Method for predicting click rate and conversion rate of advertisement
By constructing advertising knowledge graphs and knowledge graph representation learning, and combining multi-layer perceptrons to process advertising and user feature information, the problem that existing models fail to effectively integrate advertising features is solved, and the prediction accuracy of advertising click-through rate and conversion rate is improved.
Patent Information
- Application Number
- CN202411413039.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-11
- Publication Date
- 2024-11-08
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing advertising click-through rate and conversion rate prediction models fail to effectively integrate the ad's own characteristic information, resulting in insufficient prediction accuracy.
The KGC module is used to build an advertising knowledge graph, and the knowledge graph representation is learned through the GE module, an embedded representation with advertising feature information is obtained, and a multi-layer perceptron processing is carried out in combination with user feature information to predict click-through rate and conversion rate.
By considering more dimensions of ad feature information, the prediction accuracy of click-through rate and conversion rate is improved.
Smart Images

Figure CN118917894A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of Internet advertising, and in particular to a method for predicting the click rate and conversion rate of an advertisement. Background Art
[0002] Click-through rate prediction and conversion rate prediction models have made significant progress in the past few years, but the current models only consider the advertising information and user information when pushed to users, ignoring the characteristic information of the advertisement itself and the connection between advertisements. For example, the attributes of the advertisement itself, such as the province where the advertisement is placed and the restrictions on the advertisement position, should be better expressed in the model; for similar advertisements, they should be closer in the vector space, but the traditional model fails to effectively capture this semantic connection. Therefore, how to integrate the characteristic information of the advertisement itself into the model and improve the accuracy of predicting click-through rate and conversion rate is the focus of this invention. Summary of the invention
[0003] The purpose of the present invention is to provide a method for predicting the click-through rate and conversion rate of an advertisement, which overcomes the above-mentioned defects and can integrate advertisement feature information in the process of predicting the click-through rate and conversion rate to improve the accuracy of predicting the click-through rate and conversion rate.
[0004] To achieve the above object, the solution of the present invention is: A method for predicting the click-through rate and conversion rate of an advertisement comprises a KGC module, a GE module and a P-CTR-CVR module, and comprises the following steps: S10: the KGC module extracts the advertisement feature information of an advertisement ad to form a triple with ad_advertisement id as a head entity, and constructs each triple to form an advertisement knowledge graph; S20: the GE module comprises a knowledge graph representation learning model, and the GE module inputs these triples with ad_advertisement id as a head entity into the knowledge graph representation learning model, thereby obtaining an embedded representation T of the advertisement id with the advertisement feature information. ad , and embed the ad ID with ad feature information into T ad Input to the P-CTR-CVR module; S30: The P-CTR-CVR module includes a PCTR module and a PCVR module. The PCTR module is used to predict the click-through rate, and the PCVR module is used to predict the conversion rate. The PCTR module has a click-through rate multilayer perceptron, and the PCVR module has a conversion rate multilayer perceptron. The PCTR module can receive the embedded representation T of the advertising ID with advertising feature information. ad , the embedded representation of the advertising features transmitted by users in real scenarios T an And the embedding representation T of the user features transmitted by the user in the real scene um , the embedded representation of advertisement features transmitted by multiple users in real scenarios T anand the embedded representation T of the ad ID with ad feature information ad The feature dataset T that constitutes the advertisement AD ,Right now , the embedding representation T of user features transmitted by multiple users in real scenarios um The feature dataset T that constitutes the user USER ,Right now , embedding the ad ID with ad feature information into T ad The embedded representation T of advertisement features transmitted by multiple users in real scenarios an The advertisement embedding representation a is obtained by fusion in a summation manner. ctr , embedding the user features transmitted by multiple users in real scenarios into a representation T um The user’s embedding representation u is obtained by fusion in a summation manner ctr , the embedded representation of the advertisement is a ctr and the user’s embedding representation u ctr The fusion operator is used to fuse the user comprehensive information embedding representation s ctr , embed the user comprehensive information into representation s ctr After being processed by the click rate multi-layer perceptron, the output data H ctr , data H ctr After Sigmoid function processing, the click rate prediction probability is obtained ctr ; The parameters of the PCVR module and the PCTR module are shared. The PCVR module embeds the advertisement ID with advertisement feature information into a representation T ad and the embedded representation of advertising features transmitted by multiple users in real scenarios T an The advertisement embedding representation a is obtained by fusion in a summation manner. cvr , embedding the user features transmitted by multiple users in real scenarios into a representation T um The user’s embedding representation u is obtained by fusion in a summation manner cvr , the embedded representation of the advertisement is a cvr and the user’s embedding representation u cvr The fusion operator is used to fuse the user comprehensive information embedding representation s cvr , embed the user comprehensive information into representation s cvr After being processed by the conversion rate multilayer perceptron, the output data H cvr , data H cvr After Sigmoid function processing, the predicted probability of conversion rate is obtained cvr .
[0005] Furthermore, in step S20, when different advertisements have the same feature values for different advertisement features, a hyperplane is introduced and projected onto the hyperplane, and a scoring function based on the hyperplane is used to obtain an embedded representation T of the advertisement ID with advertisement feature information. ad , the scoring function based on the hyperplane is defined as follows: , in, and is the projection vector of ad id and ad features, It represents the embedded representation of the relationship. The specific formula of the projection process is as follows: , , in, is the mapping matrix of each relationship, is to perform the transpose operation.
[0006] Furthermore, the P-CTR-CVR module uses binary cross entropy loss (BCE) to measure the errors of click-through rate prediction and conversion rate prediction respectively. The loss functions are: , , P ctr and P cvr is the true label, and the final total loss is L ctr and L cvr Add, By introducing the L2 norm, the final loss function is: , in, is the regularization parameter, is the model parameter.
[0007] Furthermore, it also includes an ECPM calculation module, and the P-CTR-CVR module outputs the click rate prediction probability ctr And the predicted probability of conversion rate cvr The data is transmitted to the ECPM calculation module, and the ECPM value is calculated by the ECPM calculation module. The calculation formula is: , in, Bid for your ad.
[0008] Furthermore, in step S10, the KGC module divides the advertising feature information of the advertisement ad into three categories, namely, fixed-length features, variable-length features and mixed features. The fixed-length feature means that the advertising feature can only be linked to another unique advertising feature under a link relationship; the variable-length feature means that the advertising feature can be linked to multiple different advertising features under a link relationship; the mixed feature means that there is a constraint relationship between the advertising feature and other advertising features; different types are extracted respectively to form triples with ad_advertisement id as the head entity, and each triple is constructed to form an advertising knowledge graph.
[0009] After adopting the above scheme, the beneficial effects of the present invention are: The present invention uses the KGC module to construct a triple with ad_advertisement id as the head entity using the advertisement feature information of the advertisement ad, thereby constructing a knowledge graph. The GE module extracts these triples and inputs them into the learning model represented by the knowledge graph, thereby obtaining an embedded representation of the advertisement id with advertisement feature information. ad , and these embeddings are represented as T ad The input is sent to the P-CTR-CVR module, and the click-through rate and conversion rate are predicted respectively by the P-CTR-CVR module. In the prediction process, the embedded representation T of the ad ID with ad feature information is used. ad , so that the present invention can more comprehensively understand the characteristics of each advertisement, and can consider more dimensional factors in the prediction process of click-through rate and conversion rate, thereby being able to improve the accuracy of predicting click-through rate and conversion rate. BRIEF DESCRIPTION OF THE DRAWINGS
[0010] Figure 1 A schematic diagram of a process of the present invention. DETAILED DESCRIPTION
[0011] The present invention is described in detail below with reference to the accompanying drawings and specific embodiments.
[0012] In the claims, specification and the above-mentioned drawings of the present invention, unless otherwise explicitly defined, directional words, such as the use of terms such as "center", "lateral", "longitudinal", "horizontal", "vertical", "top", "bottom", "inside", "outside", "up", "down", "front", "back", "left", "right", "clockwise", "counterclockwise" and the like to indicate directions or positional relationships are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the referred device or element must have a specific orientation or be constructed and operated in a specific orientation, so it cannot be understood as limiting the specific protection scope of the present invention.
[0013] like Figure 1As shown, the present invention provides a method for predicting the click-through rate and conversion rate of an advertisement, which has a KGC (Knowledge Graph Construction) module, a GE (Get Eembedding) module, and a P-CTR-CVR (Predicting Click-through Rate and Conversion Rate) module, and includes the following steps: S10: The KGC module extracts the advertising feature information of the advertisement ad (advertisement) to form a triple with ad_advertisement id as the head entity, and constructs each triple to form an advertising knowledge graph; S20: The GE module includes a knowledge graph representation learning model. The GE module inputs these triples with ad_advertising id as the head entity into the knowledge graph representation learning model to obtain the embedded representation T of the advertising id with advertising feature information. ad , and embed the ad ID with ad feature information into T ad Input to P-CTR-CVR module; S30: The P-CTR-CVR module includes a PCTR module and a PCVR module. The PCTR module is used to predict the click-through rate, and the PCVR module is used to predict the conversion rate. The PCTR module has a click-through rate multilayer perceptron, and the PCVR module has a conversion rate multilayer perceptron. The PCTR module can receive the embedded representation of the advertising ID with advertising feature information. ad , the embedded representation of the advertising features transmitted by users in real scenarios T an And the embedding representation T of the user features transmitted by the user in the real scene um , the embedded representation of advertisement features transmitted by multiple users in real scenarios T an and the embedded representation T of the ad ID with ad feature information ad The feature dataset T that constitutes the advertisement AD ,Right now , the embedding representation T of user features transmitted by multiple users in real scenarios um The feature dataset T that constitutes the user USER ,Right now , embedding the ad ID with ad feature information into T ad The embedded representation T of advertisement features transmitted by multiple users in real scenarios an The advertisement embedding representation a is obtained by fusion in a summation manner. ctr , embedding the user features transmitted by multiple users in real scenarios into a representation T um The user’s embedding representation u is obtained by fusion in a summation manner ctr , the embedded representation of the advertisement is a ctrand the user’s embedding representation u ctr The fusion operator is used to fuse the user comprehensive information embedding representation s ctr , embed the user comprehensive information into representation s ctr After being processed by the click rate multi-layer perceptron, the output data H ctr , data H ctr After Sigmoid function processing, the click rate prediction probability is obtained ctr ; The parameters of the PCVR module and the PCTR module are shared. The PCVR module embeds the advertisement ID with advertisement feature information into a representation T. ad and the embedded representation of advertising features transmitted by multiple users in real scenarios T an The advertisement embedding representation a is obtained by fusion in a summation manner. cvr , embedding the user features transmitted by multiple users in real scenarios into a representation T um The user’s embedding representation u is obtained by fusion in a summation manner cvr , the embedded representation of the advertisement is a cvr and the user’s embedding representation u cvr The fusion operator is used to fuse the user comprehensive information embedding representation s cvr , embed the user comprehensive information into representation s cvr After being processed by the conversion rate multilayer perceptron, the output data H cvr , data H cvr After Sigmoid function processing, the predicted probability of conversion rate is obtained cvr .
[0014] Specifically, in step S10, the main function of the KGC (Knowledge Graph Construction) module is to construct an advertising knowledge graph. The knowledge graph is a graphical data representation model used to describe the relationship and attributes between entities. It is usually stored in the form of triples. Its format is , where nodes (h, t) represent entities, h is usually called the head entity and t is called the tail entity, and edge r represents the relationship between the head entity and the tail entity. This structure helps to organize and express complex knowledge associations, making information easier to understand and reason. Part of the advertising feature information is shown in Table 1. In order to construct an advertising knowledge graph, the present invention divides the feature information of the advertisement into three categories, namely fixed-length features, variable-length features and mixed features. The fixed-length feature is that the advertising feature can only be linked to another unique advertising feature under a link relationship; the variable-length feature is that the advertising feature can be linked to multiple different advertising features under a link relationship; the mixed feature is that there is a constraint relationship between the advertising feature and other advertising features; extraction is performed according to different types to form triples with ad_advertisement id as the head entity, and each triple is constructed to form an advertising knowledge graph. Therefore, in the construction of the advertising knowledge graph, these three features all have ad_advertisement id as the head entity.
[0015] Table 1 Advertisement feature information table
[0016] For fixed-length features, in order to accurately distinguish different features, the present invention uses the combination of the header and the data in the table as the tail entity representation, and uses the header as a relationship to form a specific triple; taking au_carrier in Table 1 as an example, according to this construction rule, the present invention can generate triples in the following forms: (ad_44795027595071400, au_carrier, au_carrier_3), (ad_80349356275929000, au_carrier, au_carrier_4), etc.; When processing indefinite-length features, in order to construct a knowledge graph more accurately, the present invention concatenates each data in the list with the header to form a new entity, and also regards the header as a relationship; taking the sce feature in Table 1 as an example, according to the construction rule, the present invention can generate triples in the following form: (ad_44795027595071400, sce, sce_3), (ad_80349356275929000, sce, sce_1), (ad_80349356275929000, sce, sce_2).
[0017] When processing mixed features, in view of the constraint relationship between the two columns of features, the present invention uses one of the features as the constraint feature and the other as the constrained feature, and processes them according to their respective construction rules. Then, they are spliced in the order of constraint feature_constrained feature to form a new entity, and the header of the constraint feature is used as a relationship representation; taking au_app and au_app_json in Table 1 as examples, au_app is the constraint feature and au_app_json is the constrained feature. According to the above construction rules, the present invention can generate triples in the following forms: (ad_44795027595071400, au_app_json, au_app_2_au_app_json_4), (ad_80349356275929000, au_app_json, au_app_3_au_app_json_5), etc.
[0018] Furthermore, in step S20, when different advertisements have the same feature values for different advertisement features, this problem can be alleviated by introducing a hyperplane: and They have different representations, and they are only known by projecting them onto a specific hyperplane, and the scoring function based on the hyperplane is defined as follows: , in, and is the projection vector of ad id and ad features, Embedding representation of ad features, is the embedding representation of advertising features under different hyperplanes, where T ad is the embedding representation of the ad id in the triple, It represents the embedded representation of the relationship. The specific formula of the projection process is as follows: , , in, is the mapping matrix of each relationship, is to transpose it, In order to distinguish positive and negative triplets, a margin-based loss function is used: , Here, γ refers to the relationship between ad features and ad ID; In order to reduce the generation of false negative examples, we use The probability of replacing the head entity is Indicates the average number of advertising features corresponding to the advertising ID under this feature. Represents the average number of ads connected to the feature value under each feature. After continuous iterative optimization, by introducing a hyperplane, the embedded representation T of the ad ID with ad feature information is obtained. ad .
[0019] Specifically, the present invention sets several different fusion operators The embedding of the advertisement is represented by a ctr and the user’s embedding representation u ctr Fusion: Additive Fusion:
[0020] Splicing and fusion: ,in Represents a splicing operation Gated Fusion:
[0021] After fusion operation After that, the present invention obtains the user comprehensive information embedded representation s ctr , after being processed by the click rate multi-layer perceptron, the output data H ctr , the process can be expressed as: , Finally, data H ctr After Sigmoid function processing, we get the predicted probability of the user clicking on this ad. ctr , the process can be expressed as: , In the conversion rate prediction process, the fusion operator used in the present invention is the same as the fusion operator used in the click rate prediction process, and the fusion process is also the same, but the parameters of the conversion rate multilayer perceptron and the click rate multilayer perceptron used are different, so the output H cvr and H ctr Different, so H cvr After Sigmoid function processing, the conversion rate prediction probability is obtained cvr Different; specifically, the click-through rate multilayer perceptron and the conversion rate multilayer perceptron are multilayer perceptrons with different parameters obtained through training. The training process is an existing technical means and will not be described in detail here.
[0022] Furthermore, the P-CTR-CVR module uses binary cross entropy loss (BCE) to measure the errors of click-through rate prediction and conversion rate prediction respectively. The loss functions are: , , P ctr and Pcvr is the true label, and the final total loss is L ctr and L cvr Add, By introducing the L2 norm, the final loss function is: , in, is the regularization parameter, is the model parameter.
[0023] Furthermore, it also includes an ECPM calculation module, and the P-CTR-CVR module outputs the click rate prediction probability ctr And the predicted probability of conversion rate cvr The data is transmitted to the ECPM calculation module, and the ECPM value is calculated by the ECPM calculation module. The calculation formula is: , in, For advertisement bidding, the present invention selects the advertisement with the maximum ECPM value and recommends it to the user.
[0024] To further illustrate the various embodiments, the present invention provides drawings. These drawings are part of the disclosure of the present invention, which are mainly used to illustrate the embodiments and can be used in conjunction with the relevant descriptions in the specification to explain the operating principles of the embodiments. With reference to these contents, a person of ordinary skill in the art should be able to understand other possible implementations and advantages of the present invention. The components in the figures are not drawn to scale, and similar component symbols are generally used to represent similar components.
[0025] The above description is only a preferred embodiment of the present invention and is not a limitation on the design of this case. Any equivalent changes made based on the design key of this case shall fall within the protection scope of this case.
Claims
1. A method for predicting the click rate and conversion rate of an advertisement, characterized in that: The method comprises a KGC module, a GE module and a P-CTR-CVR module, and includes the following steps: S10: The KGC module extracts the advertisement feature information of the advertisement ad to form a triple with ad_advertisement id as the head entity, and constructs each triple to form an advertisement knowledge graph; S20: The GE module includes a knowledge graph representation learning model. The GE module inputs these triples with ad_advertising id as the head entity into the knowledge graph representation learning model to obtain the embedded representation T of the advertising id with advertising feature information. ad , and embed the ad ID with ad feature information into T ad Input to P-CTR-CVR module; S30: The P-CTR-CVR module includes a PCTR module and a PCVR module. The PCTR module is used to predict the click-through rate, and the PCVR module is used to predict the conversion rate. The PCTR module has a click-through rate multilayer perceptron, and the PCVR module has a conversion rate multilayer perceptron. The PCTR module can receive the embedded representation of the advertising ID with advertising feature information. ad , the embedded representation of the advertising features transmitted by users in real scenarios T an And the embedding representation T of the user features transmitted by the user in the real scene um , the embedded representation of advertisement features transmitted by multiple users in real scenarios T an and the embedded representation T of the ad ID with ad feature information ad The feature dataset T that constitutes the advertisement AD ,Right now , the embedding representation T of user features transmitted by multiple users in real scenarios um The dataset T that constitutes the user's features USER ,Right now , embedding the ad ID with ad feature information into T ad The embedded representation T of advertisement features transmitted by multiple users in real scenarios an The advertisement embedding representation a is obtained by fusion in a summation manner. ctr , embedding the user features transmitted by multiple users in real scenarios into a representation T um The user’s embedding representation u is obtained by fusion in a summation manner ctr , the embedded representation of the advertisement is a ctr and the user’s embedding representation u ctr The fusion operator is used to fuse the user comprehensive information embedding representation s ctr , embed the user comprehensive information into representation s ctr After being processed by the click rate multi-layer perceptron, the output data H ctr , data H ctr After Sigmoid function processing, the click rate prediction probability is obtained ctr ; The parameters of the PCVR module and the PCTR module are shared. The PCVR module embeds the advertisement ID with advertisement feature information into a representation T. ad and the embedded representation of advertising features transmitted by multiple users in real scenarios T an The advertisement embedding representation a is obtained by fusion in a summation manner. cvr , embedding the user features transmitted by multiple users in real scenarios into a representation T um The user’s embedding representation u is obtained by fusion in a summation manner cvr , the embedded representation of the advertisement is a cvr and the user’s embedding representation u cvr The fusion operator is used to fuse the user comprehensive information embedding representation s cvr , embed the user comprehensive information into representation s cvr After being processed by the conversion rate multilayer perceptron, the output data H cvr , data H cvr After Sigmoid function processing, the predicted probability of conversion rate is obtained cvr .
2. The method for predicting the click rate and conversion rate of an advertisement according to claim 1, characterized in that: In step S20, when different advertisements have the same feature values for different advertisement features, a hyperplane is introduced and projected onto the hyperplane, and a scoring function based on the hyperplane is used to obtain an embedded representation T of the advertisement ID with advertisement feature information. ad , the scoring function based on the hyperplane is defined as follows: , in, and is the projection vector of ad id and ad features, It represents the embedded representation of the relationship. The specific formula of the projection process is as follows: , , in, is the mapping matrix of each relationship, is to perform the transpose operation.
3. The method for predicting the click rate and conversion rate of an advertisement according to claim 1, characterized in that: The P-CTR-CVR module uses binary cross entropy loss (BCE) to measure the errors of click-through rate prediction and conversion rate prediction respectively. The loss functions are: , , P ctr and P cvr is the true label, and the final total loss is L ctr and L cvr Add, By introducing the L2 norm, the final loss function is: , in, is the regularization parameter, is the model parameter.
4. The method for predicting the click rate and conversion rate of an advertisement according to claim 1, characterized in that: It also includes the ECPM calculation module, and the P-CTR-CVR module outputs the click rate prediction probability ctr And the predicted probability of conversion rate cvr The data is transmitted to the ECPM calculation module, and the ECPM value is calculated by the ECPM calculation module. The calculation formula is: , in, Bid for your ad.
5. The method for predicting the click rate and conversion rate of an advertisement according to claim 1, characterized in that: In step S10, the KGC module divides the advertisement feature information of the advertisement ad into three categories, namely fixed-length features, variable-length features and mixed features. The fixed-length features are that the advertisement feature can only be linked to another unique advertisement feature under a link relationship; the variable-length features are that the advertisement feature can be linked to multiple different advertisement features under a link relationship; The mixed feature is that there is a constraint relationship between the advertising feature and other advertising features; they are extracted according to different types to form triples with ad_advertising id as the head entity, and each triple is constructed to form an advertising knowledge graph.
