Advertisement picture putting method based on user behavior data

Abstract

The invention relates to the technical field of advertisement putting, and discloses a user behavior data-based advertisement picture putting method, which comprises the following steps of: acquiring behavior data of a target user, determining associated interests of the target user according to the behavior data of the target user, and determining an interest probability of the target user according to interest information of the target user; determining an associated interest of the target advertisement picture, and determining an interest probability of the target advertisement picture according to the associated interest of the target advertisement picture; according to the interest probability of the target user and the target advertisement picture, obtaining the matching degree of the target user and the target advertisement picture; constructing a game model based on the matching degree of the target user and the target advertisement picture, wherein the game model comprises participants, a strategy space and a revenue function; and determining a revenue balance point of the revenue function of the advertiser and the revenue function of the user according to a finite element accumulation optimization algorithm, and determining target advertisement picture putting according to the revenue balance point.

Description

Technical Field

[0001] The present invention relates to the field of advertisement delivery, and more specifically, to an advertisement picture delivery method based on user behavior data. Background Art

[0002] In the current Internet advertising delivery system, there is a situation where the advertiser is the commissioned individual of the advertiser. The advertiser delivers targeted advertisements, and when users click on the advertisements delivered by the advertiser, the advertiser pays the advertiser for the user's clicks.

[0003] However, existing advertising delivery methods for advertisers generally have the following problems:

[0004] 1. Most methods are based only on rough user behavior labels, resulting in a poor match between ads and user interests and low user click-through rates.

[0005] 2. The frequency of ad display is usually set manually by advertisers, which results in excessive ad display affecting user experience or too low display frequency affecting advertisers' revenue. Summary of the invention

[0006] The present invention provides an advertisement picture delivery method based on user behavior data, which solves the technical problems raised in the background technology.

[0007] A method for delivering advertisement images based on user behavior data, comprising:

[0008] Step 1, collecting the behavior data of the target user, determining the associated interests of the target user according to the behavior data of the target user, and determining the probability of the target user's interest in each according to the associated interests of the target user;

[0009] Step 2, determining the associated interests of the target advertising image, and determining the interest probability of the target advertising image according to the associated interests of the target advertising image;

[0010] Step 3, according to the interest probability of the target user and the target advertising picture, obtain the matching degree between the target user and the target advertising picture;

[0011] Step 4: construct a game model based on the matching degree between the target user and the target advertisement image. The game model includes: participants, strategy space and profit function;

[0012] Step 5: Determine the profit equilibrium point of the profit function of the advertiser and the profit function of the user based on the finite element accumulation optimization algorithm, and determine the target advertising image delivery according to the profit equilibrium point;

[0013] Step 6: If the target advertisement image is not delivered, style migration is performed on the target advertisement image according to the behavior data of the target user to obtain a migrated style advertisement, and the delivery of the migrated style advertisement to the target user is determined.

[0014] Further, determine related interests, including:

[0015] Determine the associated interests of target users as follows:

[0016] Preprocessing several groups of behavior data of target users in a preset time period to obtain several groups of standard behavior data;

[0017] Constructing behavior feature vectors from several sets of standard behavior data respectively, inputting the behavior feature vectors into an interest recognition tool, and the interest recognition tool outputs the associated interests of the target user;

[0018] Determine the associated interests of the target ad image as follows:

[0019] By identifying the image content of the target advertising image, the image text is mapped into a word vector through a word vector tool, and the image style is mapped into a style vector through an image recognition tool. The word vector and style vector are fused to construct an image feature vector, which is input into the interest matching tool. The interest matching tool outputs the associated interests of the target advertising image.

[0020] Furthermore, determining the interest probability distribution of the target user according to the interest information of the target user includes:

[0021] The target user's preference coefficient for the i-th interest is obtained through the target user's behavior data, which includes: , represents the total number of interactions of the target user on the i-th interest; , represents the total browsing time of the target user's related content of interest to the i-th item; , represents the total number of clicks of the target user on the relevant content of interest to the i-th item;

[0022] Divide the behavior data into K parts in chronological order, and obtain the number of interactions of the target user with the i-th interest, the total browsing time of related content, and the number of clicks on related content in each time period;

[0023] The target user's preference coefficient for the i-th area of ​​interest at the k-th time point The calculation formula is as follows: ;in, , and They represent the interaction weight coefficient, browsing weight coefficient and click weight coefficient respectively. =1, , and All are not 0, represents the number of interactions of the target user with the i-th interest in the k-th time period, It indicates the total browsing time of the target user for the i-th interest-related content in the k-th time period. The total number of clicks by the target user on the i-th interest-related content in the k-th time period, where K represents the total number of collection time periods. The index representing the total number of collection time periods. , and Both represent the linear attenuation coefficient, , and The first loss ratio, the second loss ratio, and the third loss ratio are the total number of interactions of the target user with the ith interest, the total browsing time of related content, or the total number of clicks on related content at the k-1th moment and the kth moment, respectively. represents the total number of interactions of the target user with the i-th interest in K time periods, represents the total browsing time of the target user for the i-th interest in K time periods, represents the total number of clicks of the target user on the i-th interest in K time periods;

[0024] The probability of the target user's interest in the i-th interest category is expressed as , the calculation formula is as follows: ;in represents the initial probability of the target user's interest in the i-th item, represents the target user's preference coefficient for the i-th region of interest, represents the Gaussian function used to describe the behavior density decay; ; ; ;

[0025] in and represents the first parameter of interest and the second parameter of interest, and The sum of is 1, and and All are not 0, , and They represent the total number of interactions of the target user with the jth interest, the total browsing time of related content, and the total number of clicks on related content;

[0026] According to the target user's initial probability for each interest, the target user's interest probability for each interest is obtained by normalizing the maximum and minimum values.

[0027] Further, determining the interest probability of the target advertising image according to the associated interest of the target advertising image includes:

[0028] The associated interest of the target advertising image is expressed as ;in represents the Kth interest of the target advertising image;

[0029] The initial probability of the target advertising image for the i-th interest category is expressed as , the calculation formula is as follows: ;in represents the initial probability of the target user's interest in the i-th item, represents the target user's preference coefficient for the i-th region of interest, represents the weight of the i-th interest on the target advertising image, represents the cosine similarity between the i-th interest and the j-th interest, Indicates the total number of interests associated with the target ad image. An index representing the total number of interests, represents the influence coefficient of the target advertising image on the i-th interest;

[0030] According to the initial probability of the target advertising image for each interest, the interest probability of the target advertising image for each interest is obtained by maximum and minimum value normalization.

[0031] Furthermore, the matching degree between the target user and the target advertising image includes:

[0032] Determine the overlapping interests of target users and target advertising images;

[0033] The matching degree between target users and target advertising images The formula is as follows: ;in Indicates the matching degree between the target user and the target advertising image. represents the probability of the target user's interest in the sth overlapping interest, represents the interest probability of the target advertising image for the jth overlapping interest, Q represents the number of overlapping interests between the target user and the target advertising image, i represents the index of the number of overlapping interests between the target user and the target advertising image, represents the first exponential coefficient, Represents the second exponential coefficient.

[0034] Furthermore, the participants, strategy space and profit function include:

[0035] Participants include: advertisers and target users;

[0036] The advertising provider's strategy space includes: delivering target advertising images and not delivering target advertising images;

[0037] The target user’s strategy space includes: clicking on the target ad image and not clicking on the target ad image;

[0038] The revenue function of the advertiser is as follows: ;in represents the revenue function of the advertiser, It refers to the revenue earned by the advertiser after a user clicks on the target ad image once. It indicates the cost for an advertiser to place a target advertising image once;

[0039] The target user’s revenue function is as follows: ;in represents the revenue function of the target user, Indicates the value obtained by users clicking on the target ad image. Represents the loss of users clicking on the target ad image.

[0040] Furthermore, the value obtained by a user clicking on a target ad image and the loss caused by a user clicking on a target ad image include:

[0041] The value obtained by users clicking on the target ad image represents the matching degree between the ad and the user's interests, as well as the actual benefit of the target ad image to the user. The formula is as follows: ;in and represents a first weight parameter and a second weight parameter;

[0042] The loss of users clicking on the target ad image represents the inverse measure of the matching degree between the ad content and the user's interests, and the formula is as follows: ;in and represents the third weight parameter and the fourth weight parameter, Indicates the number of times the target ad image is displayed to the target users per unit time.

[0043] Furthermore, the profit equilibrium point of the profit function of the advertiser and the profit function of the user is determined based on the optimization algorithm of finite element accumulation, including:

[0044] Determine the targeting strategy for advertisers and users;

[0045] Advertisers’ target strategies: ; Where max represents the max function;

[0046] User's target strategy: ; 1 means click, 0 means no click;

[0047] Randomly generate an initial population with R individuals that meet the constraints, and initialize the current iteration number Time to 1;

[0048] Encode the individuals: ;in Each element in , It represents the cost of delivering a target ad image once. represents the dimensions of individual coding;

[0049] The objective function is expressed as: , An index representing the dimension of individual coding;

[0050] Constraints: ; ; It represents the minimum cost for an advertiser to place a target ad image once. It indicates the maximum cost for an advertiser to place a target advertising image once;

[0051] By linearly accumulating the number of target user impressions per unit time of the target ad image, Update and iterate;

[0052] The number of times any target ad image is displayed to a target user per unit time is Update the preset number of iterations; Updates include:

[0053] Output the fitness function value of each individual;

[0054] Update individuals whose fitness function values ​​are lower than the preset fitness threshold as parents;

[0055] After determining that the current number of iterations = the preset number of iterations, the iteration is stopped; the individual with the highest fitness function value is output, and the fitness function value of the individual is used as the maximum revenue of the corresponding target advertising image under the number of display times to the target user per unit time;

[0056] The number of times the target advertisement image is displayed to the target user per unit time is linearly accumulated. When the fitness function value is less than the fitness function value of the previous iteration, the linear accumulation of the number of times the target advertisement image is displayed to the target user per unit time is stopped.

[0057] Determine the maximum number of times the target advertising image is displayed to the target user per unit time, and obtain the maximum revenue under the maximum number of times the target advertising image is displayed to the target user per unit time;

[0058] The maximum number of display times of the target advertisement image to the target user per unit time and the maximum revenue under the maximum number of display times of the target advertisement image to the target user per unit time are the revenue equilibrium points of the revenue function of the advertiser and the revenue function of the user.

[0059] Furthermore, the target advertising image delivery is determined based on the profit equilibrium point, including:

[0060] The number of times the target advertising image is displayed to the target user per unit time is evenly distributed according to the unit time, so as to obtain the interval delivery time of the target advertising image within the unit time.

[0061] Furthermore, the style of the target advertisement image is transferred according to the behavior data of the target user, and the transferred style advertisement is delivered to the target user, including:

[0062] Selecting multiple keywords corresponding to each interest category from a preset interest category library according to the interest probability of the target user for each interest;

[0063] Weighting the selected keywords to form a keyword combination that matches the user's interest distribution;

[0064] The target advertisement image text content is used as the constraint word;

[0065] The constraint words and keywords are combined as prompt words for the image generation model, and the target advertisement image is used as the style transfer object of the image generation model to generate several transfer style advertisements.

[0066] Calculate the loss of each transferred style ad and the target ad image. The loss function is as follows: ;in represents the visual features extracted from the transfer style ad or target ad image, represents the calculation of the Euclidean distance between the visual features in the transferred style ad and the target ad image. Represents the text content extracted from the transfer style advertisement or the target advertisement image. Indicates calculating the Euclidean distance between the text content in the transferred style ad and the target ad image;

[0067] By encoding the weighted coefficient of the keyword, the value range of the weighted coefficient is between zero and one; the weighted coefficient is encoded into a number of encoded individuals; according to the calculated gradient of the loss function, the encoded individuals are iteratively updated to approximately obtain the final weighted coefficient of each keyword;

[0068] A transfer-style advertisement is generated based on the final weighted coefficient of each keyword, and the transfer-style advertisement is used to replace the target advertisement image and delivered to the target users.

[0069] The beneficial effects of the present invention are: by combining user behavior data with interest matching and game models, and combining optimization algorithms to dynamically adjust the frequency of advertising display and delivery strategies, a balance is achieved between advertiser revenue and user experience; by finely analyzing user interests and determining the degree of advertising matching, the click-through rate of advertising is improved, and the automated process efficiently optimizes the effect of advertising delivery. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 It is a flow chart of a method for delivering advertisement pictures based on user behavior data of the present invention;

[0071] Figure 2 It is a module diagram of a training system for an AI discrimination model based on user behavior data of the present invention. DETAILED DESCRIPTION

[0072] The subject matter described herein will now be discussed with reference to example implementations. It should be understood that the discussion of these implementations is only to enable those skilled in the art to better understand and implement the subject matter described herein, and the functions and arrangements of the elements discussed may be changed without departing from the scope of protection of the contents of this specification. Various examples may omit, replace, or add various processes or components as needed. In addition, the features described in some examples may also be combined in other examples.

[0073] like Figure 1-2 As shown, a method for delivering advertisement images based on user behavior data includes:

[0074] Step 1, collecting the behavior data of the target user, determining the associated interests of the target user according to the behavior data of the target user, and determining the probability of the target user's interest in each according to the associated interests of the target user;

[0075] Step 2, determining the associated interests of the target advertising image, and determining the interest probability of the target advertising image according to the associated interests of the target advertising image;

[0076] Step 3, according to the interest probability of the target user and the target advertising picture, obtain the matching degree between the target user and the target advertising picture;

[0077] Step 4: construct a game model based on the matching degree between the target user and the target advertisement image. The game model includes: participants, strategy space and profit function;

[0078] Step 5: Determine the profit equilibrium point of the profit function of the advertiser and the profit function of the user based on the finite element accumulation optimization algorithm, and determine the target advertising image delivery according to the profit equilibrium point;

[0079] Step 6: If the target advertisement image is not delivered, style migration is performed on the target advertisement image according to the behavior data of the target user to obtain a migrated style advertisement, and the delivery of the migrated style advertisement to the target user is determined.

[0080] In one embodiment of the present invention, determining the associated interests includes:

[0081] Determine the associated interests of target users as follows:

[0082] Preprocessing several groups of behavior data of target users in a preset time period to obtain several groups of standard behavior data;

[0083] Constructing behavior feature vectors from several sets of standard behavior data respectively, inputting the behavior feature vectors into an interest recognition tool, and the interest recognition tool outputs the associated interests of the target user;

[0084] Determine the associated interests of the target ad image as follows:

[0085] By identifying the image content of the target advertising image, the image text is mapped into a word vector through a word vector tool, and the image style is mapped into a style vector through an image recognition tool. The word vector and style vector are fused to construct an image feature vector, which is input into the interest matching tool. The interest matching tool outputs the associated interests of the target advertising image.

[0086] It should be noted that the collected behavioral data are grouped according to the time interval of the behavioral data. If the interval between group a and group b behavioral data is greater than a preset time threshold, then group a and group b behavioral data are considered independent of each other; otherwise, group a and group b behavioral data are in the same group.

[0087] It should be noted that TesseractOCR is used to extract the text in the target advertising image; Word2Vec is used to map the text content of the target advertising image into word vectors; VGGNet is used to extract the image style vector; NumPy is used to concatenate the word vector and the style vector; and Scikit-learn is used to build an interest matching tool.

[0088] In one embodiment of the present invention, Scikit-learn is used to construct an interest matching tool, including:

[0089] Data preparation: Use the word vector tool Word2Vec to map the text content extracted from the target advertisement image into a word vector, and use the image feature extraction tool VGGNet to extract the image style vector. The word vector and style vector are merged to form a complete image feature vector.

[0090] Interest tag annotation: manually annotate corresponding interest tags for each target advertising image based on user behavior data.

[0091] Construct a training set and use the labeled image feature vectors and corresponding interest tags to construct training samples and sample labels; the image feature vectors are used as input features, and the interest tags are used as the expected output of Scikit-learn.

[0092] Select and train the classification model, use the classification algorithm support vector machine (SVM) provided by Scikit-learn, select a suitable classification model for training; adjust the model parameters through cross-validation and other methods to obtain the best matching effect.

[0093] The implementation of the interest matching tool uses the trained classification model as the interest matching tool. When a new target advertising image is input, the image feature vector is generated through the same feature extraction and fusion process, and is input into the interest matching tool to output the interest category related to the advertising image.

[0094] In one embodiment of the present invention, determining the interest probability distribution of the target user according to the interest information of the target user includes:

[0095] The target user's preference coefficient for the i-th interest is obtained through the target user's behavior data, which includes: , represents the total number of interactions of the target user on the i-th interest; , represents the total browsing time of the target user's related content of interest to the i-th item; , represents the total number of clicks of the target user on the relevant content of interest to the i-th item;

[0096] Divide the behavior data into K parts in chronological order, and obtain the number of interactions of the target user with the i-th interest, the total browsing time of related content, and the number of clicks on related content in each time period;

[0097] The target user's preference coefficient for the i-th area of ​​interest at the k-th time point The calculation formula is as follows: ;in, , and They represent the interaction weight coefficient, browsing weight coefficient and click weight coefficient respectively. =1, , and All are not 0, represents the number of interactions of the target user with the i-th interest in the k-th time period, It indicates the total browsing time of the target user for the i-th interest-related content in the k-th time period. The total number of clicks by the target user on the i-th interest-related content in the k-th time period, where K represents the total number of collection time periods. The index representing the total number of collection time periods. , and Both represent the linear attenuation coefficient, , and The first loss ratio, the second loss ratio, and the third loss ratio are the total number of interactions of the target user with the ith interest, the total browsing time of related content, or the total number of clicks on related content at the k-1th moment and the kth moment, respectively. represents the total number of interactions of the target user with the i-th interest in K time periods, represents the total browsing time of the target user for the i-th interest in K time periods, represents the total number of clicks of the target user on the i-th interest in K time periods;

[0098] The probability of the target user's interest in the i-th interest category is expressed as , the calculation formula is as follows: ;in represents the initial probability of the target user's interest in the i-th item, represents the target user's preference coefficient for the i-th region of interest, represents the Gaussian function used to describe the behavior density decay; ; ; ;

[0099] in and represents the first parameter of interest and the second parameter of interest, and The sum of is 1, and and All are not 0, , and They represent the total number of interactions of the target user with the jth interest, the total browsing time of related content, and the total number of clicks on related content;

[0100] According to the target user's initial probability for each interest, the target user's interest probability for each interest is obtained by normalizing the maximum and minimum values.

[0101] When calculating the target user's preference coefficient for a specific interest, we introduced multi-dimensional behavioral data such as the number of user interactions, browsing time, and number of clicks, and combined it with the time decay factor. , and , more accurately reflecting the changes in the intensity of users' interests in different time periods. Through the careful processing of behavioral data, the accuracy of matching prediction can be improved, ensuring that the recommended ads are more targeted.

[0102] Combines the number of user interactions in the kth time period , browsing time and clicks This multi-dimensional fusion can more comprehensively reflect the real interests of users, avoid the deviation caused by relying solely on single behavioral data, and improve the diversity and accuracy of advertising recommendations.

[0103] In one embodiment of the present invention, determining the interest probability of the target advertising picture according to the associated interest of the target advertising picture includes:

[0104] The associated interest of the target advertising image is expressed as ;in represents the Kth interest of the target advertising image;

[0105] The initial probability of the target advertising image for the i-th interest category is expressed as , the calculation formula is as follows: ;in represents the initial probability of the target user's interest in the i-th item, represents the target user's preference coefficient for the i-th region of interest, represents the weight of the i-th interest on the target advertising image, represents the cosine similarity between the i-th interest and the j-th interest, Indicates the total number of interests associated with the target ad image. An index representing the total number of interests, represents the influence coefficient of the target advertising image on the i-th interest;

[0106] According to the initial probability of the target advertising image for each interest, the interest probability of the target advertising image for each interest is obtained by maximum and minimum value normalization.

[0107] By integrating advertisements with multi-dimensional information of user interests, accurate matching of advertisements with users in multiple points of interest is achieved. Compared with traditional single-dimensional matching, this formula can more comprehensively capture users' potential needs and significantly improve the relevance and personalization of advertising recommendations.

[0108] The cosine similarity between multiple interest points is introduced to measure the correlation between different interest points in the advertisement. This means that if there is a high similarity between certain interest points in the advertisement, the matching degree will be further enhanced due to this similarity. This processing method can make full use of the correlation of interest points within the advertisement and improve users' acceptance of related advertisements.

[0109] In one embodiment of the present invention, the matching degree between the target user and the target advertisement image includes:

[0110] Determine the overlapping interests of target users and target advertising images;

[0111] The matching degree between target users and target advertising images The formula is as follows: ;in Indicates the matching degree between the target user and the target advertising image. represents the probability of the target user's interest in the sth overlapping interest, represents the interest probability of the target advertising image for the jth overlapping interest, Q represents the number of overlapping interests between the target user and the target advertising image, i represents the index of the number of overlapping interests between the target user and the target advertising image, represents the first exponential coefficient, Represents the second exponential coefficient.

[0112] By introducing and These two exponential coefficients respectively affect the probability of user interest and Ad Interest Probability This exponential amplification process is performed. This exponential amplification makes the matching degree more sensitive to high interest probabilities, that is, when the user's preference probability for a certain interest is high, the matching degree will be further amplified, which can effectively highlight the advertising content that is highly matched with the user's interests.

[0113] By using the user's interest area preference coefficient As the denominator, and taking the square of the reverse weight processing, the formula can make more delicate adjustments to the distribution between the user's points of interest. When the user has a low preference for a certain area, the preference coefficient will have a greater impact on the matching degree, weakening the contribution of the matching degree. For the user's high preference area, the matching degree will be correspondingly amplified. This reverse weight processing improves the accuracy of advertising display and ensures that the displayed ads are more in line with the user's area of ​​interest.

[0114] The matching degree calculation method based on probability distribution and overlapping interests can accurately quantify the interest correlation between users and ads, effectively filter out irrelevant interests, improve matching accuracy, and facilitate the processing of complex and diverse user data. Compared with simple tag matching or keyword matching, this method can better reflect the intensity and specific differences of interests.

[0115] In one embodiment of the present invention, the participants, strategy space and profit function include:

[0116] Participants include: advertisers and target users;

[0117] The advertising provider's strategy space includes: delivering target advertising images and not delivering target advertising images;

[0118] The target user’s strategy space includes: clicking on the target ad image and not clicking on the target ad image;

[0119] The revenue function of the advertiser is as follows: ;in represents the revenue function of the advertiser, It refers to the revenue earned by the advertiser after a user clicks on the target ad image once. It indicates the cost for an advertiser to place a target advertising image once;

[0120] The target user’s revenue function is as follows: ;in represents the revenue function of the target user, Indicates the value obtained by users clicking on the target ad image. Represents the loss of users clicking on the target ad image.

[0121] The strategies and benefits of advertisers and target users are clearly quantified. By matching ads with users, advertisers can maximize their advertising revenue, while target users decide whether to click on ads after weighing the value of the click and potential losses.

[0122] In one embodiment of the present invention, the value obtained by a user when clicking on a target advertisement image and the loss caused by a user when clicking on a target advertisement image include:

[0123] The value obtained by users clicking on the target ad image represents the matching degree between the ad and the user's interests, as well as the actual benefit of the target ad image to the user. The formula is as follows: ;in and represents a first weight parameter and a second weight parameter;

[0124] The loss of users clicking on the target ad image represents the inverse measure of the matching degree between the ad content and the user's interests, and the formula is as follows: ;in and represents the third weight parameter and the fourth weight parameter, Indicates the number of times the target ad image is displayed to the target users per unit time.

[0125] It should be noted that the value that a user obtains by clicking on an advertisement and obtaining a target advertisement image is generally obtained from the cost of the advertiser placing a target advertisement image once. At the same time, the advertiser can control the cost of placing a target advertisement image once within a certain range, thereby controlling the benefits obtained by the user by clicking on the advertisement.

[0126] The value and loss of a user clicking on an ad are quantified through matching and weight parameters to help optimize ad delivery strategies. The value of a user's click is determined by the matching degree between the ad and the user's interests and the actual benefits brought by the ad, while the loss is based on the inverse measurement of the matching degree and the frequency of ad display.

[0127] In one embodiment of the present invention, the profit equilibrium point of the profit function of the advertiser and the profit function of the user is determined based on the optimization algorithm of finite element accumulation, including:

[0128] Determine the targeting strategy for advertisers and users;

[0129] Advertisers’ target strategies: ; Where max represents the max function;

[0130] User's target strategy: ; 1 means click, 0 means no click;

[0131] Randomly generate an initial population with R individuals that meet the constraints, and initialize the current iteration number Time to 1;

[0132] Encode the individuals: ;in Each element in , It represents the cost of delivering a target ad image once. represents the dimensions of individual coding;

[0133] The objective function is expressed as: , An index representing the dimension of individual coding;

[0134] Constraints: ; ; It represents the minimum cost for an advertiser to place a target ad image once. It indicates the maximum cost for an advertiser to place a target advertising image once;

[0135] By linearly accumulating the number of target user impressions per unit time of the target ad image, Update and iterate;

[0136] The number of times any target ad image is displayed to a target user per unit time is Update the preset number of iterations; Updates include:

[0137] Output the fitness function value of each individual;

[0138] Update individuals whose fitness function values ​​are lower than the preset fitness threshold as parents;

[0139] After determining that the current number of iterations = the preset number of iterations, the iteration is stopped; the individual with the highest fitness function value is output, and the fitness function value of the individual is used as the maximum revenue of the corresponding target advertising image under the number of display times to the target user per unit time;

[0140] The number of times the target advertisement image is displayed to the target user per unit time is linearly accumulated. When the fitness function value is less than the fitness function value of the previous iteration, the linear accumulation of the number of times the target advertisement image is displayed to the target user per unit time is stopped.

[0141] Determine the maximum number of times the target advertising image is displayed to the target user per unit time, and obtain the maximum revenue under the maximum number of times the target advertising image is displayed to the target user per unit time;

[0142] The maximum number of display times of the target advertisement image to the target user per unit time and the maximum revenue under the maximum number of display times of the target advertisement image to the target user per unit time are the revenue equilibrium points of the revenue function of the advertiser and the revenue function of the user.

[0143] A method for determining the equilibrium point of the profit function of advertisers and users through an optimization algorithm. First, a population of eligible individuals is randomly generated and iterated. In each iteration, the strategy that maximizes the profit is selected by updating the number of ad impressions. When the iteration reaches the set number of times or the fitness function no longer improves, the optimal range of ad impressions is determined, and finally the maximum profit and ad impression number of the advertiser are obtained, thereby achieving a balance between the profits of both parties.

[0144] It should be noted that the number of times the target ad image is displayed to the target user per unit time is a finite element, that is, when the number of times the target ad image is displayed to the target user per unit time exceeds a certain number, it cannot meet the requirement. ; Therefore, by gradually stacking the target advertising images, the number of times they are displayed to the target users per unit time;

[0145] When the target advertisement image is displayed to the target user once per unit time, the maximum profit of the advertiser is obtained after the genetic algorithm is iterated for a preset number of times;

[0146] When the target advertisement image is displayed to the target user n times per unit time, the maximum profit of the advertiser is obtained after the genetic algorithm is iterated for a preset number of times;

[0147] When the maximum revenue of the advertiser when the target advertisement image is displayed to the target user n+1 times per unit time is less than the maximum revenue of the advertiser when the target advertisement image is displayed to the target user n times per unit time, then the number of displays of the target advertisement image to the target user per unit time is stopped from being gradually added; the number of displays of the target advertisement image to the target user per unit time is determined to be 1 to n times;

[0148] Determine the maximum revenue for the nth time, and the number of times the target advertising image is displayed to the target user per unit time is n times.

[0149] It should be noted that when the target advertising image is displayed to the target user once per unit time, after the genetic algorithm iterates a preset number of times, the maximum profit of the advertiser is a plural number, which means that the target user does not match the target advertising image, and the advertising to the target user is abandoned.

[0150] In one embodiment of the present invention, style migration is performed on a target advertisement image according to the behavior data of the target user, and the migrated style advertisement is delivered to the target user, including:

[0151] Selecting multiple keywords corresponding to each interest category from a preset interest category library according to the interest probability of the target user for each interest;

[0152] Weighting the selected keywords to form a keyword combination that matches the user's interest distribution;

[0153] The target advertisement image text content is used as the constraint word;

[0154] The constraint words and keywords are combined as prompt words for the image generation model, and the target advertisement image is used as the style transfer object of the image generation model to generate several transfer style advertisements.

[0155] Calculate the loss of each transferred style ad and the target ad image. The loss function is as follows: ;in represents the visual features extracted from the transfer style ad or target ad image, represents the calculation of the Euclidean distance between the visual features in the transferred style ad and the target ad image. Represents the text content extracted from the transfer style advertisement or the target advertisement image. Indicates calculating the Euclidean distance between the text content in the transferred style ad and the target ad image;

[0156] By encoding the weighted coefficient of the keyword, the value range of the weighted coefficient is between zero and one; the weighted coefficient is encoded into a number of encoded individuals; according to the calculated gradient of the loss function, the encoded individuals are iteratively updated to approximately obtain the final weighted coefficient of each keyword;

[0157] A transfer-style advertisement is generated based on the final weighted coefficient of each keyword, and the transfer-style advertisement is used to replace the target advertisement image and delivered to the target users.

[0158] It should be noted that visual features include color histogram, texture feature analysis and edge information; color histogram can be obtained by calculating the color distribution of each color channel in the image; texture features can be obtained through gray-level co-occurrence matrix texture analysis method; edge information can be extracted using Canny edge detection image processing method.

[0159] The text content includes brand names and advertising slogans, which are extracted through OCR technology.

[0160] The text content of the target ad image is used as a constraint word and combined with the weighted keywords selected based on user interests to form the prompt words of the generative model. This design enables the generated style transfer ad to keep the text information of the ad intact, enhancing the accuracy of the generated ad in content communication.

[0161] The visual features and text content of the target ad image are extracted and Euclidean distance is calculated to construct a comprehensive loss function, which can accurately measure the visual and textual differences between the generated transfer style ad and the target ad. The design of the loss function effectively balances style consistency and text communication, so that the generated ad image can better meet user interests and advertising needs.

[0162] The weighted coefficients of keywords are encoded and iteratively updated, and the weighted coefficients of each keyword are gradually adjusted using the gradient information of the loss function, thereby optimizing the generated prompt words. This gradient-based optimization method enables the generated ads to match user interests more accurately and improves the personalized effect of style transfer.

[0163] Based on the optimization of weighted keywords, the prompt words of the generation model can be dynamically adjusted to make the style of the generated advertisement more in line with user preferences while ensuring the accurate delivery of the advertisement text information.

[0164] The above describes an embodiment of the present embodiment, but the present embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present embodiment, ordinary technicians in this field can also make many forms, all of which are within the protection of the present embodiment.

Claims

1. A method for delivering advertising images based on user behavior data, characterized in that: include: Step 1, collecting the behavior data of the target user, determining the associated interests of the target user according to the behavior data of the target user, and determining the probability of the target user's interest in each according to the associated interests of the target user; Step 2, determining the associated interests of the target advertising image, and determining the interest probability of the target advertising image according to the associated interests of the target advertising image; Step 3, according to the interest probability of the target user and the target advertising picture, obtain the matching degree between the target user and the target advertising picture; Step 4: construct a game model based on the matching degree between the target user and the target advertisement image. The game model includes: participants, strategy space and profit function; Step 5: Determine the profit equilibrium point of the profit function of the advertiser and the profit function of the user based on the finite element accumulation optimization algorithm, and determine the target advertising image delivery according to the profit equilibrium point; Step 6: If the target advertisement image is not delivered, style migration is performed on the target advertisement image according to the behavior data of the target user to obtain a migrated style advertisement, and the delivery of the migrated style advertisement to the target user is determined.

2. The method for delivering advertisement images based on user behavior data according to claim 1, characterized in that: Identify relevant interests, including: Determine the associated interests of target users as follows: Preprocessing several groups of behavior data of target users in a preset time period to obtain several groups of standard behavior data; Constructing behavior feature vectors from several sets of standard behavior data respectively, inputting the behavior feature vectors into an interest recognition tool, and the interest recognition tool outputs the associated interests of the target user; Determine the associated interests of the target ad image as follows: By identifying the image content of the target advertising image, the image text is mapped into a word vector through a word vector tool, and the image style is mapped into a style vector through an image recognition tool. The word vector and style vector are fused to construct an image feature vector, which is input into the interest matching tool. The interest matching tool outputs the associated interests of the target advertising image.

3. The method for delivering advertisement pictures based on user behavior data according to claim 1, characterized in that: Determine the interest probability distribution of the target user according to the interest information of the target user, including: The target user's preference coefficient for the i-th interest is obtained through the target user's behavior data, which includes: , represents the total number of interactions of the target user on the i-th interest; , represents the total browsing time of the target user's related content of interest to the i-th item; , represents the total number of clicks of the target user on the relevant content of interest to the i-th item; Divide the behavior data into K parts in chronological order, and obtain the number of interactions of the target user with the i-th interest, the total browsing time of related content, and the number of clicks on related content in each time period; The target user's preference coefficient for the i-th area of ​​interest at the k-th time point The calculation formula is as follows: ;in, , and They represent the interaction weight coefficient, browsing weight coefficient and click weight coefficient respectively. =1, , and All are not 0, represents the number of interactions of the target user with the i-th interest in the k-th time period, It indicates the total browsing time of the target user for the i-th interest-related content in the k-th time period. The total number of clicks by the target user on the i-th interest-related content in the k-th time period, where K represents the total number of collection time periods. The index representing the total number of collection time periods. , and Both represent the linear attenuation coefficient, , and The first loss ratio, the second loss ratio, and the third loss ratio are the total number of interactions of the target user with the ith interest, the total browsing time of related content, or the total number of clicks on related content at the k-1th moment and the kth moment, respectively. represents the total number of interactions of the target user with the i-th interest in K time periods, represents the total browsing time of the target user for the i-th interest in K time periods, represents the total number of clicks of the target user on the i-th interest in K time periods; The probability of the target user's interest in the i-th interest category is expressed as , the calculation formula is as follows: ;in represents the initial probability of the target user's interest in the i-th item, represents the target user's preference coefficient for the i-th region of interest, represents the Gaussian function used to describe the behavior density decay; ； ； ； in and represents the first parameter of interest and the second parameter of interest, and The sum of is 1, and and All are not 0, , and They represent the total number of interactions of the target user with the jth interest, the total browsing time of related content, and the total number of clicks on related content; According to the target user's initial probability for each interest, the target user's interest probability for each interest is obtained by normalizing the maximum and minimum values.

4. The method for delivering advertisement pictures based on user behavior data according to claim 1, characterized in that: Determining the interest probability of the target advertising image according to the associated interests of the target advertising image includes: The associated interest of the target advertising image is expressed as ;in represents the Kth interest of the target advertising image; The initial probability of the target advertising image for the i-th interest category is expressed as , the calculation formula is as follows: ;in represents the initial probability of the target user's interest in the i-th item, represents the target user's preference coefficient for the i-th region of interest, represents the weight of the i-th interest on the target advertising image, represents the cosine similarity between the i-th interest and the j-th interest, Indicates the total number of interests associated with the target ad image. An index representing the total number of interests, represents the influence coefficient of the target advertising image on the i-th interest; According to the initial probability of the target advertising image for each interest, the interest probability of the target advertising image for each interest is obtained by maximum and minimum value normalization.

5. The method for delivering advertisement pictures based on user behavior data according to claim 1, characterized in that: The matching degree between the target user and the target advertising image, including: Determine the overlapping interests of target users and target advertising images; The matching degree between target users and target advertising images The formula is as follows: ;in Indicates the matching degree between the target user and the target advertising image. represents the probability of the target user's interest in the sth overlapping interest, represents the interest probability of the target advertising image for the jth overlapping interest, Q represents the number of overlapping interests between the target user and the target advertising image, i represents the index of the number of overlapping interests between the target user and the target advertising image, represents the first exponential coefficient, Represents the second exponential coefficient.

6. The method for delivering advertisement images based on user behavior data according to claim 1, characterized in that: Participants, strategy spaces, and payoff functions, including: Participants include: advertisers and target users; The advertising provider's strategy space includes: delivering target advertising images and not delivering target advertising images; The target user’s strategy space includes: clicking on the target ad image and not clicking on the target ad image; The revenue function of the advertiser is as follows: ;in represents the revenue function of the advertiser, It refers to the revenue earned by the advertiser after a user clicks on the target ad image once. It indicates the cost for an advertiser to place a target advertising image once; The target user’s revenue function is as follows: ;in represents the revenue function of the target user, Indicates the value obtained by users clicking on the target ad image. Represents the loss of users clicking on the target ad image.

7. The method for delivering advertisement pictures based on user behavior data according to claim 6, characterized in that: The value gained by users clicking on the target ad image and the loss of users clicking on the target ad image include: The value obtained by users clicking on the target ad image represents the matching degree between the ad and the user's interests, as well as the actual benefit of the target ad image to the user. The formula is as follows: ;in and represents a first weight parameter and a second weight parameter; The loss of users clicking on the target ad image represents the inverse measure of the matching degree between the ad content and the user's interests, and the formula is as follows: ;in and represents the third weight parameter and the fourth weight parameter, Indicates the number of times the target ad image is displayed to the target users per unit time.

8. The method for delivering advertisement images based on user behavior data according to claim 1, characterized in that: The optimization algorithm based on finite element accumulation determines the profit equilibrium point of the profit function of the advertiser and the profit function of the user, including: Determine the targeting strategy for advertisers and users; Advertisers’ target strategies: ; Where max represents the max function; User's target strategy: ; 1 means click, 0 means no click; Randomly generate an initial population with R individuals that meet the constraints, and initialize the current iteration number Time to 1; Encode the individuals: ;in Each element in , It represents the cost of delivering a target ad image once. represents the dimensions of individual coding; The objective function is expressed as: , An index representing the dimension of individual coding; Constraints: ; ; It indicates the minimum cost for an advertiser to place a target advertising image once. It indicates the maximum cost for an advertiser to place a target advertising image once; By linearly accumulating the number of target user impressions per unit time of the target ad image, Update and iterate; The number of times any target ad image is displayed to a target user per unit time is Update the preset number of iterations; Updates include: Output the fitness function value of each individual; Individuals whose fitness function values ​​are lower than the preset fitness threshold are updated as parents; After determining that the current number of iterations = the preset number of iterations, the iteration is stopped; the individual with the highest fitness function value is output, and the fitness function value of the individual is used as the maximum revenue of the corresponding target advertising image under the number of display times to the target user per unit time; The number of times the target advertisement image is displayed to the target user per unit time is linearly accumulated. When the fitness function value is less than the fitness function value of the previous iteration, the linear accumulation of the number of times the target advertisement image is displayed to the target user per unit time is stopped. Determine the maximum number of times the target advertising image is displayed to the target user per unit time, and obtain the maximum revenue under the maximum number of times the target advertising image is displayed to the target user per unit time; The maximum number of display times of the target advertisement image to the target user per unit time and the maximum revenue under the maximum number of display times of the target advertisement image to the target user per unit time are the revenue equilibrium points of the revenue function of the advertiser and the revenue function of the user.

9. The method for delivering advertisement images based on user behavior data according to claim 8, characterized in that: Determine the target advertising image delivery based on the profit balance point, including: The number of times the target advertising image is displayed to the target user per unit time is evenly distributed according to the unit time, so as to obtain the interval delivery time of the target advertising image per unit time.

10. The method for delivering advertisement images based on user behavior data according to claim 1, characterized in that: Perform style migration on target advertising images based on target user behavior data, and deliver migration-style advertising to target users, including: Selecting multiple keywords corresponding to each interest category from a preset interest category library according to the interest probability of the target user for each interest; Weighting the selected keywords to form a keyword combination that matches the user's interest distribution; The target advertisement image text content is used as the constraint word; The constraint words and keywords are combined as prompt words for the image generation model, and the target advertisement image is used as the style transfer object of the image generation model to generate several transfer style advertisements. Calculate the loss of each transferred style ad and the target ad image. The loss function is as follows: ;in represents the visual features extracted from the transfer style ad or target ad image, represents the calculation of the Euclidean distance between the visual features in the transferred style ad and the target ad image. Represents the text content extracted from the transfer style advertisement or the target advertisement image. Indicates calculating the Euclidean distance between the text content in the transferred style ad and the target ad image; By encoding the weighted coefficient of the keyword, the value range of the weighted coefficient is between zero and one; the weighted coefficient is encoded into a number of encoded individuals; according to the calculated gradient of the loss function, the encoded individuals are iteratively updated to approximately obtain the final weighted coefficient of each keyword; A transfer-style advertisement is generated based on the final weighted coefficient of each keyword, and the transfer-style advertisement is used to replace the target advertisement image and delivered to the target users.

Images