) information to obtain the feature vector of the sample. A model prior module is used to predict the CTR value from the feature vector. A selection unit is configured to select samples from the sample set. An input unit is configured to input the sample into the CTR prediction network to obtain the CTR prediction value output by the network. A calculation unit is configured to calculate the feature loss value of the feature prior layer, the model loss value of the feature processing layer, and the model prior module, respectively. A result unit is configured to use the CTR prediction network as the CTR prediction model in response to the CTR prediction network meeting the training completion condition.

[0012] In some embodiments, the feature processing layer includes a discretization layer and a high-dimensional embedding layer; the discretization layer is used to discretize the prior click-through rate information; the high-dimensional embedding layer is used to extract high-dimensional features of the discretized prior click-through rate information to obtain a first high-dimensional feature vector, and the first high-dimensional feature vector is used as a feature vector.

[0013] In some embodiments, the feature processing layer further includes: an initial embedding layer, used to extract high-dimensional features of each feature in the sample to obtain a second high-dimensional feature vector, and to connect the first high-dimensional feature vector and the second high-dimensional feature vector to obtain a feature vector.

[0014] In some embodiments, the above-mentioned computing unit is further configured to: construct the loss function of the feature prior layer using the maximum likelihood estimation method; and obtain the feature loss value of the feature prior layer based on the loss function of the feature prior layer.

[0015] In some embodiments, the above-mentioned computing unit is further configured to: construct the cross-entropy loss function of the feature processing layer and the model prior module; and obtain the model loss value of the feature processing layer and the model prior module based on the cross-entropy loss function.

[0016] In some embodiments, the above-mentioned computing unit is further configured to: construct the cross-entropy loss function of the feature processing layer and the model prior module; construct the distance function between the current parameter distribution and the historical parameter distribution of the feature processing layer and the model prior module; and obtain the model loss value of the feature processing layer and the model prior module based on the cross-entropy loss function and the distance function.

[0017] Fourthly, embodiments of this disclosure provide a click-through rate (CTR) prediction apparatus, comprising: a data acquisition unit configured to acquire user data and product data to be tested; a processing unit configured to obtain user features and product features based on the user data and product data; and a prediction unit configured to input the user features and product features into a CTR prediction model generated by a CTR prediction model training apparatus according to any embodiment of the third aspect, and obtain a CTR prediction value output by the CTR prediction model.

[0018] Fifthly, embodiments of this disclosure provide an electronic device comprising: one or more processors; a storage device having one or more programs stored thereon; and, when the one or more programs are executed by the one or more processors, causing the one or more processors to implement the method described in any embodiment of the first or second aspect.

[0019] In a sixth aspect, embodiments of the present disclosure provide a computer-readable medium having a computer program stored thereon that, when executed by a processor, implements the methods described in any of the embodiments of the first or second aspect.

[0020] The click-through rate (CTR) prediction model training method and apparatus provided in the embodiments of this disclosure firstly acquire a preset sample set, which includes at least one sample, including user features and product features; secondly, acquire a pre-constructed CTR prediction network, which includes a feature prior module and a model prior module, wherein the feature prior module includes a feature prior layer and a feature processing layer, the feature prior layer is used to acquire prior CTR information of each feature in the sample, the feature processing layer obtains the feature vector of the sample based on the prior CTR information; the model prior module is used to predict the CTR value by the feature vector; thirdly, select a sample from the sample set; fourthly, input the sample into the CTR prediction network to obtain the CTR prediction value output by the CTR prediction network; fifthly, calculate the feature loss value of the feature prior layer, the model loss value of the feature processing layer and the model prior module respectively; finally, in response to the CTR prediction network meeting the training completion condition, use the CTR prediction network as the CTR prediction model. Therefore, by constructing a feature prior layer, we can provide prior click changes for different features, thereby enhancing the expressive power of each feature. The feature prior layer is optimized together with the overall click-through rate prediction network, enabling the model to learn the expression of each feature and improving the robustness and accuracy of the click prediction model. Attached Figure Description

[0021] Other features, objects, and advantages of this disclosure will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings:

[0022] Figure 1 This is an exemplary system architecture diagram to which one embodiment of this disclosure can be applied;

[0023] Figure 2 This is a flowchart of an embodiment of the click-through rate prediction model training method according to the present disclosure;

[0024] Figure 3 This is a schematic diagram of the structure of this publicly available click-through rate prediction network;

[0025] Figure 4 This is a flowchart of an embodiment of the click-through rate prediction method according to this disclosure;

[0026] Figure 5 This is a schematic diagram of the structure of an embodiment of a click-through rate prediction model training apparatus according to the present disclosure;

[0027] Figure 6 This is a schematic diagram of a structure of an embodiment of the click-through rate prediction device according to the present disclosure;

[0028] Figure 7 This is a schematic diagram of the structure of an electronic device suitable for implementing embodiments of the present disclosure. Detailed Implementation

[0029] The present disclosure will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings.

[0030] It should be noted that, unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other. This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0031] Figure 1 An exemplary system architecture 100 is shown that can be applied to the click-through rate prediction model training method or click-through rate prediction method of this disclosure.

[0032] like Figure 1 As shown, system architecture 100 may include terminals 101 and 102, network 103, database server 104, and server 105. Network 103 serves as the medium for providing communication links between terminals 101 and 102, database server 104, and server 105. Network 103 may include various connection types, such as wired or wireless communication links or fiber optic cables, etc.

[0033] User 110 can use terminals 101 and 102 to interact with server 105 via network 103 to receive or send messages, etc. Various client applications can be installed on terminals 101 and 102, such as model training applications, image recognition applications, shopping applications, payment applications, web browsers, and instant messaging tools.

[0034] The terminals 101 and 102 here can be either hardware or software. When terminals 101 and 102 are hardware, they can be various electronic devices with displays, including but not limited to smartphones, tablets, e-book readers, MP3 players (Moving Picture Experts Group Audio Layer III), laptops, and desktop computers. When terminals 101 and 102 are software, they can be installed in the electronic devices listed above. They can be implemented as multiple software programs or software modules (e.g., to provide distributed services) or as a single software program or software module. No specific limitations are set here.

[0035] Database server 104 can be a database server that provides various services. For example, the database server can store a sample set. The sample set contains a large number of samples, each of which is different, and the samples can include user characteristics and product characteristics. User 110 can also select samples from the sample set stored in database server 104 through terminals 101 and 102.

[0036] Server 105 can also be a server providing various services, such as a backend server supporting various applications displayed on terminals 101 and 102. The backend server can use samples from the sample set sent by terminals 101 and 102 to train a click-through rate (CTR) prediction model, and can send the trained CTR prediction model back to terminals 101 and 102. In this way, users can apply the generated CTR prediction model to predict click-through rates.

[0037] The database server 104 and server 105 here can be either hardware or software. When they are hardware, they can be implemented as a distributed server cluster consisting of multiple servers, or as a single server. When they are software, they can be implemented as multiple software programs or software modules (for example, to provide distributed services), or as a single software program or software module. No specific limitations are made here.

[0038] It should be noted that the click-through rate (CTR) prediction model training method or CTR prediction method provided in the embodiments of this disclosure are generally executed by the server 105. Accordingly, the CTR prediction model training device or CTR prediction device is also generally located in the server 105.

[0039] It should be noted that if server 105 can perform the relevant functions of database server 104, database server 104 may not be set up in system architecture 100.

[0040] It should be understood that Figure 1 The number of terminals, networks, database servers, and servers shown is merely illustrative. Depending on implementation needs, any number of terminals, networks, database servers, and servers can be included.

[0041] In traditional recommendation system scenarios, the main approach is to use the previous model as a teacher network to guide the current model's learning, or to use reinforcement learning based on the output of previous models.

[0042] These methods are all sample-level based, treating each sample as a whole to stabilize the model's parameter distribution. This involves learning the joint information of features from each sample to enhance the model's expressive power. For example, one method uses the prediction result of each sample in the previous model as a guidance signal and employs cross-entropy to learn this signal. Another method uses the prediction result of each sample in the previous model as a baseline and uses hinge loss to supervise whether the current model's prediction is better than the baseline. In real-world scenarios, the sample distribution is very sparse, which means that the confidence level of predictions made by sample-level methods is not necessarily reliable, limiting their applicability.

[0043] To address the aforementioned issues, this disclosure provides a click-through rate (CTR) prediction model training method that learns the individual information of each feature in each sample, making the sample distribution richer and more stable, and improving the robustness of model parameter learning. Figure 2 The diagram illustrates a flow 200 of an embodiment of a click-through rate (CTR) prediction model training method according to the present disclosure, which includes the following steps:

[0044] Step 201: Obtain the preset sample set.

[0045] In this embodiment, the sample set is a set of samples collected in advance for training the click-through rate prediction model. The samples in the sample set can be obtained by collecting sample data from the Internet and extracting features from the sample data (e.g., by using a feature extractor). The sample set includes at least one sample, and each sample includes user features and product features.

[0046] In this embodiment, user features and product features can be discrete feature vectors after discretization processing, and the dataset... The sample set representing the incremental training data for the current model, where This indicates that the discrete feature vector after One-Hot encoding contains user features and product features, etc. dimensional features, Indicates whether the user has clicked on the current product. Indicates the current moment.

[0047] In this embodiment, product characteristics can be the characteristics of products within the specified content of the same website or application. Product characteristics include: product color, product type, product origin, etc. User characteristics are the characteristics of individuals who are interested in the products within the specified content of the aforementioned website or application during a certain period of time. User characteristics include: user attribute characteristics and user behavioral characteristics. User attribute characteristics include: user age group, user interests, etc., while user behavioral characteristics include: items the user follows, content the user operates on, etc.

[0048] In this embodiment, the execution entity of the click-through rate prediction model training method (e.g., Figure 1 The server 105 shown can obtain the sample set in various ways. For example, the executing entity can obtain it from a database server (e.g., via a wired or wireless connection). Figure 1 The sample set stored in the database server 104 shown can be retrieved. For example, a user can access the sample set via a terminal (e.g., ...). Figure 1 The terminals 101 and 102 shown are used to collect samples. In this way, the executing entity can receive the samples collected by the terminals and store these samples locally, thereby generating a sample set.

[0049] Step 202: Obtain the pre-built click-through rate prediction network.

[0050] In this embodiment, the click-through rate (CTR) prediction network includes a feature prior module and a model prior module. The feature prior module includes a feature prior layer and a feature processing layer. The feature prior layer is used to obtain the prior CTR information of each feature in the sample. The feature processing layer obtains the feature vector of the sample based on the prior CTR information. The model prior module is used to predict the CTR value by the feature vector.

[0051] In this embodiment, the prior click-through rate information includes: the average estimated click-through rate obtained by predicting the features. The average estimated click-through rate is the average of the predicted values ​​of the click-through rate (CTR), which refers to the ratio of the number of times a specified piece of content on a website or application is clicked to the number of times it is displayed. Click-through rate is usually an important indicator for measuring recommendation efficiency in recommendation systems. Because feature data is more aggregated than sample-level data, the distribution of CTR values ​​at the feature granularity is more stable than the distribution at the instance level. The prior features ultimately serve as auxiliary feature information and provide more stable learning for model updates, which is beneficial for optimizing long-tail terms.

[0052] Generally, click-through rate (CTR) can be calculated by dividing the actual number of clicks on a product on a webpage or application (which is also the number of times the product reaches the target page) by the number of times the product is displayed on the webpage or application (show content).

[0053] In this embodiment, compared to traditional click-through rate (CTR) prediction models, the CTR prediction network of this disclosure has a feature prior layer. This feature prior layer can learn the CTR of each feature in the sample, making the information of each feature richer and more stable. The feature prior layer obtains the prior CTR information for each feature. ,in This represents the user and product characteristics in the sample. Indicates the first Feature prior layer of each feature This represents the total number of eigenvalues.

[0054] When training the click-through rate (CTR) prediction network, it is necessary to adjust the parameters of the feature prior layer simultaneously through the loss function of the feature prior layer. After the loss of the feature prior layer converges and the loss function of the CTR prediction network other than the feature prior layer converges, the CTR prediction network is considered to have converged, and the training is considered complete.

[0055] In this embodiment, the feature processing layer is a network layer that processes prior click-through rate information. Through the feature processing layer, the prior click-through rate information can be transformed into feature vectors suitable for analysis by the model's prior module.

[0056] In this embodiment, the model prior module can use an NN (Neural Network) network or a network transformed from an NN network. The model prior module analyzes the feature vectors related to users and products to predict the click-through rate of users clicking on products.

[0057] Step 203: Select samples from the sample set.

[0058] In this embodiment, the executing entity can select samples from the sample set obtained in step 201 and execute the training steps from 204 to 206. The method and number of samples selected are not limited in this application. For example, at least one sample can be randomly selected.

[0059] Step 204: Input the sample into the click-through rate prediction network to obtain the click-through rate prediction value output by the click-through rate prediction network.

[0060] In this embodiment, the samples include user features and product features. The samples selected from the sample set are input into the click-through rate (CTR) prediction network. The CTR prediction network analyzes the user features and product features to determine the CTR prediction value corresponding to the current user features and product features. The CTR prediction value can be used to measure the recommendation efficiency of each product in the recommendation system, thereby providing an effective means for recommending and displaying products in the recommendation system of web pages and applications.

[0061] In this embodiment, the click-through rate (CTR) prediction network is used to characterize the correspondence between samples and CTR prediction values. The samples include user features and product features, and the CTR prediction value is the estimated value of the user's click-through rate on the product under the current user features and product features.

[0062] Step 205: Calculate the feature loss value of the feature prior layer in the click-through rate prediction network, and the model loss value of the feature processing layer and the model prior module in the click-through rate prediction network.

[0063] In this embodiment, the model parameters of the click-through rate prediction network include: the first parameter of the feature prior layer. The second parameter of the click-through rate prediction network, excluding the feature prior layer. When training the click-through rate prediction network, the first and second parameters need to be adjusted so that the click-through rate prediction network can converge and meet the training completion conditions.

[0064] In this embodiment, multiple iterations of training are required in the training steps of the click-through rate (CTR) prediction network in order to converge the overall loss of the CTR prediction network. Each iteration executes steps 203-205 in sequence. After calculating the feature loss value and model loss value in step 205, the parameters of the CTR estimation network are adjusted before the next iteration of training is performed.

[0065] In this embodiment, loss functions for the feature prior layer, the feature processing layer, and the model prior module can be set separately. The feature loss value of the feature prior layer during each training iteration is calculated using the loss function of the feature prior layer, and the first parameter of the feature prior layer is adjusted based on the direction of change of the feature loss value. The model loss value during each training iteration is calculated using the loss functions of the feature processing layer and the model prior module in the click-through rate prediction network, and the second parameter of the feature processing layer and the model prior module is adjusted based on the direction of change of the model loss value.

[0066] It should be noted that in the training step of the current iteration, if the click-through rate estimation network has met the training completion condition, the execution subject will no longer input samples into the click-through rate estimation network, and the click-through rate estimation network is the click-through rate estimation model that has been trained.

[0067] Step 206: In response to the click-through rate prediction network meeting the training completion conditions, the click-through rate prediction network is used as the click-through rate prediction model.

[0068] In this embodiment, the training completion condition includes at least one of the following: the number of training iterations of the click-through rate (CTR) prediction network reaches a predetermined iteration threshold; the rate of change of the model parameters of the CTR prediction network is less than a predetermined threshold, wherein the model parameters include the first parameter and the second parameter mentioned above. The CTR prediction network is determined to meet the training completion condition when the rates of change of both the first and second parameters of the CTR prediction network are less than the predetermined threshold. For example, the training iterations reach 5,000. The rate of change of the model parameters of the CTR prediction network is less than 0.05. In this embodiment, setting the training completion condition can accelerate the model convergence speed.

[0069] In step 206, if the model parameters do not change significantly and are less than a predetermined threshold, the loss value of the click-through rate prediction network is determined to have converged, and the click-through rate prediction model training is complete.

[0070] In some alternative implementations of this disclosure, in response to the click-through rate prediction network not meeting the training completion conditions, the first parameter and / or the second parameter are adjusted to make the loss value of the click-through rate prediction network converge, and the above steps 203-205 are continued.

[0071] In this embodiment, if training is not complete, adjusting the first parameter and / or the second parameter to converge the loss value of the click-through rate (CTR) prediction network includes: keeping the first parameter unchanged, adjusting the second parameter, and repeatedly executing steps 203-205, adjusting the first parameter to gradually reduce the loss value of the CTR prediction network until it stabilizes. Optionally, adjusting the first parameter and / or the second parameter to converge the loss value of the CTR prediction network includes: keeping the second parameter unchanged, adjusting the first parameter, and repeatedly executing steps 203-205, adjusting the second parameter to gradually reduce the loss value of the CTR prediction network until it stabilizes. This optional implementation of the CTR prediction model training method, when the training completion condition is not met, adjusts the model parameters, enabling the CTR prediction model to be trained quickly and ensuring that the trained CTR prediction model is the optimal model.

[0072] The click-through rate (CTR) prediction model training method provided in the embodiments of this disclosure firstly obtains a preset sample set, which includes at least one sample, including user features and product features; secondly, it obtains a pre-constructed CTR prediction network, which includes a feature prior module and a model prior module. The feature prior module includes a feature prior layer and a feature processing layer. The feature prior layer is used to obtain prior CTR information for each feature in the sample, and the feature processing layer obtains the feature vector of the sample based on the prior CTR information. The model prior module is used to predict the CTR value by the feature vector; thirdly, it selects a sample from the sample set; fourthly, it inputs the sample into the CTR prediction network to obtain the CTR prediction value output by the CTR prediction network; fifthly, it calculates the feature loss value of the feature prior layer, the model loss value of the feature processing layer, and the model prior module, respectively; finally, in response to the CTR prediction network meeting the training completion condition, it uses the CTR prediction network as the CTR prediction model. Therefore, by constructing a feature prior layer, we can provide prior click changes for different features, thereby enhancing the expressive power of each feature. The feature prior layer is optimized together with the overall click-through rate prediction network, enabling the model to learn the expression of each feature and improving the robustness and accuracy of the click prediction model.

[0073] In some embodiments of this disclosure, the aforementioned feature processing layer includes a discretization layer and a high-dimensional embedding layer; the discretization layer is used to discretize the prior click-through rate information; the high-dimensional embedding layer is used to extract high-dimensional features of the discretized prior click-through rate information to obtain a first high-dimensional feature vector, and the first high-dimensional feature vector is used as a feature vector.

[0074] In this optional implementation, the feature prior information for continuous values Discrete values ​​of prior information for each feature are obtained through a discretization layer. ,in Represents the dimension after discretization. This refers to a one-hot encoding function, which primarily uses an N-bit state register to encode N states. Each state has its own independent register bit, and only one bit is active at any given time. Since click-through rate estimates are generally around 0... To make the model focus more on the predicted values ​​that vary around 0, the prior click-through rate information is discretized and then embedded using a high-dimensional embedding layer. Obtain the high-dimensional feature vector of the feature prior information, i.e., the first high-dimensional feature vector. .

[0075] In this optional implementation, the high-dimensional embedding layer, relative to the M×N matrix, can ensure that the discretized prior click-through rate information remains consistent across the high-dimensional feature vector.

[0076] The feature processing layer provided by this optional implementation includes a discretization layer and a high-dimensional embedding layer. Through the discretization processing of the discretization layer and the high-dimensional feature processing of the high-dimensional embedding layer, the reliability of prior click-through rate information processing is guaranteed.

[0077] In another embodiment of this disclosure, the feature processing layer further includes: an initial embedding layer, used to extract high-dimensional features of each feature in the sample to obtain a second high-dimensional feature vector, and to connect the first high-dimensional feature vector and the second high-dimensional feature vector to obtain a feature vector.

[0078] like Figure 3 As shown, the feature processing layer includes: a discretization layer, a high-dimensional embedding layer, and an initial embedding layer. The discretization layer and the high-dimensional embedding layer are two layers sequentially arranged with the feature prior layer. The feature prior layer is used to obtain the prior click-through rate information for each feature in the sample. The initial embedding layer is a layer parallel to the feature prior layer. The high-dimensional dense feature vectors of each feature in the sample can be directly extracted to obtain the second high-dimensional feature vector. in It is the total number of eigenvalues. It is the dimension of high-dimensional dense features. By merging and concatenating the second high-dimensional feature vector and the first high-dimensional feature vector, we can obtain the feature vector that can be directly input into the prior module of the model.

[0079] In this embodiment, the first high-dimensional feature vector It is the encoding of prior click-through rate information of the features, the second high-dimensional feature vector. It encodes the features of the sample by merging the second high-dimensional feature vector with the first high-dimensional feature vector to obtain the final feature vector. ,in This represents a merge operation.

[0080] The feature processing layer provided by this optional implementation includes an initial embedding layer, a discretization layer, and a high-dimensional embedding layer. Through the feature processing of the sample by the initial embedding layer, the discretization processing by the discretization layer, and the high-dimensional feature processing by the high-dimensional embedding layer, the information contained in the obtained feature vector can be more comprehensive, thereby improving the comprehensiveness of the information in the feature vector.

[0081] In some embodiments of this disclosure, the above-mentioned calculation of the feature loss value of the feature prior layer includes: constructing the loss function of the feature prior layer using the maximum likelihood estimation method; and obtaining the feature loss value of the feature prior layer based on the loss function of the feature prior layer.

[0082] In this optional implementation, in order to enable the feature prior layer to learn together with the model prior module and the feature processing layer, the loss function of the feature prior layer is obtained by maximizing the likelihood estimation of the click rate prediction value of each feature, as shown in Equation (1):

[0083] (1)

[0084] In equation (1), Labels representing samples, Represents incremental data. The first parameter represents the feature prior layer. Based on By utilizing the SGD (Stochastic Gradient Descent) optimizer, the feature prior layer can be optimized together with the model prior module and the feature processing layer as a whole.

[0085] The optional implementation provides a method for calculating the loss function of the feature prior layer. By employing the maximum likelihood estimation method to construct the loss function of the feature prior layer, it provides a reliable means for calculating the feature loss value of the feature prior layer and ensures the reliability of the feature loss value calculation.

[0086] In some embodiments of this disclosure, the above-mentioned calculation of the model loss value of the feature processing layer and the model prior module includes: constructing the cross-entropy loss function of the feature processing layer and the model prior module; and obtaining the model loss value of the feature processing layer and the model prior module based on the cross-entropy loss function.

[0087] In this optional implementation, the feature processing layer and the model prior module are treated as a whole, and a cross loss function of the whole is constructed. The model loss value is calculated through the cross loss function of the whole, providing a reliable means to obtain the model loss value of the feature processing layer and the model prior module.

[0088] One problem addressing the overfitting of the model on the dataset is that model updates only consider the likelihood function on the incremental data for optimization. To mitigate this issue, this disclosure provides posterior parameter estimates on the full dataset:

[0089]

[0090] Through Bayesian derivation, given the full dataset... Down Maximum a posteriori estimation is equivalent to optimizing the likelihood of incremental data. Lag estimation under old data .

[0091] Indicates time All previous datasets. The parameters of the feature processing layer and the model prior module in the model at the previous time step are denoted as... The task to be solved by this disclosure is: in a given and In this case, obtain the first parameter of the current optimal model (click-through rate prediction model). Second parameter .

[0092] In some embodiments of this disclosure, the above-mentioned calculation of the model loss value of the feature processing layer and the model prior module includes: constructing the cross-entropy loss function of the feature processing layer and the model prior module; constructing the distance function between the current parameter distribution and the historical parameter distribution of the feature processing layer and the model prior module; and obtaining the model loss value of the feature processing layer and the model prior module based on the cross-entropy loss function and the distance function.

[0093] In this optional implementation, the likelihood of incremental data... By assuming labels in the sample Following a Bernoulli distribution, the cross-entropy loss function of the feature processing layer and the model prior module can be derived. As shown in equation (2):

[0094] (2)

[0095] In equation (2) It's a tag. For incremental data, Indicates the second parameter The sum of the prediction operations performed on the samples by the lower feature processing layer and the model prior module. ,in It is the second parameter of the current model. Represents all operations. Represents the Sigmoid activation function. This represents the model's prior module. It relates to the second parameter based on past data. The posterior estimation is computationally difficult due to the large amount of data. One approach is to use the Laplace approximation, employing the distance between the current parameter distribution and the old model's parameter distribution as a substitute. Furthermore, normalizing the model's output can enhance learning stability and yield better results. Based on the above algorithm, the final posterior estimate for the old data is... Distance function As shown in equation (3):

[0096] (3)

[0097] Combination and The final objective function is shown in equation (4):

[0098] (4)

[0099] In equation (4), the cross-entropy loss function is combined with preset coefficients (as shown in equation (4)). The loss functions of the feature processing layer and the model prior module are obtained by adding the distance functions of multiples of 1.5. Based on the loss functions of the feature processing layer and the model prior module, the model loss values ​​of the feature processing layer and the model prior module are calculated, and the second parameter is optimized using the Adam optimizer.

[0100] This optional implementation provides a computational feature processing layer and a model prior module. Based on Ye Bayes' rule, it designs a model prior using cross-entropy. This prior approximates the posterior estimate on the complete data by maximizing the likelihood function on incremental data and minimizing the distance between the current model and the prior model. In practice, the output of previous models can be easily integrated into this framework to implement the model prior, where the output of previous models is used to supervise the current model. Thus, by maximizing the model's likelihood on incremental data and minimizing the distance between the current model and previous models, the goal of learning model parameters on the full data distribution is achieved.

[0101] In summary, the click-through rate prediction model training method disclosed herein can simultaneously enhance the robustness of model parameter learning at both the sample and feature levels. At the feature level, by constructing a feature prior module, it can provide the prior click probabilities for different features, enhancing the expressive power of features on incremental data, especially on long-tailed data distributions. At the sample level, by constructing the objective function of the model prior module, and based on Bayes' rule, it theoretically derives the goal of learning model parameters on the full data distribution by maximizing the model's likelihood on incremental data and minimizing the distance between the model and previous models.

[0102] Please see Figure 4 The diagram illustrates a flow 400 of an embodiment of the click-through rate (CTR) prediction method provided in this disclosure, which may include the following steps:

[0103] Step 401: Obtain the user data and product data to be tested.

[0104] In this embodiment, the execution entity running the click-through rate prediction method can communicate with a terminal (such as...). Figure 1 The system communicates with terminals 101 and 102 to obtain the user data and product data to be tested sent by the terminals.

[0105] In this embodiment, user data and product data can be related data from user interactions with products within the same webpage or application. For example, user data includes user attribute data and user behavior data. User attribute data includes user age, gender, etc. User behavior data includes the sequence of products clicked by the user in a historical period. Product data includes product type, product color, product origin, etc.

[0106] Step 402: Based on user data and product data, obtain user characteristics and product characteristics.

[0107] In this embodiment, after obtaining user data and product data, a feature extractor is used to extract features from the user data and product data to obtain user features and product features.

[0108] Step 403: Input user features and product features into the click-through rate prediction model generated by the click-through rate prediction model training method to obtain the click-through rate prediction value output by the click-through rate prediction model.

[0109] In this embodiment, the click-through rate (CTR) prediction model is trained based on the CTR prediction model training method of this embodiment. The CTR prediction model is generated using the CTR prediction model training method described above. The specific generation process of the CTR prediction model can be found in [link to relevant documentation]. Figure 2 The relevant descriptions of the embodiments will not be repeated here.

[0110] It should be noted that the click-through rate (CTR) prediction method in this embodiment can be used to test the CTR prediction models generated in the above embodiments. Furthermore, the CTR prediction model can be continuously optimized based on the test results. This method can also be a practical application of the CTR prediction models generated in the above embodiments. Using the CTR prediction models generated in the above embodiments to identify entity types in test sentences helps improve entity recognition efficiency.

[0111] The click-through rate (CTR) prediction method provided in this disclosure first acquires user data and product data to be tested; second, based on the user data and product data, user features and product features are obtained; finally, the user features and product features are input into a CTR prediction model generated by a CTR prediction model training method to obtain the CTR prediction value output by the CTR prediction model. Therefore, by using a pre-trained CTR prediction model to predict user data and product data, an accurate and rapid CTR prediction value can be obtained.

[0112] Further reference Figure 5 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of a click-through rate prediction model training device, which is similar to... Figure 2 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0113] like Figure 5 As shown, embodiments of this disclosure provide a click-through rate (CTR) prediction model training device 500, which includes: a sample acquisition unit 501, a network acquisition unit 502, a selection unit 503, an input unit 504, a calculation unit 505, and an acquisition unit 506. The sample acquisition unit 501 can be configured to acquire a preset sample set, which includes at least one sample, including user features and product features. The network acquisition unit 502 can be configured to acquire a pre-constructed CTR prediction network, which includes a feature prior module and a model prior module. The feature prior module includes a feature prior layer and a feature processing layer. The feature prior layer acquires prior CTR information for each feature in the sample, and the feature processing layer obtains the feature vector of the sample based on the prior CTR information. The model prior module predicts the CTR value from the feature vector. The selection unit 503 can be configured to select samples from the sample set. The input unit 504 can be configured to input the sample into the click-through rate (CTR) prediction network to obtain the CTR prediction value output by the network. The calculation unit 505 can be configured to calculate the feature loss value of the feature prior layer, the model loss value of the feature processing layer, and the model prior module, respectively. The obtaining unit 506 can be configured to use the CTR prediction network as the CTR prediction model in response to the CTR prediction network meeting the training completion condition.

[0114] In this embodiment, the specific processing of the sample acquisition unit 501, network acquisition unit 502, selection unit 503, input unit 504, calculation unit 505, and acquisition unit 506 in the click-through rate prediction model training device 500, and the resulting technical effects, can be found in the following references: Figure 2 The corresponding steps in the embodiment are 201, 202, 203, 204, 205, and 206.

[0115] In some embodiments, the feature processing layer includes a discretization layer and a high-dimensional embedding layer; the discretization layer is used to discretize the prior click-through rate information; the high-dimensional embedding layer is used to extract high-dimensional features of the discretized prior click-through rate information to obtain a first high-dimensional feature vector, and the first high-dimensional feature vector is used as a feature vector.

[0116] In some embodiments, the feature processing layer further includes: an initial embedding layer, used to extract high-dimensional features of each feature in the sample to obtain a second high-dimensional feature vector, and to connect the first high-dimensional feature vector and the second high-dimensional feature vector to obtain a feature vector.

[0117] In some embodiments, the above-mentioned computing unit 505 is further configured to: construct the loss function of the feature prior layer using the maximum likelihood estimation method; and obtain the feature loss value of the feature prior layer based on the loss function of the feature prior layer.

[0118] In some embodiments, the above-mentioned computing unit 505 is further configured to: construct the cross-entropy loss function of the feature processing layer and the model prior module; and obtain the model loss value of the feature processing layer and the model prior module based on the cross-entropy loss function.

[0119] In some embodiments, the above-mentioned calculation unit 505 is further configured to: construct the cross-entropy loss function of the feature processing layer and the model prior module; construct the distance function between the current parameter distribution and the historical parameter distribution of the feature processing layer and the model prior module; and obtain the model loss value of the feature processing layer and the model prior module based on the cross-entropy loss function and the distance function.

[0120] The click-through rate (CTR) prediction model training apparatus provided in the embodiments of this disclosure firstly includes a sample acquisition unit 501 acquiring a preset sample set, which includes at least one sample, comprising user features and product features. Secondly, a network acquisition unit 502 acquires a pre-constructed CTR prediction network, which includes a feature prior module and a model prior module. The feature prior module includes a feature prior layer and a feature processing layer. The feature prior layer is used to acquire prior CTR information for each feature in the sample, and the feature processing layer obtains the CTR based on the prior CTR information. The feature vector is used to predict the click-through rate (CTR) value. Next, the selection unit 503 selects samples from the sample set. Then, the input unit 504 inputs these samples into the CTR prediction network to obtain the CTR prediction value output by the network. Next, the calculation unit 505 calculates the feature loss value of the feature prior layer, the model loss value of the feature processing layer, and the model prior module, respectively. Finally, the acquisition unit 506, in response to the CTR prediction network meeting the training completion condition, uses the CTR prediction network as the CTR prediction model. Thus, by constructing a feature prior layer, different features can be provided with their own prior CTR values, enhancing the expressive power of each feature. The feature prior layer is optimized along with the overall CTR prediction network, enabling the model to learn the expression of each feature and improving the robustness and accuracy of the click prediction model.

[0121] Further reference Figure 6 As an implementation of the methods shown in the above figures, this disclosure provides an embodiment of a click-through rate prediction device, which is similar to... Figure 4 Corresponding to the method embodiments shown, this device can be specifically applied to various electronic devices.

[0122] like Figure 6 As shown, an embodiment of this disclosure provides a click-through rate (CTR) prediction device 600, which includes a data acquisition unit 601, a processing unit 602, and a prediction unit 603. The data acquisition unit 601 can be configured to acquire user data and product data to be tested. The processing unit 602 can be configured to obtain user features and product features based on the user data and product data. The prediction unit 603 can be configured to input the user features and product features into a CTR prediction model generated by a CTR prediction model training device to obtain a CTR prediction value output by the CTR prediction model.

[0123] In this embodiment, the click-through rate prediction model is trained using a click-through rate prediction model training device.

[0124] In this embodiment, the specific processing of the data acquisition unit 601, the processing unit 602, and the prediction unit 603, and the resulting technical effects, can be found by referring to [the relevant documentation]. Figure 4The corresponding steps are 401, 402, and 403 in the embodiment.

[0125] The following is for reference. Figure 7 It shows a schematic diagram of the structure of an electronic device 700 suitable for implementing embodiments of the present disclosure.

[0126] like Figure 7 As shown, the electronic device 700 may include a processing unit (e.g., a central processing unit, a graphics processor, etc.) 701, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 702 or a program loaded from a storage device 708 into a random access memory (RAM) 703. The RAM 703 also stores various programs and data required for the operation of the electronic device 700. The processing unit 701, ROM 702, and RAM 703 are interconnected via a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0127] Typically, the following devices can be connected to I / O interface 705: input devices 706 including, for example, touchscreens, touchpads, keyboards, mice, etc.; output devices 707 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 708 including, for example, magnetic tapes, hard disks, etc.; and communication devices 709. Communication device 709 allows electronic device 700 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 7 An electronic device 700 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively. Figure 7 Each box shown can represent a device or multiple devices as needed.

[0128] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 709, or installed from storage device 708, or installed from ROM 702. When the computer program is executed by processing device 701, it performs the functions defined in the methods of embodiments of this disclosure.

[0129] It should be noted that the computer-readable medium in the embodiments of this disclosure may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium may be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In the embodiments of this disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In the embodiments of this disclosure, a computer-readable signal medium may include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium may be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (Radio Frequency), etc., or any suitable combination thereof.

[0130] The aforementioned computer-readable medium may be included in the aforementioned server; or it may exist independently and not assembled into the server. The aforementioned computer-readable medium carries one or more programs. When the server executes one or more of these programs, the server causes the server to: acquire a pre-set sample set, the sample set including at least one sample, the sample including user features and product features; acquire a pre-constructed click-through rate (CTR) prediction network, the CTR prediction network including a feature prior module and a model prior module, wherein the feature prior module includes a feature prior layer and a feature processing layer, the feature prior layer being used to acquire prior CTR information for each feature in the sample, the feature processing layer obtaining the feature vector of the sample based on the prior CTR information; the model prior module being used to predict the CTR value from the feature vector; and perform the following training steps: select a sample from the sample set; input the sample into the CTR prediction network to obtain the CTR prediction value output by the CTR prediction network; calculate the feature loss value of the feature prior layer, the model loss value of the feature processing layer, and the model prior module respectively; and, in response to the CTR prediction network meeting the training completion condition, use the CTR prediction network as the CTR prediction model.

[0131] Computer program code for performing the operations of embodiments of this disclosure can be written in one or more programming languages ​​or a combination thereof. Programming languages ​​include object-oriented programming languages—such as Java, Smalltalk, and C++—and conventional procedural programming languages—such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a local area network (LAN) or a wide area network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0132] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0133] The units described in the embodiments of this disclosure can be implemented in software or hardware. The described units can also be housed in a processor; for example, it can be described as: a processor including a sample acquisition unit, a network acquisition unit, a selection unit, an input unit, a calculation unit, and an acquisition unit. The names of these units do not necessarily limit the unit itself; for example, the sample acquisition unit can also be described as a unit "configured to acquire a preset sample set, the sample set including at least one sample, the sample including: user features and product features".

[0134] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. A method for training a click-through rate (CTR) prediction model, the method comprising: Obtain a preset sample set, wherein the sample set includes at least one sample, and the sample includes: user features and product features; A pre-constructed click-through rate (CTR) prediction network is obtained. The CTR prediction network includes a feature prior module and a model prior module. The feature prior module includes a feature prior layer and a feature processing layer. The feature prior layer is used to obtain prior CTR information for each feature in the sample. The feature processing layer obtains the feature vector of the sample based on the prior CTR information. The model prior module is used to predict the CTR value from the feature vector. The following training steps are performed: Samples are selected from the sample set; the samples are input into the click-through rate (CTR) prediction network to obtain the CTR prediction value output by the network; the feature loss value of the feature prior layer, the model loss value of the feature processing layer, and the model prior module are calculated respectively, wherein the cross-entropy loss function of the feature processing layer and the model prior module is derived by assuming that the labels in the samples follow a Bernoulli distribution; the distance function of the posterior estimate under the old data is derived using the distance between the current parameter distribution and the old model parameter distribution through Laplace approximation; the model loss value of the feature processing layer and the model prior module is obtained based on the cross-entropy loss function and the distance function; in response to the CTR prediction network meeting the training completion condition, the CTR prediction network is used as the CTR prediction model.

2. The method according to claim 1, wherein the feature processing layer comprises: Discretization layer and high-dimensional embedding layer; The discretization layer is used to discretize the prior click-through rate information; the high-dimensional embedding layer is used to extract the high-dimensional features of the discretized prior click-through rate information to obtain a first high-dimensional feature vector, and the first high-dimensional feature vector is used as the feature vector.

3. The method according to claim 2, wherein the feature processing layer further comprises: An initial embedding layer is used to extract high-dimensional features of each feature in the sample to obtain a second high-dimensional feature vector. The first high-dimensional feature vector and the second high-dimensional feature vector are then connected to obtain the feature vector.

4. The method according to claim 2 or 3, wherein calculating the feature loss value of the feature prior layer includes: The loss function of the feature prior layer is constructed using the maximum likelihood estimation method. Based on the loss function of the feature prior layer, the feature loss value of the feature prior layer is obtained.

5. The method according to claim 2 or 3, wherein calculating the model loss value of the feature processing layer and the model prior module includes: Construct the cross-entropy loss function for the feature processing layer and the model prior module; Based on the cross-entropy loss function, the model loss values ​​of the feature processing layer and the model prior module are obtained.

6. The method according to claim 2 or 3, wherein calculating the model loss value of the feature processing layer and the model prior module comprises: Construct the cross-entropy loss function for the feature processing layer and the model prior module; Construct a distance function between the current parameter distribution and the historical parameter distribution of the feature processing layer and the model prior module; Based on the cross-entropy loss function and the distance function, the model loss values ​​of the feature processing layer and the model prior module are obtained.

7. A click-through rate (CTR) prediction method, the method comprising: Obtain the user data and product data to be tested; Based on the user data and product data, user characteristics and product characteristics are obtained; The user features and product features are input into the click-through rate prediction model generated by the click-through rate prediction model training method according to any one of claims 1-6, and the click-through rate prediction value output by the click-through rate prediction model is obtained.

8. A click-through rate prediction model training device, the device comprising: The sample acquisition unit is configured to acquire a preset sample set, the sample set including at least one sample, the sample including: user features and product features; The network acquisition unit is configured to acquire a pre-constructed click-through rate (CTR) prediction network, which includes a feature prior module and a model prior module. The feature prior module includes a feature prior layer and a feature processing layer. The feature prior layer acquires prior CTR information for each feature in the sample. The feature processing layer obtains a feature vector for the sample based on the prior CTR information. The model prior module predicts the CTR value from the feature vector. The selection unit is configured to select samples from the sample set; The input unit is configured to input the sample into the click-through rate prediction network to obtain the click-through rate prediction value output by the click-through rate prediction network; The computation unit is configured to calculate the feature loss value of the feature prior layer, the model loss value of the feature processing layer, and the model prior module, respectively. Specifically, the cross-entropy loss function of the feature processing layer and the model prior module is derived by assuming that the labels in the samples follow a Bernoulli distribution; the distance function of the posterior estimate under the old data is derived using the distance between the current parameter distribution and the old model parameter distribution through Laplace approximation; and the model loss values ​​of the feature processing layer and the model prior module are obtained based on the cross-entropy loss function and the distance function. The obtained unit is configured to use the click-through rate prediction network as a click-through rate prediction model in response to the click-through rate prediction network meeting the training completion condition.

9. A click-through rate (CTR) prediction device, the device comprising: The data acquisition unit is configured to acquire user data and product data to be tested. The processing unit is configured to obtain user characteristics and product characteristics based on the user data and product data; The prediction unit is configured to input the user features and the product features into the click-through rate prediction model generated by the click-through rate prediction model training device according to claim 8, and obtain the click-through rate prediction value output by the click-through rate prediction model.

10. An electronic device, comprising: One or more processors; A storage device on which one or more programs are stored; When the one or more programs are executed by the one or more processors, the one or more processors implement the method as described in any one of claims 1-7.

11. A computer-readable medium having a computer program stored thereon, wherein, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.

Images