A Machine Learning-Based Advertising Transaction Simulation and Prediction Method and Device

Abstract

The present application discloses a method and apparatus for advertising transaction simulation and prediction based on machine learning. The method includes: creating a virtual media with specific features and user behavior patterns; creating a virtual advertising budget party with specific behavior patterns; generating an advertising request for the virtual media and sending the advertising request to the AD Exchange system; the AD Exchange system selects a matching advertisement for display according to the received advertising request in combination with the specific behavior pattern of the virtual advertising budget party; using a two-tower model to predict the matching relationship between the virtual media and the virtual advertising budget party; simulating the bidding result of the AD Exchange system; returning the bidding result to the virtual media and using a click-through rate prediction model to predict the click-through rate of the advertisement on the virtual media. The present application can utilize the existing infrastructure of the AD Exchange system, reduce hardware investment, and also reduce the costs of system maintenance and upgrade; greatly reduce the technical risks and costs of developing new algorithms from scratch, and reduce technical input.

Description

Technical Field

[0001] This application relates to the field of digital advertising technology, and particularly to a method and device for simulating and predicting advertising transactions. Background Art

[0002] The rapid growth of the digital advertising industry is closely related to the popularization of the Internet. With the rapid development and popularization of Internet technology, especially mobile Internet technology, the number of Internet users worldwide has increased sharply, and online activities have become the norm in daily life. This transformation has triggered the rapid expansion of the digital advertising market, making it an important part of the advertising industry. Enterprises and advertisers are increasingly shifting their marketing budgets from traditional media to online platforms in order to reach their target audiences more effectively. As more consumers spend time online, digital advertising provides brands with unprecedented opportunities to increase their visibility and influence.

[0003] At the same time, the rise of the AD Exchange system marks an important turning point in the field of digital advertising. As a platform for real-time buying and selling of advertising slots, the AD Exchange system enables advertisers to quickly and effectively purchase advertising space through the real-time bidding (RTB) mechanism. The implementation of this system has greatly improved the efficiency and accuracy of advertising placement, enabling advertisers to make advertising placement decisions based on instant user data. For example, when a user visits a web page, the AD Exchange system can complete the auction of the advertising slot in an instant, ensuring that the most relevant advertisement is shown to the user. This technological advancement not only optimizes the allocation of advertising resources but also provides users with a more relevant and personalized advertising experience.

[0004] Data plays a crucial role in digital advertising. In this data-driven era, understanding user behavior, predicting market trends, and optimizing advertising strategies all rely on the analysis and processing of large amounts of data. Every online interaction of users generates data, providing valuable insights for advertisers to understand consumers' interests, preferences, and purchasing behaviors. These data can not only be used to improve the targeting of advertisements but also to predict future market trends and consumer behaviors, helping brands make more informed marketing decisions.

[0005] With the rapid development of data science and machine learning technologies, the ability and accuracy of simulating and predicting advertising transactions have been continuously improved, which is of great significance for the development of the advertising industry. In the field of digital advertising, real-time bidding and precise targeting are core elements, and these are exactly the problems that machine learning technologies are good at dealing with. Using big data analysis and machine learning algorithms, the simulation system can deeply understand market dynamics and effectively predict the effectiveness of advertising placement.

[0006] These simulation systems can process and analyze large amounts of historical data to reproduce and understand market behavior. Through in-depth analysis of historical data, these systems can identify the complex relationships between market trends, user preferences, and advertising effectiveness. In addition, they can test new advertising strategies and algorithms in a controlled environment, which is crucial for the optimization and innovation of advertising strategies. By applying machine learning models, the simulation systems can not only predict the performance of specific advertisements but also help optimize advertising placement strategies, such as determining the best bid, selecting the most effective ad types, and targeting the most suitable audience. This is of great value to advertisers as it can significantly improve the ROI (return on investment) of advertising.

[0007] The inventors recognize that although advertising transaction simulation systems based on machine learning have great potential, their current scope of application and practicality are quite limited. On the one hand, these systems have been more widely applied in the academic research field, but they usually cannot access real-time data streams. This means that although these systems can use historical data for in-depth analysis and research, their ability to predict future market dynamics and respond in real time is limited. Therefore, although these simulation systems have significant value in theoretical research and methodological innovation, their application potential in the actual business environment has not been fully explored. This is mainly because they lack interaction with real-time market data, which is crucial for predicting market trends and guiding actual advertising placement decisions.

[0008] On the other hand, although some large enterprises have started to use these simulation systems internally, small and medium-sized enterprises have difficulty accessing and utilizing these advanced technologies. This limitation mainly stems from two aspects: First, the development and maintenance of such advanced simulation systems require expensive technical inputs and professional knowledge, which is a significant financial burden for many small and medium-sized enterprises; second, these systems usually require large amounts of data and powerful computing capabilities, and small and medium-sized enterprises often lack the necessary resources and basic hardware facilities to support such systems. Therefore, although these simulation systems theoretically provide a powerful tool for optimizing advertising placement strategies, in practical applications, their high costs and technical requirements make only large enterprises with sufficient resources able to afford them. Summary of the Invention

[0009] This application provides an advertising transaction simulation and prediction method, aiming to solve the technical problem that existing simulation systems require expensive capital and technical inputs.

[0010] In a first aspect, an advertising transaction simulation and prediction method based on machine learning includes:

[0011] S1, creating a virtual media with specific characteristics and user behavior patterns;

[0012] S2. Create a virtual advertising budget party with a specific behavior pattern;

[0013] S3. Generate an advertisement request for the virtual media and send the advertisement request to the AD Exchange system; the AD Exchange system selects a matching advertisement for display according to the received advertisement request and in combination with the specific behavior pattern of the virtual advertising budget party;

[0014] S4. Use a two-tower model to predict the matching relationship between the virtual media and the virtual advertising budget party;

[0015] S5. Simulate the bidding result of the AD Exchange system;

[0016] S6. Return the bidding result to the virtual media and use a click-through rate prediction model to predict the click-through rate of the advertisement on the virtual media.

[0017] In the above solution, optionally, step S1 includes:

[0018] Collect historical media advertisement data;

[0019] Determine the key metrics in the historical media advertisement data that affect the performance of the virtual media;

[0020] Conduct exploratory data analysis on the key metrics, and discover the specific characteristics and user behavior patterns of different types of media according to the key metrics;

[0021] Use the key metrics and the specific characteristics and user behavior patterns of the corresponding different types of media to train a first machine learning model to obtain a virtual media model;

[0022] Use the virtual media model to generate virtual media with specific characteristics and user behavior patterns.

[0023] In the above solution, optionally, step S2 includes:

[0024] Collect historical advertisement placement data;

[0025] Analyze the historical advertisement placement data, determine the behavior patterns of different advertising budget parties, and identify the behavior trends of advertising budget parties under different market conditions;

[0026] Use the historical advertisement placement data and the behavior patterns of the corresponding different advertising budget parties to train a second machine learning model to obtain a virtual advertising budget party model;

[0027] Use the virtual advertising budget party model to generate a virtual advertising budget party with a specific behavior pattern.

[0028] In the above solution, optionally, generating an advertisement request for the virtual media and sending the advertisement request to the AD Exchange system in step S3 includes:

[0029] Simulate the behavior of virtual users on the virtual media according to the user behavior pattern of the virtual media, and simulate the interaction between the virtual users and the website content according to the predefined behavior pattern of the virtual users;

[0030] When the simulated user behavior triggers the display of an advertisement, generate an advertisement request, and the advertisement request contains the interaction information between the virtual user and the website content;

[0031] Extract and encapsulate the key features of the advertisement request, and send the encapsulated advertisement request to the ADExchange system.

[0032] In the above solution, optionally, step S4 includes:

[0033] Establish the architecture of the dual-tower model, which includes a budget-side tower model and a media tower model. The budget-side tower model is used to learn the features of the advertisement budget side, and the media tower model is used to learn the features of the media;

[0034] Obtain advertisement budget-side data and media data, and perform cleaning and normalization processing on the advertisement budget-side data and media data;

[0035] Extract useful features from the advertisement budget-side data and media data respectively, and generate negative samples based on the advertisement budget-side data and media data using the negative sampling technique;

[0036] Use the advertisement budget-side data and media data after feature extraction and the negative samples to train the budget-side tower model and the media tower model simultaneously, so that the budget-side tower model and the media tower model respectively learn the high-dimensional representations of the advertisement budget side and the media;

[0037] Adopt the contrastive loss method to optimize the budget-side tower model and the media tower model;

[0038] Input the data of the virtual advertisement budget side into the budget-side tower model, and input the data of the virtual media into the media tower model;

[0039] Calculate the similarity between the output vector of the budget-side tower model and the output vector of the media tower model;

[0040] According to the similarity score, match the virtual advertisement budget side with the virtual media with the highest similarity score for each virtual advertisement budget side.

[0041] In the above solution, optionally, step S5 includes:

[0042] Set up a virtual AD Exchange environment, where the AD Exchange environment includes the characteristics of ad spaces available for bidding, a list of virtual ad budget parties, and bidding rules;

[0043] Collect data required for bidding and use the collected data to train a third machine learning model to obtain a bid prediction model;

[0044] When an ad space is available, simulate bid requests from different virtual ad budget parties and match appropriate bid requests from virtual ad budget parties for the ad space available for bidding according to the characteristics of the ad space available for bidding and the target audiences of the virtual ad budget parties;

[0045] Use the bid prediction model to predict the bids of each matched virtual ad budget party for the ad space available for bidding;

[0046] Determine the winner among the matched virtual ad budget parties according to the bidding rules to obtain the bid result of the winner.

[0047] In the above solution, optionally, the click-through rate prediction model is a recurrent neural network based on LSTM.

[0048] In the above solution, further optionally, the click-through rate prediction model includes an input layer, a hidden layer, and an output layer, and the hidden layer includes an input gate, a forget gate, a cell unit, and an output gate;

[0049] The input of the input gate consists of three input vectors, and the three input vectors are the layer nodes of the output vector the output vector of the units of the first hidden layer the information retained by the unit at the previous time point The input vector of the input gate at time t is:

[0050]

[0051] Through the activation function f of the input gate, the output vector at time t is:

[0052]

[0053] The input of the forget gate also consists of three input vectors, and these three vectors have the same sources as those of the input gate. The input vector of the forget gate at time t is:

[0054]

[0055] Through the activation function f of the forget gate, the output vector obtained at time t is:

[0056]

[0057] The input of the cell unit consists of two parts. One part is the input vector of the input layer The other part is the output of the first hidden layer of the output gate The input vector of the cell unit at time t is:

[0058]

[0059] Determine whether to retain the past information value according to the forget gate The formula is:

[0060]

[0061] The input of the output gate consists of three parts, including the output vector of the input layer, the hidden output vector of the cell layer, and the information retained by the cell. The input vector of the output gate at time t is:

[0062]

[0063] Obtain the output vector at time t through the activation function f of the output gate is:

[0064]

[0065] The output vector of the cell unit is:

[0066]

[0067] The output vector of the cell unit That is, the output vector of the hidden layer, which is used as the input vector of the output layer. The formula is:

[0068]

[0069] The result vector output by the output layer is:

[0070]

[0071] The weight w between the i node and the j node ij is updated to:

[0072]

[0073] where η represents the learning step size is expressed as:

[0074]

[0075] Among them The residual node represented as j The output vector node represented as j, w ij Simplified to:

[0076]

[0077] In the above solution, further optionally, the loss function of the click-through rate prediction model is:

[0078]

[0079] Where y i Represents the true value of the i-th click, p i Represents the value estimated by the model for the i-th click.

[0080] In a second aspect, an advertisement trading simulation and prediction device based on machine learning includes:

[0081] A virtual media creation module for creating virtual media with specific characteristics and user behavior patterns;

[0082] A virtual advertising budget party creation module for creating a virtual advertising budget party with a specific behavior pattern;

[0083] An advertisement request generation module for generating advertisement requests for the virtual media and sending the advertisement requests to the ADExchange system; the AD Exchange system selects matching advertisements for display according to the received advertisement requests and in combination with the specific behavior pattern of the virtual advertising budget party;

[0084] A matching relationship prediction module for predicting the matching relationship between the virtual media and the virtual advertising budget party by using a two-tower model;

[0085] A bid result simulation module for simulating the bid results of the AD Exchange system;

[0086] A click-through rate prediction module for returning the bid results to the virtual media and predicting the click-through rate of the advertisements on the virtual media by using a click-through rate prediction model.

[0087] Compared with the prior art, the present application has at least the following beneficial effects:

[0088] In the advertising transaction simulation and prediction method provided by the embodiments of the present application, by creating virtual media with specific characteristics and user behavior patterns, creating virtual advertising budget parties with specific behavior patterns, generating advertising requests for the virtual media, and sending the advertising requests to the AD Exchange system, and then receiving the feedback from the AD Exchange system, using a dual-tower model to predict the matching relationship between the virtual media and the virtual advertising budget parties, simulating the bidding results of the AD Exchange system, and using a click-through rate prediction model to predict the click-through rate of the advertisements on the virtual media, it can be combined with the existing AD Exchange system, utilize the infrastructure of the existing AD Exchange system, and realize advertising transaction simulation without the need to additionally build large-scale computing and storage facilities, reducing hardware investment and also reducing the costs of system maintenance and upgrade; compared with other simulation platforms that need to independently establish a complete system, the present application greatly reduces the overall capital investment by parallelly utilizing the existing system; in addition, the present application uses currently mature machine learning algorithm models for bidding result simulation and click-through rate prediction, and these models have been verified in the advertising field, so only a small amount of customization modification is required to adapt to the new system, greatly reducing the technical risks and costs of developing new algorithms from scratch and reducing the technical investment. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] Figure 1 It is a schematic flowchart of a machine learning-based advertising transaction simulation and prediction method provided by an embodiment of the present application;

[0090] Figure 2 It is a schematic diagram of a recurrent neural network structure in an embodiment of the present application;

[0091] Figure 3 It is a schematic diagram of a long short-term memory neural network structure in an embodiment of the present application;

[0092] Figure 4 It is a schematic diagram of a recurrent neural network structure based on LSTM in an embodiment of the present application;

[0093] Figure 5 It is a schematic flowchart of the functions of a simulated advertising transaction system in an embodiment of the present application;

[0094] Figure 6 It is a block diagram of the module architecture of a machine learning-based advertising transaction simulation and prediction device provided by an embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0095] To make the objectives, technical solutions and advantages of this application clearer, the following further elaborates on this application in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely for explaining this application and not for limiting it.

[0096] In the description of this application: Unless otherwise specified, "a plurality of" means two or more. Terms such as "first", "second", "third", etc. in this application are intended to distinguish the objects being referred to and do not have special significance in terms of technical connotations (for example, they should not be understood as emphasizing importance or order, etc.). Expressions such as "including", "comprising", "having", etc. also mean "not limited to" (certain units, components, materials, steps, etc.).

[0097] In one embodiment, as Figure 1 shown, a machine learning-based advertising transaction simulation and prediction method is provided, and this method includes the following steps:

[0098] S1. Create virtual media with specific characteristics and user behavior patterns.

[0099] Specifically, step S1 includes:

[0100] Collect historical media advertisement data;

[0101] Determine the key metrics in the historical media advertisement data that affect the performance of virtual media;

[0102] Conduct exploratory data analysis on the key metrics, and discover the specific characteristics and user behavior patterns of different types of media based on the key metrics;

[0103] Use the key metrics and the specific characteristics and user behavior patterns of the corresponding different types of media to train a first machine learning model to obtain a virtual media model;

[0104] Generate virtual media with specific characteristics and user behavior patterns using the virtual media model.

[0105] In other words, the first key step of the method is the creation of virtual media. By analyzing historical advertisement data, including user behavior, advertisement effects, media characteristics, etc., the system can generate a batch of virtual media with specific characteristics and behavior patterns. These virtual media represent different types of advertising platforms, such as news websites, social media, or blogs.

[0106] The goal of virtual media modeling is to create a virtual media model that can accurately reflect the attributes of different types of media and user behavior.

[0107] First, data collection is required, and the collection content includes:

[0108] Traffic statistics: including page views, number of visits, sources of visits, etc.;

[0109] User population characteristics: age, gender, geographical location, hobbies, etc.;

[0110] Historical advertising performance: ad click-through rate, conversion rate, advertising revenue, etc.

[0111] Then define the key metrics, and determine the key metrics that affect the performance of virtual media, such as user engagement, content types, audience coverage, etc.

[0112] Finally, perform data processing, which can be carried out in the large database of the AD Exchange itself. On this basis, exploratory data analysis is carried out, which includes: performing statistical analysis and visualization to discover trends and patterns in the data; using machine learning algorithms to create representative models of the media, and these models can reflect the characteristics of different types of virtual media.

[0113] S2. Create a virtual advertising budget party with specific behavior patterns.

[0114] Specifically, step S2 includes:

[0115] Collect historical advertising placement data;

[0116] Analyze the historical advertising placement data, determine the behavior patterns of different advertising budget parties, and identify the behavior trends of advertising budget parties under different market conditions;

[0117] Use the historical advertising placement data and the corresponding behavior patterns of different advertising budget parties to train a second machine learning model to obtain a virtual advertising budget party model;

[0118] Use the virtual advertising budget party model to generate a virtual advertising budget party with specific behavior patterns.

[0119] In other words, the second key step is to establish the virtual advertising budget party (DSP). These virtual DSPs simulate the budget allocation and advertising bidding strategies of real advertisers based on historical data and market trends. The design of this part enables the system to generate various types and scales of advertising budget parties, thus increasing the diversity and authenticity of the simulation environment.

[0120] The main goal of virtual budget party modeling is to create a model that can accurately reflect the behavior of real advertising budget parties, which includes budget allocation and bidding strategies. It is established through the following steps:

[0121] (1) Data analysis

[0122] Collect historical advertising data: This includes the size of the advertising budget, details of the bidding strategy, the effectiveness of the advertising placement, etc. This data can be obtained from historical advertising records, advertising management systems, or market research reports.

[0123] In-depth analysis: Conduct an in-depth analysis of this data to understand the behavior patterns of different advertising budget parties. The key points of the analysis include the preferences of the budget parties for different types of advertisements, changes in bidding behavior, budget allocation strategies, etc.

[0124] Trend identification: Identify the behavior trends of advertising budget parties under different market conditions, such as the bidding strategy in a highly competitive market environment.

[0125] (2) Strategy simulation

[0126] Simulate different behaviors: Based on the above analysis, simulate the behaviors of different types of advertising budget parties. This includes bidding strategies under different market conditions and dynamic adjustment of the budget.

[0127] Scenario analysis: Construct different market and advertising scenarios to evaluate the effects of various bidding strategies and budget allocation methods in different situations.

[0128] Establish behavior patterns: Create behavior patterns representing different types of budget parties, such as price-sensitive and brand-oriented.

[0129] (3) Model application

[0130] Select a suitable machine learning model: According to the characteristics and requirements of the data, select a suitable machine learning model. In this case, a regression analysis plus decision tree model is adopted.

[0131] Training and validation: Use historical data to train the model, optimize and validate the model to ensure its accuracy and reliability.

[0132] Prediction and application: Apply these models to predict bidding behavior and budget allocation under different conditions, providing insights into the strategies that budget parties may adopt.

[0133] S3, generate an advertising request for virtual media and send the advertising request to the AD Exchange system; the AD Exchange system selects a matching advertisement for display according to the received advertising request and in combination with the specific behavior pattern of the virtual advertising budget party.

[0134] Furthermore, generating an advertising request for virtual media and sending the advertising request to the AD Exchange system in step S3 includes:

[0135] Simulate the behavior of virtual users on virtual media according to the user behavior patterns of virtual media, and simulate the interaction between virtual users and website content according to the predefined behavior patterns of virtual users;

[0136] When the simulated user behavior triggers the display of an advertisement, generate an advertisement request, which includes the interaction information between the virtual user and the website content;

[0137] Extract and encapsulate the key features of the advertisement request, and send the encapsulated advertisement request to the AD Exchange system.

[0138] This part is the simulation request step. After the virtual media and the virtual advertising budget party are established, the system can simulate the advertisement request process of the virtual media. This includes generating advertisement requests, simulating the interaction between users and the media, and sending advertisement requests to the AD Exchange system according to the characteristics of the virtual media and the user behavior patterns. The specific steps are as follows:

[0139] (1) Generate advertisement requests

[0140] Simulate the activities of users on the media according to the user behavior patterns of the virtual media (such as access frequency, page stay time, content preference, etc.). This may involve simulating behaviors such as users browsing web pages and clicking on content.

[0141] Whenever the simulated user behavior triggers the display of an advertisement, the system generates an advertisement request. This request contains information related to user behavior and media characteristics, such as user demographic data, types of pages browsed, historical behavior data, etc.

[0142] (2) Characteristics and formatting of advertisement requests

[0143] Extract and encapsulate the key features of the advertisement request, such as user geographical location, device type, category of browsed content, etc. Format these features into a format that the AD Exchange system can understand. This usually follows specific protocols or standards to ensure the correct transmission of information.

[0144] (3) Simulate the interaction between users and the media

[0145] Simulate the interaction with website content according to the predefined behavior patterns of virtual users, such as reading articles, watching videos, clicking on web pages, etc.; then feedback these interactions into the advertisement request generation process to ensure that the advertisement request can accurately reflect the current user behavior and interests.

[0146] (4) Send advertisement requests to AD Exchange

[0147] Advertising requests are sent to the AD Exchange system in a simulated real - time manner. This requires considering the limitations of network latency and system processing capabilities. Meanwhile, in the AD Exchange system, these requests are used to match suitable advertising content and budget providers to determine which advertisements will be shown to users.

[0148] (5) Processing and response of advertising requests

[0149] Based on the received requests, the AD Exchange system selects suitable advertisements for display by combining the advertising inventory and bidding strategies of the virtual DSP. The system simulates showing the selected advertisements to virtual users and records the corresponding user reactions, such as clicking on the advertisement and other behaviors.

[0150] S4. Use the two - tower model to predict the matching relationship between virtual media and virtual advertising budget providers.

[0151] Furthermore, step S4 includes:

[0152] Establish the architecture of the two - tower model. The two - tower model includes a budget - provider tower model and a media tower model. The budget - provider tower model is used to learn the characteristics of advertising budget providers, and the media tower model is used to learn the characteristics of media.

[0153] Obtain the pre - collected advertising budget - provider data and media data, and perform cleaning and normalization processing on the advertising budget - provider data and media data.

[0154] Extract useful features from the advertising budget - provider data and media data respectively, and generate negative samples based on the advertising budget - provider data and media data using negative sampling technology.

[0155] Use the advertising budget - provider data and media data after feature extraction and the negative samples to train the budget - provider tower model and the media tower model simultaneously, so that the budget - provider tower model and the media tower model respectively learn the high - dimensional representations of advertising budget providers and media.

[0156] Adopt the contrastive loss method to optimize the budget - provider tower model and the media tower model.

[0157] Input the data of the virtual advertising budget provider into the budget - provider tower model, and input the data of the virtual media into the media tower model.

[0158] Calculate the similarity between the output vector of the budget - provider tower model and the output vector of the media tower model.

[0159] According to the similarity score, match each virtual advertising budget provider with the virtual media with the highest similarity score.

[0160] That is to say, predicting the matching relationship between the media and the budget side is to use an optimization algorithm to find the best matching combination. The algorithm model adopted in this application is the two-tower model in deep learning. The specific steps are as follows:

[0161] (1) Establish the two-tower model architecture

[0162] User tower (budget side tower): A neural network "tower" is specifically used to learn the characteristics of the advertising budget side (DSP). This tower processes data related to the budget side, such as historical bidding behavior, advertising budget scale, preferred advertising types, etc.

[0163] Item tower (media tower): Another "tower" is used to learn the characteristics of the media. This includes information such as the media's traffic data, user population characteristics, content types, etc.

[0164] (2) Feature extraction and representation learning

[0165] Data preprocessing: Clean and normalize the data of the media and the budget side.

[0166] Feature engineering: Extract useful features from the original data, such as using an embedding layer to convert categorical data into dense vectors.

[0167] Deep learning network: Design a deep neural network to learn complex non-linear feature representations.

[0168] (3) Train the two-tower model

[0169] Negative sampling: To improve the efficiency and effect of training, use negative sampling technology to select negative samples.

[0170] Joint training: Train the two towers simultaneously so that they can respectively learn the high-dimensional representations of the budget side and the media.

[0171] Optimization objective: Adopt the method of contrastive loss to optimize the model. The goal is to make the matching media and budget side close in the feature space and far away from the unmatched ones.

[0172] (4) Prediction and matching

[0173] Similarity calculation: After the model training is completed, evaluate the matching degree between them by calculating the similarity between the output vectors of the budget side tower and the media tower. Commonly used similarity measurement methods include dot product, cosine similarity, etc.

[0174] Best match selection: Select the most matching media for each budget side according to the similarity score.

[0175] S5. Simulate the bidding results of the AD Exchange system.

[0176] Further, step S5 includes:

[0177] Set up a virtual AD Exchange environment, where the AD Exchange environment includes the characteristics of the ad slots available for bidding, a list of virtual advertising budget parties, and bidding rules;

[0178] Collect data required for bidding, and use the collected data to train a third machine learning model to obtain a bid prediction model;

[0179] When the ad slot is available, simulate bid requests from different virtual advertising budget parties, and based on the characteristics of the ad slots available for bidding and the target audiences of the virtual advertising budget parties, match the appropriate bid requests of the virtual advertising budget parties for the ad slots available for bidding;

[0180] Use the bid prediction model to predict the bids of each matched virtual advertising budget party for the ad slot available for bidding;

[0181] According to the bidding rules, determine the winner among the matched virtual advertising budget parties to obtain the bid result of the winner.

[0182] This part is the simulated bidding step. Immediately following is the bidding process of the simulated AD Exchange system. The system uses a machine learning model to simulate the Real-Time Bidding (RTB) process. This process involves complex algorithms and strategies, aiming to perform dynamic price calculations based on the bids and budgets of virtual DSPs, as well as the characteristics of the ad slots of virtual media. The steps of the simulated bidding are as follows:

[0183] (1) Define the bidding environment

[0184] First, set up a virtual AD Exchange environment, including the characteristics of the ad slots available for bidding (such as location, size, page type, etc.), the list of available virtual DSPs and their budget limits; second, define the bidding rules, such as the highest bidder wins, second-price auction, etc.

[0185] (2) Collect data required for bidding

[0186] Collect data on each ad slot, such as traffic statistics, user population characteristics, page content, etc. At the same time, analyze the historical bidding strategies and budget allocations of virtual DSPs, and use the collected data to train a bid prediction model to predict their possible bidding behaviors.

[0187] (3) Simulate bid requests

[0188] When the ad space is available, the system simulates and generates bid requests from different virtual DSPs, and matches appropriate DSP bid requests according to the characteristics of the ad space and the target audience of the DSP.

[0189] (4) Predict matching relationships

[0190] Build a prediction model using historical data to estimate the click-through rate of a specific ad on a specific medium. Integrate the key features affecting the click-through rate, including the ad's position, content, target user profile (such as age, gender, interests, etc.), page context, and the user's historical behavior data.

[0191] (5) Use a machine learning model for bidding (generate a bidding strategy)

[0192] Use a machine learning model to predict the bids of each DSP for a specific ad space. These models may be based on historical data, taking into account the characteristics of the ad space and the bidding behavior of the DSP, and use the prediction model to dynamically determine the price of the ad space. This may involve complex algorithms such as dynamic pricing strategies based on user behavior and market demand.

[0193] (6) Predict the final bid

[0194] According to the bidding rules, compare the bids of different DSPs and determine the highest bidder. Then conduct the winner selection: select the DSP with the highest bid to obtain the ad space, or determine the winner according to other rules (such as the second-highest bidder wins).

[0195] S6, return the bidding result to the virtual media, and use the click-through rate prediction model to predict the click-through rate of the ad on the virtual media.

[0196] This part is the step of predicting the ad effect. Finally, the system simulates the final bidding result and returns it to the virtual media, and estimates the click-through rate of the media. This process takes into account multiple factors such as the ad's position, content, user profile, etc. to predict the ad effect, such as click-through rate and user engagement. The system helps users understand the potential effects of different ad strategies through this simulation and provides guidance for actual ad placement. Here, a click-through rate prediction model is used.

[0197] Specifically, the click-through rate prediction model is a recurrent neural network based on LSTM to predict the click-through rate of the media for a certain type of ad.

[0198] The present invention uses a recurrent neural network to train the click-through rate prediction model based on historical data. The following is an introduction to the improved RNN model based on LSTM developed by the present invention in the key item of "predicting the click-through rate of the media for a certain type of ad".

[0199] The click-through rate prediction model includes an input layer, a hidden layer, and an output layer. The hidden layer includes an input gate, a forget gate, a cell unit, and an output gate. The model has three hidden layers, each with 256 nodes. The model is used to simulate the click behavior of users and estimate the click-through rate of advertisements. The following are the model definition, the model training process, and the evaluation function.

[0200] (I) Model Definition

[0201] The input gate determines whether to allow the signal from the input layer to enter the hidden layer node. When the gate is open, the output signal from the input layer is allowed to enter; when the gate is closed, the signal is rejected. The output gate determines whether the output value of the current node is output to the next layer. When the gate is open, the signal from the hidden layer node is allowed to be output; when the gate is closed, the signal is rejected. The forget gate determines whether to retain the historical information stored in the current hidden layer node. When the gate is open, the historical information of the hidden layer node is retained; when the gate is closed, the historical information of the hidden layer node is not retained.

[0202] S represents the information value stored at time t. The input and output layers of the model are the same as those of the RNN model, as Figure 2 shown. Where w represents the weight between the hidden layer and the input gate unit. In the next part of this article, different representations of w represent the weights of different nodes.

[0203] The long short-term memory structure is as Figure 3 shown, and its hidden layer nodes are replaced. Therefore, the structure of the recurrent neural network model based on LSTM is as Figure 4 shown.

[0204] (II) Model Training

[0205] Different from the two value sources of the general recurrent neural network as input, the input of the input gate consists of three values (three input vectors), and the three input vectors are the output vector of the layer node, the cell output vector of the first hidden layer the information retained by the unit at the previous time point The input vector of the input gate at time t is as follows:

[0206]

[0207] In the formula, w represents the corresponding weight;

[0208] Through the activation function f of the input gate, the output vector at time t is as follows:

[0209]

[0210] The input of the forget gate also consists of three input vectors, which are from the same sources as those of the input gate. The input vector of the forget gate at time t is as follows:

[0211]

[0212] In the formula, w represents the corresponding weights;

[0213] The formula for the output vector obtained at time t through the activation function f of the forget gate is as follows:

[0214]

[0215] It can be seen from the cell unit in the long short-term memory structure diagram that its input consists of two parts. One part is the input vector of the input layer and the other part is the output of the first hidden layer of the output gate The input vector of the cell unit at time t is as follows:

[0216]

[0217] In the formula, w represents the corresponding weights;

[0218] Determine whether to retain the past information value according to the forget gate The formula is as follows:

[0219]

[0220] The input of the output gate consists of three parts, including the output vector of the input layer, the hidden output vector of the cell layer, and the information retained by the cell. The input vector of the output gate at time t is as follows:

[0221]

[0222] In the formula, w represents the corresponding weights;

[0223] By using the activation function f of the output gate unit at time t, the output vector obtained at time t is:

[0224]

[0225] The output vector of the cell unit is as follows:

[0226]

[0227] The output vector of the cell unit That is, the output vector of the hidden layer, as the input vector of the output layer, is as follows:

[0228]

[0229] The result vector output by the output layer is as follows:

[0230]

[0231] The weight w between node i and node j ij is updated to:

[0232]

[0233] where η represents the learning step size, is expressed as follows:

[0234]

[0235] where represents the residual node of j, represents the output vector node of j. Therefore, formula (12) can be simplified using formula (13), and w ij can be simplified to:

[0236]

[0237] (III) Loss function and evaluation function

[0238] A criterion is needed to judge whether the model is good or bad. The area under the ROC curve (AUC) is a common criterion. However, AUC focuses on the ranking of CTR estimation, while log loss focuses on the accuracy of CTR estimation. When the click-through rate rises by a certain percentage, it does not cause a change in AUC. However, it will cause a change in log loss. Log loss is a reflection of the difference between the click-through rate estimated by the model and the true click-through rate. The smaller the log loss, the more accurate the estimated result of the advertisement CTR.

[0239] This solution uses log loss in scikit-learn, and the log loss function is defined as follows:

[0240]

[0241] where y i represents the true value of the i-th click, and p i represents the value of the i-th click estimated by the model.

[0242] Exemplarily, the training process of the click-through rate prediction model may include:

[0243] (1) Data collection: First, a large amount of advertising data needs to be collected, including the characteristics of the advertisements (such as advertisement text, images, videos, etc.) and the corresponding click-through rates. This data can be obtained from historical advertising data or collected through real-time monitoring.

[0244] (2) Data preprocessing: Preprocess the collected advertising data, including data cleaning, feature selection, handling missing values, data normalization, etc. This step aims to convert the original data into a format suitable for model input.

[0245] (3) Feature engineering: According to the characteristics of the advertisements and the prediction target, perform feature engineering. For example, text features, image features, time features, etc. can be extracted, and operations such as feature encoding and dimensionality reduction can be carried out to enhance the expressive ability of the model.

[0246] (4) Model training: Divide the preprocessed dataset into a training set and a validation set, and use the training set to train the model. During the training process, continuously adjust the model parameters through the backpropagation algorithm to minimize the error between the predicted value and the true click-through rate.

[0247] (5) Model evaluation: Use the validation set to evaluate the trained model, calculate the performance metrics of the model, such as the logarithmic loss function, to judge the accuracy and generalization ability of the model.

[0248] (6) Model optimization: If the model performs poorly, try to adjust the model architecture, adjust the hyperparameters, or add more feature engineering steps to improve the model performance.

[0249] (7) Model application: When the model reaches the expected performance, it can be deployed to actual applications to predict the click-through rate of a certain type of advertisement by the media.

[0250] In the method provided by the embodiments of this application, a new system architecture is proposed for constructing a system that simulates an advertising transaction, and this system works in parallel with the actually operating AD Exchange system. The core function of this system is to use historical data and advanced machine learning models to construct a series of virtual media and virtual advertising budget parties (Demand-Side Platforms, DSP). The design of this system aims to comprehensively simulate all aspects of digital advertising transactions, provide an almost real transaction environment, thereby helping users deeply understand the dynamics of advertising transactions and test different advertising strategies. The functional process design of the system is as Figure 5 shown.

[0251] In this system, the key points are: virtual media modeling, virtual budget party modeling, and using machine learning models to predict the matching relationship between media and budget parties, and predict the click-through rate of a certain type of advertisement by the media.

[0252] Overall, this system provides a comprehensive framework for simulating and analyzing various aspects of digital advertising transactions. It combines data analysis, machine learning, and simulation techniques to provide in-depth insights for advertisers, publishers, and market analysts, helping them optimize advertising strategies and improve return on investment. Through this highly realistic simulation environment, users can safely test different advertising models and strategies without taking risks in the actual market.

[0253] In view of the current situation, the method provided by the embodiments of this application has two purposes: one is to be able to process historical data and online real-time data simultaneously, and the other is to significantly reduce the input of technology and computing resources; it aims to make up for the above-mentioned deficiencies and expand its application scope.

[0254] First of all, this application integrates the ability to process historical data and real-time data. Research-oriented simulation systems usually rely only on historical data, which limits their ability to predict future trends and adapt to market changes. The model of this application can analyze user behavior and market dynamics in real time by integrating advanced data stream processing technologies, providing more accurate and timely predictions. This real-time data processing ability enables the model not only to learn past trends from historical data, but also to adapt to the real-time changes in the current market, thus providing more dynamic and accurate insights.

[0255] Secondly, the present invention focuses on reducing the input of technology and computing resources in its design. By optimizing algorithms and data processing processes, the model of the present invention reduces the dependence on high-performance computing resources.

[0256] In addition, the present invention adopts a series of techniques for simplifying the model, such as feature selection and model compression, which not only improve the operation efficiency, but also reduce the demand for storage and computing resources. These optimizations make this method applicable not only to large enterprises with rich resources, but also to small and medium-sized enterprises with limited resources.

[0257] In the method provided by the embodiments of this application, the core advantage is that it can be combined with the existing AD Exchange system to realize its functions. This method avoids the need for a large amount of repetitive computing power resources and storage resources. At the same time, it has significant cost-effective advantages compared with similar advertising transaction simulation systems, which are mainly reflected in the following aspects:

[0258] 1. Optimization of resource utilization

[0259] Efficient resource utilization: By utilizing the infrastructure of the existing AD Exchange system, the present invention can realize its functions without the need to build large-scale computing and storage facilities additionally. This not only reduces hardware investment, but also lowers the costs of system maintenance and upgrade.

[0260] Reduced duplicate investment: By avoiding the repeated construction of similar system architectures, the present invention significantly reduces redundant investment in computing power and storage, enabling resources to be utilized more efficiently.

[0261] 2. Cost-effectiveness

[0262] Savings in capital costs: Compared with other simulation platforms that require the independent establishment of a complete system, the present invention significantly reduces the overall capital investment, especially the expenditure on hardware and infrastructure, by leveraging the existing system in parallel.

[0263] Optimization of operating costs: Relying on the existing system reduces the additional operation and maintenance requirements, thereby achieving savings in operating costs.

[0264] 3. Technology investment and innovation

[0265] Application of mature algorithms: The present invention employs currently mature machine learning algorithm models. These models have been verified in the advertising field, so only a small amount of customization is required to adapt to the new system, greatly reducing the technical risks and costs of developing new algorithms from scratch.

[0266] Rapid deployment and iteration: Due to the maturity of the algorithm models, the present invention can be rapidly deployed and iterated, accelerating the time to market of the product and reducing the technical costs during the R & D cycle.

[0267] It should be understood that although Figure 1 the steps in the flowchart are shown sequentially in the direction of the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise clearly stated in this article, there is no strict order restriction for the execution of these steps, and these steps can be executed in other orders. Moreover, Figure 1 at least a part of the steps in

[0268] In one embodiment, as Figure 6 shown, a machine learning-based advertising transaction simulation and prediction device is provided, including the following program modules:

[0269] A virtual media creation module 601 for creating virtual media with specific characteristics and user behavior patterns;

[0270] A virtual advertising budget party creation module 602 for creating virtual advertising budget parties with specific behavior patterns;

[0271] An advertisement request generation module 603, configured to generate advertisement requests for virtual media and send the advertisement requests to the ADExchange system; the AD Exchange system selects a matching advertisement for display according to the received advertisement requests in combination with the specific behavior patterns of virtual advertisement budget parties;

[0272] A matching relationship prediction module 604, configured to use a dual tower model to predict the matching relationship between virtual media and virtual advertisement budget parties;

[0273] A bid result simulation module 605, configured to simulate the bid results of the AD Exchange system;

[0274] A click-through rate prediction module 606, configured to return the bid results to the virtual media and use a click-through rate prediction model to predict the click-through rate of advertisements on the virtual media.

[0275] The specific implementation content of each module can be referred to the definition of a machine learning-based advertisement transaction simulation and prediction method in the above text, which will not be elaborated here.

[0276] In one embodiment, a computer-readable storage medium is further provided, on which a computer program is stored, which relates to all or part of the processes in the method of the above embodiment.

[0277] The technical features of the above embodiments can be combined arbitrarily. For the sake of concise description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope recorded in this specification.

Claims

1. An advertising transaction simulation and prediction method based on machine learning, characterized in that, Including: S1, creating virtual media with specific characteristics and user behavior patterns; S2, creating virtual advertising budget parties with specific behavior patterns; S3, generating advertising requests for the virtual media and sending the advertising requests to the AD Exchange system; the AD Exchange system selects a matching advertisement for display according to the received advertising requests and in combination with the specific behavior patterns of the virtual advertising budget parties; S4, predicting the matching relationship between the virtual media and the virtual advertising budget parties by using a two-tower model; S5, simulating the bidding results of the AD Exchange system; S6, returning the bidding results to the virtual media and predicting the click-through rate of the advertisements on the virtual media by using a click-through rate prediction model; Step S5 includes: Setting up a virtual AD Exchange environment, which includes the characteristics of the available ad slots for bidding, a list of virtual advertising budget parties, and bidding rules; Collecting the data required for bidding and training a third machine learning model by using the collected data to obtain a bidding prediction model; When the ad slot is available, simulating and generating bidding requests from different virtual advertising budget parties, and matching the bidding requests of the appropriate virtual advertising budget parties to the available ad slots for bidding according to the characteristics of the available ad slots for bidding and the target audience of the virtual advertising budget parties; Using the bidding prediction model to predict the bids of each matched virtual advertising budget party for the available ad slots for bidding; Determining the winner among the matched virtual advertising budget parties according to the bidding rules to obtain the bidding results of the winner; The click-through rate prediction model is a recurrent neural network based on LSTM; Step S1 includes: Collecting historical media advertising data; Determining the key metrics that affect the performance of the virtual media in the historical media advertising data; Performing exploratory data analysis on the key metrics, and discovering the specific characteristics and user behavior patterns of different types of media according to the key metrics; Training a first machine learning model by using the key metrics and the specific characteristics and user behavior patterns of the corresponding different types of media to obtain a virtual media model; Generating virtual media with specific characteristics and user behavior patterns by using the virtual media model.

2. The method for simulating and predicting advertising transactions based on machine learning according to claim 1, characterized in that Step S2 includes: Collecting historical advertising placement data; Analyzing the historical advertising placement data, determining the behavior patterns of different advertising budget parties, and identifying the behavior trends of advertising budget parties under different market conditions; Training a second machine learning model by using the historical advertising placement data and the corresponding behavior patterns of different advertising budget parties to obtain a virtual advertising budget party model; Generating virtual advertising budget parties with specific behavior patterns by using the virtual advertising budget party model.

3. The machine learning-based advertising transaction simulation and prediction method according to claim 1, characterized in that In step S3, the generating of the advertising requests for the virtual media and sending the advertising requests to the AD Exchange system includes: According to the user behavior patterns of the virtual media, simulating the behavior of virtual users on the virtual media, and simulating the interaction between virtual users and website content according to the predefined behavior patterns of the virtual users; When the simulated user behavior triggers the display of an advertisement, an advertisement request is generated, and the advertisement request contains the interaction information between the virtual user and the website content; Extract and encapsulate the key features of the advertisement request, and send the encapsulated advertisement request to the ADExchange system.

4. The machine learning-based advertising transaction simulation and prediction method according to claim 1, characterized in that Step S4 includes: Establish the architecture of a dual-tower model, which includes a budget-side tower model and a media tower model. The budget-side tower model is used to learn the features of the advertisement budget side, and the media tower model is used to learn the features of the media; Obtain advertisement budget-side data and media data, and perform cleaning and normalization processing on the advertisement budget-side data and media data; Extract useful features from the advertisement budget-side data and media data respectively, and generate negative samples based on the advertisement budget-side data and media data using negative sampling techniques; Simultaneously train the budget-side tower model and the media tower model using the advertisement budget-side data and media data after feature extraction and the negative samples, so that the budget-side tower model and the media tower model respectively learn the high-dimensional representations of the advertisement budget side and the media; Adopt a contrastive loss method to optimize the budget-side tower model and the media tower model; Input the data of the virtual advertisement budget side into the budget-side tower model, and input the data of the virtual media into the media tower model; Calculate the similarity between the output vector of the budget-side tower model and the output vector of the media tower model; According to the similarity score, match the virtual advertisement budget side with the virtual media having the highest similarity score for each virtual advertisement budget side.

5. An advertising transaction simulation and prediction device based on machine learning, characterized in that, Includes: A virtual media creation module for creating virtual media with specific features and user behavior patterns; A virtual advertisement budget-side creation module for creating virtual advertisement budget sides with specific behavior patterns; An advertisement request generation module for generating advertisement requests for virtual media and sending the advertisement requests to the ADExchange system; the AD Exchange system selects and displays matching advertisements according to the received advertisement requests and the specific behavior patterns of the virtual advertisement budget sides; A matching relationship prediction module for predicting the matching relationship between the virtual media and the virtual advertisement budget sides using the dual-tower model; A bid result simulation module for simulating the bid results of the AD Exchange system; A click-through rate prediction module for returning the bid results to the virtual media and predicting the click-through rate of the advertisements on the virtual media using a click-through rate prediction model; The simulation of the bid results of the AD Exchange system includes: Set up a virtual AD Exchange environment, which includes the characteristics of the available ad slots for bidding, a list of virtual advertisement budget sides, and bidding rules; Collect the data required for bidding, and use the collected data to train a third machine learning model to obtain a bid prediction model; When the ad slot is available, simulate and generate bid requests from different virtual advertisement budget sides, and match the appropriate bid requests of the virtual advertisement budget sides for the available ad slots according to the characteristics of the available ad slots for bidding and the target audience of the virtual advertisement budget sides; Use a bid prediction model to predict the bids of each matched virtual advertising budget party for the ad slots available for bidding; Determine the winner among the matched virtual advertising budget parties according to the bidding rules to obtain the bid result of the winner; The click-through rate prediction model is a recurrent neural network based on LSTM.

Images