System and method for multi-dimensional knowledge transfer for predicting click-through rate

Abstract

A multidimensional knowledge transfer model for predicting and calculating ad click-through rates (CTRs). The multidimensional knowledge transfer model includes a preprocessor for constructing an ad group node graph based on similarities between ad group nodes, constructing an ad campaign node graph by merging ad group node graphs, and constructing an ad account node graph by merging ad campaign node graphs. The multidimensional knowledge transfer model also includes a multi-knowledge click-through rate (CTR) prediction model for each layer: the ad account, ad campaign, and ad group layers. The multi-knowledge CTR prediction model predicts the CTR of each node based on the ad account node graph, ad campaign node graph, or ad group node graph, audience characteristics, characteristics of nodes for which CTRs have been predicted, and the hidden vector of its parent node extracted from the upper-layer multi-knowledge CTR prediction model.

Description

Technical Field

[0001] The present invention relates to calculating click-through rate (CTR) prediction in advertisements, and in particular to a technique for performing CTR prediction using machine learning (ML) and deep learning (DL) models. Background Art

[0002] CTR prediction is the task of predicting the probability that some artificial content, such as text, images, video clips, or sound clips, which typically represent advertisements and are displayed on websites or any widely accessible online electronic user interface, will be clicked or visited when displayed to viewers. Advertisements typically contain a link called a Uniform Resource Locator (URL); the primary goal of such ads is to attract visitors to a destination or landing page, or to direct online traffic to or use a specific website or online electronic user interface by clicking on or visiting the ad.

[0003] CTR prediction is particularly important for advertisers when planning targeted advertising when planning computational advertising campaigns. A computational advertising plan includes, at a minimum, the creation of ad content, the selection of searchable keywords, ad placement, budget, and other parameters tailored to the target audience. Therefore, CTR prediction is crucial for properly adjusting advertising plans to optimize campaign performance.

[0004] Various ML and DL models have been developed for CTR prediction. These machine learning models rely on a large amount of historical data collected from various advertising channels when calculating prediction results. Some ML and DL models used in this field include Support Vector Machines (SVM), which construct a hyperplane or a set of hyperplanes in high-dimensional or infinite-dimensional space for classification and regression; factorization models, which are a class of collaborative filtering algorithms that decompose the user-item interaction matrix into the product of two low-dimensional rectangular matrices; Factorization Machines (FM), which combine the advantages of SVMs with factorization models for universal predictors with any real-valued feature vectors; and FM-based DL models.

[0005] However, the accuracy of these ML prediction models is often hampered by data imbalance and the cold start problem during their training. Data imbalance occurs when there is abundant data on certain aspects of a campaign, such as the type of product being promoted or specific audience demographics, but very little data on others. The cold start problem occurs when attempting to predict CTR in a new campaign with few, if any, known parameters and / or parameter values; thus, there is a lack of historical data to properly train the ML prediction model. Summary of the Invention

[0006] The object of the present invention is to solve the above-mentioned imbalanced data problem and cold start problem by providing a system and method for CTR prediction using one or more knowledge transfer models, wherein the knowledge transfer model includes one or more of a hierarchical knowledge transfer model, a horizontal knowledge transfer model and a multidimensional knowledge transfer model.

[0007] The hierarchical knowledge transfer model addresses the data imbalance problem by considering the hierarchical structure of typical functionality and data organization in computational advertising. This can be viewed as a top-level structure consisting of one or more ad account nodes, followed by a middle-level structure consisting of one or more ad campaign nodes, and then a bottom-level structure consisting of one or more ad group nodes. Each ad account node has a hierarchical relationship with one or more ad campaign nodes below it. Each ad campaign node has a hierarchical relationship with one ad account node above it and one or more ad group nodes below it. Each ad group node has a hierarchical relationship with one ad campaign node above it. The higher the node's level, the more data it possesses. Therefore, in the hierarchical knowledge transfer model, a CTR prediction model for the ad account node is first trained to generate a representation vector for the ad account node. The representation vector of the ad account node is then embedded into the CTR prediction model for the ad campaign node in the next layer, generating a representation vector for the ad campaign node. Finally, the same steps are repeated to transfer the representation knowledge of the ad campaign node downward to the ad group node. This hierarchical knowledge transfer model can leverage knowledge from the upper layers to improve click-through rate prediction in the lower layers, thereby alleviating the data imbalance problem.

[0008] The use of a horizontal knowledge transfer model addresses the cold start problem. Under the horizontal knowledge transfer model, a node graph is constructed, representing the data structure of each layer of nodes, based on node characteristics such as keywords from landing pages associated with ads, user-searchable keywords associated with ads, and similarities between nodes. Based on this constructed node graph, knowledge learned from existing and past nodes (thus, historical data) is propagated to any new parts of the graph added by adding new nodes, thereby facilitating CTR prediction for new nodes with little or no training data.

[0009] Finally, the multidimensional knowledge transfer model combines the hierarchical knowledge transfer model and the horizontal knowledge transfer model to establish a holistic learning framework for CTR prediction, which can transfer learned knowledge along multiple dimensions and solve the data imbalance problem and cold start problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] Embodiments of the present invention are described in more detail below with reference to the accompanying drawings, in which:

[0011] Figure 1 A logical block diagram showing the logical data structure of computational advertising;

[0012] Figure 2 A schematic diagram showing a hierarchical knowledge transfer model according to an embodiment of the present invention;

[0013] Figure 3 A logic block diagram showing a node diagram under a horizontal knowledge transfer model according to an embodiment of the present invention;

[0014] Figure 4 Schematic diagram showing the horizontal knowledge transfer model;

[0015] Figure 5A A schematic diagram showing a multi-dimensional knowledge transfer model according to an embodiment of the present invention; and

[0016] Figure 5B Further schematic diagram showing the multidimensional knowledge transfer model. DETAILED DESCRIPTION

[0017] In the following description, as a preferred example, a system and method for predicting the probability that artificial content displayed on a website or online electronic user interface will be visited or clicked when presented to an audience group is presented. It will be apparent to those skilled in the art that modifications, including additions and / or substitutions, can be made without departing from the scope and spirit of the present invention. Specific details may be omitted so as not to obscure the present invention; however, the disclosure of the present invention is intended to enable those skilled in the art to practice the teachings herein without undue experimentation.

[0018] In computational advertising, an online advertisement (ad) is artificial content displayed on a website or online electronic user interface, or is a search keyword used to return search results from an Internet search engine, with an embedded URL link for the viewer to click or visit, thereby directing the viewer to a destination or landing page, or directing online traffic to or use of a specific website or online electronic user interface by clicking or visiting the advertisement.

[0019] Please refer to the following description Figure 1In the art, the different aspects of online advertising are generally categorized as ad accounts, ad campaigns, ad groups, and ads. An ad can be text, an image, a video clip, a sound clip, or a bid representing a specific product or service, or aspects thereof, provided by an advertiser.

[0020] An ad group contains one or more ads that share the same target. Other parameters of an ad group include a theme, a target location, a target language, and the product or service being advertised. In the context of ML and DL modeling, an ad group can be represented as an ad group node 101, whose setup features can include one or more search keywords, an ad theme, a target location, a target language, one or more advertised products, and / or one or more ad services for one or more ads in the ad group node 101. The ad group node also collects performance data for its ads, which can have some performance features, such as the number of user clicks / visits received on its ads and the cost of the ads. In an exemplary embodiment, the ad group node includes the setup features and performance features provided in Table 1 below.

[0021]

[0022]

[0023] Table 1

[0024] An ad campaign is a collection of one or more ad groups. Ad campaigns are also often used to organize the categories of products or services offered by advertisers. In the context of ML and DL modeling, an ad campaign can be represented as an ad campaign node 102, whose settings and performance characteristics are the collection of ad group nodes belonging to it.

[0025] An ad account is a collection of one or more ad campaigns. An advertiser can have one or more ad accounts. In the context of ML and DL modeling, an ad account can be represented as an ad account node 103, whose settings and performance characteristics are the collection of ad campaign nodes belonging to it.

[0026] One or more ad account nodes 103 form the top layer; one or more ad campaign nodes 102 form the middle layer; and one or more ad group nodes 101 form the bottom layer. Each ad group node 101 has an upward hierarchical relationship 111 with an ad campaign node 102, indicating that the ad group node 101 belongs to the ad campaign node 102. Each ad campaign node 102 has one or more downward hierarchical relationships 111 (each hierarchical relationship 111 is connected to an ad group node 101), representing all ad group nodes 101 belonging to the ad campaign node 102. Each ad campaign node 102 also has an upward hierarchical relationship 112 with an ad account node 103, indicating that the ad campaign node 102 belongs to the ad account node 103. Each ad account node 103 has one or more downward hierarchical relationships 112 (each hierarchical relationship 112 is connected to an ad campaign node 102), representing all ad campaign nodes 102 belonging to the ad account node 103. Therefore, the nodes organized into layers and their hierarchical relationships form a node network.

[0027] In the context of ML and DL modeling, audiences have many characteristics; for example, audience ID, age, gender, placement type (online assets, i.e. channels, social media sites, applications, etc., where the advertisements are served and viewed).

[0028] To simplify the description of the various inventive concepts of the present invention, the embodiments described herein assume that the methods and systems are implemented based on a logical data structure having the aforementioned layers of ad account nodes, ad campaign nodes, and ad group nodes. However, this assumption should not be construed as limiting the present invention. Those skilled in the art can readily implement the embodiments of the present invention in systems having different logical data structures without undue experimentation or departing from the spirit of the present invention.

[0029] According to one embodiment of the present invention, a CTR prediction system is provided that utilizes one or more knowledge transfer models, including one or more of a hierarchical knowledge transfer model, a horizontal knowledge transfer model, and a multidimensional knowledge transfer model. To establish its knowledge transfer model, the CTR prediction system first identifies and accepts as input a node network consisting of one or more ad account nodes, one or more ad campaign nodes, and one or more ad group nodes.

[0030] Please refer to the following description Figure 2. The hierarchical knowledge transfer model includes an ad account CTR prediction model 201, an ad campaign CTR prediction model 202, and an ad group CTR prediction model 203. At runtime, the ad account CTR prediction model 201 takes the audience characteristics 204 and the settings and performance characteristics 205 of the ad account node as input to generate a predicted CTR 206 for the ad account node. The ad account CTR prediction model 201 also extracts a latent vector 207 from the ad account node. The extracted latent vector 207 is then appended to the settings and performance characteristics 208 of the ad campaign node and taken as input by the ad campaign CTR prediction model 202 together with the audience characteristics 204 to generate a predicted CTR 209 for the ad campaign node. The ad campaign CTR prediction model 202 also extracts a latent vector 210 from the ad campaign node. The extracted latent vector 210 is then appended to the settings and performance features 211 of the ad group node and taken as input by the ad group CTR prediction model 203 together with the audience features 204 to generate the predicted CTR 212 of the ad group node.

[0031] Under the hierarchical knowledge transfer model, an ad account CTR prediction model is first trained using the settings and performance data of ad account nodes. The performance data for each ad account is the set of all ad campaign nodes belonging to that ad account, and the performance data for each ad campaign is the set of all ad group nodes belonging to that ad campaign. Then, the hidden vectors of the ad account nodes are extracted. The hidden vectors of the ad account nodes are embedded into the campaign CTR prediction model by appending each hidden vector of the ad account node to the features of each of its child ad campaign nodes and inputting them into the campaign CTR prediction model. Then, the hidden vectors of the ad campaign nodes are extracted. Finally, the hidden vectors of the ad campaign nodes are embedded into the ad group CTR prediction model by appending each hidden vector of the ad campaign node to the features of each of its child ad group nodes and inputting them into the ad group CTR prediction model.

[0032] The loss functions of the ad account CTR prediction model, ad campaign CTR prediction model, and ad group CTR prediction model can be expressed as:

[0033]

[0034] Where S represents the observed data, y is the true CTR value, x is the input of the model, Θ is the model parameter set, is the predicted CTR value based on x and Θ, λθ∈R + is the regularization value of the model parameters θ∈Θ.

[0035] In one embodiment, each of the ad account CTR prediction model, ad campaign CTR prediction model, and ad group CTR prediction model is implemented as a support vector machine (SVM) model. In this case, the latent vector extracted for an ad account node, ad campaign node, or ad group node is a feature mapping function of the feature vector of each node, which can be expressed as:

[0036]

[0037] in is the feature map of the SVM model; nodeid is the ID of the ad account node, ad campaign node, or ad group node whose hidden vector is extracted.

[0038] In another embodiment, each of the ad account CTR prediction model, the ad campaign CTR prediction model, and the ad group CTR prediction model is implemented as a factorization machine (FM). In this case, the latent vector extracted for the ad account node, ad campaign node, or ad group node is a parameter vector of the fields of each node, which can be expressed as:

[0039] hidden vector node id = parameter vector(field node id );

[0040] Where node id is the ID of the ad account node, ad campaign node, or ad group node whose latent vector is extracted.

[0041] In another embodiment, each of the ad account CTR prediction model, the ad campaign CTR prediction model, and the ad group CTR prediction model is implemented as a parallel structure DL model. In this case, when predicting the CTR of an ad account node, an ad campaign node, or an ad campaign node (whose hidden vector has been extracted), the extracted hidden vector of the ad account node, the ad campaign node, or the ad group node is extracted from the embedding layer of the parallel structure DL model. The parallel structure DL model can be, but is not limited to, a generalized and deep learning model or a deep FM model. Other parallel structure DL models can also be easily adopted without excessive experimentation or deviating from the spirit of the present invention.

[0042] In another embodiment, each of the advertising account CTR prediction model, the advertising campaign CTR prediction model, and the advertising group CTR prediction model is implemented as a serial structure DL model. In this case, when predicting the CTR of an advertising account node, an advertising campaign node, or an advertising group node (its latent vector has been extracted), the extracted latent vector of the advertising account node, the advertising campaign node, or the advertising group node is extracted from the serial structure DL model feature interaction. The serial structure DL model can be, but is not limited to, a factorization machine-supported neural network (FNN) model or a product-based neural network (PNN) model. Other serial structure DL models can also be easily adopted without excessive experimentation or deviating from the spirit of the present invention.

[0043] In another embodiment, each of the advertising account CTR prediction model, the advertising campaign CTR prediction model, and the advertising group CTR prediction model is implemented as a general interest structure DL model (General Interest-Structure DL model). In this case, when predicting the CTR of an advertising account node, an advertising campaign node, or an advertising group node (whose hidden vector has been extracted), the extracted hidden vector of the advertising account node, the advertising campaign node, or the advertising group node is extracted from the feature embedding layer of the general interest structure DL model. The general interest structure DL model can be, but is not limited to, a deep interest network (DIN) model or a model based on a deep neural network (YouTubeNet) recommended by YouTube. Other interest structure DL models can also be easily adopted without excessive experimentation or deviating from the spirit of the present invention.

[0044] Please refer to the following description Figure 3 and Figure 4 The horizontal knowledge transfer model includes an ad group node graph 301 constructed by a logical preprocessor based on similarities between ad group nodes 303a, 303b, 303c, 303d, and 303e represented by edges 302 of the ad group node graph 301. In one embodiment, the similarity is calculated based on one or more keywords that can be extracted from destinations or landing web pages associated with the ads in each ad group node 303a, 303b, 303c, 303d, and 303e. In other embodiments, the similarity is calculated based on one or more set features of each ad group node 303a, 303b, 303c, 303d, and 303e.

[0045] The horizontal knowledge transfer model also includes a graph convolutional network (GCN) 402. At runtime, GCN 402 takes as input the ad group node graph 401 and the settings and performance features 405 of the ad group nodes to generate ad group embedding vectors 406 for the ad group nodes, where ad group nodes 405 represent one or more ads whose CTRs are predicted and contain the settings features of these ads. Each neural network layer l of GCN 402 can be represented by a rectifier activation function:

[0046]

[0047] in is the adjacency matrix of the ad group node graph G with self-connections added, I N is the identity matrix, is a trainable weight matrix for a specific layer, H (l) is the activation matrix of layer l.

[0048] The horizontal knowledge transfer model also includes a regression artificial neural network (ANN) 403, which takes the ad group embedding vector 406 and the audience characteristics 407 as input to generate the predicted CTR 408 of the ad group node. Each neural network layer l of the regression ANN 403 can be represented by a rectifier activation function:

[0049]

[0050] in and b (l) is the parameter of the neural network layer l, a (l) is the activation value of layer l.

[0051] Its output:

[0052]

[0053] This means that the activation a of the last layer L L ∈R + The CTR value used as the prediction

[0054] And its loss function:

[0055]

[0056] Where S represents the observed data, y is the true CTR value, x is the input of the model, Θ is the model parameter set, is the predicted CTR value based on x and Θ, λ θ ∈R + is the regularization value of the model parameters θ∈Θ.

[0057] During training, the ad group node graph of one or more ad group nodes, the settings and performance characteristics of the ad group nodes, the characteristics of the audience, and the past CTR values of pre-existing ad group nodes are used as training datasets for training the GCN and the regression ANN, and the regression ANN is trained only using the past CTR values of pre-existing ad group nodes.

[0058] For the campaign node layer and the ad account node layer, the components of the horizontal knowledge transfer model described above are replicated. For the campaign node layer, the GCN takes as input the campaign node graph 311, which is created by merging the nodes belonging to each campaign node 312a, 312b, 312c, 312d, and 312e in the ad group node graph 301 using a logical preprocessor. For example, ad group node 303a goes into campaign node 312a; ad group node 303b goes into campaign node 312b. The GCN also takes as input the settings and performance characteristics of the campaign node, and together with the campaign node graph 311, generates a campaign embedding vector. The campaign embedding vector is then input to the regression ANN; along with the audience characteristics as input, the regression ANN generates a predicted CTR for the campaign node.

[0059] During training, a campaign node graph of one or more campaign nodes, settings and performance characteristics of the campaign nodes, characteristics of the audience, and past CTR values of pre-existing campaign nodes are used as training datasets for training the GCN and the regression ANN, and the regression ANN is trained only using past CTR values of pre-existing campaign nodes.

[0060] Similarly, for the ad account node layer, the GCN takes as input the ad account node graph 321, which is created by the logic preprocessor by merging the ad campaign nodes belonging to each ad account node 322a, 322b, and 322c in the ad campaign node graph 311. For example, ad campaign node 312a enters ad account node 322a; ad campaign nodes 312b and 312c enter ad account node 322b. The GCN also takes as input the settings and performance characteristics of the ad account nodes and, together with the ad account node graph 321, generates an ad account embedding vector. The ad account embedding vector is then input to the regression ANN; together with the audience characteristics as input, the regression ANN generates a predicted CTR for the ad account node.

[0061] During training, an ad account node graph of one or more ad account nodes, the settings and performance characteristics of the ad account nodes, the characteristics of the audience, and the past CTR values of pre-existing ad account nodes are used as training datasets for training the GCN and the regression ANN. The regression ANN is trained only using the past CTR values of pre-existing ad account nodes.

[0062] The following description refers to Figure 5A and 5B . The multidimensional knowledge transfer model combines the hierarchical knowledge transfer model and the horizontal knowledge transfer model. The multidimensional knowledge transfer model includes an advertising account multi-knowledge CTR prediction model 501, an advertising campaign multi-knowledge CTR prediction model 502, and an advertising group multi-knowledge CTR prediction model 503. Each multi-knowledge CTR prediction model includes a GCN 504 and a CTR prediction model 505 of the horizontal knowledge transfer model, which are the same as the CTR prediction model 201, 202, or 203 corresponding to the advertising account node layer, the advertising campaign node layer, or the advertising group node layer in the hierarchical knowledge transfer model. The multidimensional knowledge transfer model also includes an advertising account node graph 508, an advertising campaign node graph 513, and an advertising group node graph 517 constructed by the logic preprocessor of the above-mentioned horizontal knowledge transfer model during the pre-run process.

[0063] At runtime, the ad account multi-knowledge CTR prediction model 501 takes as input audience characteristics 506, the settings and performance characteristics 507 of one or more target ad account nodes with predicted CTRs, and an ad account node graph 508 to generate a predicted CTR 510 for each target ad account node. Each target ad account node can represent a newly created ad account or an existing ad account whose characteristics have been modified and / or whose child ad campaign nodes and / or ad group nodes have been modified. In the ad account multi-knowledge CTR prediction model 501, the settings and performance characteristics 507 of the target ad account nodes and the ad account node graph 508 are input into its GCN to generate an embedding vector 509, which is then input into its CTR prediction model along with the audience characteristics 506 to generate a predicted CTR 510 for each target ad account node. The CTR prediction model of the ad account multi-knowledge CTR prediction model 501 also extracts a latent vector 511 from each target ad account node.

[0064] At the campaign node level, the extracted latent vector 511 is then appended to the target campaign node's set features 512 and, along with the audience features 506 and the target campaign node graph 513, used as input by the campaign multi-knowledge CTR prediction model 502 to generate a predicted CTR 514 for the target campaign node. The target campaign node can represent a newly created campaign, or an existing campaign whose features have been modified and / or whose child ad group nodes have been modified, and which belongs to one of the target advertising accounts. The campaign multi-knowledge CTR prediction model 502 also extracts a latent vector 515 from the target campaign node. Within the campaign multi-knowledge CTR prediction model 502, its GCN and CTR prediction models function identically to those in the ad group multi-knowledge CTR prediction model 501.

[0065] Finally, at the ad group node layer, the extracted latent vector 515 is then appended to the target ad group node's set features 516. This, along with the audience features 506 and the target ad group node graph 517, is used as input by the ad group CTR prediction model 503 to generate the target ad group node's predicted CTR 518. The target ad group node can represent a newly created ad group or a modified existing ad group belonging to the target ad campaign. Similarly, within the ad group multi-knowledge CTR prediction model 503, its GCN and CTR prediction model function identically to those in the ad group multi-knowledge CTR prediction model 501.

[0066] The logical functional units, modules, processors, and preprocessors of the prediction and knowledge transfer ML and DL models according to the embodiments disclosed herein can be implemented using computing devices, computer processors, or electronic circuits, including but not limited to application-specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), and other programmable logic devices configured or programmed according to the teachings of this disclosure. Based on the teachings of this disclosure, those skilled in the art of software or electronics can easily prepare computer instructions or software codes that run in computing devices, computer processors, or programmable logic devices.

[0067] All or part of the method according to an embodiment of the present invention may be executed in one or more computing devices, including server computers, personal computers, laptop computers, and mobile computing devices such as smartphones and tablet computers.

[0068] Embodiments of the present invention include computer storage media, transient and non-transient storage devices having computer instructions or software code stored therein, which can be used to program a computer or microprocessor to perform any of the processes of the present invention. The storage media, transient and non-transient storage devices may include, but are not limited to, floppy disks, optical disks, Blu-ray disks, DVDs, CD-ROMs, magneto-optical disks, ROMs, RAMs, flash memory devices, or any other type of medium or device suitable for storing instructions, code, and / or data.

[0069] Each functional unit and module of the various embodiments may also be implemented in a distributed computing environment and / or a cloud computing environment, where all or part of the machine instructions are executed in a distributed manner by one or more processing devices interconnected by a communication network, including, for example, an intranet, a wide area network (WAN), a local area network (LAN), the Internet, and other forms of data transmission media.

[0070] The above description of the present invention is provided for the purpose of illustration and description. It is not intended to be exhaustive or to limit the invention to the precise form disclosed. For those skilled in the art, many modifications and variations will be apparent.

[0071] The embodiments were chosen and described in order to best explain the principles of the invention and its practical application, thereby enabling others skilled in the art to understand the invention for various embodiments and with various modifications as are suited to the particular use contemplated.

Claims

1. A system for predicting the probability that a computational advertisement displayed on a website or online electronic user interface will be visited when presented to an audience, the system comprising: A multidimensional knowledge transfer model, implemented by one or more processors, includes: The logic preprocessor is configured to build the following: an ad group node graph consisting of a plurality of ad group nodes, wherein the ad group node graph is constructed based on one or more feature similarities between the ad group nodes; an ad campaign node graph consisting of one or more ad campaign nodes, the ad campaign node graph being formed by merging the ad group nodes in the ad group node graph belonging to each ad campaign node; and an advertising account node graph consisting of one or more advertising account nodes, wherein the advertising account node graph is formed by merging the advertising campaign nodes in the advertising campaign node graph belonging to each advertising account node; an advertising account multi-knowledge click-through rate (CTR) prediction model, which is trained to predict the CTR of an advertising account, wherein the CTR of the advertising account is predicted based on the advertising account node graph, characteristics of the audience group, characteristics of the advertising account node for which the CTR has been predicted, and characteristics of other advertising account nodes; a multi-knowledge CTR prediction model for an advertising campaign, which is trained to predict the CTR of an advertising campaign, wherein the CTR of the advertising campaign is predicted based on the advertising campaign node graph, characteristics of the audience group, characteristics of the advertising campaign nodes for which the CTR has been predicted, an ad account node hidden vector, and characteristics of other advertising campaign nodes; wherein the advertising campaign with a predicted CTR belongs to the advertising account with a predicted CTR; and wherein, when predicting the CTR of the advertising account whose CTR has been predicted, the advertising account node hidden vector is extracted from the advertising account multi-knowledge CTR prediction model; and an ad group multi-knowledge prediction model, which is trained to predict the CTR of an ad group, wherein the CTR of the ad group is predicted from the ad group node graph, the characteristics of the audience group, the characteristics of the ad group node for which the CTR has been predicted, and an additional hidden vector of an ad campaign node and the characteristics of other ad group nodes; wherein the ad group for which CTR has been predicted belongs to the ad campaign for which CTR has been predicted; and wherein, when predicting the CTR of an advertising campaign whose CTR has already been predicted, the hidden vector of the advertising campaign node is extracted from the advertising campaign multi-knowledge CTR prediction model; The advertising group with predicted CTR includes computational advertising, and the predicted CTR of the advertising group represents the probability that the computational advertising will be visited. In the computational advertising, online advertising is artificial content displayed on a website or an online electronic user interface, or is a search keyword used to return search results of an Internet search engine, wherein there is an embedded URL link for viewers to click or visit, thereby guiding the viewers to a destination or landing on a web page, or directing online traffic to or use of a specific website or online electronic user interface by clicking or visiting the advertisement.

2. The system according to claim 1, in, The logic pre-processor is further configured to construct the ad group node graph based on feature similarities between the ad group nodes, the feature similarities including one or more keywords of landing pages associated with one or more ads in each of the ad group nodes.

3. The system according to claim 1, in, Each of the advertising account multi-knowledge CTR prediction model, the advertising campaign multi-knowledge CTR prediction model, and the advertising group multi-knowledge CTR prediction model includes a horizontal knowledge transfer model and a hierarchical knowledge transfer model.

4. The system according to claim 1, in, During training, the advertising account multi-knowledge CTR prediction model is first trained using the advertising account settings and performance data of multiple advertising accounts; wherein the advertising account performance data is a collection of all advertising campaign performance data of all advertising campaigns belonging to the advertising account; and Each of the advertising campaign performance data is a collection of all advertising group performance data of all advertising groups belonging to the advertising campaign.

5. The system according to claim 4, in, The advertising campaign multi-knowledge CTR prediction model is trained after the advertising account multi-knowledge CTR prediction model is trained; The hidden vector of the advertising account node is embedded in the advertising campaign multi-knowledge CTR prediction model by appending each hidden vector of the advertising account node to the features of each child advertising campaign node of the advertising account node, and inputting the hidden vector into the advertising campaign multi-knowledge CTR prediction model.

6. The system according to claim 5, in, The advertising group multi-knowledge CTR prediction model is trained after the advertising campaign multi-knowledge CTR prediction model is trained; The hidden vector of the advertising campaign node is embedded in the advertising group multi-knowledge CTR prediction model by appending each hidden vector of the advertising campaign node to the features of each child advertising group node of the advertising campaign node and inputting it into the advertising group multi-knowledge CTR prediction model.

7. The system of claim 1 , wherein during training: Using an ad group node graph of one or more ad group nodes, settings and performance characteristics of the ad group nodes, characteristics of one or more audience groups, and past CTR values of pre-existing ad group nodes as a training dataset for training the ad group multi-knowledge CTR prediction model; Using an advertising campaign node graph of one or more advertising campaign nodes, settings and performance characteristics of the advertising campaign nodes, characteristics of the audience group, and past CTR values of pre-existing advertising campaign nodes as a training dataset for training the advertising campaign multi-knowledge CTR prediction model; An advertising account node graph of one or more advertising account nodes, settings and performance characteristics of the advertising account nodes, characteristics of the audience group, and past CTR values of pre-existing advertising account nodes are used as a training data set to train the advertising account multi-knowledge CTR prediction model.

8. The system according to claim 1, wherein: Features of the ad group node of the ad group whose CTR has been predicted and other ad group nodes include: Setup features include: account ID, campaign ID, ad group ID, ad network type, engagement type, target location, and target language; and Performance characteristics include: clicks, impressions, click-through rate (CTR), cost, interactions, conversions, engaged video views, active view measurability, active view impressions, active view measurable cost, active view measurable impressions, active view viewability, active CTR, active view cost per thousand impressions (CPM), average cost per click, average CPM, average cost per engagement (CPE), and average cost per view (CPV).

9. The system according to claim 1, wherein the characteristics of the audience include audience ID, age, gender, and delivery type.

10. A computer-implemented method for predicting the probability that a computational advertisement displayed on a website or online electronic user interface will be visited when presented to an audience, the method comprising: Executing a multidimensional knowledge transfer model to predict the probability that the computational advertisement will be accessed, the multidimensional knowledge transfer model comprising: The logic preprocessor is configured to build the following: an ad group node graph consisting of a plurality of ad group nodes, wherein the ad group node graph is constructed based on one or more feature similarities between the ad group nodes; an ad campaign node graph consisting of one or more ad campaign nodes, the ad campaign node graph being formed by merging the ad group nodes in the ad group node graph belonging to each ad campaign node; and an advertising account node graph consisting of one or more advertising account nodes, wherein the advertising account node graph is formed by merging the advertising campaign nodes in the advertising campaign node graph belonging to each advertising account node; an advertising account multi-knowledge click-through rate (CTR) prediction model, which is trained to predict the CTR of an advertising account, wherein the CTR of the advertising account is predicted based on the advertising account node graph, characteristics of the audience group, characteristics of the advertising account node for which the CTR has been predicted, and characteristics of other advertising account nodes; a multi-knowledge CTR prediction model for an advertising campaign, which is trained to predict the CTR of an advertising campaign, wherein the CTR of the advertising campaign is predicted based on the advertising campaign node graph, characteristics of the audience group, characteristics of the advertising campaign nodes for which the CTR has been predicted, an ad account node hidden vector, and characteristics of other advertising campaign nodes; wherein the advertising campaign with a predicted CTR belongs to the advertising account with a predicted CTR; and wherein, when predicting the CTR of the advertising account having predicted CTR, the advertising account node hidden vector is extracted from the advertising account multi-knowledge CTR prediction model; and an ad group multi-knowledge prediction model, which is trained to predict the CTR of an ad group, wherein the CTR of the ad group is predicted from the ad group node graph, the characteristics of the audience group, the characteristics of the ad group node for which the CTR has been predicted, and an additional hidden vector of an ad campaign node and the characteristics of other ad group nodes; wherein the ad group for which CTR has been predicted belongs to the ad campaign for which CTR has been predicted; and wherein, when predicting the CTR of an advertising campaign whose CTR has been predicted, the hidden vector of the advertising campaign node is extracted from the advertising campaign multi-knowledge CTR prediction model; The advertising group with predicted CTR includes computational advertising, and the predicted CTR of the advertising group represents the probability that the computational advertising will be visited. In the computational advertising, online advertising is artificial content displayed on a website or an online electronic user interface, or is a search keyword used to return search results of an Internet search engine, wherein there is an embedded URL link for viewers to click or visit, thereby guiding the viewers to a destination or landing on a web page, or directing online traffic to or use of a specific website or online electronic user interface by clicking or visiting the advertisement.

11. The method according to claim 10, in, The logic pre-processor is further configured to construct the ad group node graph based on feature similarities between the ad group nodes, the feature similarities including one or more keywords of landing pages associated with one or more ads in each of the ad group nodes.

12. The method according to claim 10, in, Each of the advertising account multi-knowledge CTR prediction model, the advertising campaign multi-knowledge CTR prediction model, and the advertising group multi-knowledge CTR prediction model includes a horizontal knowledge transfer model and a hierarchical knowledge transfer model.

13. The method according to claim 10, in, During training, the advertising account multi-knowledge CTR prediction model is first trained using the advertising account settings and performance data of multiple advertising accounts; wherein the advertising account performance data is a collection of all advertising campaign performance data of all advertising campaigns belonging to the advertising account; and Each of the advertising campaign performance data is a collection of all advertising group performance data of all advertising groups belonging to the advertising campaign.

14. The method according to claim 13, in, The advertising campaign multi-knowledge CTR prediction model is trained after the advertising account multi-knowledge CTR prediction model is trained; The hidden vector of the advertising account node is embedded in the advertising campaign multi-knowledge CTR prediction model by appending each hidden vector of the advertising account node to the features of each child advertising campaign node of the advertising account node, and inputting the hidden vector into the advertising campaign multi-knowledge CTR prediction model.

15. The method according to claim 14, in, The advertising group multi-knowledge CTR prediction model is trained after the advertising campaign multi-knowledge CTR prediction model is trained; The hidden vector of the advertising campaign node is embedded in the advertising group multi-knowledge CTR prediction model by appending each hidden vector of the advertising campaign node to the features of each child advertising group node of the advertising campaign node and inputting it into the advertising group multi-knowledge CTR prediction model.

16. The method according to claim 10, wherein during training, Using an ad group node graph of one or more ad group nodes, settings and performance characteristics of the ad group nodes, characteristics of one or more audience groups, and past CTR values of pre-existing ad group nodes as a training dataset for training the ad group multi-knowledge CTR prediction model; Using an advertising campaign node graph of one or more advertising campaign nodes, settings and performance characteristics of the advertising campaign nodes, characteristics of the audience group, and past CTR values of pre-existing advertising campaign nodes as a training dataset for training the advertising campaign multi-knowledge CTR prediction model; An advertising account node graph of one or more advertising account nodes, settings and performance characteristics of the advertising account nodes, characteristics of the audience group, and past CTR values of pre-existing advertising account nodes are used as a training data set to train the advertising account multi-knowledge CTR prediction model.

17. The method according to claim 10, wherein Features of the ad group node of the ad group whose CTR has been predicted and other ad group nodes include: Setup features include: account ID, campaign ID, ad group ID, ad network type, engagement type, target location, and target language; and Performance characteristics include: clicks, impressions, click-through rate (CTR), cost, interactions, conversions, engaged video views, active view measurability, active view impressions, active view measurable cost, active view measurable impressions, active view viewability, active CTR, active view cost per thousand impressions (CPM), average cost per click, average CPM, average cost per engagement (CPE), and average cost per view (CPV). The method according to claim 10 , wherein the characteristics of the audience group include audience group ID, age, gender, and delivery type.

Images