A method for advertising push based on big data

Abstract

The present application relates to the field of advertising push technology, and in particular to an advertising push method based on big data, including: constructing a rating matrix based on each user's feedback behavior on each historical advertisement, decomposing the rating matrix into a user feature matrix and an advertisement feature matrix using the alternating least squares method; calculating a loss function based on the advertisement features in the advertisement feature matrix to train a feature extraction model; inputting a newly invested advertisement into the feature extraction model, multiplying the user feature matrix with the advertisement features of the newly invested advertisement, obtaining each user's predicted rating for the newly invested advertisement, and pushing the newly invested advertisement to users whose predicted rating is greater than a rating threshold. The technical solution of the present application can accurately obtain the feature vectors of each advertisement and realize accurate advertisement push.

Description

Technical Field

[0001] The present application relates to the technical field of advertisement push, and in particular to an advertisement push method based on big data. Background Art

[0002] With the rapid development of e-commerce and online advertising, providing users with personalized advertising content has become the core goal of advertising push. Most advertising push systems rely mainly on users' purchase history or browsing history to push ads, which requires obtaining users' feedback on ads such as purchase history or browsing history. However, for a newly launched ad, it is impossible to collect any user's feedback on the newly launched ad, which leads to inaccurate push.

[0003] At present, a patent application document with publication number CN117495458A discloses an online advertising push method based on user portraits, the method comprising: constructing a user portrait based on user data; constructing a user feature vector representing user interest preferences based on the user portrait; using the user feature vector as training data to train a preset alternating least squares model, obtain a pre-trained model and perform prediction, and generate an advertising recommendation ranking result; constructing an enhanced deep learning model based on the pre-trained model and the advertising recommendation ranking result; and in response to user browsing behavior feedback, performing online advertising recommendation updates through the enhanced deep learning model.

[0004] Among them, the alternating least squares model is often used in the field of product recommendation. It is necessary to collect the click rate of users on each advertisement to calculate the user rating matrix, and then use the alternating least squares method to decompose the user rating matrix into the product of the user feature matrix and the advertisement feature matrix. The user feature matrix and the advertisement feature matrix can be used to predict the rating of any user for any advertisement.

[0005] The above method pushes advertisements that may be of interest to target users by building user portraits. However, the above method relies on the user's browsing behavior of advertisements. For a newly launched advertisement or any other advertisement, it is impossible to collect the user's rating of the advertisement, and it is impossible to accurately obtain the feature vector of the advertisement, which leads to inaccurate advertisement push. Summary of the invention

[0006] In order to solve the technical problem of inaccurate advertising push, the present application provides an advertising push method based on big data, which can accurately obtain the feature vector of each advertisement and realize accurate advertising push.

[0007] The present application provides an advertisement push method based on big data, the push method comprising: constructing a rating matrix according to each user's feedback behavior on each historical advertisement, decomposing the rating matrix into a user feature matrix and an advertisement feature matrix by using the alternating least squares method; calculating a loss function according to the advertisement features in the advertisement feature matrix to train a feature extraction model, the feature extraction model being used to extract features from any advertisement; inputting a newly invested advertisement into the feature extraction model to obtain advertisement features, multiplying the user feature matrix by the advertisement features of the newly invested advertisement to obtain a predicted score of each user for the newly invested advertisement, and pushing the newly invested advertisement to users whose predicted scores are greater than a score threshold; the loss function for: , is the user feature matrix, is the historical advertisement in the advertisement feature matrix The advertising features For each user in the scoring matrix, The real rating, is an exponential function with base e, is the number of historical advertisements, For feature extraction model on historical ads The feature extraction results.

[0008] The above-mentioned big data-based advertising push method provided in the embodiment of the present application collects the feedback behaviors of each user on multiple historical advertisements for which feedback behaviors of each user can be collected, collects the feedback behaviors of each user on each historical advertisement to construct a scoring matrix, and uses the alternating least squares method to decompose the scoring matrix into a user feature matrix and an advertising feature matrix, wherein the user feature matrix includes the user features of all users, and the advertising feature matrix includes the advertising features of all historical advertisements; the advertising features in the advertising feature matrix are used to calculate the loss function to train the feature extraction model, wherein the loss function is used to constrain the feature extraction results output by the feature extraction model to be close to the advertising features, in order to enable the output feature extraction results to be accurately calculated Each user's predicted score for the advertisement is calculated. During the process of training the feature extraction model, a sample attention is assigned to each historical advertisement, and the sample attention is positively correlated with the accuracy of the predicted score corresponding to the advertisement feature, so that the trained feature extraction model can extract features for any advertisement. Furthermore, after the newly-launched advertisement is input into the feature extraction model, the user feature matrix is ​​multiplied by the advertisement feature of the newly-launched advertisement to obtain the predicted score of each user for the newly-launched advertisement, and the newly-launched advertisement is pushed to users whose predicted score is greater than the score threshold. Without collecting feedback from each user on the newly-launched advertisement, the feature vector of each advertisement can be accurately obtained to achieve accurate advertisement push.

[0009] Preferably, the feedback behavior includes the number of clicks, browsing time, number of shares and / or number of purchases, and the rating matrix includes the real ratings of each user for each historical advertisement; each feedback behavior corresponds to a preset score, and the user About historical advertising The true rating is equal to the user About historical advertising The sum of the products of the values ​​of each feedback behavior and the preset score.

[0010] Assign a preset score to each feedback behavior and count the users About historical advertising The values ​​of each feedback behavior are then quantified to obtain the user About historical advertising The real rating.

[0011] Preferably, decomposing the rating matrix into a user feature matrix and an advertisement feature matrix by using the alternating least squares method comprises: initializing the user feature matrix and the advertisement feature matrix; constructing a decomposition loss, wherein the decomposition loss for: , For users About historical advertising The real rating, For users User characteristics, Advertisement for History The advertising features For users About historical advertising The prediction score of is the regularization factor, for ; Use alternating least squares method to update the user feature matrix and the advertising feature matrix until the decomposition loss is less than the preset loss.

[0012] The scoring matrix is ​​decomposed using the alternating least squares method to obtain the user characteristics of each user and the advertising characteristics of each historical advertisement. User characteristics and historical advertising Advertising features Available users About historical advertising Prediction score of .

[0013] Preferably, The confidence level is: , Advertisement for History To users Number of pushes, Advertisement for History Number of pushes to all users.

[0014] Taking into account the different number of times historical advertisements are pushed to users, the real scores in the rating matrix cannot accurately reflect the users’ preference for historical advertisements. Therefore, in the process of decomposing the rating matrix using the alternating least squares method, it is also necessary to judge the confidence of each real score based on the number of times historical advertisements are pushed to users. This confidence is used to reflect the degree of credibility that the real score accurately reflects the users’ preference for historical advertisements.

[0015] Preferably, The confidence level is is equal to 1.

[0016] Preferably, the structure of the feature extraction model is related to the advertisement type of the historical advertisement; in response to the historical advertisement being a picture advertisement, the feature extraction model is a convolutional neural network; in response to the historical advertisement being a text advertisement or a video advertisement, the feature extraction model is a recurrent neural network.

[0017] Preferably, the training method of the feature extraction model includes: inputting historical advertisements into the feature extraction model, obtaining feature extraction results, and calculating the loss function; updating the feature extraction model using the gradient descent method; iteratively updating the feature extraction model until the value of the loss function is less than the loss threshold, or the number of iterations is greater than a preset number, to obtain a trained feature extraction model.

[0018] The advertising features in the advertising feature matrix are used as labels to complete the training of the feature extraction model. The trained feature extraction model can directly output the advertising features of any advertisement.

[0019] Preferably, the user The predicted score for the newly-added ad is equal to the user User characteristics Ad features of new ads The product of .

[0020] The technical solution of this application has the following beneficial technical effects:

[0021] The above-mentioned big data-based advertising push method provided in the embodiment of the present application, for multiple historical advertisements that can collect feedback behaviors of each user, collects the feedback behaviors of each user on each historical advertisement to construct a scoring matrix, and uses the alternating least squares method to decompose the scoring matrix into a user feature matrix and an advertising feature matrix, the user feature matrix includes user features of all users, and the advertising feature matrix includes advertising features of all historical advertisements; the advertising features in the advertising feature matrix are used to calculate the loss function to train the feature extraction model, and the loss function can constrain the output result of the feature extraction model to be equal to the advertising features, so that the trained feature extraction model can extract features for any advertisement; after the newly invested advertisement is input into the feature extraction model, the user feature matrix is ​​multiplied by the advertising features of the newly invested advertisement to obtain the predicted score of each user for the newly invested advertisement, and the newly invested advertisement is pushed to users whose predicted score is greater than the scoring threshold, without collecting the feedback behaviors of each user on the newly invested advertisement, the feature vector of each advertisement is accurately obtained to achieve accurate advertising push. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] By reading the detailed description below with reference to the accompanying drawings, the above and other purposes, features and advantages of the exemplary embodiments of the present application will become easy to understand. In the accompanying drawings, several embodiments of the present application are shown in an exemplary and non-limiting manner, and the same or corresponding reference numerals represent the same or corresponding parts, wherein:

[0023] Figure 1 It is a flowchart of an advertising push method based on big data according to an embodiment of the present application. DETAILED DESCRIPTION

[0024] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0025] It should be understood that when the application uses the terms "first", "second", etc., they are only used to distinguish different objects, rather than to describe a specific order. The terms "include" and "comprise" used in the application indicate the presence of the described features, wholes, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or their collections.

[0026] This application provides an advertisement push method based on big data, which is used to push any advertisement to target users, realize accurate advertisement push, and save advertisement resources. In an application scenario, it can be applied to any online mall to push advertisements to all users of the online mall.

[0027] See also Figure 1 FIG. 1 is a flowchart of a method for pushing advertisements based on big data according to an embodiment of the present application. According to different requirements, the order of the steps in the flowchart can be changed, and some steps can be omitted.

[0028] S11, constructing a rating matrix according to each user's feedback behavior on each historical advertisement, and decomposing the rating matrix into a user feature matrix and an advertisement feature matrix using an alternating least squares method.

[0029] In one embodiment, the user's feedback behavior on each historical advertisement is collected, and the feedback behavior includes the number of clicks, browsing time, number of shares and / or number of purchases; one feedback behavior corresponds to a preset score, and the user About historical advertising The true rating is equal to the user About historical advertising The sum of the products of the values ​​of each feedback behavior and the preset scores. The rating matrix includes the real ratings of each user for each historical advertisement. The size of the rating matrix is OK List, is the number of users, is the number of historical advertisements, the OK The column value is the user About historical advertising The real rating.

[0030] For example, the preset score of the number of clicks is recorded as 1, the preset score of the browsing time is recorded as 2, the preset score of the number of shares is recorded as 3, and the preset score of the number of purchases is recorded as 4; if the user About historical advertising The feedback behaviors include 5 clicks, 3 minutes of browsing time, 1 share, and 3 purchases; then the user About historical advertising The actual score is 26.

[0031] In one embodiment, after obtaining the rating matrix, the rating matrix can be decomposed using the alternating least squares method. Taking into account the different times that historical advertisements are pushed to users, the real ratings in the rating matrix cannot accurately reflect the users' preference for the historical advertisements. Therefore, in the process of decomposing the rating matrix using the alternating least squares method, it is also necessary to judge the confidence of each real rating based on the number of times the historical advertisements are pushed to users. The confidence is used to reflect the degree of credibility that the real ratings accurately reflect the users' preference for the historical advertisements.

[0032] For example, user 1 and user 2 have historical ads The actual ratings are , but historical advertising The ads were pushed to user 1 10 times and to user 2 30 times. Since the number of pushes to user 2 is greater than that to user 1, user 2 has a higher impression of the historical ads. The real rating can more truly and accurately reflect the user 2's historical advertising degree of preference.

[0033] Specifically, using the alternating least squares method to decompose the rating matrix into a user feature matrix and an advertisement feature matrix includes: initializing the user feature matrix and the advertisement feature matrix; constructing a decomposition loss, wherein the decomposition loss for: , For users About historical advertising The real rating, For users User characteristics, Advertisement for History The advertising features For users About historical advertising The prediction score of is the regularization factor, for Confidence; Use alternating least squares method to update the user feature matrix and the advertising feature matrix until the decomposition loss is less than the preset loss; express The Euclidean distance of express The Euclidean distance of .

[0034] in, The confidence level is: , Advertisement for History To users Number of pushes, Advertisement for History Number of pushes to all users.

[0035] Among them, the size of the user feature matrix is OK Columns, including User features of users, each user feature is 1 row columns; the size of the ad feature matrix is OK Columns, including The advertising features of historical advertisements, each of which is Row 1 column, where Can be pre-set, in this application The value of is 20. Regularization factor The value of is 0.2; the default loss value is 0.01.

[0036] In other embodiments, Confidence is 1, that is, the confidence of all true ratings in the rating matrix is ​​1.

[0037] In this way, by decomposing the rating matrix, we can obtain the user characteristics of each user and the advertising characteristics of each historical advertisement. By calculating the product of user characteristics and advertising characteristics, we can accurately predict the user's true rating of historical advertisements.

[0038] S12, calculating a loss function based on the advertisement features in the advertisement feature matrix to train a feature extraction model, wherein the feature extraction model is used to extract features from any advertisement.

[0039] In one embodiment, the advertisement features in the advertisement feature matrix can accurately represent the information of the corresponding historical advertisements. The advertisement features are used as labels of the corresponding historical advertisements to train the feature extraction model. The trained feature extraction model can output the advertisement features of any advertisement.

[0040] Specifically, the structure of the feature extraction model is related to the advertisement type of the historical advertisement, which includes picture advertisements, text advertisements and video advertisements; in response to the historical advertisement being a picture advertisement, the feature extraction model is a convolutional neural network; in response to the historical advertisement being a text advertisement or a video advertisement, the feature extraction model is a recurrent neural network.

[0041] The convolutional neural network is the feature extraction part of ResNet, ShuffleNet or VGGNet, and the recurrent neural network is the feature extraction part of LSTM or Transformer.

[0042] In one embodiment, during the training of the feature extraction model, the loss function satisfy:

[0043] , is the user feature matrix, is the historical advertisement in the advertisement feature matrix The advertising features For each user in the scoring matrix, The real rating, is an exponential function with base e, is the number of historical advertisements, For feature extraction model on historical ads The feature extraction results of express The Euclidean distance of .

[0044] The loss function is used to constrain the feature extraction results output by the feature extraction model to be close to the advertising features. After inputting any advertisement into the feature extraction model, in order to enable the output feature extraction results to accurately calculate the predicted score of each user for the advertisement, during the training of the feature extraction model, a sample attention is assigned to each historical advertisement, and the sample attention is positively correlated with the accuracy of the predicted score corresponding to the advertising features. For example, for OK List, for Row 1 column, then for A vector with 1 row and 1 column, containing all the user's historical ads The prediction score of Likewise A vector with 1 row and 1 column, containing all the users' historical ads in the rating matrix The real rating, Ability to characterize historical advertising The prediction score accuracy corresponding to the advertising features of The smaller it is, the greater the accuracy of the prediction score is, and it should be the historical advertising Allocate a larger sample size.

[0045] In one embodiment, the training method of the feature extraction model includes: inputting historical advertisements into the feature extraction model, obtaining feature extraction results, and calculating a loss function; updating the feature extraction model using a gradient descent method; iteratively updating the feature extraction model until the value of the loss function is less than a loss threshold, or the number of iterations is greater than a preset number, thereby obtaining a trained feature extraction model.

[0046] Among them, the loss threshold is 0.01, and the preset number of times is 300.

[0047] In this way, the advertising features in the advertising feature matrix are used as labels to complete the training of the feature extraction model. The trained feature extraction model can obtain the advertising features of any advertisement, and the user feature matrix is ​​multiplied by the advertising features output by the feature extraction model to obtain the predicted score of each user for the advertisement.

[0048] S13, input the newly launched advertisement into the feature extraction model to obtain advertisement features, multiply the user feature matrix by the advertisement features of the newly launched advertisement to obtain the predicted score of each user for the newly launched advertisement, and push the newly launched advertisement to users whose predicted scores are greater than the score threshold.

[0049] In one embodiment, since it is impossible to collect feedback from each user on the newly launched advertisement, it is impossible to accurately obtain the advertisement features of the newly launched advertisement according to the alternating least squares method. The newly launched advertisement is input into the trained feature extraction model to obtain the advertisement features of the newly launched advertisement.

[0050] The user feature matrix is ​​multiplied by the advertising feature of the newly launched advertisement to obtain the predicted score of each user for the newly launched advertisement. The advertising feature of the newly launched advertisement is: Row 1 column, OK The user feature matrix of the columns and the advertising features of the newly launched advertisements Multiply them together and you get A vector with 1 row and 1 column, is the number of all users, and the vector includes the predicted score of each user for the newly invested advertisement.

[0051] The score threshold is set to 15, and the newly invested advertisement is pushed to all users whose predicted scores are greater than 15, so as to achieve accurate push of advertisements.

[0052] In this way, it is possible to obtain the predicted scores of each user for the newly launched advertisements without collecting the feedback behaviors of each user for the newly launched advertisements, thereby achieving accurate push of advertisements.

[0053] The above-mentioned big data-based advertising push method provided in the embodiment of the present application, for multiple historical advertisements that can collect feedback behaviors of each user, collects the feedback behaviors of each user on each historical advertisement to construct a scoring matrix, and uses the alternating least squares method to decompose the scoring matrix into a user feature matrix and an advertising feature matrix, the user feature matrix includes user features of all users, and the advertising feature matrix includes advertising features of all historical advertisements; the advertising features in the advertising feature matrix are used to calculate the loss function to train the feature extraction model, and the loss function can constrain the output result of the feature extraction model to be equal to the advertising features, so that the trained feature extraction model can extract features for any advertisement; after the newly invested advertisement is input into the feature extraction model, the user feature matrix is ​​multiplied by the advertising features of the newly invested advertisement to obtain the predicted score of each user for the newly invested advertisement, and the newly invested advertisement is pushed to users whose predicted score is greater than the scoring threshold, without collecting the feedback behaviors of each user on the newly invested advertisement, the feature vector of each advertisement is accurately obtained to achieve accurate advertising push.

[0054] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0055] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the patent application. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application.

Claims

1. An advertisement push method based on big data, characterized in that: The push method includes: Constructing a rating matrix according to each user's feedback behavior on each historical advertisement, and decomposing the rating matrix into a user feature matrix and an advertisement feature matrix using an alternating least squares method; Calculating a loss function based on the advertisement features in the advertisement feature matrix to train a feature extraction model, wherein the feature extraction model is used to extract features from any advertisement; Input the newly-invested advertisement into the feature extraction model to obtain advertisement features, multiply the user feature matrix by the advertisement features of the newly-invested advertisement to obtain the predicted score of each user for the newly-invested advertisement, and push the newly-invested advertisement to users whose predicted scores are greater than a score threshold; Loss Function for: , is the user feature matrix, is the historical advertisement in the advertisement feature matrix The advertising features For each user in the scoring matrix, The real rating, is an exponential function with base e, is the number of historical advertisements, For feature extraction model on historical ads The feature extraction results of The structure of the feature extraction model is related to the advertisement type of the historical advertisement; in response to the historical advertisement being a picture advertisement, the feature extraction model is a convolutional neural network; in response to the historical advertisement being a text advertisement or a video advertisement, the feature extraction model is a recurrent neural network.

2. The method for pushing advertisements based on big data as claimed in claim 1, characterized in that: The feedback behavior includes the number of clicks, browsing time, number of shares and / or number of purchases, and the rating matrix includes the real ratings of each user for each historical advertisement; Each feedback behavior corresponds to a preset score. About historical advertising The true rating is equal to the user About historical advertising The sum of the products of the values ​​of each feedback behavior and the preset score.

3. The method for pushing advertisements based on big data as claimed in claim 1, characterized in that: Decomposing the rating matrix into a user feature matrix and an advertisement feature matrix using the alternating least squares method includes: Initialize the user feature matrix and the advertising feature matrix; construct a decomposition loss, the decomposition loss for: , For users About historical advertising The real rating, For users User characteristics, Advertisement for History The advertising features For users About historical advertising The prediction score of is the regularization factor, for Confidence level; The user feature matrix and the advertising feature matrix are updated using the alternating least squares method until the decomposition loss is less than the preset loss.

4. The method for pushing advertisements based on big data as claimed in claim 3, characterized in that: The confidence level is: , Advertisement for History To users Number of pushes, Advertisement for History Number of pushes to all users.

5. The method for pushing advertisements based on big data as claimed in claim 3, characterized in that: The confidence level is Equal to 1.

6. The method for pushing advertisements based on big data according to claim 1, characterized in that: The training methods for feature extraction models include: Input historical advertisements into the feature extraction model, obtain feature extraction results, and calculate the loss function; Updating the feature extraction model using a gradient descent method; The feature extraction model is iteratively updated until the value of the loss function is less than the loss threshold, or the number of iterations is greater than a preset number, to obtain a trained feature extraction model.

7. The method for pushing advertisements based on big data as claimed in claim 6, characterized in that: user The predicted score for the newly-added ad is equal to the user User characteristics Ad features of new ads The product of .

Images