Click rate prediction method, device, electronic device and medium

Abstract

The disclosed embodiments provide a click-through rate prediction method, a click-through rate prediction device, an electronic device, and a computer-readable medium; and relate to the field of big data processing technology. The click-through rate prediction method includes: collecting advertisement exposure data as sample data, and determining the label of the sample data according to the click field contained in the advertisement exposure data; extracting the explicit interaction features and implicit interaction features of the sample data, and obtaining a click-through rate prediction model based on the label of the sample data and combining the explicit interaction features and the implicit interaction features; predicting the click-through rate of the advertisement to be delivered through the click-through rate prediction model, and delivering the advertisement to be delivered according to the click-through rate. The technical solution of the disclosed embodiments can obtain richer features in the advertisement exposure data from different angles, thereby improving the accuracy of the click-through rate prediction model.

Description

Technical Field

[0001] The present disclosure relates to the technical field of big data processing, and in particular to a click rate prediction method, a click rate prediction device, an electronic device, and a computer-readable medium. Background Art

[0002] The emergence of online advertising has led traditional advertising industries such as television and newspapers to develop towards emerging Internet forms such as live broadcasts and short videos. As a result, a large amount of behavioral data has been accumulated, making it possible to mine detailed user interests from the data. Faced with a large number of users and advertisements, how to push an online advertisement to the right users to maximize the number of clicks is a core issue that needs to be studied in online advertising.

[0003] Ad click-through rate prediction technology can predict the probability that users may click on ads. On the one hand, it can push ads that users are interested in and reduce the time users spend searching. On the other hand, it can also provide advertisers with valuable reference information to improve the effectiveness of ads. Features play a core role in ad prediction. Input features are usually sparse and high-dimensional. Since the original features rarely produce the best results, effective predictions usually rely on high-order combination features. For example, young working women are more interested in cosmetics ads, and school boys are more interested in sports ads. Therefore, data scientists spend a lot of time and energy manually interacting with the original features to obtain the best prediction results. The generation cost of traditional manual features is high, requiring data scientists to have rich experience. In addition, the interactive features do not include some implicit and invisible features, resulting in insufficient extraction of effective interactive features. In recent years, machine learning and deep learning have made amazing achievements in various fields such as natural language processing and image processing, which has led the industry to study and explore how to apply machine learning and deep learning technologies to ad click-through rate prediction scenarios.

[0004] Currently, many models use factorization machines as the generator of embedded features. However, factorization machines only maintain one feature vector for each feature domain, which will lose too much information and cannot fully express the features. For example, the meanings of men interacting with advertising categories and men interacting with age are different, but the factorization machine treats them as the same, which affects the prediction accuracy of the model.

[0005] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore may include information that does not constitute the prior art known to ordinary technicians in the field. Summary of the invention

[0006] The purpose of the embodiments of the present disclosure is to provide a click-through rate prediction method, a click-through rate prediction device, an electronic device and a computer-readable medium. By extracting explicit interaction features and implicit interaction features, richer features in the advertising exposure data can be obtained from different angles, thereby improving the accuracy of the click-through rate prediction model.

[0007] Other features and advantages of the present disclosure will become apparent from the following detailed description, or may be learned in part by the practice of the present disclosure.

[0008] According to a first aspect of an embodiment of the present disclosure, a click-through rate prediction method is provided, comprising: collecting advertisement exposure data as sample data, and using a target feature contained in the advertisement exposure data as a label of the sample data; wherein the advertisement exposure data comprises a plurality of features corresponding to an advertisement, and the target feature is one of the plurality of features; the label is used to characterize whether the advertisement is clicked; performing explicit feature interaction on the plurality of features in the sample data to obtain a display interaction feature, and performing implicit feature interaction on the plurality of features in the sample data to obtain an implicit interaction feature; based on the label of the sample data, a click-through rate prediction model is obtained in combination with the explicit interaction feature and the implicit interaction feature; wherein the click-through rate prediction model is used to predict the click-through rate of an advertisement to be delivered, and the click-through rate is used to determine whether to deliver the advertisement to be delivered.

[0009] In an exemplary embodiment of the present disclosure, performing explicit feature interaction on multiple features in the sample data to obtain explicit interaction features, and performing implicit feature interaction on multiple features in the sample data to obtain implicit interaction features include: generating embedding vectors of multiple feature domains of the sample data; performing outer product operations on the embedding vectors of different feature domains to obtain explicit interaction features; and performing dot product operations on the embedding vectors of different feature domains to obtain implicit interaction features.

[0010] In an exemplary embodiment of the present disclosure, the embedding vector includes a first embedding vector and a second embedding vector; the generating of the embedding vector of the multiple feature domains of the sample data includes: generating the first embedding vector of the multiple feature domains of the sample data by a factor decomposition machine; and obtaining the second embedding vector of the multiple feature domains of the sample data by a field-aware decomposition machine.

[0011] In an exemplary embodiment of the present disclosure, the method of obtaining a click-through rate prediction model by combining the explicit interaction features and the implicit interaction features includes: determining a first effective weight of the explicit interaction features and a second effective weight of the implicit interaction features through an attention mechanism; inputting the displayed interaction features, the implicit interaction features, and the first effective weight and the second effective weight into the output layer of the click-through rate prediction model, wherein the output layer takes a value range of 0 to 1.

[0012] In an exemplary embodiment of the present disclosure, after collecting the advertisement exposure data as sample data, the method further includes: encoding the sample data to convert the sample data into a sparse feature vector, and using the sparse feature vector as an embedding vector of the sample data.

[0013] In an exemplary embodiment of the present disclosure, the collecting of advertising exposure data as sample data includes: extracting sample data and test data according to the timestamp of the advertising exposure data, wherein the timestamp of the sample data is before the test data; the test data is used to test the click-through rate prediction model after the click-through rate prediction model is obtained.

[0014] In an exemplary embodiment of the present disclosure, the click-through rate is used to determine whether to deliver the advertisement to be delivered, including: if the click-through rate output by the click-through rate prediction model for the advertisement to be delivered exceeds a preset threshold, the advertisement to be delivered is delivered, wherein the preset threshold value ranges from 0 to 1.

[0015] According to a second aspect of an embodiment of the present disclosure, a click rate prediction device is provided, which may include a data acquisition module, a model acquisition module and an advertisement prediction module.

[0016] Among them, the data acquisition module is used to collect advertising exposure data as sample data, and use the target feature contained in the advertising exposure data as the label of the sample data; wherein the advertising exposure data includes multiple features corresponding to the advertisement, and the target feature is one of the multiple features; the label is used to characterize whether the advertisement is clicked; the feature extraction module is used to perform explicit feature interaction on multiple features in the sample data to obtain display interaction features, and perform implicit feature interaction on multiple features in the sample data to obtain implicit interaction features; the model acquisition module is used to obtain a click-through rate prediction model based on the label of the sample data, combined with the explicit interaction features and the implicit interaction features; wherein the click-through rate prediction model is used to predict the click-through rate of the advertisement to be delivered, and the click-through rate is used to determine whether to deliver the advertisement to be delivered.

[0017] In an exemplary embodiment of the present disclosure, the feature extraction module may include an embedding vector generation unit, an explicit interaction feature acquisition unit, and an implicit interaction feature acquisition unit.

[0018] Among them, the embedding vector generation unit is used to generate embedding vectors of multiple feature domains of the sample data; the explicit interaction feature acquisition unit is used to perform outer product operations on the embedding vectors of different feature domains to obtain explicit interaction features; the implicit interaction feature acquisition unit is used to perform dot product operations on the embedding vectors of different feature domains to obtain implicit interaction features.

[0019] In an exemplary embodiment of the present disclosure, the embedding vector generation unit can be configured to: generate a first embedding vector of multiple feature domains of the sample data through a factor decomposition machine; obtain a second embedding vector of multiple feature domains of the sample data through a field-aware decomposition machine, and the embedding vector includes a first embedding vector and a second embedding vector.

[0020] In an exemplary embodiment of the present disclosure, the model acquisition module may include an attention mechanism unit and an output layer unit.

[0021] Among them, the attention mechanism unit is used to determine the first effective weight of the explicit interaction feature and the second effective weight of the implicit interaction feature through the attention mechanism; the output layer unit is used to input the displayed interaction feature, the implicit interaction feature, the first effective weight and the second effective weight into the output layer of the click-through rate prediction model, wherein the value range of the output layer is 0 to 1.

[0022] In an exemplary embodiment of the present disclosure, the click rate prediction device further includes a sample encoding module for encoding the sample data to convert the sample data into a sparse feature vector, and using the sparse feature vector as an embedding vector of the sample data.

[0023] In an exemplary embodiment of the present disclosure, the data acquisition module may be configured to: extract sample data and test data according to the timestamp of the advertising exposure data, wherein the timestamp of the sample data is before the test data; and test the click-through rate prediction model using the test data.

[0024] In an exemplary embodiment of the present disclosure, the data acquisition module may be configured to deliver the advertisement if the click rate output by the click rate prediction model for the advertisement exceeds a preset threshold, wherein the preset threshold ranges from 0 to 1.

[0025] According to a third aspect of an embodiment of the present disclosure, an electronic device is provided, comprising: one or more processors; a storage device for storing one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors implement the click rate prediction method as described in the first aspect of the above embodiment.

[0026] According to a fourth aspect of an embodiment of the present disclosure, a computer-readable medium is provided, on which a computer program is stored. When the program is executed by a processor, the click rate prediction method as described in the first aspect of the above embodiment is implemented.

[0027] The click-through rate prediction method, click-through rate prediction device, electronic device and computer-readable medium provided in the embodiments of the present disclosure can avoid manual labeling and save manpower and time costs by using the target features contained in the advertising exposure data as labels for the sample data; and extract the display interaction features and implicit interaction features of the sample data. By combining explicit and implicit methods, the relationship between the features can be fully considered to obtain more effective features and improve the accuracy of the model; in addition, there is no need to manually extract interaction features, which can improve the dimension and granularity of the feature combination.

[0028] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] The drawings herein are incorporated into the specification and constitute a part of the specification, showing embodiments consistent with the present disclosure, and together with the specification, are used to explain the principles of the present disclosure. Obviously, the drawings described below are only some embodiments of the present disclosure, and for ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:

[0030] Figure 1 A schematic diagram of an exemplary system architecture of a click rate prediction method or a click rate prediction device applied to an embodiment of the present disclosure is schematically shown;

[0031] Figure 2 A flowchart of a click rate prediction method according to an embodiment of the present disclosure is schematically shown;

[0032] Figure 3 The schematic diagram of the structure of the click rate prediction model in one embodiment of the present disclosure is shown;

[0033] Figure 4 A flowchart of a click rate prediction method according to another embodiment of the present disclosure is schematically shown;

[0034] Figure 5A block diagram of a click rate prediction device according to an embodiment of the present disclosure is schematically shown;

[0035] Figure 6 A schematic diagram of the structure of a computer system suitable for implementing an electronic device of an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0036] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this disclosure will be more comprehensive and complete and will fully convey the concept of the example embodiments to those skilled in the art.

[0037] In addition, the described features, structures or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present disclosure. However, those skilled in the art will appreciate that the technical solutions of the present disclosure may be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other cases, known methods, devices, implementations or operations are not shown or described in detail to avoid blurring the various aspects of the present disclosure.

[0038] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0039] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to actual conditions.

[0040] In this specification, the terms "a", "an", "the", "said" and "at least one" are used to indicate the presence of one or more elements / components / etc.; the terms "comprising", "including" and "having" are used to express an open-ended inclusion, and mean that there may be additional elements / components / etc. in addition to the listed elements / components / etc.; the terms "first", "second", "third", etc. are used only as labels and are not intended to limit the quantity of their objects.

[0041] The exemplary embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0042] Figure 1A schematic diagram showing a system architecture of an exemplary application environment that can be applied to the click rate prediction method or click rate prediction device of the embodiments of the present disclosure.

[0043] like Figure 1 As shown, the system architecture 100 may include one or more of terminal devices 101, 102, 103, a network 104, and a server 105. The network 104 is used to provide a medium for communication links between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired, wireless communication links, or optical fiber cables, etc.

[0044] Users can use terminal devices 101, 102, 103 to interact with server 105 through network 104 to receive or send messages, etc. Terminal devices 101, 102, 103 can be various electronic devices with display screens and supporting web browsing, including but not limited to desktop computers, portable computers, smart phones and tablet computers, wearable devices, virtual reality devices, smart homes, etc.

[0045] The server 105 may be a server that provides various services, such as a background management server that provides support for devices operated by users using the terminal devices 101, 102, and 103. The background management server may analyze and process the received request data, and feed back the processing results to the terminal device.

[0046] For example, the server 105 may collect advertising exposure data as sample data, and determine the label of the sample data according to the target features contained in the advertising exposure data; the server 105 may perform explicit feature interaction on multiple features in the sample data to obtain display interaction features, and perform implicit feature interaction on multiple features in the sample data to obtain implicit interaction features, and obtain a click-through rate prediction model by combining the explicit interaction features and the implicit interaction features; obtain the advertisement to be delivered, predict the click-through rate of the advertisement to be delivered by using the click-through rate prediction model, and deliver the advertisement to be delivered to the client 101 according to the click-through rate.

[0047] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only for illustration. According to the implementation requirements, there may be any number of terminal devices, networks and servers. For example, the server 105 may be a server cluster composed of multiple servers.

[0048] The click rate prediction method provided in the embodiment of the present disclosure is generally executed by the server 105, and accordingly, the click rate prediction device is generally set in the server 105. However, it is easy for those skilled in the art to understand that the click rate prediction method provided in the embodiment of the present disclosure can also be executed by the terminal devices 101, 102, and 103, and accordingly, the click rate prediction device can also be set in the terminal devices 101, 102, and 103, which is not particularly limited in this exemplary embodiment.

[0049] Based on this, the embodiments of the present disclosure provide a technical solution for a click-through rate prediction method, which can extract explicit interaction features and implicit interaction features from advertising exposure data, thereby obtaining richer and more effective combined features and improving the accuracy of the click-through rate prediction model.

[0050] like Figure 2 As shown, a click rate prediction method provided by an embodiment of the present disclosure may include step S21, step S22 and step S23, wherein the execution subjects of steps S21, S22 and S23 may be the server 105 or the terminals 101, 102, 103, and each step may be executed by the same execution subject or by different execution subjects respectively.

[0051] In step S21, advertising exposure data is collected as sample data, and a target feature contained in the advertising exposure data is used as a label of the sample data; wherein the advertising exposure data includes multiple features corresponding to the advertisement, and the target feature is one of the multiple features; and the label is used to indicate whether the advertisement is clicked.

[0052] Advertisement exposure data can be obtained by collecting client log records. Log records can not only record the advertisements displayed by the client and the user's click operations on the advertisements, but also accurately record the timestamp of the advertisement display. Therefore, the advertisement exposure data can include user information and advertisement information, such as user ID, advertisement ID, etc. In addition, other information can be further queried from the database using the obtained user information and advertisement information, such as the user's gender, age, occupation, user registration time, etc.; for example, advertisement size, advertisement type, advertiser, advertisement corresponding delivery group, etc. The advertisement exposure data can also include information and timestamps of the advertisement position, such as the APP where the advertisement is displayed. In addition, the advertisement exposure data can also include other fields, such as the type of network used to display the advertisement, client information, etc., which are not specifically limited in this embodiment.

[0053] The advertisement exposure data may include a click field used to characterize whether the user clicks on the advertisement. The click field can be used as a target feature, and the value of the click field can be used to distinguish whether the user clicks on the advertisement. For example, when an advertisement is exposed to a user, an advertisement exposure data is recorded in the client corresponding to the user. If the user clicks on the advertisement, the click field of the advertisement exposure data can be 1, and if the user does not click on the advertisement, the click field of the advertisement exposure data can be 0. The click field in the advertisement exposure data can be used to determine the label of each advertisement exposure data. For example, if the value of the click field is 1, the label of the advertisement exposure data is 1, and if the value of the click field is 2, the label of the advertisement exposure data is 2.

[0054] The ad exposure data collected from each client can be used as sample data for training the click-through rate prediction model. In addition, after obtaining the sample data, the original sample data can also be processed. For example, the data can be cleaned to eliminate useless data, such as removing outliers, such as removing data where the user's age is greater than 150 years old; filling in missing values, such as filling in missing user gender according to the average value, or uniformly filling missing fields with special values, etc.; normalizing data values ​​to between 0 and 1, keeping the weights of each field in the sample data consistent, etc.

[0055] In an exemplary embodiment, the original sample data can be encoded to convert the sample data into a vector. For example, for the "gender" field in the sample data, it can be divided into two cases: male and female. If the value of the field is male, it can be encoded as "10" accordingly, and if it is female, it can be encoded as "01" accordingly. Exemplarily, the encoding method can be one-hot encoding. A sparse feature vector is a sparse representation of a vector, but when the feature domain is large, using a sparse representation of a vector can greatly save space. For example, for the feature v = [0, 0, 0, 0, 1, 0, 3, 0, 0, 0], its sparse representation is (10, [4, 6], [1, 3]).

[0056] In an exemplary embodiment, the collected advertising exposure data can be divided into sample data and test data according to the timestamp of the advertising exposure data. In order to simulate the actual advertising delivery scenario, the advertising exposure data can be sorted according to the time order of the timestamp, and then a certain proportion of the data with a time earlier can be extracted as sample data. Compared with the sample data, the timestamp of the test data is later, that is, the timestamp of the test data is closer to the current time than the sample data. For example, after all the advertising exposure data are sorted by time, the time range of the timestamp is from January 1, 2000 to October 10, 2000, then it can be divided from September 1, 2000, and the data before September 1, 2000 is used as sample data, and the data after September 1, 2000 is used as test data, so as to be consistent with the actual prediction situation. Exemplarily, the sample data and the test data can be divided according to a certain ratio, for example, the ratio can be 9:1, or 8:1, 7.6:1, etc., and the present embodiment is not limited to this.

[0057] Among them, the test data can be used to test the click-through rate prediction model after the click-through rate prediction model is obtained. For example, after obtaining the click-through rate prediction model, the test data can be used as input, and the prediction result corresponding to the test data can be determined by the click-through rate prediction model, and then the model prediction result can be judged whether it is correct by the label of the test data. If the model's prediction result for the test data is the same as the label of the test data, the model prediction is correct; if the prediction result is different from the label, it can be determined that the model prediction is wrong. Furthermore, the prediction accuracy of the click-through rate prediction model can be calculated through the test data; or the click-through rate prediction model can be further optimized using the test data whose prediction result is inconsistent with the label.

[0058] In step S22, explicit feature interaction is performed on multiple features in the sample data to obtain explicit interaction features, and implicit feature interaction is performed on multiple features in the sample data to obtain implicit interaction features.

[0059] Positive samples and negative samples can be determined according to the labels of the sample data. For example, when the label is 1, it indicates that the user clicked on the advertisement, and the sample data can be used as a positive sample. When the label is 0, it indicates that the user did not click on the advertisement, and the sample data can be used as a negative sample. The features corresponding to each type of label can be determined by the different labels of the sample data. The label can be used as the output of the click-through rate prediction model, and the sample data can be used as the input, so that the click-through rate prediction model learns the mapping relationship between input and output, label and feature, and determines the various parameters of the model. The click-through rate prediction model can include an explicit interaction feature part and an implicit interaction feature part. The explicit interaction feature part can explicitly perform feature interaction on the sample data, while the implicit interaction feature part is responsible for implicitly performing feature interaction on the sample data. Among them, the explicit interaction feature and the implicit interaction feature refer to the features obtained by cross-combining the features of different feature domains. For example, the sample data has three feature domains, namely: gender, age, and occupation. The combination can obtain personality age features, personality occupation features, and age occupation features. The feature domain can be understood as a set of features of the same data type, such as a set of positive integers in the age feature domain, a set of "male" and "female" characters in the gender domain, etc. Specifically, the method of extracting explicit interaction features and implicit interaction features may include: generating embedding vectors of multiple feature domains of sample data; performing outer product operations on the embedding vectors to obtain explicit interaction features; and performing dot product operations on the embedding vectors to obtain implicit interaction features.

[0060] First, it is necessary to generate an embedding vector for the sample data in order to interact with the features of the sample data. The sparse feature vector after encoding the sample data can be directly used as an embedding vector. Or after converting the sample data into a sparse feature vector, an embedding vector can be generated by FM (Factorization Machine). FM can generate a one-dimensional embedding vector for each feature, and the number of feature domains is the same as the total number of generated embedding vectors; FFM (Field Perceptual Decomposition Machine) can also be used to generate a multi-dimensional embedding vector for each feature domain. In this embodiment, after the sample data is converted into a vector through the input layer, a first embedding vector can be generated by FM, and a second embedding vector can be generated by FFM, and the first embedding vector and the second embedding vector can be used together as the embedding vector.

[0061] Then, the outer product operation is performed on the embedding vectors to obtain explicit interaction features. For example, if the embedding vector of one feature domain is (a1, b1, c1) and the embedding vector of another feature domain is (a2, b2, c2), then the form of interaction is (w*(a1*a2, b1*b2, c1*c2)), where w is a learnable parameter of the click rate prediction model. For implicit interaction features, based on the generated embedding vector, the embedding vectors are used to implicitly interact with each other between features of different feature domains, so the features are crossed by dot product operation. The dot product operation refers to extracting features in the dimension of points with each point in the vector as a unit, that is, the features of each point in the vector are learned through parameters. For example, if there are two embedding vectors (a1, b1, c1) and (a2, b2, c2), then the form of interaction is (w1*a1*a2, w2*b1*b2, w3*c1*c2), where w1, w2, and w3 are learnable parameters of the click prediction model. Therefore, explicit interaction features have good interpretability for the click-through rate prediction model, while implicit interaction features can extract the implicit information behind the data, thereby improving the accuracy of the click-through rate prediction model.

[0062] In step S23, based on the label of the sample data, a click-through rate prediction model is obtained in combination with the explicit interaction features and the implicit interaction features; wherein the click-through rate prediction model is used to predict the click-through rate of the advertisement to be delivered, and the click-through rate is used to determine whether to deliver the advertisement to be delivered.

[0063] In an exemplary embodiment, the structure of the click rate prediction model can be as follows: Figure 3 As shown, the process of generating explicit interaction features and implicit interaction features is as follows:

[0064] refer to Figure 3 As shown, after the sample data is converted into a feature vector through the input layer 310, it can be output to FM320 and FFM 330 at the same time. The feature vector is converted into a first embedding vector using the FM structure of 320, and the feature vector is converted into a second embedding vector using the FFM structure of 330. Since the generated embedding vectors FM and FFM both have a first-order linear part, in order to avoid redundancy, the first-order linear part can be separated to optimize the model. The linear part in the feature vector is uniformly extracted through the linear part 370. The formula of the linear part can be:

[0065]

[0066] Among them, m is the number of feature domains corresponding to a sample data, w i Represents input x i The corresponding parameter, b is the bias parameter.

[0067] Then, the explicit interaction feature part 340 can be combined with the FM 320 to generate an explicit interaction feature. Specifically, the explicit interaction feature part can be a CIN structure, based on the first embedding vector generated by the FM as the input of the CIN, and the CIN structure includes multiple convolutional layers, each of which can interact with the input first embedding vector. For example, the first embedding vector generated by the FM is represented by X 0 , then X 0 ∈R m*k , where m is the number of feature domains and k is the embedding vector dimension. 0 Finally, the intermediate layer generated by the CIN structure is combined with X 0 Perform outer product operation at the vector level, the formula is as follows:

[0068]

[0069] in, Represents X i The j-th dimension of the embedding vector is k, is the original feature X 0 The lth dimension in is the k-th embedding vector, and · is represented as the outer product. This can generate a three-dimensional feature map, on which multiple one-dimensional convolution kernels can extract rich explicit interaction features from multiple angles, as shown in the following formula:

[0070]

[0071] Among them, H i represents the number of convolution kernels in the i-th layer, W represents the parameters of the convolution kernel, represents the j-th dimension embedding vector corresponding to the i-th layer, Represents the l-th dimension embedding vector corresponding to the original feature. Then perform the sum pooling maximum pooling operation to compress the vector into a one-dimensional vector, as shown in the following formula:

[0072]

[0073] Among them, k represents the interactive dimension of the embedding vector, i+1 represents the i+1th layer of the CIN structure, and j represents the output generated by the jth convolution kernel. Finally, the one-dimensional vectors generated by each layer are concatenated as the output of the explicit interactive feature part, and the output of the explicit interactive feature part is denoted as P.

[0074] The second embedding vector extracted by FFM 330 is input into the implicit interaction feature part 350, which may be a DNN structure. The input second embedding vector is processed by the DNN structure 350 to generate an implicit interaction feature. Exemplarily, the two-dimensional embedding vector generated by FFM is output to the DNN structure for higher-order interaction between features. The DNN may use a three-layer fully connected layer, and the activation function may be a linear rectifier function relu, as shown in the following formula:

[0075] l=f(W3f(W2f(W1I+b1)+b2)+b3),

[0076] Among them, f represents the relu activation function, W1, W2, W3 represent the fully connected weights, and b1, b2, b3 are bias parameters.

[0077] Since the explicit interaction features and implicit interaction features of the embedded vector are extracted by CIN and DNN, rich combination features are obtained, but the generated high-order combination features are not effectively selected, and the importance of different features is also different. For example, the intersection of gender features and age features, and the intersection between gender features and item category features, the intersection between gender features and item category features is more important. Therefore, the click-through rate prediction model provided by this embodiment can also include an attention mechanism part, which can be connected to the implicit interaction feature part and the explicit interaction feature part respectively, so as to perform feature selection on the output implicit interaction features and explicit interaction features, and extract useful features. For example, through the SENET structure, the model can automatically learn the importance between different interaction features, increase the weight of important features, and reduce the weight of useless features, so that the model is more in line with the scenario of click-through rate prediction. In this embodiment, the first effective weight of the displayed interaction feature and the second effective weight of the implicit interaction feature can be determined by the attention mechanism.

[0078] The attention mechanism part 360 can be divided into three steps. The first step is to compress the features to obtain global features, the second step is to learn the importance of the features, and the third step is to recalibrate the original features. Specifically, for the explicit interaction feature part, the output X of 340 i , first use the formula: Compression is performed, and on this basis, the excitation operation is used to obtain the importance of features. The formula is: i =sigmoid(W2relu(W1sq i ), where relu and sigmoid are activation functions, and W1 and W2 are full connection parameters. Finally, the scale operation is performed to obtain the feature X of each original output. iMultiply by a first effective weight to represent the importance of this feature, expressed as: X i =ex i *X i Similarly, for the implicit interaction feature part, the second-order interaction features generated by FFM are also recalibrated with the SENET structure to screen important cross-features. Specifically, for the squeeze operation, the formula is expressed as The one-dimensional vector is represented as sq, and then the excitation step is performed, the formula is ex=sigmoid(W2relu(W1sq)), where relu and sigmoid are activation functions, W1 and W2 are full connection parameters, and the full connection parameters can be different from the full connection parameters of the explicit interaction feature part. Finally, the scale operation is performed to determine the second effective weight for the implicit interaction feature as shown in the following formula: I=I*ex, so that the SENET module is introduced into the model to play the role of screening important features.

[0079] In an exemplary embodiment, after obtaining the extracted explicit interaction features and implicit interaction features, all the obtained features can be used to determine the final prediction result through the output layer 380. The output layer 380 of the click-through rate prediction model can use a sigmoid activation function, the value range of which is 0 to 1, and the closer the value is to 1, the greater the probability that the sample data is clicked. The formula of the output layer can be:

[0080]

[0081] in, is the linear part, l, w l is the output of the implicit interaction feature part and the corresponding parameters, P, W P is the output and corresponding parameters of the explicit interaction feature part. Since the label corresponding to the sample data is the value of the output layer y, the value of y and the linear part, explicit interaction feature part, and implicit interaction feature part extracted from the sample data are used to learn and determine the unknown parameters of each part, thereby completing the model training and obtaining the click-through rate prediction model.

[0082] It should be understood that each layer in the click-through rate prediction model has a corresponding activation function. The activation function can be understood as a function expression of the mapping relationship between input and output. In other embodiments of the present disclosure, the click-through rate prediction model can be constructed by other activation functions, such as custom functions, etc., which also belong to the protection scope of the present disclosure. After obtaining the trained click-through rate prediction model, the information of the advertisement to be placed can be obtained, such as the advertisement size, advertisement type, advertiser, target group type of advertisement placement, etc. of the advertisement to be placed; and the information of the client user can be obtained, such as the user's gender, age, occupation, etc. The information of the advertisement to be placed and the user's information obtained are used as the input of the click-through rate prediction model, so that the predicted click-through rate of the advertisement to be placed can be determined, that is, the probability that the advertisement to be placed is clicked by the user. If the click-through rate is large, the advertisement to be placed can be pushed to the user, which can maximize the delivery effect of the advertisement to be placed. Exemplarily, it is determined whether the click-through rate exceeds the preset value. If the click-through rate exceeds the preset value, the advertisement to be placed is placed. If it does not exceed the preset value, the advertisement to be placed may not be placed, thereby more accurately controlling the cost of placing advertisements and the effect of placing advertisements. The preset value may be 0.5, 0.6, etc., or other values ​​may be set according to actual needs, and this implementation does not impose any special limitation on this.

[0083] In an exemplary embodiment, the above click rate prediction method may include steps S41 to S46. Figure 4 shown.

[0084] In step S41, collect advertising logs to obtain advertising exposure data. In step S42, perform data cleaning and normalization on the advertising exposure data. In step S43, divide the processed advertising exposure data into a training set and a test set according to the timestamp; specifically, the data with an earlier timestamp can be the training set, and the data with a timestamp after the training set can be divided into the test set. In step S44, construct a click-through rate prediction model, and train the model using the training set; use the data contained in the training set as sample data to train the click-through rate prediction model. In step S45, test the trained click-through rate prediction model using the test set; use each piece of advertising exposure data in the test set as input to determine the prediction result of the click-through rate prediction model, and use the label of the advertising exposure data to determine the accuracy of the click-through rate model prediction; if the accuracy of the click-through rate prediction model does not meet the requirements, more advertising exposure data can be obtained again to update the training set, and the click-through rate prediction model can be trained again. In step S46, the trained click rate prediction model is used to predict the newly exposed advertisement to be delivered to the user, and the predicted click rate of the advertisement to be delivered is determined. The higher the predicted click rate, the greater the probability that the user will click on the advertisement. Based on the predicted click rate, it can be determined whether the user needs to be exposed to the advertisement.

[0085] The following describes an embodiment of the device disclosed in the present invention, which can be used to execute the click rate prediction method described above in the present invention. Figure 5 The click rate prediction device 50 provided in the embodiment of the present disclosure may include: a data acquisition module 51, a model acquisition module 52 and an advertisement prediction module 53.

[0086] Among them, the data acquisition module 51 is used to collect advertising exposure data as sample data, and use the target feature contained in the advertising exposure data as a label of the sample data; wherein the advertising exposure data includes multiple features corresponding to the advertisement, and the target feature is one of the multiple features; the label is used to indicate whether the advertisement is clicked.

[0087] The model acquisition module 52 is used to perform explicit feature interaction on multiple features in the sample data to obtain explicit interaction features, and to perform implicit feature interaction on multiple features in the sample data to obtain implicit interaction features.

[0088] The advertisement prediction module 53 is used to obtain a click rate prediction model based on the label of the sample data and in combination with the explicit interaction features and the implicit interaction features; wherein the click rate prediction model is used to predict the click rate of the advertisement to be delivered, and the click rate is used to determine whether to deliver the advertisement to be delivered.

[0089] In an exemplary embodiment of the present disclosure, the feature extraction module 52 may include an embedding vector generation unit, an explicit interaction feature acquisition unit, and an implicit interaction feature acquisition unit.

[0090] Among them, the embedding vector generation unit is used to generate embedding vectors of multiple feature domains of the sample data; the explicit interaction feature acquisition unit is used to perform outer product operations on the embedding vectors of different feature domains to obtain explicit interaction features; the implicit interaction feature acquisition unit is used to perform dot product operations on the embedding vectors of different feature domains to obtain implicit interaction features.

[0091] In an exemplary embodiment of the present disclosure, the embedding vector generation unit can be configured to: generate a first embedding vector of multiple feature domains of the sample data through a factor decomposition machine; obtain a second embedding vector of multiple feature domains of the sample data through a field-aware decomposition machine, and the embedding vector includes a first embedding vector and a second embedding vector.

[0092] In an exemplary embodiment of the present disclosure, the model acquisition module 53 may include an attention mechanism unit and an output layer unit.

[0093] Among them, the attention mechanism unit is used to determine the first effective weight of the explicit interaction feature and the second effective weight of the implicit interaction feature through the attention mechanism; the output layer unit is used to input the displayed interaction feature, the implicit interaction feature, the first effective weight and the second effective weight into the output layer of the click-through rate prediction model, wherein the activation function of the output layer has a value range of 0 to 1.

[0094] In an exemplary embodiment of the present disclosure, the click rate prediction device further includes a sample encoding module for encoding the sample data to convert the sample data into a sparse feature vector, and using the sparse feature vector as an embedding vector of the sample data.

[0095] In an exemplary embodiment of the present disclosure, the data acquisition module 51 may be configured to: extract sample data and test data according to the timestamp of the advertising exposure data, wherein the timestamp of the sample data is before the test data; and test the click-through rate prediction model using the test data.

[0096] In an exemplary embodiment of the present disclosure, the data acquisition module 51 may be configured to deliver the advertisement if the click rate output by the click rate prediction model for the advertisement exceeds a preset threshold, wherein the preset threshold ranges from 0 to 1.

[0097] Since the various functional modules of the click rate prediction device of the example embodiment of the present disclosure correspond to the steps of the example embodiment of the click rate prediction method described above, for details not disclosed in the embodiment of the device of the present disclosure, please refer to the embodiment of the click rate prediction method described above in the present disclosure.

[0098] Reference below Figure 6 , which shows a schematic diagram of the structure of a computer system 600 suitable for implementing an electronic device of an embodiment of the present disclosure. Figure 6 The computer system 600 of the electronic device shown is only an example and should not bring any limitation to the functions and scope of use of the embodiments of the present disclosure.

[0099] like Figure 6 As shown, the computer system 600 includes a central processing unit (CPU) 601, which can perform various appropriate actions and processes according to the program stored in the read-only memory (ROM) 602 or the program loaded from the storage part 608 to the random access memory (RAM) 603. In the RAM 603, various programs and data required for system operation are also stored. The CPU 1201, ROM 602 and RAM 603 are connected to each other through a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0100] The following components are connected to the I / O interface 605: an input section 606 including a keyboard, a mouse, etc.; an output section 607 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 608 including a hard disk, etc.; and a communication section 609 including a network interface card such as a LAN card, a modem, etc. The communication section 609 performs communication processing via a network such as the Internet. A drive 610 is also connected to the I / O interface 605 as needed. A removable medium 611, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 610 as needed, so that a computer program read therefrom is installed into the storage section 608 as needed.

[0101] In particular, according to an embodiment of the present disclosure, the process described above with reference to the flowchart can be implemented as a computer software program. For example, an embodiment of the present disclosure includes a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program includes a program code for executing the method shown in the flowchart. In such an embodiment, the computer program can be downloaded and installed from a network through a communication part 609, and / or installed from a removable medium 611. When the computer program is executed by a central processing unit (CPU) 601, the above-mentioned functions defined in the system of the present application are executed.

[0102] It should be noted that the computer-readable medium shown in the present disclosure may be a computer-readable signal medium or a computer-readable storage medium or any combination of the above two. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of computer-readable storage media may include, but are not limited to: an electrical connection with one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present disclosure, a computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in combination with an instruction execution system, device or device. In the present disclosure, a computer-readable signal medium may include a data signal propagated in a baseband or as part of a carrier wave, in which a computer-readable program code is carried. This propagated data signal may take a variety of forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. Computer-readable signal media may also be any computer-readable medium other than computer-readable storage media, which may send, propagate or transmit a program for use by or in conjunction with an instruction execution system, apparatus or device. The program code contained on the computer-readable medium may be transmitted using any appropriate medium, including but not limited to: wireless, wire, optical cable, RF, etc., or any suitable combination of the above.

[0103] The flow charts and block diagrams in the accompanying drawings illustrate the possible architecture, functions and operations of the systems, methods and computer program products according to various embodiments of the present disclosure. In this regard, each box in the flow chart or block diagram can represent a module, a program segment, or a part of a code, and the above-mentioned module, program segment, or a part of a code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in a different order from the order marked in the accompanying drawings. For example, two boxes represented in succession can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram or flow chart, and the combination of the boxes in the block diagram or flow chart can be implemented with a dedicated hardware-based system that performs a specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0104] The units involved in the embodiments described in the present disclosure may be implemented by software or hardware, and the units described may also be arranged in a processor. The names of these units do not, in some cases, constitute limitations on the units themselves.

[0105] As another aspect, the present application also provides a computer-readable medium, which may be included in the electronic device described in the above embodiment; or may exist independently without being assembled into the electronic device. The above computer-readable medium carries one or more programs, and when the above one or more programs are executed by an electronic device, the electronic device implements the click rate prediction method described in the above embodiment.

[0106] For example, the electronic device can implement Figure 2 As shown in: Step S21, collecting advertising exposure data as sample data, and taking the target feature contained in the advertising exposure data as the label of the sample data; wherein the advertising exposure data includes multiple features corresponding to the advertisement, and the target feature is one of the multiple features; the label is used to characterize whether the advertisement is clicked; Step S22, performing explicit feature interaction on the multiple features in the sample data to obtain display interaction features, and performing implicit feature interaction on the multiple features in the sample data to obtain implicit interaction features; Step S23, based on the label of the sample data, combining the explicit interaction features and the implicit interaction features to obtain a click-through rate prediction model; wherein the click-through rate prediction model is used to predict the click-through rate of the advertisement to be delivered, and the click-through rate is used to determine whether to deliver the advertisement to be delivered.

[0107] For example, the electronic device can achieve Figure 4 The steps shown.

[0108] It should be noted that, although several modules or units of the device for action execution are mentioned in the above detailed description, this division is not mandatory. In fact, according to the embodiments of the present disclosure, the features and functions of two or more modules or units described above can be embodied in one module or unit. On the contrary, the features and functions of one module or unit described above can be further divided into multiple modules or units to be embodied.

[0109] Through the description of the above implementation, it is easy for those skilled in the art to understand that the example implementation described here can be implemented by software, or by software combined with necessary hardware. Therefore, the technical solution according to the implementation of the present disclosure can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (which can be a CD-ROM, a USB flash drive, a mobile hard disk, etc.) or on a network, including several instructions to enable a computing device (which can be a personal computer, a server, a touch terminal, or a network device, etc.) to execute the method according to the implementation of the present disclosure.

[0110] 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 that are not disclosed in the present disclosure. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present disclosure are indicated by the following claims.

[0111] It should be understood that the present disclosure is not limited to the exact structures that have been described above and shown in the drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims

1. A click rate prediction method, characterized in that: include: Collecting advertisement exposure data as sample data, and using a target feature contained in the advertisement exposure data as a label of the sample data; wherein the advertisement exposure data includes a plurality of features corresponding to an advertisement, and the target feature is one of the plurality of features; and the label is used to indicate whether the advertisement is clicked; Performing outer product operations on the embedding vectors of different feature domains to obtain explicit interaction features, and performing dot product operations on the embedding vectors of different feature domains to obtain implicit interaction features; Wherein, the embedding vector is a multidimensional embedding vector generated for each feature domain by a field-aware factorization machine FFM; or, when the embedding vector includes a first embedding vector and a second embedding vector, the sample data is simultaneously output to the factorization machine FM and the FFM, the first embedding vector is an embedding vector of multiple feature domains of the sample data generated by the FM, and the total number of the first embedding vectors is the same as the number of feature domains, and the second embedding vector is an embedding vector of multiple feature domains of the sample data obtained by the FFM; Based on the labels of the sample data, a click-through rate prediction model is obtained in combination with the explicit interaction features and the implicit interaction features; wherein the click-through rate prediction model is used to predict the click-through rate of the advertisement to be delivered, and the click-through rate is used to determine whether to deliver the advertisement to be delivered.

2. The method according to claim 1, characterized in that The step of combining the explicit interaction features and the implicit interaction features to obtain a click rate prediction model includes: Determining a first effective weight of the explicit interaction feature and a second effective weight of the implicit interaction feature through an attention mechanism; The explicit interaction feature, the implicit interaction feature, the first effective weight and the second effective weight are input into the output layer of the click rate prediction model, wherein the value range of the output layer is 0 to 1.

3. The method according to claim 1, characterized in that After collecting the advertisement exposure data as sample data, the method further includes: The sample data is encoded to convert the sample data into a sparse feature vector, and the sparse feature vector is used as an embedding vector of the sample data.

4. The method according to claim 1, characterized in that The collecting of advertisement exposure data as sample data includes: Extracting sample data and test data according to the timestamp of the advertisement exposure data, wherein the timestamp of the sample data is before the timestamp of the test data; The test data is used to test the click rate prediction model after the click rate prediction model is obtained.

5. The method according to claim 2, characterized in that: The click rate is used to determine whether to place the advertisement to be placed, including: If the click rate output by the click rate prediction model for the advertisement to be delivered exceeds a preset threshold, the advertisement to be delivered is delivered, wherein the preset threshold has a value range of 0 to 1.

6. A click rate prediction device, characterized in that: include: A data acquisition module, used to collect advertisement exposure data as sample data, and use the target feature contained in the advertisement exposure data as a label of the sample data; wherein the advertisement exposure data includes multiple features corresponding to the advertisement, and the target feature is one of the multiple features; the label is used to indicate whether the advertisement is clicked; A feature extraction module, used for simultaneously outputting the sample data to a factorization machine FM and a field perception factorization machine FFM, and generating first embedding vectors of multiple feature domains of the sample data through the FM, and obtaining second embedding vectors of multiple feature domains of the sample data through the FFM; using the first embedding vector and the second embedding vector as embedding vectors, performing outer product operations on the embedding vectors of different feature domains to obtain explicit interaction features; performing dot product operations on the embedding vectors of different feature domains to obtain implicit interaction features; wherein, extracting linear parts from the first embedding vector and the second embedding vector to optimize the click rate prediction model based on the linear parts; A model acquisition module is used to acquire a click-through rate prediction model based on the label of the sample data in combination with the explicit interaction features and the implicit interaction features; wherein the click-through rate prediction model is used to predict the click-through rate of the advertisement to be delivered, and the click-through rate is used to determine whether to deliver the advertisement to be delivered.

7. An electronic device, characterized in that: include: one or more processors; a storage device for storing one or more programs, When the one or more programs are executed by the one or more processors, the one or more processors implement the click rate prediction method according to any one of claims 1 to 5.

8. A computer readable medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the click rate prediction method according to any one of claims 1 to 5 is implemented.

Images