A method for intelligently optimizing advertisement delivery

Abstract

The present invention proposes an intelligent optimization advertising delivery method, which belongs to the technical field of advertising channel delivery strategy. It enhances the interaction between users and advertisements through multi-dimensional user feature extraction, dual-tower model matching, and multi-head self-attention mechanism, accurately calculates the advertising matching score, and combines the multi-touch attribution model to evaluate the contribution of advertisements to GMV; predicts the advertising conversion rate through the logistic regression binary classification model, and evaluates the user's payment ability in combination with the user lifetime value model, and constructs an ROI prediction model to optimize advertising budget allocation; monitors the advertising effect and ROI in real time, and dynamically adjusts the delivery strategy to ensure the efficiency and sustainability of advertising delivery. The present invention realizes accurate user matching, efficient cost control and global optimization, improves the advertiser's return on investment and the long-term value of users, ensures the flexibility and adaptability of advertising delivery, and provides advertisers with a transparent and explainable delivery solution.

Description

Technical Field

[0001] The invention relates to an intelligent optimization advertisement delivery method, belonging to the technical field of advertisement delivery channel strategies. Background Art

[0002] With the rapid development of the Internet advertising industry, advertisers have increasing requirements for the accuracy and effectiveness of advertising. Traditional advertising methods often rely on simple user portraits and fixed delivery strategies, which leads to serious waste of advertising resources and difficulty in quantifying and optimizing advertising effects. The core challenge of advertising is how to accurately predict user needs and behaviors, so as to achieve personalized recommendations and precise delivery of advertising content. Existing technologies usually use machine learning models based on user historical behavior data to predict user attributes and then select target user groups for advertising.

[0003] For example, the Chinese invention patent "A method and device for detecting advertising effects" with announcement number CN107871244B discloses a method for detecting advertising effects, including: obtaining the advertisements to be delivered provided by the advertiser; inputting the advertisements to be delivered into the user attribute prediction model for prediction, and outputting the target user attributes corresponding to the advertisements to be delivered, wherein the user attribute prediction model is established by machine learning training using the user behavior data set in the network data source; selecting a user set matching the target user attributes from the network platform according to the target user attributes, and delivering the advertisements to be delivered to the selected user set; collecting user feedback information generated by users in the user set on the delivered advertisements, and determining the user group matching the delivered advertisements according to the user feedback information. However, the above patent has many problems in terms of the complexity of the model and the dynamics of the delivery strategy.

[0004] First, traditional machine learning models are prone to overfitting problems when dealing with high-dimensional and complex user features, resulting in insufficient generalization capabilities of the model. The user attribute prediction model used in the above-mentioned patent is relatively simple, and only trains user behavior data through machine learning methods, without considering the complex interactive relationship of high-dimensional features; in addition, there is a lack of real-time monitoring and dynamic adjustment mechanism for advertising effects during the advertising delivery process, and it is impossible to optimize the delivery strategy in a timely manner according to the actual effect of the advertisement, which will lead to a series of problems such as waste of resources, unsatisfactory advertising effects, missed market opportunities, reduced user experience, reduced ROI, delayed decision-making, and difficulty in responding to emergencies. These problems will not only affect advertisers' short-term profits, but may also have a negative impact on the long-term development and market competitiveness of the brand.

[0005] Therefore, there is an urgent need for an intelligent optimization advertising delivery method that can combine high-dimensional feature processing, complex interaction modeling, real-time monitoring and dynamic optimization mechanism to achieve global optimization of advertising delivery, improve advertisers' return on investment, and ensure the sustainability and accuracy of advertising effects. Summary of the invention

[0006] In order to solve the above problems existing in the prior art, the present invention proposes an intelligent optimization advertisement delivery method.

[0007] The technical solution of the present invention is as follows:

[0008] The present invention provides an intelligent optimization advertising delivery method, the method comprising:

[0009] Collect user data, including user attributes, behavior data, interest tags and real-time data, perform feature extraction on the user data, and obtain multi-dimensional user features;

[0010] Constructing a matching model, taking the user features and the advertisement features corresponding to the advertisements to be placed as input, outputting the matching scores of the advertisements to be placed, sorting the matching scores, and placing the advertisements to be placed in order from high to low;

[0011] Obtain the user's historical conversion data, and based on the historical conversion data, use a multi-touch attribution model to calculate the contribution of the delivered advertisements to the user's conversion, and calculate the contribution of the delivered advertisements to the GMV based on the contribution of the delivered advertisements to the user's conversion;

[0012] Constructing a conversion rate prediction model, training the conversion rate prediction model using user characteristics, advertising characteristics of delivered advertisements, and historical conversion data to obtain a final conversion rate prediction model, and using the final conversion rate prediction model to calculate the predicted conversion rate of delivered advertisements for users;

[0013] Evaluate the advertising effectiveness of delivered ads based on the predicted conversion rate of delivered ads to users and their contribution to GMV; construct a user value model for evaluating the lifetime value of each user, and predict each user's ability to pay for ads based on the user value model;

[0014] Based on the predicted conversion rate and advertising effect of the delivered advertisements for users and the users' ability to pay for the advertisements, a return on investment (ROI) prediction model is constructed. The predicted ROI of the delivered advertisements is obtained using the ROI prediction model. Based on the ROI target preset by the advertiser and the predicted ROI, a corresponding optimized delivery strategy is generated.

[0015] As a preferred implementation mode of the present invention, the matching model is a dual-tower model, including a user tower, an advertising tower, an interaction layer and a fully connected network, wherein:

[0016] The user tower is used to encode user features and obtain the user's embedding vector, which is expressed as:

[0017] ;

[0018] ;

[0019] In the formula, is the user’s embedding vector; is the mapping function of the user tower; are the parameters of the user tower; For users User characteristics; , , and They are user attribute features, behavior data features, interest tag features, and real-time features;

[0020] The advertising tower is used to encode the advertising features and obtain the embedded vector of the advertisement, which is expressed as follows:

[0021] ;

[0022] ;

[0023] In the formula, is the embedding vector of the advertisement; is the mapping function of the advertising tower; The parameters of the advertising tower; For the Ad features of the ads to be placed; , , and They are advertising attribute characteristics, advertising content characteristics, advertising audience positioning characteristics and advertising real-time status characteristics.

[0024] As a preferred embodiment of the present invention, the interaction layer adopts a multi-head self-attention mechanism to expand the user's embedding vector and the advertisement's embedding vector into a multi-head form, which can be expressed as follows:

[0025] ;

[0026] ;

[0027] In the formula, and They are user extension embedding and ad extension embedding respectively; is the number of attention heads; and are the dimensions of the embedding vectors for users and ads, respectively; and They are and A real matrix of ; is the embedding vector of user features Output of heads; is the embedding vector of the advertising feature Output of heads;

[0028] Perform a linear transformation on the user extended embedding, expressed as:

[0029] ;

[0030] ;

[0031] In the formula, For user extensions embedded in The query vector of the attention head; For user extensions embedded in The key vector of the attention heads; Extending the user embedding in the multi-head attention mechanism The representation of an attention head; and is the weight matrix corresponding to the pre-set learnable linear transformation, with dimension ;

[0032] The linear transformation of ad extension embedding is expressed as:

[0033] ;

[0034] ;

[0035] In the formula, For ad extensions embedded in The key vector of the attention heads; For ad extensions embedded in The value vector of the attention heads; Extending the embedding for ads in the multi-head attention mechanism The representation of an attention head; and is the weight matrix corresponding to the pre-set learnable linear transformation, with dimension ;

[0036] The key vector is transformed into Align to query vector The dimension is expressed as:

[0037] ;

[0038] In the formula, Expand the aligned ad embed The key vector of the attention heads; The weight matrix corresponding to the preset learnable linear transformation;

[0039] For each attention head, the attention score is calculated by dot product, which is expressed as:

[0040] ;

[0041] In the formula, For the The attention score of each attention head; is the transpose operation; is the activation function;

[0042] The attention scores of each attention head are concatenated to obtain the total attention score, which can be expressed as:

[0043] ;

[0044] In the formula, for total attention score; is the connection function.

[0045] As a preferred implementation of the present invention, the matching score of the advertisement to be placed is output as follows:

[0046] The total attention score is extracted through a fully connected network to extract abstract features and calculate the matching score, which can be expressed as:

[0047] ;

[0048] ;

[0049] In the formula, is an abstract feature; is the implicit function of the fully connected network; Score for the match; The corresponding weight vector for the preset learnable fully connected network; is the corresponding bias term of the fully connected network.

[0050] As a preferred implementation mode of the present invention, the historical conversion data includes user At the advertising touchpoints where ads have been served Completed conversions on and users Through advertising touchpoints where ads have been served The conversion path formed ;

[0051] The multi-touch attribution model is a U-shaped attribution model, and the contribution of delivered ads to user conversion is calculated using the formula:

[0052] ;

[0053] In the formula, For ads that have been delivered to users Contribution in transformation; Advertising touchpoints that have delivered ads For users The weight of the conversion; Index variable used to traverse users All advertising touchpoints encountered in the conversion path; is the total number of advertising touch points;

[0054] Calculate the contribution of delivered ads to GMV based on the contribution of delivered ads in user conversions , expressed as:

[0055] .

[0056] As a preferred implementation of the present invention, the conversion rate prediction model is a binary classification model based on logistic regression. The final conversion rate prediction model is used to calculate the predicted conversion rate of the delivered advertisement to the user, which is expressed as follows:

[0057] ;

[0058] In the formula, For the Ads served to users Predicted conversion rate; is the Sigmoid function; is the weight vector after training; is the bias term after training; is the transpose operation.

[0059] As a preferred implementation of the present invention, the advertising effect of the delivered advertisements is evaluated based on the predicted conversion rate of the delivered advertisements to the users and the contribution to the GMV, which is expressed as follows:

[0060] ;

[0061] In the formula, For the The advertising performance of the ads that have been placed; For the Ads served to users Predicted conversion rate; For the Contribution of delivered ads to GMV.

[0062] As a preferred implementation mode of the present invention, the user value model is expressed as follows:

[0063] ;

[0064] In the formula, For the The lifetime value of a user; For in time Neidi The amount of consumption by each user; For in time Neidi The repurchase rate of each user; For in time Neidi Conversion rate of each user; is the discount rate; is the total length of the forecast period;

[0065] The payment ability of each user for advertisement is predicted based on the user value model and expressed as:

[0066] ;

[0067] In the formula, For the The payment capability of individual users; The average user life cycle.

[0068] As a preferred implementation mode of the present invention, the ROI prediction model is expressed as follows:

[0069] ;

[0070] In the formula, For the Predicted ROI of ads that have been served; is the total number of advertisements served; is the total number of users; For the direct costs of advertising that was placed; For the Indirect costs of advertising that has been placed; For the Ads served to users Predicted conversion rate; For the The payment capability of individual users; For the The advertising performance of the ads that have been placed;

[0071] Generating a corresponding optimization delivery strategy based on the ROI target preset by the advertiser and the predicted ROI specifically includes comparing the predicted ROI with the ROI target preset by the advertiser, and performing optimization if the predicted ROI is lower than the target ROI, including:

[0072] Keep ads with predicted ROI above the preset ROI target, stop advertising ads with predicted ROI below the preset ROI target, and prioritize advertising to users with high paying ability and high conversion rate;

[0073] Optimize advertising budget allocation and objective function based on predicted ROI and advertising cost The formula is:

[0074] ;

[0075] In the formula, For ads to be optimized budget proportion; For ads to be optimized Predicted ROI; For ads to be optimized direct costs; For ads to be optimized indirect costs; To maximize operations;

[0076] The corresponding constraints include budget constraints and budget ratio constraints, where:

[0077] The budget constraint is expressed as:

[0078] ;

[0079] In the formula, for the total budget;

[0080] The allocation ratio constraint is expressed as:

[0081] ;

[0082] The linear programming algorithm is used to solve and obtain the optimal budget ratio, and the advertising budget is adjusted based on the budget ratio to achieve optimization.

[0083] As a preferred implementation mode of the present invention, the method further includes, during the actual delivery process, real-time monitoring of the advertising effect and ROI of the delivered advertisements, and dynamically adjusting the delivery strategy according to the real-time data.

[0084] The present invention has the following beneficial effects:

[0085] 1. The present invention is an intelligent optimization advertising delivery method, which extracts user features from multiple dimensions such as user attributes, behavior data, interest tags and real-time data to ensure that advertising delivery can cover the diverse needs of users; the user features and advertising features are respectively encoded through the user tower and advertising tower in the matching model to generate an embedding vector, and the multi-head self-attention mechanism is used to enhance the interaction between user and advertising features, so as to more accurately calculate the matching score between advertisements and users, and prioritize the delivery of advertisements that are highly matched with user interests through the matching score ranking output by the matching model, effectively capturing the complex relationship between user interests and advertising content, and improving the accuracy of advertising delivery;

[0086] 2. The present invention is an intelligent optimization advertising delivery method, which uses a multi-touch attribution model to accurately calculate the contribution of each advertising touch point in the user conversion process, thereby evaluating the contribution of advertising to the overall GMV, solving the problem of difficulty in quantifying advertising effects in traditional advertising delivery, helping advertisers to better understand the return on advertising investment, and more accurately evaluating the role of advertising in the user conversion path by learning from the user's historical conversion data, avoiding simple reliance on single indicators such as click-through rate or exposure;

[0087] 3. The present invention is an intelligent optimization advertising delivery method, which predicts the ROI of advertising through the direct cost, indirect cost, predicted conversion rate and user payment ability of advertising, and generates an optimized advertising delivery strategy based on the ROI target preset by the advertiser to ensure the sustainability and efficiency of advertising delivery. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0089] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0090] It should be understood that the step numbers used in this document are only for convenience of description and are not intended to limit the order in which the steps are executed.

[0091] It should be understood that the terms used in the present specification are only for the purpose of describing specific embodiments and are not intended to limit the present invention. As used in the present specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0092] The terms “include” and “comprising” indicate the presence of described features, integers, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or combinations thereof.

[0093] The term "and / or" means and includes any and all possible combinations of one or more of the associated listed items.

[0094] Embodiment 1:

[0095] See also Figure 1 This embodiment provides an intelligent optimization advertising delivery method, including the following steps:

[0096] S1. Collect user data, including user attributes, behavior data, interest tags and real-time data, perform feature extraction on the user data, and obtain multi-dimensional user features, where:

[0097] The user attributes include age, gender, geographic location and device type;

[0098] The behavioral data includes click records, browsing history and search queries;

[0099] The interest tags are obtained by extracting user behavior and context information;

[0100] The real-time data includes the user's current online status, time and weather;

[0101] Preprocess the user data to ensure data quality and consistency. The data preprocessing includes cleaning data, standardizing data, and coding and classifying data. Furthermore, the data cleaning includes removing invalid or missing data and processing abnormal values. The standardizing data includes standardizing or normalizing numerical data.

[0102] Feature extraction is performed on the preprocessed user data to obtain user features, which include user attribute features, behavior data features, interest tag features, and real-time features, where:

[0103] S11. User attribute characteristics:

[0104] Age is discretized into different age groups, such as teenagers, adolescents, young adults, middle-aged people, and elderly people. In this embodiment, the teenager is under 12 years old, the adolescent is 13-17 years old, the young adult is 18-35 years old, the middle-aged is 36-55 years old, and the elderly is over 66 years old. Each age group is converted into a binary vector using one-hot encoding, where the teenager is [1, 0, 0, 0, 0], the adolescent is [0, 1, 0, 0, 0], the young adult is [0, 0, 1, 0, 0], the middle-aged is [0, 0, 0, 1, 0], and the elderly is [0,0, 0, 0, 1];

[0105] The gender characterization vector includes male and female. In this embodiment, the encoding is [male, female]. If it is male, the corresponding one-hot encoding is [1, 0], and if it is female, the corresponding one-hot encoding is [0, 1];

[0106] The geographic location is the longitude and latitude coordinates. The characteristic vector of the geographic location is obtained by dividing the longitude and latitude coordinates into grids and mapping the geographic location of the user to the corresponding grid number. In this embodiment, there are 36 grids for longitude (one grid every 10 degrees) and 18 grids for latitude (one grid every 10 degrees). Each grid number corresponds to a one-bit code. For example, the longitude and latitude coordinates of user A are (116.4074, 39.9042). The longitude 116.4074 of user A is in the 12th grid (110-120 degrees) and the latitude 39.9042 is in the 4th grid (30-40 degrees). The grid code of user A is: [12, 4], and the corresponding one-hot code is [0, 0, 0, ..., 1, ..., 0, 0, ..., 1, ...] (the 12th and 4th bits are 1, and the rest are 0);

[0107] The device types include mobile phones, tablets, and computers, which are encoded as [mobile phone, tablet, computer]. For example, when the user's device type is a mobile phone, the corresponding one-hot encoding is [1,0,0];

[0108] The above-mentioned user attribute characterization vectors are concatenated to obtain the user User attribute characteristics ;

[0109] S12. Behavioral data characteristics:

[0110] The characterization vector of the click record is a numerical representation of the number of times the user clicks;

[0111] The browsing history is encoded using multiple tags, and the tags are pre-set based on the theme, keywords, and metadata of the website. Furthermore, technology-related websites and content, such as computers, software, the Internet, and artificial intelligence, are divided into technology tags; entertainment-related websites and content, such as movies, music, videos, and games, are divided into entertainment tags; sports-related websites and content, such as football, basketball, athletics, and fitness, are divided into sports tags; news-related websites and content, such as current affairs, international news, and local news, are divided into news tags; finance-related websites and content, such as stock market, investment, and financial news, are divided into finance tags; education-related websites and content, such as online courses, academic papers, and education news, are divided into education tags; health-related websites and content, such as medical information, healthy life, and mental health, are divided into health tags; daily life-related websites and content, such as food, travel, and home, are divided into life tags. Therefore, in this embodiment, the multiple tags are encoded as [technology, entertainment, sports, news, finance, education, health, and life];

[0112] The search query is a keyword searched by a user, which is converted into a vector through TF-IDF or word embedding;

[0113] The characterization vectors of the above behavior data are concatenated to obtain the user Behavioral data characteristics ;

[0114] S13, interest tag features;

[0115] The interest tag feature uses multi-tag coding. In this embodiment, the multi-tag coding of interest tags is consistent with the multi-tag coding of browsing history. By analyzing user behavior and context information, users are assigned corresponding interest tags, and the user's interest tags are obtained through multi-tag coding. Behavioral data characteristics ;

[0116] S14. Real-time features:

[0117] The current online status of the user is represented by a binary classification, coded as [online, offline];

[0118] The time is coded in hours, for example, 3:00 p.m. is 15;

[0119] The weather code is [sunny, cloudy, snowy, rainy]; for example, if the current weather of user A is sunny, the corresponding one-hot code is [1,0,0,0];

[0120] The characterization vectors of real-time data are concatenated to obtain the user Real-time features ;

[0121] S15, vectorize the user attribute features, behavior data features, interest tag features and real-time features to obtain the user User feature vector ;

[0122] S2. Build a matching model, take the user features and the advertisement features corresponding to the advertisements to be placed as input, output the matching scores of the advertisements to be placed, sort the matching scores, and place the advertisements to be placed in order from high to low;

[0123] S21. The advertisement features corresponding to the advertisement to be placed include advertisement attribute features, advertisement content features, advertisement audience positioning features, and advertisement real-time status features, wherein:

[0124] The advertising attribute features include advertising type, advertising format, advertising duration and advertising cost. Specifically, the advertising type includes pictures, videos and texts, which are encoded as [picture, video, text]. For example, when the advertising type is a picture, the corresponding unique hot encoding is [1,0,0]; the advertising format includes dynamic advertising, static advertising and interactive advertising, which are encoded as [dynamic, static, interactive]; the advertising duration is a numerical feature, for example, the video duration is 15 seconds, and the corresponding feature is 15; the advertising cost is a numerical feature, for example, the advertising cost is 100 yuan, and the corresponding feature is 100; the advertising type, advertising format, advertising duration and advertising cost are vectorized to obtain the first Advertisement attribute characteristics of ads to be delivered ;

[0125] The advertising content features include the ad title vectorization, ad description vectorization, ad image features and ad video features. Furthermore, the ad title and ad text description are converted into vectors using TF-IDF or word embedding methods to obtain ad title vectorization and ad description vectorization; ad image features are extracted using convolutional neural networks to obtain ad image features; key frame features are extracted from video ads to obtain ad video features; ad title vectorization, ad description vectorization, ad image features and ad video features are vectorized and spliced ​​to obtain the first Ad content features of ads to be delivered ;

[0126] The advertising audience positioning features include audience age group, audience gender, audience geographic location and audience device type. Furthermore, the audience age group is consistent with the age feature extraction method of the user attribute in step S1, the audience gender is consistent with the gender feature extraction method of the user attribute in step S1, the audience geographic location is consistent with the geographic location feature extraction method of the user attribute in step S1, and the audience device type is consistent with the device type feature extraction method of the user attribute in step S1; the audience age group, audience gender, audience geographic location and audience device type are vectorized to obtain the first Ad content features of ads to be delivered ;

[0127] The real-time features of the advertisement include the advertisement delivery time and the advertisement remaining budget. Furthermore, the advertisement delivery time is coded in hours, for example, 15 is coded for 3 pm. The advertisement remaining budget is coded in numerical values, for example, if the remaining budget is 50 yuan, the corresponding feature is 50. The advertisement delivery time and the advertisement remaining budget are vectorized to obtain the first Ad content features of ads to be delivered ;

[0128] The advertisement attribute features, advertisement content features, advertisement audience positioning features and advertisement real-time status features are vectorized to obtain the first Ad features of ads to be served ;

[0129] S22. The matching model is a dual-tower model, including a user tower, an advertising tower, an interactive layer, and a fully connected network, wherein:

[0130] S221. The user characteristics are encoded through user towers and expressed as follows:

[0131] ;

[0132] In the formula, is the user’s embedding vector; is the mapping function of the user tower; are the parameters of the user tower;

[0133] The advertising features are encoded by the advertising tower and expressed as follows:

[0134] ;

[0135] In the formula, is the embedding vector of the advertisement; is the mapping function of the advertising tower; The parameters of the advertising tower;

[0136] S222, using multi-head self-attention mechanism as the interaction layer;

[0137] The embedding vectors of user features and advertising features are expanded into multi-head forms respectively, and expressed as follows:

[0138] ;

[0139] ;

[0140] In the formula, and They are user extension embedding and ad extension embedding respectively; is the number of attention heads; and are the dimensions of the embedding vectors of user features and ad features respectively; and They are and A real matrix of ; is the embedding vector of user features Output of heads; is the embedding vector of the advertising feature Output of heads;

[0141] Perform a linear transformation on the user extended embedding, expressed as:

[0142] ;

[0143] ;

[0144] In the formula, For user extensions embedded in The query vector of the attention head; For user extensions embedded in The key vector of the attention heads; Extending the user embedding in the multi-head attention mechanism The representation of an attention head; and is the weight matrix corresponding to the pre-set learnable linear transformation, with dimension ;

[0145] The linear transformation of ad extension embedding is expressed as:

[0146] ;

[0147] ;

[0148] In the formula, For ad extensions embedded in The key vector of the attention heads; For ad extensions embedded in The value vector of the attention heads; Extending the embedding for ads in the multi-head attention mechanism The representation of an attention head; and is the weight matrix corresponding to the pre-set learnable linear transformation, with dimension ;

[0149] The key vector is transformed into Align to query vector The dimension is expressed as:

[0150] ;

[0151] In the formula, Expand the aligned ad embed The key vector of the attention heads; The weight matrix corresponding to the preset learnable linear transformation;

[0152] For each attention head, the attention score is calculated by dot product, which is expressed as:

[0153] ;

[0154] In the formula, For the The attention score of each attention head; is the transpose operation; is the activation function;

[0155] The attention scores of each attention head are concatenated to obtain the total attention score, which can be expressed as:

[0156] ;

[0157] In the formula, for total attention score; is the connection function;

[0158] S223. After further processing the total attention score through the fully connected network, a higher level of abstract features are extracted to calculate the matching score, which is expressed as:

[0159] ;

[0160] ;

[0161] In the formula, is an abstract feature; is the implicit function of the fully connected network; Score for the match; The corresponding weight vector for the preset learnable fully connected network; is the corresponding bias term of the fully connected network;

[0162] S3. Obtain the historical conversion data of the user, and based on the historical conversion data, use a multi-touch attribution model to calculate the contribution of the delivered advertisements to the user conversion, and calculate the contribution of the delivered advertisements to GMV based on the contribution of the delivered advertisements to the user conversion; construct a conversion rate prediction model, train the conversion rate prediction model using user characteristics, advertising characteristics of the delivered advertisements, and historical conversion data to obtain a final conversion rate prediction model, and use the final conversion rate prediction model to calculate the predicted conversion rate of the delivered advertisements for the user; evaluate the advertising effect of the delivered advertisements based on the predicted conversion rate of the delivered advertisements for the user and their contribution to GMV;

[0163] S31, the historical conversion data includes user At the advertising touchpoints where ads have been served Completed conversions on and users Through advertising touchpoints where ads have been served The conversion path formed ;

[0164] S32. In this embodiment, the multi-touch attribution model is a U-shaped attribution model, and the contribution of the delivered advertisements to user conversion is calculated using the formula:

[0165] ;

[0166] In the formula, For ads that have been delivered to users Contribution in transformation; Advertising touchpoints that have delivered ads For users The weight of the conversion; Index variable used to traverse users All advertising touchpoints encountered in the conversion path; is the total number of advertising touch points;

[0167] Calculate the contribution of delivered ads to GMV based on the contribution of delivered ads in user conversions , expressed as:

[0168] ;

[0169] S33. In this embodiment, the conversion rate prediction model is a binary classification model based on logistic regression. The conversion rate prediction model concatenates user features and advertising features to obtain a concatenated vector, performs a linear transformation on the concatenated vector, and maps the linearly transformed vector to the [0,1] interval through a Sigmoid function to obtain a predicted conversion rate.

[0170] The conversion rate prediction model is trained using user characteristics, advertising characteristics of delivered advertisements, and historical conversion data as follows:

[0171] Acquire the advertisement features corresponding to the delivered advertisements and the user features corresponding to the historical conversion data. The acquisition method is based on the same principle as steps S1 and S2 of this embodiment and will not be described in detail here.

[0172] Obtaining a conversion tag based on the historical conversion data, wherein the conversion tag is used to indicate whether the user has converted the delivered advertisement. If the user has completed the conversion (i.e., has made a purchase, registered, or other conversion behavior) at a certain advertisement touch point, the conversion tag of the advertisement touch point may be defined as 1; if the user has not completed the conversion at a certain advertisement touch point, the conversion tag of the advertisement touch point may be defined as 0;

[0173] The conversion rate prediction model is trained using the advertising features corresponding to the delivered advertisements, the user features and conversion labels corresponding to the users in the historical conversion data, and the final conversion rate prediction model is obtained;

[0174] The final conversion rate prediction model is used to predict the conversion rate of the delivered advertisements to the users. The formula is:

[0175] ;

[0176] In the formula, For the Ads served to users Predicted conversion rate; is the Sigmoid function; is the weight vector after training; is the bias term after training;

[0177] The conversion rate prediction model is trained by user characteristics, ad characteristics and historical conversion data to accurately predict the user's conversion rate for ads. This prediction capability can estimate the effect of ads in advance and help advertisers make more informed decisions before placing ads.

[0178] S34. Evaluate the advertising effect of the delivered advertisements based on the predicted conversion rate of the delivered advertisements to users and their contribution to GMV, expressed as follows:

[0179] ;

[0180] In the formula, For the The advertising performance of the ads that have been placed; For the Contribution of the ads placed to GMV;

[0181] S4. constructing a user value model for evaluating the lifetime value of each user, and predicting each user's ability to pay for advertisements based on the user value model;

[0182] S41. The user value model is expressed as follows:

[0183] ;

[0184] In the formula, For the The lifetime value of a user; For in time Neidi The amount of consumption by each user; For in time Neidi The repurchase rate of each user; For in time Neidi Conversion rate of each user; is the discount rate; is the total length of the forecast period;

[0185] By using indicators such as consumption, repurchase rate and conversion rate, a user lifetime value model is constructed to predict each user's ability to pay for advertising. This model helps advertisers identify high-value users and deliver personalized advertising to these users to improve advertising conversion effects. In addition, considering the value of time, a discount rate is used to discount the user's future consumption, making the prediction of user lifetime value more accurate and reasonable.

[0186] S42: predicting each user's ability to pay for advertisements based on the user value model, expressed as follows:

[0187] ;

[0188] In the formula, For the The payment capability of individual users; is the average user life cycle;

[0189] S5. Based on the predicted conversion rate and advertising effect of the delivered advertisements and the user's payment ability, a ROI prediction model is constructed, and the predicted ROI of the delivered advertisements is obtained by using the ROI prediction model, and a corresponding optimized delivery strategy is generated based on the ROI target preset by the advertiser and the predicted ROI;

[0190] S51. ROI is an important indicator for measuring the effectiveness of advertising, and is calculated by the ratio of the revenue brought by advertising to the advertising cost. In this embodiment, the ROI prediction model is expressed as follows:

[0191] ;

[0192] In the formula, For the Predicted ROI of ads that have been served; is the total number of advertisements served; is the total number of users; For the direct costs of advertising that was placed; For the Indirect costs of advertising that has been placed;

[0193] S52, generating a corresponding optimized delivery strategy based on the ROI target preset by the advertiser and the predicted ROI, specifically comparing the predicted ROI with the ROI target preset by the advertiser, and performing optimization if the predicted ROI is lower than the target ROI, including:

[0194] S521, retaining advertisements with predicted ROIs above the preset ROI target, stopping the delivery of advertisements with predicted ROIs below the preset ROI target, and giving priority to delivering advertisements to users with high paying ability and high conversion rate; further, if the predicted ROI is above the preset ROI target or there are too many advertisements of the same type, sorting the delivered advertisements according to the predicted ROI, giving priority to retaining advertisements with higher ROIs, and reducing or stopping the delivery of advertisements with lower ROIs;

[0195] S522. Give priority to advertising to users with high paying ability and high conversion rate;

[0196] S523. Optimize advertising budget allocation and objective function based on predicted ROI and advertising cost The formula is:

[0197] ;

[0198] In the formula, For ads to be optimized budget proportion; For ads to be optimized Predicted ROI; For ads to be optimized direct costs; For ads to be optimized indirect costs; To maximize operations;

[0199] The corresponding constraints include budget constraints and budget ratio constraints, where:

[0200] The budget constraint is expressed as:

[0201] ;

[0202] In the formula, for the total budget;

[0203] The allocation ratio constraint is expressed as:

[0204] ;

[0205] Using a linear programming algorithm to solve and obtain the optimal budget ratio, and adjusting the advertising budget based on the budget ratio to achieve optimization;

[0206] This embodiment optimizes the advertising budget allocation through a linear programming algorithm to ensure that the ROI of advertising is maximized under the total budget constraint, helping advertisers to reasonably allocate resources and avoid resource waste;

[0207] Through the direct cost, indirect cost, predicted conversion rate and user payment ability of advertising, as well as the predicted ROI of advertising, a globally optimized advertising delivery strategy is generated, which not only focuses on the short-term effect of advertising, but also considers the long-term value of users to ensure the long-term sustainability of advertising. Based on the ROI target preset by the advertiser, an optimized advertising delivery strategy is generated to ensure that the advertising delivery target is consistent with the advertiser's strategic goal. If the predicted ROI is lower than the target, the advertising delivery is automatically adjusted to ensure the sustainability and efficiency of advertising.

[0208] S6. During the actual delivery process, monitor the advertising effect and ROI of the delivered advertisements in real time, and dynamically adjust the delivery strategy based on real-time data to ensure that the advertising delivery can adapt to market changes and changes in user behavior. For example: if the real-time ROI of an advertisement is lower than expected, reduce the delivery frequency of the advertisement or stop delivery; if the real-time ROI of an advertisement is higher than expected, increase the delivery budget of the advertisement or expand the delivery scope.

[0209] In summary, this embodiment provides an intelligent optimization advertising delivery method, which realizes the optimization of the entire process of advertising delivery through precise user matching, multi-touch attribution model, conversion rate prediction, user lifetime value prediction, ROI prediction and optimization delivery strategy. The method described in this embodiment can not only help advertisers improve the conversion rate and ROI of advertisements, but also take into account the long-term value of users and ensure the sustainability and efficiency of advertising delivery. At the same time, through real-time monitoring and dynamic adjustment, it can quickly respond to market changes and improve the flexibility and adaptability of advertising delivery.

[0210] In the embodiments of the present invention, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B may represent the existence of A alone, the existence of A and B at the same time, and the existence of B alone. A and B may be singular or plural. The character " / " generally indicates that the associated objects before and after are in an "or" relationship. "At least one of the following" and similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one of a, b and c may represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, c may be single or multiple.

[0211] Those skilled in the art will appreciate that the various units and algorithm steps described in the embodiments disclosed herein can be implemented in a combination of electronic hardware, computer software, and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of the present invention.

[0212] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0213] In several embodiments provided by the present invention, if any function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art or the part of the technical solution, can be embodied in the form of a software product, which is stored in a storage medium and includes several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory; hereinafter referred to as: ROM), random access memory (Random Access Memory; hereinafter referred to as: RAM), disk or optical disk and other media that can store program codes.

[0214] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. An intelligent optimization advertising delivery method, characterized in that: The method comprises: Collect user data, including user attributes, behavior data, interest tags and real-time data, perform feature extraction on the user data, and obtain multi-dimensional user features; Constructing a matching model, taking the user features and the advertisement features corresponding to the advertisements to be placed as input, outputting the matching scores of the advertisements to be placed, sorting the matching scores, and placing the advertisements to be placed in order from high to low; Obtain the user's historical conversion data, and based on the historical conversion data, use a multi-touch attribution model to calculate the contribution of the delivered advertisements to the user's conversion, and calculate the contribution of the delivered advertisements to the GMV based on the contribution of the delivered advertisements to the user's conversion; Constructing a conversion rate prediction model, training the conversion rate prediction model using user characteristics, advertising characteristics of delivered advertisements, and historical conversion data to obtain a final conversion rate prediction model, and using the final conversion rate prediction model to calculate the predicted conversion rate of delivered advertisements for users; Evaluate the advertising effectiveness of delivered ads based on the predicted conversion rate of delivered ads to users and their contribution to GMV; construct a user value model for evaluating the lifetime value of each user, and predict each user's ability to pay for ads based on the user value model; Based on the predicted conversion rate and advertising effect of the delivered advertisements for users and the users' ability to pay for the advertisements, a return on investment (ROI) prediction model is constructed. The predicted ROI of the delivered advertisements is obtained using the ROI prediction model. Based on the ROI target preset by the advertiser and the predicted ROI, a corresponding optimized delivery strategy is generated.

2. The intelligent optimization advertising delivery method according to claim 1, characterized in that: The matching model is a dual-tower model, including a user tower, an advertising tower, an interaction layer, and a fully connected network, wherein: The user tower is used to encode user features and obtain the user's embedding vector, which is expressed as follows: ； ； In the formula, is the user’s embedding vector; is the mapping function of the user tower; are the parameters of the user tower; For users User characteristics; , , and They are user attribute features, behavior data features, interest tag features, and real-time features; The advertising tower is used to encode the advertising features and obtain the embedded vector of the advertisement, which is expressed as follows: ； ； In the formula, is the embedding vector of the advertisement; is the mapping function of the advertising tower; The parameters of the advertising tower; For the Ad features of the ads to be placed; , , and They are advertising attribute characteristics, advertising content characteristics, advertising audience positioning characteristics and advertising real-time status characteristics.

3. The intelligent optimization advertising delivery method according to claim 2, characterized in that: The interaction layer adopts a multi-head self-attention mechanism to expand the user's embedding vector and the advertisement's embedding vector into multi-head forms, which can be expressed as: ； ； In the formula, and They are user extension embedding and ad extension embedding respectively; is the number of attention heads; and are the dimensions of the embedding vectors for users and ads, respectively; and They are and A real matrix of ; is the embedding vector of user features Output of heads; is the embedding vector of the advertising feature Output of heads; Perform a linear transformation on the user extended embedding, expressed as: ； ； In the formula, For user extensions embedded in The query vector of the attention head; For user extensions embedded in The key vector of the attention heads; Extending the user embedding in the multi-head attention mechanism The representation of an attention head; and is the weight matrix corresponding to the pre-set learnable linear transformation, with dimension ; The linear transformation of ad extension embedding is expressed as: ； ； In the formula, For ad extensions embedded in The key vector of the attention heads; For ad extensions embedded in The value vector of the attention heads; Extending the embedding for ads in the multi-head attention mechanism The representation of an attention head; and is the weight matrix corresponding to the pre-set learnable linear transformation, with dimension ; The key vector is transformed into Align to query vector The dimension is expressed as: ； In the formula, Expand the aligned ad embed The key vector of the attention heads; The weight matrix corresponding to the preset learnable linear transformation; For each attention head, the attention score is calculated by dot product, which is expressed as: ； In the formula, For the The attention score of each attention head; is the transpose operation; is the activation function; The attention scores of each attention head are concatenated to obtain the total attention score, which can be expressed as: ； In the formula, for total attention score; is the connection function.

4. The intelligent optimization advertising delivery method according to claim 3, characterized in that: Output the matching score of the ad to be delivered as follows: The total attention score is extracted through a fully connected network to extract abstract features and calculate the matching score, which can be expressed as: ； ； In the formula, is an abstract feature; is the implicit function of the fully connected network; Score for the match; The corresponding weight vector for the preset learnable fully connected network; is the corresponding bias term of the fully connected network.

5. The intelligent optimization advertising delivery method according to claim 1, characterized in that: The historical conversion data includes user At the advertising touchpoints where ads have been served Completed conversions on and users Through advertising touchpoints where ads have been served The conversion path formed ; The multi-touch attribution model is a U-shaped attribution model, and the contribution of delivered ads to user conversion is calculated using the formula: ； In the formula, For ads that have been delivered to users Contribution in transformation; Advertising touchpoints that have delivered ads For users The weight of the conversion; Index variable used to traverse users All advertising touchpoints encountered in the conversion path; is the total number of advertising touch points; Calculate the contribution of delivered ads to GMV based on the contribution of delivered ads in user conversions , expressed as: 。 6. The intelligent optimization advertising delivery method according to claim 1, characterized in that: The conversion rate prediction model is a binary classification model based on logistic regression. The final conversion rate prediction model is used to calculate the predicted conversion rate of the delivered advertisements to the users, which is expressed as follows: ； In the formula, For the Ads served to users Predicted conversion rate; is the Sigmoid function; is the weight vector after training; is the bias term after training; is the transpose operation.

7. The intelligent optimization advertising delivery method according to claim 1, characterized in that: The advertising effect of the delivered ads is evaluated based on the predicted conversion rate of the delivered ads to users and their contribution to GMV, which can be expressed as: ； In the formula, For the The advertising performance of the ads that have been placed; For the Ads served to users Predicted conversion rate; For the Contribution of delivered ads to GMV.

8. The intelligent optimization advertising delivery method according to claim 1, characterized in that: The user value model is expressed in the formula: ； In the formula, For the The lifetime value of a user; For in time Neidi The amount of consumption by each user; For in time Neidi The repurchase rate of each user; For in time Neidi Conversion rate of each user; is the discount rate; is the total length of the forecast period; The payment ability of each user for advertisement is predicted based on the user value model and expressed as: ； In the formula, For the The payment capability of individual users; The average user life cycle.

9. The intelligent optimization advertising delivery method according to claim 1, characterized in that: The ROI prediction model is expressed as: ； In the formula, For the Predicted ROI of ads that have been served; is the total number of advertisements served; is the total number of users; For the direct costs of advertising that was placed; For the Indirect costs of advertising that has been placed; For the Ads served to users Predicted conversion rate; For the The payment capability of individual users; For the The advertising performance of the ads that have been placed; Generating a corresponding optimization delivery strategy based on the ROI target preset by the advertiser and the predicted ROI specifically includes comparing the predicted ROI with the ROI target preset by the advertiser, and performing optimization if the predicted ROI is lower than the target ROI, including: Keep ads with predicted ROI above the preset ROI target, stop advertising ads with predicted ROI below the preset ROI target, and prioritize advertising to users with high paying ability and high conversion rate; Optimize advertising budget allocation and objective function based on predicted ROI and advertising cost The formula is: ； In the formula, For ads to be optimized budget proportion; For ads to be optimized Predicted ROI; For ads to be optimized direct costs; For ads to be optimized indirect costs; To maximize operations; The corresponding constraints include budget constraints and budget ratio constraints, where: The budget constraint is expressed as: ； In the formula, for the total budget; The allocation ratio constraint is expressed as: ； The linear programming algorithm is used to solve and obtain the optimal budget ratio, and the advertising budget is adjusted based on the budget ratio to achieve optimization.

10. The intelligent optimization advertising delivery method according to claim 1, characterized in that: The method also includes monitoring the advertising effect and ROI of the delivered advertisements in real time during the actual delivery process, and dynamically adjusting the delivery strategy according to the real-time data.

Images