Method and system for automatic bidding of advertisements based on predictive placement planning

Abstract

The application relates to the technical field of advertisement delivery, and in particular to an advertisement automatic bidding method and system based on predictive delivery planning, which comprises the following steps: constructing a strategic delivery scene composed of multi-dimensional advertisement activity metric vectors based on historical advertisement delivery data, and extracting high-quality strategic delivery scenes; encoding the delivery demand of an advertiser into a standardized business intention code; constructing a forward disturbance and conditional reverse recovery process training conditional scene synthesizer by using the high-quality strategic delivery scene and the business intention code; starting from an initial strategic delivery scene in an actual bidding stage, gradually recovering and generating an optimal strategic delivery scene meeting the delivery demand through the conditional scene synthesizer; and constructing and training a non-backtracking bidding command deducer based on the optimal strategic delivery scene to generate a bidding command in real time. The application can meet the long-term delivery demand of an advertiser with multiple targets and multiple constraints, so that a stable, flexible and optimal advertisement bidding strategy is difficult to achieve as a whole.

Description

Technical Field

[0001] This application relates to the field of advertising delivery technology, and in particular to an automatic bidding method and system for advertising based on predictive delivery planning. Background Technology

[0002] With the rapid development of digital marketing, online advertising has become a crucial means for businesses to acquire user traffic and improve conversion rates. Online advertising platforms typically offer advertisers ad placement bidding services. Advertisers can set budgets and strategies based on their marketing goals, and the platform makes specific bidding and display decisions during the advertising campaign. To improve advertising efficiency and reduce manual intervention costs, platforms have gradually introduced automatic bidding technology. This technology uses algorithmic models to calculate and execute optimal bidding instructions in real time, and has become a fundamental capability module in intelligent ad placement systems.

[0003] Currently, online advertising platforms typically employ real-time responsive models to intelligently optimize advertisers' bidding behavior. Specifically, during the advertising campaign, the current state parameters of the advertising account are obtained at fixed time intervals, such as remaining budget, amount spent, real-time return on investment (ROI), and average cost per click (CPC). Based on this current state information, the model calculates the corresponding bid adjustment instructions and executes them in real time. This process is repeated continuously throughout the campaign, forming a bidding mechanism that continuously makes decisions based on the current state. The core feature of this type of model is its real-time response to the state at each moment, and its ability to make locally optimal bidding decisions accordingly.

[0004] However, the decision-making process of instant-response models primarily revolves around the current state, using instant-response logic for progressive bidding control. As the advertising campaign period lengthens and the number of decision points increases, their limitations become increasingly apparent. First, advertising campaigns typically last 24 hours or longer, encompassing hundreds of decision points. The optimal decision often requires anticipating future global conditions. For example, a conservative bidding strategy in the morning might reserve sufficient budget for the more competitive afternoon sessions. Instant-response models lack the ability to plan globally across the entire campaign period, failing to capture these pervasive, non-local dependencies, leading to suboptimal campaign performance. Second, the online advertising environment is highly random; the supply and demand structure of ad impressions, user behavior characteristics, and competitor strategies all exhibit strong fluctuations over time. In this context, even small deviations in the current decision can be amplified in subsequent decision chains, causing significant fluctuations in actual campaign performance. This results in unstable performance of locally optimized strategies in actual campaigns, reducing the robustness of the campaign system. Finally, advertisers typically set long-term campaign goals with multiple objectives and constraints, such as maximizing total gross merchandise volume (GMV) within a limited budget and CPC, ensuring even budget spending, or improving the smoothness of the campaign curve. Since real-time response models cannot plan the path to achieving these goals globally, they require extensive engineering configuration to forcibly break down complex objectives into the bidding logic at each moment. This increases system complexity and reduces the ability to adapt to advertisers' personalized needs.

[0005] In summary, existing real-time responsive models, which make local decisions based only on the current state, lack the ability to plan global bidding paths for the entire campaign cycle. They are also unable to withstand the instability of strategies caused by random fluctuations in the advertising environment, and cannot meet the long-term campaign needs of advertisers with multiple objectives and constraints. Therefore, they are unable to achieve a stable, flexible and optimal advertising bidding strategy as a whole. Summary of the Invention

[0006] This application provides an automatic advertising bidding method and system based on predictive delivery planning, which can meet the long-term delivery needs of advertisers with multiple objectives and constraints, thus making it difficult to achieve a stable, flexible, and optimal advertising bidding strategy overall. This application provides the following technical solution:

[0007] In a first aspect, this application provides an automatic advertising bidding method based on predictive delivery planning, the method comprising:

[0008] Construct a strategic placement landscape based on historical advertising data, consisting of multi-dimensional advertising campaign metric vectors, and extract high-quality strategic placement landscapes;

[0009] Encode advertisers' campaign needs into standardized business intent codes;

[0010] The extracted high-quality strategic placement landscape and the advertiser's business intent encoding are used to construct a forward perturbation and conditional inverse recovery process to train a preset conditional landscape synthesizer. In the actual bidding stage, starting from a random initial strategic placement landscape, the conditional landscape synthesizer is used to gradually recover and generate the optimal strategic placement landscape that meets the placement requirements.

[0011] Based on the optimal strategic deployment scenario, a non-backtracking bid command inferrer is constructed and trained. The non-backtracking bid command inferrer generates bid commands in real time by combining historical trends and target status.

[0012] In one specific implementation scheme, the construction of a strategic placement landscape composed of multi-dimensional advertising campaign metric vectors based on historical advertising placement data includes:

[0013] Collect historical advertising data, and for each historical advertising data point, discretize it on the time axis using a complete advertising cycle as the time unit. A series of observation nodes, at each observation node Using a multidimensional advertising campaign metric vector This represents the operational status of the ad placement at that moment;

[0014] Collected in sequence The advertising activity metric vectors at each observation node are used to construct a data matrix representing the dynamic evolution of the entire cycle, denoted as the strategic placement landscape. :

[0015] ;

[0016] in, This represents a strategic advertising campaign overview spanning a complete advertising campaign cycle. Indicates the first The advertising activity metric vector recorded at each observation node; This represents the total number of observation nodes obtained after discretizing the advertising campaign period. This indicates the number of feature dimensions contained in each advertising campaign's metric vector. Represent a OK A column of real numbers.

[0017] In one specific implementation, encoding the advertiser's campaign needs into a standardized business intent code includes:

[0018] Advertisers' campaign needs are uniformly coded into standardized business intent codes, denoted as... ,in Indicates the number of dimensions in the encoding. Represents the set of real numbers;

[0019] Business Intent Coding It includes core optimization intent, boundary constraints, and process form preferences; where core optimization intent expresses the advertiser's main goal, boundary constraints describe the constraints that must be met, and process form preferences reflect the advertiser's preference for the form of the campaign.

[0020] In one specific implementation scheme, the step of using the extracted high-quality strategic placement landscape and the advertiser's business intent encoding to construct a forward perturbation and conditional inverse recovery process to train a preset conditional landscape synthesizer includes:

[0021] During the forward perturbation process, the extracted high-quality strategic deployment scenario is presented. Gaussian perturbations of preset strength are gradually injected to obtain image representations of different perturbation levels. The execution of the perturbation is based on the set maximum number of perturbation steps. At every step The current disturbance pattern The disturbance pattern from the previous moment can be used to determine the nature of the scene. The perturbation pattern at any given time can be generated in one step using a closed-form function, and its mathematical form is as follows:

[0022] ;

[0023] in, It is a standard Gaussian random matrix. It is preset, random Varying signal-to-noise ratio control parameters Let be the perturbation function;

[0024] In the conditional inverse recovery process, a deep neural network is constructed. Neural Networks The input includes a picture under the current level of disturbance. Current perturbation time step and business intent coding Its task is to predict the original perturbation injected in this step given a triple input. Neural Networks The optimization objective of training is to minimize the following expected error. :

[0025] ;

[0026] in, This represents the expected value for all samples in the training data. It is real Gaussian noise injected during the forward perturbation process. This represents the model's predicted value for the noise. This represents the Frobenius norm of the matrix.

[0027] In one specific implementation scheme, the step of starting from a random initial strategic deployment scenario during the actual bidding phase and gradually recovering and generating an optimal strategic deployment scenario that meets the deployment requirements through the conditional scenario synthesizer includes:

[0028] In the actual bidding phase, starting from a completely random initial strategic deployment scenario... Starting from a given advertiser's business intent encoding Guided by [the program], the already trained deep neural network is invoked. ,by As input, execute sequentially. Reverse recovery operation, that is, at each time step Deep neural networks Receive current disturbance pattern Current perturbation time step and business intent coding By inferring backwards, the original perturbation in the current scene is estimated. And calculate the previous step's picture based on the estimation results. This process iterates continuously, gradually making the image clearer and more stable from the initial Gaussian noise, at each time step. At that time, generate the optimal strategic deployment scenario. .

[0029] In one specific implementation scheme, the step of constructing and training a non-backtracking bid order inferring based on the optimal strategic deployment scenario, wherein the non-backtracking bid order inferring generates bid orders in real time by combining historical trends and target states, includes:

[0030] Construct a non-backtracking bid order derivator. Its task is to derive, in real time, the optimal bid order that drives the system toward the target state, by combining the current and recent historical observations with the target state at the next moment, during system operation. ;

[0031] The design goal of this derivator is to answer the following question: Given observed historical metric sequences... In order to make the next state of the system as close as possible to the planned target state What bid order should be executed? ;

[0032] Model the derivator as a function Its mathematical form is as follows:

[0033] ;

[0034] in, For the current moment The bid order to be executed. This indicates that the current time is the endpoint and the length is... Historical measurement observation windows, in which each Indicates the system at time 10:00 The operational metric status; Indicates the first [item] in the current scene The target state of each observation node;

[0035] The non-backtracking bid order derivative is invoked in real time at each observation node to collect actual operational metrics over the past H time steps. And in combination with the current picture, the first Target state of each observation node This input is fed into the derivator, which performs feature extraction and deep modeling on the input, and outputs the bid command that should be executed at the current moment. .

[0036] In a specific feasible implementation, the non-backtracking bid order inferrer is trained using a standard supervised learning method, specifically as follows: A large-scale training sample pair is constructed from historical advertising delivery data. The input for each sample It is a combination of historical metric windows and the actual next metric vector reached, that is, the historical observation sequence at a certain moment and the preset next target state in the picture, label. The bid order actually executed within that history window. ;

[0037] The training objective of the non-backtracking bid order inferrer is to minimize the distance between the predicted order and the true order, using the L2 loss function. as follows:

[0038] ;

[0039] in, Represents the expectation over all training sample pairs. This represents the bid order that was actually executed in history. The model represents the input The predicted bid order.

[0040] Secondly, this application provides an automatic advertising bidding system based on predictive delivery planning, employing the following technical solution:

[0041] An automatic advertising bidding system based on predictive delivery planning includes:

[0042] The high-quality strategic placement image extraction module is used to construct a strategic placement image composed of multi-dimensional advertising campaign metric vectors based on historical advertising placement data, and to extract high-quality strategic placement images.

[0043] The ad placement demand coding module is used to encode advertisers' ad placement demands into standardized business intent codes;

[0044] The optimal scenario generation module is used to construct a forward perturbation and conditional inverse recovery process by utilizing the extracted high-quality strategic placement scenario and the advertiser's business intent encoding to train a preset conditional scenario synthesizer. In the actual bidding stage, starting from a random initial strategic placement scenario, the optimal strategic placement scenario that meets the placement requirements is gradually recovered and generated through the conditional scenario synthesizer.

[0045] The bid command generation module is used to construct and train a non-backtracking bid command inferrer based on the optimal strategic deployment scenario. The non-backtracking bid command inferrer generates bid commands in real time by combining historical trends and target status.

[0046] Thirdly, this application provides an electronic device, the device including a processor and a memory; the memory stores a program, the program being loaded and executed by the processor to implement an automatic advertising bidding method based on predictive delivery planning as described in the first aspect.

[0047] Fourthly, this application provides a computer-readable storage medium storing a program that, when executed by a processor, is used to implement an automatic advertising bidding method based on predictive delivery planning as described in the first aspect.

[0048] In summary, the beneficial effects of this application include at least the following:

[0049] (1) By constructing an optimal strategic placement scenario covering the entire placement cycle, a global long-term planning of advertising bidding strategies is achieved. Unlike traditional real-time response models that only make local adjustments based on the current state, the strategic placement scenario generated in this application not only reflects the structural characteristics of historical high-quality placement strategies but also integrates the advertiser's current business intentions, realizing comprehensive planning and guidance for future placement paths. This global planning capability enables the system to proactively plan future placement nodes, effectively avoiding frequent strategy adjustments caused by short-term fluctuations, thereby improving the scientific nature and predictability of the overall placement strategy and ensuring the steady achievement of advertising placement goals.

[0050] (2) By generating the globally optimal delivery scenario in one go, this application significantly reduces the problem of error accumulation caused by stepwise decision-making. Combined with a non-backtracking bidding command derivator, the system can derive reasonable bidding adjustment commands in real time and continuously based on the current and historical observation states, achieving dynamic approximation of the global planning target. This mechanism effectively suppresses the interference of random fluctuations in the advertising environment on the bidding strategy, improves the stability and robustness of the model, makes advertising bidding decisions more reliable, avoids frequent and drastic fluctuations, and ensures continuous optimization of delivery performance.

[0051] (3) By introducing commercial intent coding, advertisers' diverse campaign objectives (such as maximizing returns, controlling costs, and managing campaign pace) are uniformly expressed as standardized instruction inputs, achieving effective integration and management of multiple objectives and constraints. Based on this coding, the conditional scenario synthesizer flexibly generates strategic campaign scenarios that meet specific campaign needs, supporting adaptation and response to different business requirements and constraints. This unified framework not only enhances the model's applicability and scalability but also ensures the coordination and balance of the generated campaign strategies under multiple objectives and constraints, meeting the complex and ever-changing actual needs of advertisers.

[0052] By extracting historical high-performing strategic placement scenarios and combining them with advertisers' personalized commercial intent encoding, a conditional scenario synthesizer is constructed and trained to generate an optimal strategic placement scenario covering the entire future placement cycle, serving as the system's global operational plan. This scenario takes into account the structural characteristics of historically high-performing strategies and current placement goals, enabling long-term planning and goal guidance for the placement path. Subsequently, a non-backtracking bidding command derivator is designed to derive the optimal bidding command driving the system towards the global planning goal in real time, combining the current and historical observation states with the target states in the strategic placement scenario. This ensures that the bidding strategy possesses both dynamic adaptability and stably approximates the long-term planning trajectory. Through this method, this application effectively overcomes the short-sightedness and volatility sensitivity of immediate response models, achieving stability, flexibility, and performance optimization of advertising bidding strategies. It meets the complex placement needs of multiple objectives and constraints, improving overall advertising placement effectiveness and resource utilization efficiency.

[0053] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, the preferred embodiments of this application are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0054] Figure 1 This is a flowchart illustrating the automatic bidding method for advertising based on predictive delivery planning in the embodiments of this application.

[0055] Figure 2This is a structural block diagram of the automatic bidding system for advertising based on predictive delivery planning in the embodiments of this application.

[0056] Figure 3 This is a block diagram of an electronic device for automatic advertising bidding based on predictive delivery planning, as described in this application. Detailed Implementation

[0057] The specific embodiments of this application will be described in further detail below with reference to the accompanying drawings and examples. The following examples are used to illustrate this application, but are not intended to limit the scope of this application.

[0058] Optionally, this application uses the automatic bidding method for advertising based on predictive delivery planning provided in various embodiments as an example for application in an electronic device. The electronic device is a terminal or a server. The terminal can be a computer, tablet computer, etc. This embodiment does not limit the type of electronic device.

[0059] First, referring to Table 1, let's introduce the technical terms used in this application and their English expressions and abbreviations.

[0060]

[0061] Reference Figure 1 This is a flowchart illustrating an embodiment of an ad bidding method based on predictive delivery planning provided in this application. The method includes at least the following steps:

[0062] Step S101: Construct a strategic placement landscape composed of multi-dimensional advertising campaign metric vectors based on historical advertising placement data, and extract high-quality strategic placement landscapes.

[0063] In step S101, historical advertising data is first collected. The historical advertising data includes advertising campaign-related data recorded at different time points in multiple completed advertising campaign cycles. This data covers the time progress information during the campaign, budget usage, resource consumption speed, conversion revenue performance, unit cost fluctuations, and other business indicators closely related to the health of the campaign.

[0064] Specifically, after collecting historical advertising data, each historical advertising data point is discretized on a time axis using a complete advertising cycle (e.g., one day) as the time unit. A series of continuous observation nodes. At each observation node... Using a multidimensional advertising campaign metric vector This is used to characterize the operational status of ad delivery at a given moment. This metric vector may typically include the following dimensions: progress indicators representing the remaining percentage of the cycle (e.g., ...). The metrics include budget resource indicators (such as remaining budget and budget consumption rate), performance indicators (such as current cumulative return on investment and unit cost performance), and other characteristic indicators used to characterize the stability, continuity, or other business attributes of advertising campaigns. Each of these dimensions can be flexibly configured and expanded according to actual business needs.

[0065] Collected in sequence The advertising activity metric vectors at each observation node are used to construct a data matrix representing the dynamic evolution of the entire cycle, denoted as the strategic placement landscape. :

[0066] ;

[0067] in, A strategic advertising campaign, representing a complete advertising campaign cycle, is essentially a two-dimensional data matrix composed of multiple advertising campaign metric vectors arranged chronologically. Indicates the first The advertising campaign metric vector recorded at each observation node is used to describe the delivery status at that point in time. This represents the total number of observation nodes obtained after discretizing the advertising campaign period. This indicates the number of feature dimensions contained in each advertising campaign metric vector, i.e., the number of metrics, such as multiple dimensions such as cycle progress, budget status, and performance. Represent a OK A column of real numbers.

[0068] Strategic deployment scenario in implementation The matrix representation comprehensively expresses the joint evolution trajectory of advertising campaigns across time and feature space. Based on a large number of historical strategic campaign scenarios, evaluation criteria are established to assess the quality of each scenario, such as measuring its comprehensive performance across multiple dimensions including resource utilization efficiency, campaign stability, and target achievement rate. From this, high-performing scenario samples are selected as high-quality strategic campaign scenarios, denoted as […]. .

[0069] Optionally, this application does not impose restrictions on the specific screening requirements for high-quality strategic deployment scenarios, which shall be determined by the choices made during specific implementation.

[0070] Step S102: Encode the advertiser's placement needs into a standardized business intent code.

[0071] In step S102, the diverse and subjective advertising needs of advertisers are uniformly encoded into a standardized, machine-understandable business intent code, denoted as . ,in Indicates the number of dimensions in the encoding. This represents the set of real numbers. The design of this business intent encoding aims to explicitly quantify complex advertising goals and constraints, serving as guiding constraints for subsequently generating the optimal strategic advertising scenario.

[0072] In a preferred embodiment, business intent encoding This includes, but is not limited to, core optimization intent, boundary constraints, and process form preferences; where the core optimization intent expresses the advertiser's main objective, such as maximizing return on investment, and is normalized to a scalar value, such as... Boundary constraints describe the constraints that must be met, such as the average cost per click (CPC) needing to be below a preset threshold. These constraints are represented by binary indicator variables. 0 indicates that the constraint must be met, and 0 indicates that there is no such restriction; the process form preference is used to reflect the advertiser's preference for the form of the campaign, such as a smoother budget consumption. This preference is achieved by first calculating the corresponding quantitative indicators, and then encoding whether the indicators meet the requirements into binary variables, for example... This indicates that the smoothing requirement is met.

[0073] The aforementioned business intent encoding effectively transforms advertisers' diverse campaign needs into structured numerical vectors, facilitating input and processing by subsequent machine learning models. Optionally, in practice, advertisers' natural language requirements or options can be transformed into structured business intent encoding through rule-based methods, questionnaire mapping, or the use of predefined label feature templates. This application does not impose any restrictions on specific encoding methods.

[0074] Step S103: Using the extracted high-quality strategic placement imagery and the advertiser's business intent encoding, a forward perturbation and conditional inverse recovery process is constructed to train the preset conditional imagery synthesizer. In the actual bidding stage, starting from the random initial strategic placement imagery, the conditional imagery synthesizer gradually recovers and generates the optimal strategic placement imagery that meets the placement requirements.

[0075] In step S103, after extracting the high-quality strategic placement landscape and encoding the advertiser's personalized and diversified placement needs into standardized business intent codes, a conditional landscape synthesizer is needed to generate an optimal strategic placement landscape that conforms to the current business intent, covers the entire future placement cycle, and has a clear structure and high quality. The optimal strategic placement landscape is equivalent to a detailed operational plan that spans the entire cycle, defining the ideal state that the system should achieve at each observation node. The conditional landscape synthesizer can learn the complex data distribution of historical high-quality strategic placement landscapes and perform conditional generation under the guidance of the advertiser's business intent, thereby ensuring that the generated result not only conforms to the structural characteristics of historical high-performing strategies but also meets the specific needs of the current placement goal. Its working mechanism is based on the joint modeling and learning of a controllable random perturbation process and its reverse recovery process, and is divided into two stages: the forward perturbation process and the conditional reverse recovery process.

[0076] Specifically, during the forward perturbation process, the first step is to extract the high-quality strategic deployment scenario. The process involves progressively injecting Gaussian perturbations of preset strength to obtain scene representations at different perturbation levels. The fundamental purpose of this process is to construct a continuous sequence of perturbation trajectories to train subsequent models to inversely recover scene states under arbitrary perturbation levels, thereby understanding the deep structural features of the scene distribution. The perturbation is executed based on a set maximum number of perturbation steps. The disturbance process can be understood as a shift from a high-quality strategic deployment landscape. This clear and complete original picture is gradually injected with noise, causing it to blur or degrade at each step, eventually evolving into a picture completely dominated by noise. The training process starts from these degraded pictures of varying degrees, learning how to restore them to a high-quality picture under the guidance of business intent. At each step… The current disturbance pattern The disturbance pattern from the previous moment can be used to determine the nature of the scene. The perturbation pattern at any given time can be generated in one step using a closed-form function, and its mathematical form is as follows:

[0077] ;

[0078] in, It is a standard Gaussian random matrix. It is preset, random Varying signal-to-noise ratio control parameters This is the perturbation function. Through the aforementioned controllable noise injection, a complete set of training samples ranging from slight degradation to severe degradation can be obtained. These training samples not only contain the original, high-quality strategic deployment scenarios... It also includes strategic deployment scenarios with different degrees of degradation.

[0079] In the conditional inverse recovery process, the goal of the conditional scene synthesizer is to learn how to gradually recover the original, high-quality strategic deployment scene from these strategic deployment scenes that have been perturbed to varying degrees. The core of this process lies in constructing a deep neural network with strong expressive power. In this application, This is a transformation network suitable for matrix data, capable of effectively modeling the degradation patterns of a scene under noise perturbation, and performing back-inference based on a triple condition of perturbed scene, perturbation degree, and business intent. Specifically, the neural network... The input includes a picture under the current level of disturbance. Current perturbation time step and business intent coding Its task is to accurately predict the original perturbation injected in this step, given the aforementioned triple input. In other words, the network must learn to reconstruct the noise components corresponding to each step of the degraded image. In this way, the model essentially learns the path for the image to return to a clear state from noise, thus acquiring the ability to generate high-quality images. Neural Networks The optimization objective of training is to minimize the following expected error. :

[0080] ;

[0081] in, This represents the expected value for all samples in the training data. It is real Gaussian noise injected during the forward perturbation process. This represents the model's predicted value for the noise. This represents the Frobenius norm of the matrix. Through optimization, the neural network... It can gradually improve its ability to recover from various levels of disturbance, and ultimately has the ability to recover a strategic deployment picture with a clear structure and consistent semantics from any degraded picture under the guidance of business intentions.

[0082] In the actual bidding phase, starting from a completely random initial strategic deployment scenario... Starting from a given advertiser's business intent encoding Guided by [the program], the already trained deep neural network is invoked. ,by As input, execute sequentially. Reverse recovery operation, that is, at each time step Deep neural networks Receive current disturbance pattern Current perturbation time step and business intent coding By inferring backwards, the original perturbation in the current scene is estimated. And based on the estimation results, a clearer picture of the previous step is calculated. This process iterates continuously, causing the image to gradually become clearer and more stable from the initial Gaussian noise, eventually reaching a clearer and more stable state at time step 1. At that time, a clear, theoretically optimal strategic placement scenario is generated, tailored to the current advertiser. .

[0083] Optionally, the conditional scene synthesizer is essentially a generative model with conditional generation capabilities. This application does not limit the specific type or network structure of the conditional scene synthesizer used. As long as the model has the ability to learn distribution features from historical scenes and achieve scene generation under the guidance of business intentions, it can be used as the implementation method of this application.

[0084] Step S104: Construct and train a non-backtracking bid command inferrer based on the optimal strategic deployment scenario. The non-backtracking bid command inferrer generates bid commands in real time by combining historical trends and target status.

[0085] In step S104, an optimal strategic deployment scenario is generated. ,in Indicates the system at the 1st The desired metric state at each observation node is used to describe the system's target trajectory throughout its entire lifecycle. To ensure the system's actual operation closely approximates this trajectory, an efficient, stable, and real-time responsive mechanism is needed to transform the target states of each node in the graph into executable bidding commands. Therefore, this step constructs a non-backtracking bidding command derivator. Its task is to derive, in real-time, the optimal bidding command driving the system towards the target state, by combining the current and recent historical observation states with the target state at the next moment. .

[0086] The design goal of this derivator is to answer the following question: Given observed historical metric sequences... In order to make the next state of the system as close as possible to the planned target state What bid order should be executed? .

[0087] Therefore, it is proposed to model the derivator as a function. Its mathematical form is as follows:

[0088] ;

[0089] in, For the current moment The bid command to be executed includes the adjustment amounts for each adjustable parameter in the bidding system; This indicates that the current time is the endpoint and the length is... Historical measurement observation windows, in which each Indicates the system at time 10:00 The operational metric status; Indicates the first [item] in the current scene The target state of each observation node. The non-backtracking nature of this function means that its input depends only on the current node and the observation states before it, without involving future metric information, thus making it suitable for online real-time inference. In practical deployments, by introducing a historical window of length H instead of using only the metric state at the current moment, the inferrer can more fully capture the short-term dynamic trends of the system operation, improve the robustness of command generation, and avoid overreacting to transient fluctuations.

[0090] In actual operation, the non-backtracking bid order derivative is invoked in real time at each observation node. The system first collects the actual operational metrics over the past H time steps. And in combination with the current picture, the first Target state of each observation node This input is fed into the derivator, which performs feature extraction and deep modeling on the input, and outputs the bid command that should be executed at the current moment. .

[0091] Furthermore, preferably, a standard supervised learning method is used to train this non-backtracking bid order inferrer, specifically as follows: a large-scale training sample pair is constructed from historical advertising data. The input for each sample It is a combination of historical metric windows and the actual next metric vector reached, that is, the historical observation sequence at a certain moment and the preset next target state in the picture. The two together reflect the input context at that time, while the label This refers to the bid order actually executed within that history window. This represents the bidding operation made by the system in the context of the deployment at that time to achieve business objectives.

[0092] The training objective of this inferr is to minimize the distance between the predicted command and the true command, for example, using the L2 loss function. as follows:

[0093] ;

[0094] in, Represents the expectation over all training sample pairs. This represents the bid order that was actually executed in history. The model represents the input The predicted bid command. Through the above training process, the non-backtracking bid command inferrer can learn decision-making patterns summarized from historical operating trajectories and target states, thereby achieving adaptive generation of bidding behavior in complex delivery environments. This ensures that the automatic bidding method for advertising meets real-time requirements while possessing good robustness and convergence.

[0095] In summary, by extracting historical high-performing strategic placement scenarios and combining them with advertisers' personalized commercial intent encoding, a conditional scenario synthesizer is constructed and trained to generate an optimal strategic placement scenario covering the entire future placement cycle, serving as the system's global operational plan. This scenario considers both the structural characteristics of historically high-performing strategies and current placement goals, enabling long-term planning and goal guidance for the placement path. Subsequently, a non-backtracking bidding command derivator is designed, combining current and historical observation states with the target states in the strategic placement scenario to derive the optimal bidding command driving the system towards the global planning goal in real time. This ensures that the bidding strategy possesses both dynamic adaptability and stable approximation of the long-term planning trajectory. Through this method, this application effectively overcomes the short-sightedness and volatility sensitivity issues of immediate response models, achieving stability, flexibility, and performance optimization of advertising bidding strategies. It meets the complex placement needs of multiple objectives and constraints, improving overall advertising placement effectiveness and resource utilization efficiency.

[0096] Figure 2 This is a structural block diagram of an automatic advertising bidding system based on predictive delivery planning, provided in one embodiment of this application. The system includes at least the following modules:

[0097] The high-quality strategic placement image extraction module is used to construct a strategic placement image composed of multi-dimensional advertising campaign metric vectors based on historical advertising placement data, and to extract high-quality strategic placement images.

[0098] The ad placement demand coding module is used to encode advertisers' ad placement demands into standardized business intent codes;

[0099] The optimal scenario generation module is used to construct a forward perturbation and conditional inverse recovery process by using the extracted high-quality strategic placement scenario and the advertiser's business intent code to train a preset conditional scenario synthesizer. In the actual bidding stage, starting from a random initial strategic placement scenario, the optimal strategic placement scenario that meets the placement requirements is gradually recovered and generated through the conditional scenario synthesizer.

[0100] The bid command generation module is used to build and train a non-backtracking bid command inferrer based on the optimal strategic deployment scenario. The non-backtracking bid command inferrer generates bid commands in real time by combining historical trends and target status.

[0101] For relevant details, please refer to the above method implementation examples.

[0102] Figure 3 This is a block diagram of an electronic device provided in one embodiment of this application. The device includes at least a processor 401 and a memory 402.

[0103] Processor 401 may include one or more processing cores, such as a quad-core processor or an octa-core processor. Processor 401 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 401 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 401 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 401 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0104] Memory 402 may include one or more computer-readable storage media, which may be non-transitory. Memory 402 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in memory 402 is used to store at least one instruction, which is executed by processor 401 to implement the automatic advertising bidding method based on predictive delivery planning provided in the method embodiments of this application.

[0105] In some embodiments, the electronic device may also optionally include: a peripheral device interface and at least one peripheral device. The processor 401, memory 402, and peripheral device interface can be connected via a bus or signal line. Each peripheral device can be connected to the peripheral device interface via a bus, signal line, or circuit board. Indicatively, peripheral devices include, but are not limited to: radio frequency circuits, touch displays, audio circuits, and power supplies.

[0106] Of course, electronic devices may also include fewer or more components, and this embodiment does not limit this.

[0107] Optionally, this application also provides a computer-readable storage medium storing a program that is loaded and executed by a processor to implement the automatic advertising bidding method based on predictive delivery planning in the above method embodiments.

[0108] Optionally, this application also provides a computer product including a computer-readable storage medium storing a program that is loaded and executed by a processor to implement the automatic advertising bidding method based on predictive delivery planning described in the above method embodiments.

[0109] The technical features of the above embodiments can be combined in any way. For the sake of brevity, 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, they should be considered to be within the scope of this specification.

[0110] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for automatic advertising bidding based on predictive delivery planning, characterized in that, The method includes: Construct a strategic placement landscape based on historical advertising data, consisting of multi-dimensional advertising campaign metric vectors, and extract high-quality strategic placement landscapes. Encode advertisers' campaign needs into standardized business intent codes; The extracted high-quality strategic placement landscape and the advertiser's business intent encoding are used to construct a forward perturbation and conditional inverse recovery process to train a preset conditional landscape synthesizer. In the actual bidding stage, starting from a random initial strategic placement landscape, the conditional landscape synthesizer is used to gradually recover and generate the optimal strategic placement landscape that meets the placement requirements. Based on the aforementioned optimal strategic deployment scenario, a non-backtracking bid order inferrer is constructed and trained. This non-backtracking bid order inferrer generates bid orders in real time by combining historical trends and target states, including: Construct a non-backtracking bid order derivator. Its task is to derive, in real time, the optimal bid order that drives the system toward the target state, by combining the current and recent historical observations with the target state at the next moment, during system operation. The design goal of this derivator is to answer the following question: Given observed historical metric sequences... In order to make the next state of the system as close as possible to the planned target state What bid order should be executed? ; Model the derivator as a function Its mathematical form is as follows: ； in, For the current moment The bid order to be executed. This indicates that the current time is the endpoint and the length is... Historical measurement observation windows, in which each Indicates the system at time... The operational metric status; Indicates the first [item] in the current scene The target state of each observation node; the non-backtracking bid command derivator is invoked in real time on each observation node to collect actual operational metrics over the past H time steps. And in combination with the current picture, the first Target state of each observation node This input is fed into the derivator, which performs feature extraction and deep modeling on the input, and outputs the bid command that should be executed at the current moment. ; The non-backtracking bid order inferrer is trained using a standard supervised learning method, specifically as follows: A large-scale training sample pair is constructed from historical advertising data. The input for each sample It is a combination of historical metric windows and the actual next metric vector reached, that is, the historical observation sequence at a certain moment and the preset next target state in the picture, label. The bid order actually executed within that history window. ; The training objective of the non-backtracking bid order inferrer is to minimize the distance between the predicted order and the true order, using the L2 loss function. as follows: ； in, Represents the expectation over all training sample pairs. This represents the bid order that was actually executed in history. The model represents the input The predicted bid order.

2. The automatic advertising bidding method based on predictive delivery planning according to claim 1, characterized in that, The strategic placement landscape constructed based on historical advertising placement data, consisting of a multi-dimensional advertising campaign metric vector, includes: Collect historical advertising data, and for each historical advertising data point, discretize it on the time axis using a complete advertising cycle as the time unit. A series of observation nodes, at each observation node Using a multidimensional advertising campaign metric vector This represents the operational status of the ad placement at that moment; Collected in sequence The advertising activity metric vectors at each observation node are used to construct a data matrix representing the dynamic evolution of the entire cycle, denoted as the strategic placement landscape. : ； in, This represents a strategic advertising campaign overview spanning a complete advertising campaign cycle. Indicates the first The advertising activity metric vector recorded at each observation node; This represents the total number of observation nodes obtained after discretizing the advertising campaign period. This indicates the number of feature dimensions contained in each advertising campaign's metric vector. Represent a OK A column of real numbers.

3. The automatic advertising bidding method based on predictive delivery planning according to claim 1, characterized in that, The process of encoding advertisers' campaign needs into standardized business intent codes includes: Advertisers' campaign needs are uniformly coded into standardized business intent codes, denoted as... ,in Indicates the number of dimensions in the encoding. Represents the set of real numbers; Business Intent Coding It includes core optimization intent, boundary constraints, and process form preferences; where core optimization intent expresses the advertiser's main goal, boundary constraints describe the constraints that must be met, and process form preferences reflect the advertiser's preference for the form of the campaign.

4. The automatic advertising bidding method based on predictive delivery planning according to claim 1, characterized in that, The process of constructing a forward perturbation and conditional inverse recovery process using the extracted high-quality strategic placement landscape and the advertiser's business intent encoding to train a preset conditional landscape synthesizer includes: During the forward perturbation process, the extracted high-quality strategic deployment scenario is presented. Gaussian perturbations of preset strength are gradually injected to obtain image representations of different perturbation levels. The execution of the perturbation is based on the set maximum number of perturbation steps. At every step The current disturbance pattern The disturbance pattern from the previous moment can be used to determine the nature of the scene. The perturbation pattern at any given time can be generated in one step using a closed-form function, and its mathematical form is as follows: ； in, It is a standard Gaussian random matrix. It is preset, random Varying signal-to-noise ratio control parameters Let be the perturbation function; In the conditional inverse recovery process, a deep neural network is constructed. Neural Networks The input includes a picture under the current level of disturbance. Current perturbation time step and business intent coding Its task is to predict the original perturbation injected in this step given a triple input. Neural Networks The optimization objective of training is to minimize the following expected error. : ； in, This represents the expected value for all samples in the training data. It is real Gaussian noise injected during the forward perturbation process. This represents the model's predicted value for the noise. This represents the Frobenius norm of the matrix.

5. The automatic advertising bidding method based on predictive delivery planning according to claim 4, characterized in that, The step of starting from a random initial strategic deployment scenario during the actual bidding phase and gradually recovering and generating an optimal strategic deployment scenario that meets the deployment requirements through the conditional scenario synthesizer includes: In the actual bidding phase, starting from a completely random initial strategic deployment scenario... Starting from a given advertiser's business intent encoding Guided by [the program], the already trained deep neural network is invoked. ,by As input, execute sequentially. Reverse recovery operation, that is, at each time step Deep neural networks Receive current disturbance pattern Current perturbation time step and business intent coding By inferring backwards, the original perturbation in the current scene is estimated. And calculate the previous step's picture based on the estimation results. This process iterates continuously, gradually making the image clearer and more stable from the initial Gaussian noise, at each time step. At that time, generate the optimal strategic deployment scenario. .

6. An automatic advertising bidding system based on predictive delivery planning, characterized in that, include: The high-quality strategic placement image extraction module is used to construct a strategic placement image composed of multi-dimensional advertising campaign metric vectors based on historical advertising placement data, and to extract high-quality strategic placement images. The ad placement demand coding module is used to encode advertisers' ad placement demands into standardized business intent codes; The optimal scenario generation module is used to construct a forward perturbation and conditional inverse recovery process by utilizing the extracted high-quality strategic placement scenario and the advertiser's business intent encoding to train a preset conditional scenario synthesizer. In the actual bidding stage, starting from a random initial strategic placement scenario, the optimal strategic placement scenario that meets the placement requirements is gradually recovered and generated through the conditional scenario synthesizer. A bid order generation module is used to construct and train a non-backtracking bid order inferrer based on the optimal strategic deployment scenario. The non-backtracking bid order inferrer generates bid orders in real time by combining historical trends and target states, including: Construct a non-backtracking bid order derivator. Its task is to derive, in real time, the optimal bid order that drives the system toward the target state, by combining the current and recent historical observations with the target state at the next moment, during system operation. The design goal of this derivator is to answer the following question: Given observed historical metric sequences... In order to make the next state of the system as close as possible to the planned target state What bid order should be executed? ; Model the derivator as a function Its mathematical form is as follows: ； in, For the current moment The bid order to be executed. This indicates that the current time is the endpoint and the length is... Historical measurement observation windows, in which each Indicates the system at time... The operational metric status; Indicates the first [item] in the current scene The target state of each observation node; the non-backtracking bid command derivator is invoked in real time on each observation node to collect actual operational metrics over the past H time steps. And in combination with the current picture, the first Target state of each observation node This input is fed into the derivator, which performs feature extraction and deep modeling on the input, and outputs the bid command that should be executed at the current moment. ; The non-backtracking bid order inferrer is trained using a standard supervised learning method, specifically as follows: A large-scale training sample pair is constructed from historical advertising data. The input for each sample It is a combination of historical metric windows and the actual next metric vector reached, that is, the historical observation sequence at a certain moment and the preset next target state in the picture, label. The bid order actually executed within that history window. ; The training objective of the non-backtracking bid order inferrer is to minimize the distance between the predicted order and the true order, using the L2 loss function. as follows: ； in, Represents the expectation over all training sample pairs. This represents the bid order that was actually executed in history. The model represents the input The predicted bid order.

7. An electronic device, characterized in that, The device includes a processor and a memory; the memory stores a program that is loaded and executed by the processor to implement an automatic advertising bidding method based on predictive delivery planning as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The storage medium stores a program that, when executed by a processor, is used to implement an automatic advertising bidding method based on predictive delivery planning as described in any one of claims 1 to 5.

Images