App advertising delivery method and system

Abstract

The present invention discloses an App advertisement delivery method and system, including: obtaining App user data and social data supplementing the App user data; extracting user social feature vectors; extracting user information feature vectors and content feature vectors; obtaining user feature vectors; based on the extracted user feature vectors, calculating the similarity between App users and social users, selecting a preset number of social users with high similarity to each App user, and supplementing the corresponding data to the App user; using historical advertisement information clicked by App users as supplementary information of App users and calculating the matching degree between App users and candidate advertisements, and recommending a preset number of advertisements with high matching degrees to App users; the present invention has the advantage of achieving an ideal recommendation effect in a scenario where user data and social information are lacking.

Description

Technical Field

[0001] The present invention relates to the field of online advertising, and in particular to an App advertising delivery method and system. Background Art

[0002] Online advertising is a new advertising model that uses the Internet as a media tool to publish advertisements. Placing advertisements on apps is one of the main ways to place online advertisements. With the increasing popularity of mobile devices, the market share of app advertising is also increasing. Only by placing personalized advertisements based on the different interests of different users can advertising be converted into user consumption behavior, so that both advertisers and advertisers can get good commercial returns.

[0003] Existing advertising recommendation algorithms obtain user interest characteristics by analyzing user characteristics, and recommend information of interest to users (such as advertisements, advertising videos, etc.), such as a method and system for intelligent advertising delivery across e-commerce platforms and social platforms disclosed in Chinese Patent Publication No. CN112435067A. Although the above recommendation method has achieved good recommendation results, the premise for the implementation of the above method is that there are already a lot of social information and historical data related to users. In reality, except for social software for communication purposes such as QQ, WeChat and other dating software, which have relatively complete social relationships, other recording software, notepads, alarm clocks, etc. have almost no user social data. And these application apps also have scenarios for recommending advertisements. In the absence of user data and social information, the recommendation effect of existing recommendation algorithms is not ideal, and there is currently no effective solution. Summary of the invention

[0004] The technical problem to be solved by the present invention is that the existing advertising delivery method has unsatisfactory recommendation effect in the scenario where user data and social information are lacking.

[0005] The present invention solves the above technical problems through the following technical means: an App advertising delivery method, comprising the following steps:

[0006] Step 1: Obtain App user data and social data that supplements App user data;

[0007] Step 2: Extract user social feature vector;

[0008] Step 3: Extracting user information feature vector and content feature vector;

[0009] Step 4: Concatenate the user social feature vector, the user information feature vector and the content feature vector to obtain a user feature vector;

[0010] Step 5: Based on the extracted user feature vector, calculate the similarity between the App user and the social user, select a preset number of social users with high similarity for each App user, and supplement the corresponding data to the App user;

[0011] Step 6: Use the historical advertisement information clicked by the App user as the supplementary information of the App user and calculate the matching degree between the App user and the candidate advertisements, and recommend a preset number of advertisements with a high matching degree to the App user.

[0012] Furthermore, the step 1 comprises:

[0013] S11, extract the basic information of the App user and construct a basic information list R, where each line of the basic information list R is the basic information of a user, including nickname, email address, age, and registration time;

[0014] S12, extracting social information from the App platform and establishing a social network G, where nodes in the social network G represent users. If two users have a friend relationship, an edge is connected between the nodes corresponding to the two users;

[0015] S13, extracting the list Ah of advertisement click history records of the App user, where each line of the list Ah is a collection of information about an advertisement clicked by an App user;

[0016] S14, extracting App platform recorded data and constructing a description information list C, wherein each line in the description information list C is data recorded by an App user during the use of the App, such as events in a memo and text content in a notepad;

[0017] S15. Based on the user data of the social platform, a supplementary basic information list Rs, a supplementary social network Gs and a supplementary description information list Cs of the supplementary data are established. The process is similar to steps S11, S12 and S14, wherein each line of the supplementary description information list Cs is the description information of the user, including the Weibo, Moments, Moments, comments and likes data published or forwarded by the user.

[0018] Furthermore, the step 2 includes:

[0019] S21. Generate node walk sequences. For each node in the social network, generate n = 5 random walk sequences with a length of l = 8. Assume that the node The resulting random walk sequence is , in this sequence is from A node is randomly selected from its neighbor nodes.

[0020] S22. Use the Skip-Gram model to train the walk sequence and obtain the vector representation of the node. Representation Node Under this model, nodes in the sequence are used to predict other nodes that appear in the window. Assume that the center node of the window is , the window width is , the nodes that appear in the window are In the Skip-Gram model, given Appear The probability of , in for and The loss function of the Skip-Gram model is formula (1). Traverse each node in the current random walk sequence, calculate the loss function of the Skip-Gram model, calculate the gradient and update the embedded vector. Repeat the above process for all walk sequences until all walk sequences are traversed, so as to obtain the node embedding vector, that is, the user's social feature vector.

[0021] (1)

[0022] S23: For the App platform social network G and the supplementary social network Gs, execute steps S21 and S22 respectively to obtain vector representations of App users and social users. and ,in and Represents App users and social users respectively. For convenience, App users and social users The social feature vectors are recorded as and .

[0023] Furthermore, the step three includes:

[0024] S31. Obtain the pre-trained BERT model;

[0025] S32. Use the BERT model to extract the user information feature vector from the basic information list R and the supplementary basic information list Rs, and record the information feature vector of App user v as , the information feature vector of social user u is recorded as ;

[0026] S33, using the BERT model to extract the user content feature vector from the description information list C and the supplementary description information list Cs, the content feature vector of App user v is recorded as , the content feature vectors of social user u are recorded as .

[0027] Furthermore, the step 4 includes:

[0028] According to formula (2), the social feature vector of App user v is , information feature vector and content feature vector The feature vector of App user v is obtained by splicing; similarly, according to formula (3), the social feature vector, information feature vector and content feature vector of social user u are spliced ​​to obtain the feature vector of social user u

[0029] (2)

[0030] (3).

[0031] Furthermore, the step five includes:

[0032] S51. Calculate App users using formula (4) and social users The similarity of Represents the App user collection, Represents a collection of social users

[0033] (4)

[0034] S52, from Select the first k users with high similarity to user v, denoted as Uk;

[0035] S53. Add the description information of all users in Uk to the description information of App user v.

[0036] Furthermore, the step six includes:

[0037] S61. Use the historical advertisements clicked by the App user to supplement the description information of the App user. h The sequence of ads clicked by App user v is obtained as , the advertisement Add the description information of App user v to the description information of App user v, and perform the above steps for all App users;

[0038] S62: The relevant text description information of the advertisement is used as the description information of the advertisement. The description information of the APP user and the description information of the candidate advertisement are input into the BERT model obtained in step S31 to obtain the updated content feature vector of the user and the content feature vector of the candidate advertisement respectively. The updated content feature vector of the App user v is recorded as , the content feature vector of candidate advertisement a is recorded as ;

[0039] S63. Calculate the matching degree between App users and candidate ads using formula (5)

[0040] (5)

[0041] S64. Sort the candidate advertisements according to the values ​​of the matching degrees from large to small, and select the first k advertisements with high matching degrees to recommend to the user v.

[0042] The present invention also provides an App advertisement delivery system, comprising:

[0043] A data acquisition module, used to acquire App user data and social data that supplements App user data;

[0044] A first feature extraction module, used to extract user social feature vectors;

[0045] A second feature extraction module is used to extract a user information feature vector and a content feature vector;

[0046] A feature concatenation module is used to concatenate the user social feature vector, the user information feature vector and the content feature vector to obtain a user feature vector;

[0047] A data supplementation module is used to calculate the similarity between the App user and the social user based on the extracted user feature vector, select a preset number of social users with high similarity for each App user, and supplement the corresponding data to the App user;

[0048] The advertisement recommendation module is used to use the historical advertisement information clicked by the App user as the supplementary information of the App user and calculate the matching degree between the App user and the candidate advertisements, and recommend a preset number of advertisements with a high matching degree to the App user.

[0049] Furthermore, the data acquisition module is also used for:

[0050] S11, extract the basic information of the App user and construct a basic information list R, where each line of the basic information list R is the basic information of a user, including nickname, email address, age, and registration time;

[0051] S12, extracting social information from the App platform and establishing a social network G, where nodes in the social network G represent users. If two users have a friend relationship, an edge is connected between the nodes corresponding to the two users;

[0052] S13: Extract the App user's ad click history list Ah, ad click history list Ah Each line is a collection of information about an ad clicked by an App user;

[0053] S14, extracting App platform recorded data and constructing a description information list C, wherein each line in the description information list C is data recorded by an App user during the use of the App, such as events in a memo and text content in a notepad;

[0054] S15. Based on the user data of the social platform, a supplementary basic information list Rs, a supplementary social network Gs and a supplementary description information list Cs of the supplementary data are established. The process is similar to steps S11, S12 and S14, wherein each line of the supplementary description information list Cs is the description information of the user, including the Weibo, Moments, Moments, comments and likes data published or forwarded by the user.

[0055] Furthermore, the first feature extraction module is also used for:

[0056] S21. Generate node walk sequences. For each node in the social network, generate n = 5 random walk sequences with a length of l = 8. Assume that the node The resulting random walk sequence is , in this sequence is from A node is randomly selected from its neighbor nodes.

[0057] S22. Use the Skip-Gram model to train the walk sequence and obtain the vector representation of the node. Representation Node Under this model, nodes in the sequence are used to predict other nodes that appear in the window. Assume that the center node of the window is , the window width is , the nodes that appear in the window are In the Skip-Gram model, given Appear The probability of , in for and The loss function of the Skip-Gram model is formula (1). Traverse each node in the current random walk sequence, calculate the loss function of the Skip-Gram model, calculate the gradient and update the embedded vector. Repeat the above process for all walk sequences until all walk sequences are traversed, so as to obtain the node embedding vector, that is, the user's social feature vector.

[0058] (1)

[0059] S23: For the App platform social network G and the supplementary social network Gs, execute steps S21 and S22 respectively to obtain vector representations of App users and social users. and ,in and Represents App users and social users respectively. For convenience, App users and social users The social feature vectors are recorded as and .

[0060] Furthermore, the second feature extraction module is also used for:

[0061] S31. Obtain the pre-trained BERT model;

[0062] S32. Use the BERT model to extract the user information feature vector from the basic information list R and the supplementary basic information list Rs, and record the information feature vector of App user v as , the information feature vector of social user u is recorded as ;

[0063] S33. Use the BERT model to extract the user content feature vector from the description information list C and the supplementary description information list Cs. The content feature vector of App user v is recorded as , the content feature vectors of social user u are recorded as .

[0064] Furthermore, the feature stitching module is also used for:

[0065] According to formula (2), the social feature vector of App user v is , information feature vector and content feature vector The feature vector of App user v is obtained by splicing; similarly, according to formula (3), the social feature vector, information feature vector and content feature vector of social user u are spliced ​​to obtain the feature vector of social user u

[0066] (2)

[0067] (3).

[0068] Furthermore, the data supplement module is also used for:

[0069] S51. Calculate App users using formula (4) and social users The similarity of Represents the App user collection, Represents a collection of social users

[0070] (4)

[0071] S52, from Select the first k users with high similarity to user v, denoted as Uk;

[0072] S53. Add the description information of all users in Uk to the description information of App user v.

[0073] Furthermore, the advertisement recommendation module is also used for:

[0074] S61. Use the historical advertisements clicked by the App user to supplement the description information of the App user. h The sequence of ads clicked by App user v is obtained as , the advertisement Add the description information of App user v to the description information of App user v, and perform the above steps for all App users;

[0075] S62: The relevant text description information of the advertisement is used as the description information of the advertisement. The description information of the APP user and the description information of the candidate advertisement are input into the BERT model obtained in step S31 to obtain the updated content feature vector of the user and the content feature vector of the candidate advertisement respectively. The updated content feature vector of the App user v is recorded as , the content feature vector of candidate advertisement a is recorded as ;

[0076] S63. Calculate the matching degree between App users and candidate ads using formula (5)

[0077] (5)

[0078] S64. Sort the candidate advertisements according to the values ​​of the matching degrees from large to small, and select the first k advertisements with high matching degrees to recommend to the user v.

[0079] The advantages of the present invention are as follows: the present invention calculates the similarity between App users and social users. Similar users often have similar interests. The present invention searches for social users similar to App users and supplements the information of social users to App users. In addition, the historical advertisement information clicked by App users is also used as supplementary information of App users, thereby solving the problem of insufficient App user data. Then, the feature vectors of App users and candidate advertisements are extracted, and the matching degree between App users and candidate advertisements is calculated. The matching degree indicates the correlation between advertisements and user interests. The most matching advertisements are recommended to users, which is more likely to make users click on them, thereby achieving an ideal recommendation effect, improving the click-through rate of advertisements and thus increasing revenue. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 The present invention is a flowchart of an App advertising delivery method disclosed in an embodiment of the present invention. DETAILED DESCRIPTION

[0081] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in combination with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0082] Example 1

[0083] Existing advertising recommendation algorithms are suitable for scenarios where a large amount of user social information and network behavior data has been accumulated. However, for most App software, they do not have such user social data and network behavior data, which makes the recommendation effect of existing recommendation algorithms unsatisfactory. Based on this, the present invention provides an App advertising delivery method, which uses social data to supplement App user data and then calculates the matching degree between the advertisement to be delivered and the App user. Then, candidate advertisements with a high matching degree are delivered to App users, such as Figure 1 As shown, the method mainly includes the following steps:

[0084] S1. Get App user data D=(R, G, C, U, A h ) and social data SD=(Rs, Gs, Cs, Us) to supplement App user data. Social data SD comes from public anonymized social data, crawled / purchased social data, and social data obtained through other channels. Represents the App user collection, Represents a collection of social users; the specific process is:

[0085] S11, extract the basic information of the App user and construct a basic information list R, where each line of R is the basic information of a user, including nickname, email address, age, and registration time;

[0086] S12, extracting social information from the App platform and establishing a social network G. The nodes in G represent users. If two users have a friend relationship, an edge is connected between the nodes corresponding to the two users;

[0087] S13. Extract the ad click history list A of the App user h , A h Each line is a collection of information about an ad clicked by an App user;

[0088] S14, extracting App platform recorded data and constructing a description information list C, wherein each line in the description information list C is data recorded by an App user during the use of the App, such as events in a memo and text content in a notepad;

[0089] S15. Based on the user data of the social platform, a supplementary basic information list Rs, a supplementary social network Gs and a supplementary description information list Cs of the supplementary data are established. The process is similar to steps S11, S12 and S14, wherein each line of the supplementary description information list Cs is the description information of the user, including the Weibo, Moments, Moments, comments and likes data published or forwarded by the user.

[0090] S2. Use the DeepWalk method to extract the user social feature vector; the specific process is as follows:

[0091] S21. Generate node walk sequences. For each node in the social network, generate n = 5 random walk sequences with a length of l = 8. Assume that the node The resulting random walk sequence is , in this sequence is from A node is randomly selected from its neighbor nodes.

[0092] S22. Use the Skip-Gram model to train the walk sequence and obtain the vector representation of the node. Representation Node Under this model, nodes in the sequence are used to predict other nodes that appear in the window. Assume that the center node of the window is , the window width is , the nodes that appear in the window are In the Skip-Gram model, given Appear The probability of , in for and The loss function of the Skip-Gram model is formula (1). Traverse each node in the current random walk sequence, calculate the loss function of the Skip-Gram model, calculate the gradient and update the embedded vector. Repeat the above process for all walk sequences until all walk sequences are traversed, so as to obtain the node embedding vector, that is, the user's social feature vector.

[0093] (1)

[0094] S23: For the App platform social network G and the supplementary social network Gs, execute steps S21 and S22 respectively to obtain vector representations of App users and social users. and ,in and Represents App users and social users respectively. For convenience, App users and social users The social feature vectors are recorded as and .

[0095] S3. Use the BERT model to extract user information feature vectors and content feature vectors from basic information R / Rs and description information C / Cs respectively. The specific process is as follows:

[0096] S31. Get a pre-trained BERT model. Hugging Face, Google, etc. provide many pre-trained BERT models. Considering the high cost of training a BERT model, use a pre-trained BERT model.

[0097] S32. Use the BERT model to extract the user information feature vector from the basic information list R and the supplementary basic information list Rs. Input the user's basic information R and Rs into the pre-trained BERT to obtain the user information feature vector. For convenience, the information feature vector of App user v is recorded as , the information feature vector of social user u is recorded as ;

[0098] S33. Use the BERT model to extract the user content feature vector from the description information list C and the supplementary description information list Cs. Input the description information C and Cs into the pre-trained BERT to obtain the user's content feature vector. For convenience, the content feature vector of App user v is recorded as , the content feature vectors of social user u are recorded as .

[0099] S4. Concatenate the user social feature vector, the user information feature vector and the content feature vector to obtain a user feature vector. The specific process is as follows:

[0100] The user feature vector includes the user social feature vector, information feature vector and content feature vector. According to formula (2), the social feature vector of App user v is , information feature vector and content feature vector The feature vector of App user v is obtained by splicing; similarly, according to formula (3), the social feature vector, information feature vector and content feature vector of social user u are spliced ​​to obtain the feature vector of social user u

[0101] (2)

[0102] (3).

[0103] S5: Based on the extracted user feature vector, the similarity between the App user and the social user is calculated, a preset number of social users with high similarity are selected for each App user, and the corresponding data is supplemented to the App user; the specific process is as follows:

[0104] S51. Calculate App users using formula (4) and social users The similarity of Represents the App user collection, represents the social user set, specifically, the feature vectors of user u and user v are obtained based on step S4 and , use formula (4) to calculate the cosine similarity between user u and user v.

[0105] (4)

[0106] S52, from Select the first k users with high similarity to user v. Specifically, sort the users in Us in descending order according to the similarity between the users in Us and user v, and take the first k users, which are recorded as Uk;

[0107] S53. Add the description information of all users in Uk to the description information of App user v.

[0108] S6. Use the historical advertisement information clicked by the App user as the supplementary information of the App user and calculate the matching degree between the App user and the candidate advertisements, and recommend a preset number of advertisements with high matching degree to the App user. The specific process is as follows:

[0109] S61. Use the historical advertisements clicked by the App user to supplement the description information of the App user. h The sequence of ads clicked by App user v is obtained as , the advertisement Add the description information of App user v to the description information of App user v, and perform the above steps for all App users;

[0110] S62: Extract candidate advertisements and the content feature vector of the App user, specifically, the relevant text description information of the advertisement is used as the description information of the advertisement, the description information of the App user and the description information of the candidate advertisement are input into the BERT model obtained in step S31, and the updated content feature vector of the user and the content feature vector of the candidate advertisement are obtained respectively. For convenience, the updated content feature vector of the App user v is recorded as , the content feature vector of candidate advertisement a is recorded as ;

[0111] S63, calculating the matching degree between the App user and the candidate advertisement. Specifically, the matching degree between the App user v and the candidate advertisement is calculated by formula (5): The degree of match

[0112] (5)

[0113] S64: Select the advertisement that best matches the user v for recommendation. Specifically, the matching degree sim(v,a) between the user v and all advertisements to be delivered is calculated according to formula (5), the advertisements are sorted in descending order according to the sim(v,a) value, and the top k matching advertisements are selected and recommended to the user v.

[0114] Through the above technical scheme, the present invention obtains App user data and social data that supplements App user data. Feature vectors of App users and social users are extracted from three aspects: social relationship, basic information, and text description. Based on the extracted feature vectors, the similarity between App users and social users is calculated. Similar users often have similar interests. Social users similar to App users are found, and the information of social users is supplemented to App users. In addition, the historical advertising information clicked by App users is also used as supplementary information for App users, thereby solving the problem of less App user data. Then, the feature vectors of App users and candidate advertisements are extracted, and the matching degree between App users and candidate advertisements is calculated. The matching degree indicates the correlation between advertisements and user interests. The most matching advertisements are recommended to users, which is more likely to make users click, thereby achieving the ideal recommendation effect, improving the click-through rate of advertisements and increasing revenue.

[0115] Example 2

[0116] Based on Example 1, Example 2 of the present invention further provides an App advertising delivery system, including:

[0117] A data acquisition module, used to acquire App user data and social data that supplements App user data;

[0118] A first feature extraction module, used to extract user social feature vectors;

[0119] A second feature extraction module is used to extract a user information feature vector and a content feature vector;

[0120] A feature concatenation module is used to concatenate the user social feature vector, the user information feature vector and the content feature vector to obtain a user feature vector;

[0121] A data supplementation module is used to calculate the similarity between the App user and the social user based on the extracted user feature vector, select a preset number of social users with high similarity for each App user, and supplement the corresponding data to the App user;

[0122] The advertisement recommendation module is used to use the historical advertisement information clicked by the App user as the supplementary information of the App user and calculate the matching degree between the App user and the candidate advertisements, and recommend a preset number of advertisements with a high matching degree to the App user.

[0123] Specifically, the data acquisition module is also used for:

[0124] S11, extract the basic information of the App user and construct a basic information list R, where each line of the basic information list R is the basic information of a user, including nickname, email address, age, and registration time;

[0125] S12, extracting social information from the App platform and establishing a social network G, where nodes in the social network G represent users. If two users have a friend relationship, an edge is connected between the nodes corresponding to the two users;

[0126] S13. Extract the ad click history list A of the App user h , ad click history list A h Each line is a collection of information about an ad clicked by an App user;

[0127] S14, extracting App platform recorded data and constructing a description information list C, wherein each line in the description information list C is data recorded by an App user during the use of the App, such as events in a memo and text content in a notepad;

[0128] S15. Based on the user data of the social platform, a supplementary basic information list Rs, a supplementary social network Gs and a supplementary description information list Cs of the supplementary data are established. The process is similar to steps S11, S12 and S14, wherein each line of the supplementary description information list Cs is the description information of the user, including the Weibo, Moments, Moments, comments and likes data published or forwarded by the user.

[0129] More specifically, the first feature extraction module is further used for:

[0130] S21. Generate node walk sequences. For each node in the social network, generate n = 5 random walk sequences with a length of l = 8. Assume that the node The resulting random walk sequence is , in this sequence is from A node is randomly selected from its neighbor nodes.

[0131] S22. Use the Skip-Gram model to train the walk sequence and obtain the vector representation of the node. Representation Node Under this model, nodes in the sequence are used to predict other nodes that appear in the window. Assume that the center node of the window is , the window width is , the nodes that appear in the window are In the Skip-Gram model, given Appear The probability of , in for and The loss function of the Skip-Gram model is formula (1). Traverse each node in the current random walk sequence, calculate the loss function of the Skip-Gram model, calculate the gradient and update the embedded vector. Repeat the above process for all walk sequences until all walk sequences are traversed, so as to obtain the node embedding vector, that is, the user's social feature vector.

[0132] (1)

[0133] S23: For the App platform social network G and the supplementary social network Gs, execute steps S21 and S22 respectively to obtain vector representations of App users and social users. and ,in and Represents App users and social users respectively. For convenience, App users and social users The social feature vectors are recorded as and .

[0134] More specifically, the second feature extraction module is further used for:

[0135] S31. Obtain the pre-trained BERT model;

[0136] S32. Use the BERT model to extract the user information feature vector from the basic information list R and the supplementary basic information list Rs, and record the information feature vector of App user v as , the information feature vector of social user u is recorded as ;

[0137] S33. Use the BERT model to extract the user content feature vector from the description information list C and the supplementary description information list Cs. The content feature vector of App user v is recorded as , the content feature vectors of social user u are recorded as .

[0138] More specifically, the feature stitching module is also used for:

[0139] According to formula (2), the social feature vector of App user v is , information feature vector and content feature vector The feature vector of App user v is obtained by splicing; similarly, according to formula (3), the social feature vector, information feature vector and content feature vector of social user u are spliced ​​to obtain the feature vector of social user u

[0140] (2)

[0141] (3).

[0142] More specifically, the data supplement module is also used for:

[0143] S51. Calculate App users using formula (4) and social users The similarity of Represents the App user collection, Represents a collection of social users

[0144] (4)

[0145] S52, from Select the first k users with high similarity to user v, denoted as Uk;

[0146] S53. Add the description information of all users in Uk to the description information of App user v.

[0147] More specifically, the advertisement recommendation module is also used for:

[0148] S61. Use the historical advertisements clicked by the App user to supplement the description information of the App user. h The sequence of ads clicked by App user v is obtained as , the advertisement Add the description information of App user v to the description information of App user v, and perform the above steps for all App users;

[0149] S62: The relevant text description information of the advertisement is used as the description information of the advertisement. The description information of the APP user and the description information of the candidate advertisement are input into the BERT model obtained in step S31 to obtain the updated content feature vector of the user and the content feature vector of the candidate advertisement respectively. The updated content feature vector of the App user v is recorded as , the content feature vector of candidate advertisement a is recorded as ;

[0150] S63. Calculate the matching degree between App users and candidate ads using formula (5)

[0151] (5)

[0152] S64. Sort the candidate advertisements according to the values ​​of the matching degrees from large to small, and select the first k advertisements with high matching degrees to recommend to the user v.

[0153] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for delivering App advertisements, characterized in that: The following steps are involved: Step 1: Get App user data and social data to supplement App user data; specifically: Get App user data D = (R, G, C, U, A h ) and social data SD=(Rs, Gs, Cs, Us) to supplement App user data. Social data SD comes from public anonymized social data, crawled / purchased social data, and social data obtained through other channels. R represents the basic information list, G represents the social network, and C represents the description information list. Represents the App user set, A h represents the ad click history list, Rs represents the supplementary basic information list, Gs represents the supplementary social network, and Cs represents the supplementary description information list. Represents a collection of social users; S11, extract the basic information of the App user and construct a basic information list R, where each line of R is the basic information of a user, including nickname, email address, age, and registration time; S12, extracting social information from the App platform and establishing a social network G; the nodes in G represent users. Assuming that two users have a friend relationship, an edge is connected between the nodes corresponding to the two users; S13. Extract the ad click history list A of the App user h , A h Each line is a collection of information about an ad clicked by an App user; S14, extracting App platform recorded data and constructing a description information list C, wherein each line in the description information list C is data recorded by an App user during the use of the App; S15. Based on the user data of the social platform, a supplementary basic information list Rs, a supplementary social network Gs and a supplementary description information list Cs of supplementary data are established, wherein each line of the supplementary description information list Cs is the description information of the user, including the microblogs, friend circles, chats, comments and likes data published or forwarded by the user; Step 2: Use the DeepWalk method to extract user social feature vectors; Step 3: Use the BERT model to extract user information feature vectors and content feature vectors from the basic information R / Rs and description information C / Cs respectively; Step 4: Concatenate the user social feature vector, the user information feature vector and the content feature vector to obtain a user feature vector; Step 5: Based on the extracted user feature vector, calculate the similarity between the App user and the social user, select a preset number of social users with high similarity for each App user, and supplement the corresponding data to the App user; Step 6: Use the historical advertisement information clicked by the App user as the supplementary information of the App user and calculate the matching degree between the App user and the candidate advertisements, and recommend a preset number of advertisements with a high matching degree to the App user.

2. The method for placing App advertisements according to claim 1, characterized in that: The second step comprises: S21. Generate node walk sequences. For each node in the social network, generate n random walk sequences of length l. Assume that the node The resulting random walk sequence is , in this sequence is from A node randomly selected from the neighboring nodes of S22, use the Skip-Gram model to train the walk sequence and obtain the vector representation of the node. Representation Node Embedding vector of; Under this model, the nodes in the sequence are used to predict other nodes that appear in the window. Assume that the center node of the window is , the window width is , the nodes that appear in the window are ; In the Skip-Gram model, given Appear The probability of , in for and The data product of the Skip-Gram model is the formula (1). Each node in the current random walk sequence is traversed, the loss function of the Skip-Gram model is calculated, the gradient is calculated and the embedded vector is updated. The above process is repeated for all walk sequences until all walk sequences are traversed, thereby obtaining the embedding vector of the node, that is, the social feature vector of the user. (1) S23: For the App platform social network G and the supplementary social network Gs, execute steps S21 and S22 respectively to obtain vector representations of App users and social users. and ,in and Represents App users and social users respectively; App users and social users The social feature vectors are recorded as and .

3. The method for placing App advertisements according to claim 2, characterized in that: The step three comprises: S31. Obtain the pre-trained BERT model; S32. Use the BERT model to extract the user information feature vector from the basic information list R and the supplementary basic information list Rs, and record the information feature vector of App user v as , the information feature vector of social user u is recorded as ; S33. Use the BERT model to extract the user content feature vector from the description information list C and the supplementary description information list Cs. The content feature vector of App user v is recorded as , the content feature vectors of social user u are recorded as .

4. The method for placing App advertisements according to claim 3, characterized in that: The fourth step comprises: According to formula (2), the social feature vector of App user v is , information feature vector and content feature vector Concatenate to obtain the feature vector of App user v; According to formula (3), concatenate the social feature vector, information feature vector and content feature vector of social user u to obtain the feature vector of social user u （2） （3）。 5. The method for placing App advertisements according to claim 4, characterized in that: The step five comprises: S51. Calculate App users using formula (4) and social users The similarity of Represents the App user collection, Represents a collection of social users （4） S52, from Select the first k users with high similarity to user v, denoted as Uk; S53. Add the description information of all users in Uk to the description information of App user v.

6. The method for placing App advertisements according to claim 5, characterized in that: The step six comprises: S61. Use the historical advertisements clicked by the App user to supplement the description information of the App user, and obtain the advertisement sequence clicked by the App user v from Ah: , the advertisement Add the description information of App user v to the description information of App user v, and perform the above steps for all App users; S62: The relevant text description information of the advertisement is used as the description information of the advertisement. The description information of the APP user and the description information of the candidate advertisement are input into the BERT model obtained in step S31 to obtain the updated content feature vector of the user and the content feature vector of the candidate advertisement respectively. The updated content feature vector of the App user v is recorded as , the content feature vector of candidate advertisement a is recorded as ; S63. Calculate the matching degree between App users and candidate ads using formula (5) （5） S64. Sort the candidate advertisements according to the values ​​of the matching degrees from large to small, and select the first k advertisements with high matching degrees to recommend to the user v.

7. An App advertising delivery system, characterized in that: include: The data acquisition module is used to obtain App user data and social data that supplements App user data; specifically: obtain App user data D=(R, G, C, U, A h ) and social data SD = (Rs, Gs, Cs, Us) to supplement App user data. Social data SD comes from public anonymized social data, crawled / purchased social data, and social data obtained through other channels. R represents the basic information list, G represents the social network, and C represents the description information list. Represents the App user set, A h represents the ad click history list, Rs represents the supplementary basic information list, Gs represents the supplementary social network, and Cs represents the supplementary description information list. Represents a collection of social users; S11, extract the basic information of the App user and construct a basic information list R, where each line of R is the basic information of a user, including nickname, email address, age, and registration time; S12, extracting social information from the App platform and establishing a social network G; the nodes in G represent users. Assuming that two users have a friend relationship, an edge is connected between the nodes corresponding to the two users; S13. Extract the ad click history list A of the App user h , A h Each line is a collection of information about an ad clicked by an App user; S14, extracting App platform recorded data and constructing a description information list C, wherein each line in the description information list C is data recorded by an App user during the use of the App; S15. Based on the user data of the social platform, a supplementary basic information list Rs, a supplementary social network Gs and a supplementary description information list Cs of supplementary data are established, wherein each line of the supplementary description information list Cs is the description information of the user, including the microblogs, friend circles, chats, comments and likes data published or forwarded by the user; The first feature extraction module uses the DeepWalk method to extract user social feature vectors; The second feature extraction module uses the BERT model to extract user information feature vectors and content feature vectors from the basic information R / Rs and description information C / Cs respectively; A feature concatenation module is used to concatenate the user social feature vector, the user information feature vector and the content feature vector to obtain a user feature vector; A data supplementation module is used to calculate the similarity between the App user and the social user based on the extracted user feature vector, select a preset number of social users with high similarity for each App user, and supplement the corresponding data to the App user; The advertisement recommendation module is used to use the historical advertisement information clicked by the App user as the supplementary information of the App user and calculate the matching degree between the App user and the candidate advertisements, and recommend a preset number of advertisements with a high matching degree to the App user.

8. The App advertising delivery system according to claim 7, characterized in that: The first feature extraction module is also used for: S21. Generate node walk sequences. For each node in the social network, generate n random walk sequences of length l. Assume that the node The resulting random walk sequence is , in this sequence is from A node randomly selected from the neighboring nodes of S22, use the Skip-Gram model to train the walk sequence and obtain the vector representation of the node. Representation Node Embedding vector of; Under this model, the nodes in the sequence are used to predict other nodes that appear in the window. Assume that the center node of the window is , the window width is , the nodes that appear in the window are ; In the Skip-Gram model, given Appear The probability of , in for and The data product of the Skip-Gram model is the formula (1). Each node in the current random walk sequence is traversed, the loss function of the Skip-Gram model is calculated, the gradient is calculated and the embedded vector is updated. The above process is repeated for all walk sequences until all walk sequences are traversed, thereby obtaining the embedding vector of the node, that is, the social feature vector of the user. (1) S23: For the App platform social network G and the supplementary social network Gs, execute steps S21 and S22 respectively to obtain vector representations of App users and social users. and ,in and Represents App users and social users respectively; App users and social users The social feature vectors are recorded as and .

Images