A method for estimating basic attributes of advertising audience based on heterogeneous graph embedding technology

Abstract

This invention provides a method for predicting the basic attributes of advertising audiences based on heterogeneous graph embedding technology. The method constructs a heterogeneous graph of click behavior records for all advertising audiences based on their historical click behavior. Then, using heterogeneous graph embedding technology, it generates vector representations for each of these advertising information items (material ID, ad ID, advertiser ID), thus preserving the interrelationships between them. An LSTM neural network is built to automatically train a network model that can predict the basic attributes of advertising audiences based on their historical click behavior, given known basic attributes (age and gender). For advertising audiences with unknown basic attributes, their historical click behavior is simply input into the trained network model, allowing the model to derive these basic attributes from the audience's historical click behavior.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and more specifically, to a method for estimating basic attributes of an advertising audience based on heterogeneous graph embedding technology. Background Art

[0002] It is well known that the preferences of advertising audiences for advertising vary with their age and gender. Therefore, demographic characteristics such as the age and gender of advertising audiences are important input features of various recommendation systems, including advertising platforms. However, most advertising platforms are unable to accurately obtain basic attributes such as age and gender of all advertising audiences, which results in the advertising platform being unable to accurately push corresponding advertisements to advertising audiences that are missing basic attributes such as age and gender. In this case, the advertising platform needs to predict and complete the missing basic attributes of the advertising audience as much as possible, as well as correct the possible incorrect basic attributes of the advertising audience (age and gender), which will help improve the push effect of the advertising platform.

[0003] In the early days, advertising platforms could only rely on surveys to manually input basic attributes such as age and gender of advertising audiences, or notify advertising audiences to fill in the information themselves, which was extremely inefficient. With the continuous development of modern information technology, this information completion method has been replaced by various artificial intelligence technologies. The mainstream artificial intelligence task of predicting the age and gender attributes of advertising audiences is to directly use the advertising audience's click information (material ID, advertising ID, advertiser ID) as independent features, and use a trained tree model classifier (such as LightGBM) or neural network (such as LSTM) to estimate the basic attributes (age and gender) of advertising audiences with unknown basic attributes.

[0004] However, in actual situations, the click ad information (material id, ad id, advertiser id) of the ad audience is often not independent of each other, but related to each other. For example, a material clearly belongs to a certain ad, and an ad is clearly sent by a certain advertiser. In such scenarios, retaining the relationship between ad information (material id, ad id, advertiser id) will make the estimation of the basic attributes of the ad audience more accurate. Summary of the invention

[0005] The present invention provides a method for estimating basic attributes of an advertisement audience based on heterogeneous graph embedding technology with better prediction effect.

[0006] In order to achieve the above technical effects, the technical solution of the present invention is as follows:

[0007] A method for estimating basic attributes of advertising audience based on heterogeneous graph embedding technology includes the following steps:

[0008] S1: Based on the historical click behaviors of all advertising audiences, a heterogeneous graph of the click behavior records of advertising audiences is constructed, and based on the heterogeneous graph embedding representation technology, respective vector representations are generated for these advertising information, where the advertising information includes the material ID, advertising ID, and advertiser ID;

[0009] S2: Divide the data set into two categories according to whether the basic attributes of the advertising audience are known. The data with known basic attributes is called the training set, which is used to train the prediction model. The other data is called the test set, which is used to predict its unknown basic attributes. The basic attributes of the advertising audience are age and gender.

[0010] S3: The historical click behavior of the advertising audience in the training set is used as the input feature of the LSTM prediction model, and age and gender are used as the prediction result labels of the corresponding features. The LSTM neural network is trained to obtain a prediction model that predicts the basic attributes of the advertising audience based on their click behavior.

[0011] S4: Input the historical click behaviors of the advertising audiences in the test set into the trained prediction model to estimate the basic attributes of these advertising audiences.

[0012] Furthermore, in step S1, the process of constructing a heterogeneous graph of the click behavior records of the advertising audience is:

[0013] S11: Generate nodes of heterogeneous graph: each different id number of each type is used as a node of heterogeneous graph;

[0014] S12: Connecting nodes of heterogeneous graphs with edges: There are three types of edge relationships between the four types of nodes: the click relationship between advertising audiences and advertising materials, the subordinate relationship between advertising materials and advertisements, and the publishing relationship between advertisements and advertisers; if any of the above three relationships exists between two nodes, the two nodes are connected by an edge.

[0015] Furthermore, in step S1, the process of generating respective vector representations for these advertisement information using heterogeneous graph embedding representation technology includes heterogeneous graph sampling and SkipGram neural network training.

[0016] Furthermore, the specific process of heterogeneous graph sampling is:

[0017] Each advertising audience id is used as the starting node to walk W times along the set meta-path, with L steps each time; each node walked to is registered in turn, which is a node sequence corpus containing various node types.

[0018] Furthermore, the SkipGram neural network training process is:

[0019] The node sequence corpus is input into the SkipGram neural network, and the n-dimensional vector representation of each node in the corpus is obtained through training. The training goal is to make the vector representation of each node maintain the neighbor information of the node in the corpus. The formula for training model optimization is:

[0020]

[0021] where c t is the context node of v, and the denominator is u t ∈V t , that is, the type of node is taken into account during negative sampling; after training, an n-dimensional vector representation of each node is obtained.

[0022] Furthermore, in step S2, the process of dividing the data set is:

[0023] The original data set is divided into two categories according to whether the basic attributes of the advertising audience are known. The data with known basic attributes is called the training set. The advertising audience click behavior in the training set is used as the feature of the training model, and the basic attributes corresponding to the features are used as training labels. The features and corresponding labels are used to train the LSTM neural network to obtain a prediction model. The other type of data is called the test set, which is used to predict the unknown basic attributes of the trained prediction model.

[0024] Furthermore, in step S3, the data of the training set is processed into feature vectors and their labels that can be recognized by the subsequent LSTM. The specific steps are:

[0025] 1) For each advertising audience, arrange the three features of the advertising audience in chronological order, and the length of the three features is p time steps; if there are more than p features of an advertising audience, keep the most recent p; if there are less than p features, take the most recent ones to make up for the p;

[0026] 2) Convert the three features of each advertising audience into feature vectors, that is, replace each node in the historical time series of the three features of each advertising audience with its n-dimensional vector. The three feature vectors of each advertising audience will be used as input features for the training model;

[0027] 3) The basic attributes of each advertising audience are used as labels corresponding to the three features of the advertising audience for supervised training of the LSTM neural network.

[0028] 8. According to the method for estimating basic attributes of advertising audiences based on heterogeneous graph embedding technology described in claim 6, it is characterized in that the input of the LSTM network is 3 feature sequences of p time steps, each feature vector is n-dimensional, and the output is a sequence structure; when a model for predicting age is needed, the LSTM neural network is trained with age as a label; when a model for predicting gender is needed, the LSTM neural network is trained with gender as a label.

[0029] Furthermore, in step S4, for each advertising audience in the test set, the three features of the advertising audience are arranged in chronological order, and the length of the three features is p time steps; if there are more than p features of an advertising audience, the most recent p features are retained, and if there are less than p features, the most recent p features are taken to make up the p features;

[0030] The three features of each advertising audience are converted into feature vectors, that is, each node in the historical time series of the three features of each advertising audience is replaced by its n-dimensional vector; the three feature vectors of each advertising audience will be used as input features of the prediction model.

[0031] Furthermore, the prediction model is obtained after the LSTM neural network in step S3 is trained; if age is used as a label for training in step S3, an age estimation model of the advertising audience is obtained; if gender is used as a label for training in step S3, a gender estimation model of the advertising audience is obtained; by inputting the three feature vectors of the historical click behavior of the advertising audience into the age estimation model, their age can be estimated; by inputting the three feature vectors of the historical click behavior of the advertising audience into the gender estimation model, their gender can be estimated.

[0032] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0033] The method of the present invention uses heterogeneous graph embedding technology to represent the historical click records of advertising audiences in the form of vectors for the first time, retains the mutual connection between advertising information (advertising material id, advertising id, advertiser id), and has a better prediction effect; the method uses an LSTM neural network, retains the temporal logic of the clicking behavior of advertising audiences, and has a better prediction effect; the method completes the vectorized representation of advertising information (advertising material id, advertising id, advertiser id) offline according to the clicking behavior of advertising audiences with known basic attributes (age and gender). When it is necessary to predict the basic attributes of advertising audiences with unknown basic attributes (age and gender), it is not necessary to perform vectorized representation again. This design not only improves the accuracy, but also improves the response speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 It is the overall flow chart of the method of the present invention;

[0035] Figure 2 A simplified diagram of the heterogeneous graph of the historical click behavior of the advertising audience;

[0036] Figure 3 is the node sequence obtained by walking starting with u1;

[0037] Figure 4 The input and output structure of the LSTM neural network of this patent. DETAILED DESCRIPTION

[0038] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;

[0039] In order to better illustrate the present embodiment, some parts in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product;

[0040] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0041] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0042] like Figure 1 As shown, a method for estimating basic attributes of advertising audience based on heterogeneous graph embedding technology includes the following steps:

[0043] S1: Based on the historical click behaviors of all advertising audiences, a heterogeneous graph of the click behavior records of advertising audiences is constructed, and based on the heterogeneous graph embedding representation technology, respective vector representations are generated for these advertising information, where the advertising information includes the advertising material ID, advertising ID, and advertiser ID;

[0044] S2: Divide the data set into two categories according to whether the basic attributes (age and gender) of the advertising audience are known. The data with known basic attributes is called the training set, which is used to train the prediction model of this patent, and the other data is called the test set, which is used to predict its unknown basic attributes.

[0045] S3: The historical click behavior of the advertising audience in the training set is used as the input feature of the LSTM prediction model, and age and gender are used as the prediction result labels of the corresponding features. The LSTM neural network is trained to obtain a prediction model that can predict the basic attributes of the advertising audience based on their click behavior.

[0046] S4: Input the historical click behaviors of the advertising audiences in the test set into the trained prediction model to estimate the basic attributes (age and gender) of these advertising audiences.

[0047] (1) Formal definition of tasks

[0048] The task of estimating basic attributes of advertising audiences based on heterogeneous graph embedding technology builds a heterogeneous graph G of advertising audience historical click records based on the historical click record dataset Data of advertising audiences, and converts the data into a heterogeneous graph. Then, each node in the heterogeneous graph (i.e., each advertising information: material id, advertising id, advertiser id) is represented as an n-dimensional vector through heterogeneous graph embedding representation technology. The size of n is determined by the scale of the graph. For convenience, it is uniformly taken as 128 here.

[0049] After obtaining the vector of each advertising information, the historical click behavior of the advertising audience with known basic attributes (age and gender) is used as the features of the training model, and its basic attributes (age and gender) are used as training labels. They are input into the LSTM neural network for training, and an LSTM neural network with the ability to estimate the basic attributes of the advertising audience is obtained.

[0050] After obtaining the trained LSTM neural network, the historical click behavior vectors of the advertising audience whose basic attributes are to be predicted are input into the LSTM neural network, and finally the basic attributes (age and gender) of these advertising audiences are estimated.

[0051] (2) Unit S11: Constructing heterogeneous graphs

[0052] The historical click records of the advertising audience include four types of nodes: advertising audience, advertising material, advertising, and advertiser. Each type has a different ID number. The heterogeneous graph is constructed in the following ways:

[0053] 1) Generate nodes of heterogeneous graph: Each different ID number of each type is used as a node of the heterogeneous graph. For example, if there are 1,000 different IDs for the advertising audience type, there will be 1,000 advertising audience nodes on the heterogeneous graph.

[0054] 2) Use edges to connect nodes in heterogeneous graphs: There are three types of edge relationships between the four types of nodes: the click relationship between the advertising audience and the advertising material, the subordinate relationship between the advertising material and the advertisement, and the publishing relationship between the advertisement and the advertiser. If any of the three relationships exists between two nodes, the two nodes are connected by an edge.

[0055] After the above two steps, we can get something similar to Figure 2 Such a heterogeneous graph. In the figure, u represents the node of the advertising audience type, c represents the node of the advertising material type, a represents the node of the advertising type, and r represents the node of the advertiser type.

[0056] Heterogeneous graph embedding representation technology provides an effective method to solve the analysis problem of heterogeneous graphs. Specifically, heterogeneous graph embedding converts heterogeneous graphs into low-dimensional space vectors that preserve graph information. The input of heterogeneous graph embedding is a heterogeneous graph, and the output is a low-dimensional vector representing each node in the graph. The heterogeneous graph embedding technology of this patent consists of the following two parts: 1) Unit S12: Heterogeneous Graph Sampling; 2) Unit S13: SkipGram Neural Network.

[0057] (3) Unit S12: Heterogeneous Graph Sampling

[0058] Heterogeneous graph sampling performs random walks on heterogeneous graphs based on meta-paths to obtain a corpus of node sequences.

[0059] Design a meta-path: UCARACU (U represents the node of the advertising audience type, C represents the node of the advertising material type, A represents the node of the advertising type, and R represents the node of the advertiser type). This meta-path means that two advertising audiences clicked on the advertising material of the same advertiser. According to this meta-path, the random walk of the nodes can be evenly sampled on the heterogeneous graph.

[0060] Specifically, each advertising audience id is used as the starting node to walk W times along the set meta-path, each time walking L steps. Each node walked to is registered in turn, which is a node sequence corpus containing various node types. (W and L are set according to the complexity of the heterogeneous graph. Generally, W can be set to 1000 and L can be set to 100)

[0061] Next, Figure 2 The following is a random walk of a heterogeneous graph of , showing how to start with the U1 node and perform a random walk under the guidance of the metapath (UCARACU) designed in this patent, and obtain a node sequence starting with the U1 node. Figure 3 As shown in the figure, the metapath UCARACU specifies the node type of each step of the node sequence, for example, the first step is U, the second step is C..., and so on until step 100. There are three C-type nodes connected to u1, and one of them will be randomly selected from the three C-type nodes. After the node sequence starting with u1 has walked 100 steps, it will repeat the walk from u1 again, and repeat 1000 times.

[0062] All other u-type nodes will be walked 1000 times like u1, with 100 steps each time. After all u-type nodes are walked, a corpus of node sequences can be obtained, which retains the relationship between the nodes in the graph.

[0063] (4) Unit S13: Skip-Gram Neural Network

[0064] The node sequence corpus is input into the SkipGram neural network, and the n-dimensional vector representation of each node in the corpus is obtained through training. The training goal is to make the vector representation of each node maintain the neighbor information of the node in the corpus. The formula for training model optimization is:

[0065]

[0066] where c t is the context node of v, and the denominator is u t ∈V t , that is, the type of node is taken into account during negative sampling; after training, an n-dimensional vector representation of each node is obtained.

[0067] (5) Unit S2: Dataset Division

[0068] The original data set is divided into two categories according to whether the basic attributes (age and gender) of the advertising audience are known. The data with known basic attributes is called the training set. The advertising audience click behavior in the training set is used as the feature of the training model, and the basic attributes corresponding to the features are used as the training labels. The features and corresponding labels are used to train the LSTM neural network to obtain the prediction model. The other type of data is called the test set, which is used to predict the unknown basic attributes of the trained prediction model.

[0069] (6) Unit S31: Data processing: features, labels

[0070] The ad creative id, ad id, and advertiser id are called the three features of the ad audience, and the age and gender attributes of the ad audience are called labels. This unit aims to process the data of the training set into feature vectors and their labels that can be recognized by the subsequent LSTM. The specific steps are:

[0071] 1) For each advertising audience, the three features of the advertising audience (ad material id, advertising id, advertiser id) are arranged in chronological order, and the length of the three features is p time steps (the p value is set according to the data size, generally p = 100). If there are more than p features of an advertising audience, the most recent p are retained. If there are less than p features, the most recent ones are taken to make up the p features.

[0072] For example, for advertising audience u1: the order of the advertising material IDs they click is {c5, c10, …, c30}, the order of the corresponding advertising IDs is {a8, a9, …, a20}, and the order of the corresponding advertiser IDs is {r60, r20, …, r50}. Take 100 of each, and use the most recent one to make up for the number that is less than 100. If there are any excess numbers, keep the most recent 100.

[0073] 2) Convert the three features of each advertising audience into feature vectors, that is, replace each node in the historical time series of the three features of each advertising audience (ad material id, advertising id, advertiser id) with its n-dimensional vector. The three feature vectors of each advertising audience will be used as input features for the training model.

[0074] 3) The basic attributes of each advertising audience are used as labels corresponding to the three features of the advertising audience for supervised training of the LSTM neural network.

[0075] (7) Unit S32: LSTM neural network

[0076] LSTM (Long Short-Term Memory) is a time recurrent neural network. It is specially designed to solve the long-term dependency problem of general RNN (recurrent neural network). It is an improved recurrent neural network that can solve the problem that RNN cannot handle long-distance dependencies. It is also widely used in time series prediction problems.

[0077] The input of the LSTM network designed in this patent is 3 feature sequences (advertising material id, advertising id, advertiser id) of p time steps (consistent with the p value of unit S31), each feature vector is n-dimensional (consistent with the n value of unit S13), and the output is a sequence (age or gender) structure. When a model for predicting age is needed, the LSTM neural network is trained with age as the label; when a model for predicting gender is needed, the LSTM neural network is trained with gender as the label. The structural diagram is shown in Figure 4 As shown;

[0078] Finally, if age is used as a label for training, a model trained to predict the age of the advertising audience based on their historical click behavior is obtained; if gender is used as a label for training, a model trained to predict the gender of the advertising audience based on their historical click behavior is obtained.

[0079] (8) Unit S41: Data Processing: Features

[0080] The operation of this unit is consistent with the feature processing method of unit S31, and the specific steps are:

[0081] 1) For each advertising audience in the test set, the three features of the advertising audience (ad material id, advertising id, advertiser id) are arranged in chronological order, and the length of the three features is p time steps (consistent with the p value of unit S31). If there are more than p features of an advertising audience, the most recent p are retained. If there are less than p features, the most recent ones are taken to make up the p features.

[0082] 2) Convert the three features of each advertising audience into feature vectors, that is, replace each node in the historical time series of the three features of each advertising audience (advertising material id, advertising id, advertiser id) with its n-dimensional vector to represent (consistent with the n value of unit S13). The three feature vectors of each advertising audience will be used as input features of the prediction model.

[0083] (9) Unit S42: Trained prediction model

[0084] That is, the prediction model obtained after the LSTM neural network of unit S32 is trained. If unit S32 is trained with age as a label, an age estimation model of the advertising audience is obtained; if unit S32 is trained with gender as a label, a gender estimation model of the advertising audience is obtained.

[0085] By inputting the three feature vectors (ad material id, ad id, advertiser id) of the historical click behavior of the advertising audience into the age estimation model, their age can be estimated;

[0086] By inputting the three feature vectors (advertising material id, advertisement id, advertiser id) of the historical click behavior of the advertising audience into the gender estimation model, their gender can be estimated.

[0087] At this point, the process of estimating basic attributes of advertising audiences based on heterogeneous graph embedding technology is completed.

[0088] The same or similar reference numerals correspond to the same or similar components;

[0089] The positional relationships described in the drawings are only for illustrative purposes and should not be construed as limiting the present patent.

[0090] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A method for estimating basic attributes of advertising audience based on heterogeneous graph embedding technology, characterized in that: The following steps are involved: S1: Based on the historical click behaviors of all advertising audiences, a heterogeneous graph of the click behavior records of the advertising audiences is constructed, and respective vector representations are generated for these advertising information based on the heterogeneous graph embedding representation technology, wherein the advertising information includes the material id, the advertising id, and the advertiser id; in the step S1, the process of generating respective vector representations for these advertising information based on the heterogeneous graph embedding representation technology includes heterogeneous graph sampling and SkipGram neural network training; The specific process of heterogeneous graph sampling is: Each advertising audience ID is used as the starting node to walk W times along the set meta-path, with each walk taking L steps. Each node walked to is registered in turn, which is a node sequence corpus containing various node types. The SkipGram neural network training process is: The node sequence corpus is input into the SkipGram neural network, and the n-dimensional vector representation of each node in the corpus is obtained through training. The training goal is to make the vector representation of each node maintain the neighbor information of the node in the corpus. The formula for training model optimization is: where c t is the context node of v, and the denominator is u t ∈V t ,That is, the type of node is taken into account during negative sampling; After training, we get the n-dimensional vector representation of each node; S2: Divide the data set into two categories according to whether the basic attributes of the advertising audience are known. The data with known basic attributes is called the training set, which is used to train the prediction model. The other data is called the test set, which is used to predict its unknown basic attributes. The basic attributes of the advertising audience are age and gender. S3: The historical click behavior of the advertising audience in the training set is used as the input feature of the LSTM prediction model, and the age and gender are used as the prediction result labels of the corresponding features. The LSTM neural network is trained to obtain a prediction model that predicts the basic attributes of the advertising audience based on their click behavior. In step S3, the data of the training set is processed into feature vectors and their labels that can be recognized by the subsequent LSTM. The specific steps are: 1) For each advertising audience, arrange the three features of the advertising audience in chronological order, and the length of the three features is p time steps; if there are more than p features of an advertising audience, keep the most recent p; if there are less than p features, take the most recent ones to make up for the p; 2) Convert the three features of each advertising audience into feature vectors, that is, replace each node in the historical time series of the three features of each advertising audience with its n-dimensional vector. The three feature vectors of each advertising audience will be used as input features of the training model; 3) The basic attributes of each advertising audience are used as labels corresponding to the three features of the advertising audience for supervised training of the LSTM neural network; S4: Input the historical click behaviors of the advertising audiences in the test set into the trained prediction model to estimate the basic attributes of these advertising audiences.

2. The method for estimating basic attributes of advertising audiences based on heterogeneous graph embedding technology according to claim 1 is characterized in that: In step S1, the process of constructing a heterogeneous graph of the click behavior records of the advertising audience is: S11: Generate nodes of heterogeneous graph: Each different id number of each type is used as a node of heterogeneous graph; S12: Connecting nodes of heterogeneous graphs with edges: There are three types of edge relationships between the four types of nodes: the click relationship between advertising audiences and advertising materials, the subordinate relationship between advertising materials and advertisements, and the publishing relationship between advertisements and advertisers; if any of the above three relationships exists between two nodes, the two nodes are connected by an edge.

3. The method for estimating basic attributes of advertising audiences based on heterogeneous graph embedding technology according to claim 2 is characterized in that: In step S2, the process of data set division is: The original data set is divided into two categories according to whether the basic attributes of the advertising audience are known. The data with known basic attributes is called the training set. The advertising audience click behavior in the training set is used as the feature of the training model, and the basic attributes corresponding to the features are used as training labels. The features and corresponding labels are used to train the LSTM neural network to obtain a prediction model. The other type of data is called the test set, which is used to predict the unknown basic attributes of the trained prediction model.

4. The method for estimating basic attributes of advertising audiences based on heterogeneous graph embedding technology according to claim 3 is characterized in that: The input of the LSTM network is 3 feature sequences of p time steps, each feature vector is n-dimensional, and the output is a sequence structure; when a model is needed to predict age, the LSTM neural network is trained with age as the label; when a model is needed to predict gender, the LSTM neural network is trained with gender as the label.

5. The method for estimating basic attributes of advertising audiences based on heterogeneous graph embedding technology according to claim 4 is characterized in that: In step S4, for each advertising audience in the test set, the three features of the advertising audience are arranged in chronological order, and the length of the three features is p time steps; if there are more than p features of an advertising audience, the most recent p features are retained; if there are less than p features, the most recent p features are taken to make up the p features; The three features of each advertising audience are converted into feature vectors, that is, each node in the historical time series of the three features of each advertising audience is replaced by its n-dimensional vector; the three feature vectors of each advertising audience will be used as input features of the prediction model.

6. The method for estimating basic attributes of advertising audiences based on heterogeneous graph embedding technology according to claim 5, characterized in that: The prediction model obtained after the LSTM neural network in step S3 is trained; if age is used as a label for training in step S3, an age estimation model of the advertising audience is obtained; if gender is used as a label for training in step S3, a gender estimation model of the advertising audience is obtained; the three feature vectors of the historical click behavior of the advertising audience are input into the age estimation model, and their age can be estimated; By inputting the three feature vectors of the historical click behavior of the advertising audience into the gender estimation model, their gender can be estimated.

Images