Optimal bidding system and method in repeated online first-level price auction

Abstract

The invention provides the system, the device and the method for determining the optimal bidding strategy of which the performance is better than that of the bidding adjustment mode in a scene of repeatedly participating in the first-level price auction. The invention provides a method for predicting inventory. According to the method, an online real-time repeated auction scene can be converted into an offline optimization problem; specifically, auction is summarized through a small number of parameters, and joint distribution of the parameters is predicted based on historical data. Meanwhile, the invention further develops a method named as strategy search, and the method adopts a combined optimization means to search a bidding strategy capable of realizing the optimal performance of the advertisement activity on the premise of meeting specific constraint conditions.

Description

[0001] Cross-reference to related applications

[0002] This application claims priority to U.S. Provisional Patent Application No. 63 / 525,795, filed July 10, 2023, entitled “Optimal Bidding System and Method in Repeated Online First-Level Price Auctions,” the entire disclosure of which is incorporated herein by reference. Background Technology

[0003] Over the past decade, real-time auctions have become the mainstream method for buying and selling online advertising rights. This system is popular because it aligns the interests of both advertisers and content publishers. For advertisers, the system gives them unprecedented control over when, where, and to whom their ads are placed; for publishers, it allows them to collect fair market value for every ad impression.

[0004] In this real-time auction system, advertisers allocate their marketing budgets to participate in repeated auctions hosted by one or more platforms, such as ad exchanges or supply-side platforms (SSPs). Typically, each auction offers the opportunity to place an ad in a specific location on a specific webpage for a single user. When a user visits a webpage with an ad opportunity, relevant information is sent to the ad exchange, triggering the auction process, which must be completed within tens of milliseconds. Ad exchanges run thousands of auctions per second, so bidding must be automated by computer programs—a process known as programmatic bidding. This method of purchasing ad placement rights is called programmatic advertising.

[0005] To enable advertisers to respond quickly to each auction opportunity (typically requiring bidding within 50-100 milliseconds), advertisers generally use demand-side platforms (DSPs). DSPs provide specialized technical architectures capable of "receiving" auction bidding requests from supply-side platforms and returning bidding responses within a specified time (usually tens of milliseconds). Competition among DSPs manifests in several aspects, including the supply-side platform resources they integrate with and the sophistication of the bidding algorithms they provide to advertisers.

[0006] When advertisers participate in programmatic advertising, a key issue is determining a reasonable auction value for each bid request. In recent years, this problem has become significantly more difficult to solve because supply-side platforms have gradually shifted from a two-tier price auction model to a one-tier price auction model.

[0007] The programmed operation of primary-price auctions is more difficult because achieving the optimal bid requires bidders to anticipate the value offered by other bidders. For example, in an auction with three bidders: Bidder A bids $1, Bidder B bids $2, and Bidder C bids $2.50. In a secondary-price auction, Bidder C only needs to pay $2.01; however, in a primary-price auction, Bidder C needs to pay $2.50. This means that if Bidder C can anticipate the second-highest bid value and lower their bid accordingly, they will achieve better cost-effectiveness.

[0008] Following the switch to the primary price auction model, both supply-side platforms (PSPs) and demand-side platforms (DSPs) have introduced a feature called "bid adjustment tool" to help advertisers reduce the risk of paying excessive fees in primary price auctions. While bid adjustment is a standardized method to mitigate the risk of premiums in primary price auctions, it is not the optimal strategy in scenarios involving repeated auctions and balancing multiple objectives (such as when executing programmatic advertising campaigns).

[0009] The embodiments of the present invention are intended to address, individually or in combination, the above-mentioned and other defects of the traditional repeated first-level price auction bidding method. Summary of the Invention

[0010] As used herein, the terms "invention," "this invention," and "this disclosure" are intended to broadly refer to all subject matter disclosed in this document and its accompanying drawings, as well as the content of the claims. The inclusion of such terms does not constitute a limitation on the meaning or scope of the disclosed subject matter or the claims. The implementation reasons covered in this disclosure are defined in the claims, not the content of this invention. The content of this disclosure is a high-level overview of various aspects of this disclosure, leading to some concepts further elaborated in the following detailed description, and is not intended to identify key, necessary, or essential features of the claimed subject matter, nor can it be used alone to determine the scope of the claimed subject matter. The subject matter of this invention should be understood in conjunction with the relevant portions of this specification, all the accompanying drawings, and each claim.

[0011] Various embodiments of the present invention relate to a bidding strategy implementation system, method, and apparatus that outperform traditional bid adjustment methods in scenarios involving repeated participation in primary price auctions. In some embodiments, the innovations of this disclosure are mainly reflected in two aspects: First, a method called "inventory prediction" is proposed and developed, which can transform online real-time repeated auction scenarios into problems that can be optimized offline. In one embodiment, inventory prediction summarizes the auction using a small number of parameters and then predicts the joint distribution of these parameters based on historical data; Second, a method called "strategy search" is proposed and developed, which employs combinatorial optimization techniques to find a bidding strategy that achieves optimal performance for advertising campaigns while satisfying specific constraints.

[0012] One advantage of this publicly available approach is its ability to strategically select value to achieve specific objectives of the full-volume bidding, such as cost of action (CPA), a metric used to measure the overall performance of all winning bids. Traditional bid adjustment schemes cannot optimize for such objectives because they operate only at the level of a single bid request.

[0013] Through research and simulation verification, the inventors have demonstrated that the pricing system proposed in this application outperforms traditional bid adjustment techniques and can be applied to scenarios where bidders repeatedly participate in primary or secondary price auctions (such as programmatic advertising). Ultimately, it can improve the return on investment (ROI) for bidders who adopt this solution.

[0014] In some embodiments, the implementation method of this disclosure can be accomplished through the following steps, stages, functions, or processes: Build or call a success rate model based on historical auction data In one embodiment, the model represents the auction winning probability as a function of the bid value. For example, the model can output the shape parameter k and scale parameter λ of a Weibull distribution such that the predicted auction winning probability is the cumulative probability of the Weibull distribution corresponding to the bid value.

[0015] The Weibull distribution is suitable for parameterized distributions of positive real numbers (such as dollar amounts above the auction reserve price), where k is the shape parameter and λ is the scale parameter. Other distribution forms such as log-normal distribution and histogram fitting can also be used.

[0016] Build or call up key performance indicator (KPI) models based on historical auction data

[0017] In one embodiment, the model is used to characterize the value (or impact) of winning an auction on a specific key performance indicator. For example, the model may output a conversion probability pconv, which represents the probability that placing an advertisement at the auction item location will prompt a user to take subsequent action (such as making an online purchase or completing other goal behaviors) and obtain attribution for that action.

[0018] In addition to historical auction data, this type of model can also integrate data related to the value of the auction to the bidders. In the advertising field, such data may include: user identifiers that generate online behavior, the time when the user behavior occurred (and the time interval between when the user viewed the advertisement and when the behavior occurred), and the absolute or relative value / cost corresponding to the behavior, etc.

[0019] Generate inventory forecast results, which involves predicting the results by combining the outputs of all available auction completion rate models and KPI models using a joint histogram.

[0020] Specifically, this involves predicting the joint histogram of parameters k, λ, and pconv, which typically requires knowledge of relevant information about the target's eligibility for auction, and is achieved through the search function proposed in this disclosure.

[0021] Utilize inventory forecasting results to perform strategy search and determine bidding strategies that optimize specific key performance indicators.

[0022] The bidding strategy consists of two parts: (1) the bidding function, which determines the auction value based on the output of the transaction rate model and the KPI model, and is used to calculate the value-volume ratio of the auction; and (2) the value-volume ratio threshold, which is used to determine whether to participate in the auction.

[0023] In some embodiments, the strategy search includes an outer process and an inner process: the outer process performs a parameterized search for possible bidding functions; the inner process evaluates the performance of each bidding function using inventory forecast results.

[0024] In some embodiments, the final bidding strategy and / or model performance can be evaluated to determine whether the search process or one or more models need to be retrained or adjusted.

[0025] Deploy a defined bidding strategy

[0026] Apply the bidding function described above and determine whether to participate in the bidding based on the value-to-volume ratio threshold.

[0027] Each programmatic auction is evaluated in near real-time based on the selected strategy to determine whether to bid and the value of the bid. The decision can be submitted directly to the ad exchange platform (supply-side platform) or to the advertiser through the ad technology partner.

[0028] In some embodiments, the bidding strategy and / or model performance after deployment can be monitored and evaluated to determine whether the search process or one or more models need to be retrained or adjusted.

[0029] Other objects and advantages of the systems, apparatuses, and methods disclosed in this disclosure will become apparent to those skilled in the art upon reading the detailed embodiments and accompanying drawings. Throughout the drawings, the same reference numerals and descriptions represent similar but not necessarily identical elements. Although various modifications and substitutions are possible with the embodiments disclosed or described in this disclosure, specific embodiments are shown in the drawings by way of example and are described in detail herein. However, the embodiments of this disclosure are not limited to the exemplary forms described; rather, this disclosure covers all modifications, equivalents, and alternatives falling within the scope of the appended claims.

[0030] Brief description of the attached figures

[0031] The embodiments of this disclosure will be described in conjunction with the accompanying drawings: Figure 1 is a flowchart that exemplifies the process, operation, method, or functional steps for implementing the optimal bidding strategy in a scenario of repeatedly participating in a first-level price auction. Figure 2 is a schematic diagram illustrating the components of a computing device, server, or system configured to implement the methods, processes, functions, or operations of embodiments of the present invention. Figures 3, 4, and 5 are architectural diagrams, exemplarily illustrating a multi-tenant platform or Software as a Service (SaaS) platform architecture that can be used to implement embodiments of the systems, apparatus, and methods disclosed herein. Detailed Implementation

[0032] To meet legal requirements, this document provides a detailed description of one or more embodiments of the disclosed subject matter, but such description does not constitute a limitation on the scope of the claims. The claimed subject matter may be implemented in other forms, may include different elements or steps, and may be used in conjunction with other existing or future technologies. Unless explicitly stated that the order of execution of individual steps or the arrangement of elements is necessary, this description should not be construed as imposing any mandatory requirements on the order of execution or arrangement of the steps or elements.

[0033] This document provides a full description of embodiments of the present disclosure in conjunction with the accompanying drawings, which illustrate exemplary embodiments of the systems, apparatus, and methods that may be implemented according to the present disclosure. However, this disclosure is not limited to the embodiments set forth herein; these embodiments are provided to satisfy legal requirements and to convey the scope of protection of this disclosure to those skilled in the art.

[0034] The subject matter of this disclosure may be embodied in whole or in part as a system, one or more methods, or one or more apparatuses, and the implementation may take the form of pure hardware, pure software, or a combination of hardware and software. For example, in some embodiments, one or more operations, functions, processes, or methods disclosed and / or described in this disclosure may be executed by one or more suitable processing elements (such as processors, microprocessors, central processing units, graphics processing units, tensor processors, quantum processors, state machines, or controllers, etc.) of a client device, server, network element, remote platform (such as a SaaS platform), cloud service, or other form of computing / data processing system, apparatus, or platform.

[0035] The aforementioned processing element can be programmed to perform related functions through a set of executable instructions (such as software instructions), which are stored in one or more suitable non-volatile computer-readable data storage elements. In some embodiments, the instructions or the application executing the set of instructions can be transmitted to the user via a network (such as the Internet).

[0036] In some embodiments, the systems and methods of this disclosure can provide services or functions through a SaaS platform or a multi-tenant platform. The platform can provide access to multiple entities, each with an independent account and associated data storage area. For example, each account can correspond to an individual, customer, customer group, advertiser, demand-side platform, industry, or organization. Each account can access one or more services, some of which are instantiated within that account and used to implement one or more methods or functions disclosed and / or described in this disclosure.

[0037] In some embodiments, one or more operations, functions, processes, or methods described herein can be implemented using dedicated hardware, such as programmable gate arrays (PGAs) or application-specific integrated circuits (ASICs). It should be noted that the method embodiments disclosed and / or described in this disclosure can be embodied as applications, subroutines of large applications, plug-ins, functional extensions of data processing systems / platforms, or other suitable forms. Therefore, the following detailed descriptions should not be construed as limiting.

[0038] Advertising Campaign: Budget and Strategy

[0039] Advertisers typically collaborate with media advertising agencies to integrate their advertising efforts into various advertising campaigns. An advertising campaign usually includes designated advertising creatives (sometimes called creative assets), which are images or videos displayed as advertisements, and sets a campaign period (such as the entire month of January) and a total budget (such as $50,000).

[0040] An advertising campaign can be broken down into multiple specific campaign strategies, each representing a way to efficiently utilize the advertising budget. For example, a campaign strategy might target a specific demographic or group of people that the advertiser believes will respond to selected creative materials and advertising messages. Each campaign strategy is typically associated with one or more key performance indicators (KPIs) to measure the effectiveness of that strategy or the overall campaign. For instance, a campaign strategy might set the target attributable action cost at $40 as a KPI.

[0041] In addition to the target attribution action cost, the campaign strategy may also include other relevant key performance indicators (such as video completion rate, cost per user website visit, etc.) and different constraints. For example, advertising campaigns or campaign strategies may set a fixed number of impressions rather than a fixed budget, and the systems and methods disclosed herein can be adapted to such scenarios.

[0042] An example of a campaign strategy is that a company wants to allocate a specific budget to advertising within a specified period, keep the value of each ad spend within a certain range, and achieve the expected number of user ad response behaviors.

[0043] In programmatic advertising, a campaign strategy can be associated with a set of filtering conditions and / or rules to define which bid requests to participate in and their bid value. The complexity of these filtering conditions and / or rules varies, and in some cases may rely on machine learning models. However, for the purposes of this disclosure, a key characteristic is that these filtering conditions and rules do not utilize information about future auction participants, nor do they consider the potential bid value from other advertisers for a given ad inventory or campaign opportunity. In other words, traditional methods do not consider the characteristics of repeated bidding scenarios, nor the dynamic characteristics of primary price auctions and their impact on the optimal strategy.

[0044] To mitigate the risk of premiums in primary price auctions, traditional campaign strategies typically rely on bid adjustment technology. Once a bid value is selected, it is passed to a bid adjustment tool (usually integrated into the technical architecture of the demand-side platform or supply-side platform) before being submitted to the auction. The bid adjustment tool is the first step in the traditional bidding process to consider the dynamic characteristics of the primary price auction (such as the potential bid value from other advertisers for a given ad inventory or opportunity).

[0045] In some embodiments, this disclosure provides a system, apparatus, and method for determining the optimal bidding strategy in repeated primary price auction scenarios. In some embodiments, the method proposed in this disclosure can introduce dynamic information from the primary price auction earlier in the decision-making process of the bidding strategy, as further explained below. This approach can more effectively optimize specific key performance indicators (KPIs). Overall, embodiments of the present invention determine the optimal bidding strategy based on the following two points: (a) the likelihood that a specific bid value will have a positive impact on achieving the target KPI; and (b) the expected success rate corresponding to that bid value.

[0046] Bid Adjustment

[0047] Demand-side platforms (DSPs) and supply-side platforms (SSPs) offer bid adjustment tools within their platforms to help advertisers avoid paying excessive fees in primary price auctions. These tools reduce (i.e., adjust) the advertiser's bid value before it is submitted to the auction, aiming to decrease the bidder's payment costs. However, this operation may reduce the bidder's probability of winning the bid. The algorithms for bid adjustment vary (most are related to surplus optimization methods), but the core difference between all bid adjustment tools and the solution disclosed here is that they only lower the bid value and do not make the decision to abandon a bid in an auction.

[0048] The reason for this behavior is that bid adjustment is not part of the campaign strategy itself, but rather a service provided by the demand-side platform (DSP) / supply-side platform (SSP) to help advertisers save costs. For example, if an advertiser bids $1, the bid adjustment tool will not suggest abandoning the bid, but will suggest reducing the bid value to $0.9, without significantly affecting the probability of winning the bid. After the adjustment is completed, the tool will deduct a certain percentage of the cost savings as a service fee.

[0049] Another core feature of traditional bid adjustment tools is that they rely on the "pre-adjustment bid value" to accurately reflect the value of the auction to the advertiser. This feature has two problems: First, in repeated auction scenarios such as programmatic advertising, the value of a particular auction is necessarily related to the quality and quantity of other auctions that the advertiser can participate in; second, it is difficult to use a single value to reflect the value of the auction to all the goals of the advertising campaign, as well as the corresponding risk tolerance.

[0050] Essentially, traditional bid adjustment tools weigh bid value against the probability of winning a bid. Demand-side platforms (DSPs) and supply-side platforms typically claim their bid adjustment tools employ a conservative strategy in this trade-off, aiming to reduce bid value as little as possible without significantly impacting the success rate. While platforms rarely disclose the specific algorithms used for bid adjustment, it's clear that traditional tools fail to allow advertisers to customize this trade-off based on their campaign strategies or objectives.

[0051] Regardless of the specific bid adjustment algorithm used, this tool relies on predictions of the impact of bid value on the probability of winning an auction. These predictions are typically generated by machine learning models trained on a large amount of historical auction data. Essentially, such models (referred to herein as success rate models) integrate information from other advertisers regarding the potential bid value for a given auction. In some embodiments, the method of this disclosure may also integrate success rate models.

[0052] One reason why this disclosed solution outperforms traditional bid adjustment methods is that it advances the completion rate model and its analysis of the dynamic characteristics of primary price auctions to an earlier stage of the decision-making process. Therefore, unlike bid adjustment, this model can be used to determine whether to participate in a particular auction. Furthermore, the inventory forecasting function allows the decision-making process to comprehensively consider the full bidding behavior of a particular campaign strategy over a period of time, achieving more comprehensive optimization—specifically, by applying combinatorial optimization methods. These advantages collectively determine that this disclosed solution delivers key performance indicators (KPIs) far superior to traditional bid adjustment tools for campaign strategies.

[0053] However, introducing the transaction rate model earlier in the decision-making process also presents several challenges: for example, the number of auctions per day can reach tens of billions, and selecting the optimal bid combination to optimize key performance indicators is an extremely challenging combinatorial optimization problem; furthermore, the potential bid value of each auction is essentially a continuous value, which means that there are theoretically infinite combinations of bid values ​​in the solution space of the optimization problem; finally, auctions are conducted in real time, so this method needs to be adapted to low-latency online decision-making scenarios. The embodiments of this invention specifically address the above challenges, providing an efficient and feasible solution for determining bidding strategies under these conditions.

[0054] Transaction rate model and KPI model

[0055] To overcome the shortcomings of traditional bidding strategies and bid adjustment tools, this invention integrates two types of models in the field of advertising technology. However, unlike traditional methods, this disclosure combines these models with the disclosed inventory forecasting function to search for and determine the optimal bidding strategy. Transaction rate model Trained using historical auction results data, the output represents the auction winning probability as parameters of the value function. For simplicity without loss of generality, assume the model outputs two parameters: a shape parameter k and a scale parameter λ. The model's input consists of auction-related information, such as the size of the ad placement (e.g., its location on a webpage), the expected user type viewing the ad, and the website where the ad is placed. The model needs to be updated in real-time with the latest auction data (this can be achieved through online updates and / or model retraining).

[0056] In one embodiment, the transaction rate model is built based on the survival model approach, and the model outputs parameters of a certain probability distribution (or a mixture of probability distributions). In a specific embodiment, the model outputs two parameters k and λ of a single Weibull distribution (non-mixed distribution), whose cumulative distribution function is a function that expresses the probability of winning the bid as the value, which is one implementation of the transaction rate model.

[0057] The model is trained using the negative log-likelihood of the auction results as the loss function. In this way, the parameters output by the model can maximize the observed likelihood of historical auction data in the training set. After the model is trained, the information in the bidding requests can be converted into the prediction results of the winning probability corresponding to the bid value (for example, the winning probability of a bid of $0.1 is 3%, the winning probability of a bid of $0.2 is 5%, etc.).

[0058] KPI Model

[0059] The input is the same auction information as the conversion rate model, and the output is the improvement (i.e., optimization or achievement) of winning a particular auction on one or more specific key performance indicators (KPIs). For example, to simplify the description without loss of generality, the model can output the last-touch conversion attribution probability pconv, which is the probability that after winning the auction, the advertisement becomes the last advertisement that a user touches before making a conversion and is attributed to. Overall, the KPI model represents the benefit value of key performance indicators as a function of the winning auction.

[0060] A commonly used key performance indicator (KPI) in advertising campaigns is the cost per last-touch attributable order (CTO), which is the cost of the last ad a user views before making a purchase on a website, attributing that purchase to that ad. Advertisers can set a target value of $40 per CTO. In this example, after training with a calibration process, the KPI model can output the probability of an auction generating a CTO. If the model output is 0.0005, then according to the model, the value of that auction is $0.02 (i.e., probability × target cost).

[0061] The KPI model is trained based on historical ad impression data and attribution tag data. After training, the model can use inference to transform the information in the bid request into a prediction of the request's effect on achieving the customer's key performance indicators.

[0062] In general, bidders may have multiple optimization objectives. In this case, the KPI model can output the impact of the winning auction on each key performance indicator. Examples of other key performance indicators include cost per click, cost per website visit, and cost per video completion.

[0063] Inventory Forecast

[0064] Another core feature of this disclosure is that it transforms the real-time bidding problem into a problem that can be optimized offline as a whole. This process is achieved through the "inventory forecasting" step proposed in this disclosure. In programmatic advertising, each auction corresponds to an advertising inventory target (i.e., advertising opportunity), so the auction itself is often referred to as advertising inventory.

[0065] In some embodiments, inventory forecasting refers to the prediction of whether a given campaign strategy can participate in auctions throughout or part of the advertising campaign's lifecycle. In some embodiments of the method disclosed herein, to optimize bidding strategies, the outputs of conversion rate models and KPI models can be used to characterize and summarize the advertising inventory.

[0066] The conversion rate model integrates information about competitors' potential bids in an auction, while the KPI model integrates information about the auction's value in achieving one or more key performance indicators. Therefore, predicting whether a particular bidding strategy can participate in an auction can be mathematically represented as a joint histogram prediction of the outputs of these two models; in the example above, this means a joint histogram prediction of parameters k, λ, and pconv.

[0067] In some embodiments, the inventory forecasting process of this disclosure is achieved by sampling historical auction data: recording relevant information from historical auctions, including the screening criteria satisfied by each bidding request, and the bidding request characteristics used for inference in the completion rate model and KPI model. The demand-side platform can receive all bidding request information streams, which can provide a large number of samples for model training.

[0068] For example, if you need to design a bidding strategy for a specific campaign, you can extract auction samples from the previous week and two weeks in the historical logs. If there is a limit on the number of auctions a single user can win (this constraint is called frequency control), the sample should select all auction data generated by a subset of users. Auction data typically contains one or more anonymous user identifiers (such as cookies or device identifiers). After extracting the auction samples, you need to filter the samples according to the requirements / standards of the campaign strategy, for example, only retaining auctions for specific device types, or only retaining auctions for 300×250 size ad slots.

[0069] The transaction rate model and KPI model are applied to the sample data to generate an inventory forecast table. Each row in the table represents an auction and includes information such as: auction timestamp, associated user ID, output of the transaction rate model (representing the primary price auction dynamics of the auction), and output of the KPI model (representing the value of the auction to the bidders).

[0070] Subsequently, by calculating mutual information, the joint distribution of the transaction rate model and the KPI model outputs (such as k, λ, pconv) of the previous week and the previous two weeks is compared. By gradually expanding the sample size and repeating the sampling process, the sample size when the mutual information reaches its maximum value and tends to stabilize is selected. Finally, the joint histogram of the previous week's data under this sample size is used as the inventory forecast result for the next week.

[0071] Alternatively, a multivariate time series forecasting method can be used to integrate more historical auction data and capture the daily, weekly, monthly, and yearly time trends of the joint histogram. In this type of embodiment, a time series forecasting model (such as a long short-term memory network or an automatic Bayesian neural network) is trained based on historical auction model output data (such as k, λ, pconv) over several months or years. The model inference generates a multivariate time series of the model output within the inventory forecasting period (such as one week), and each element of this time series is used as a row of data in the inventory forecasting table.

[0072] Strategy Search

[0073] Another core feature of this invention is that it utilizes inventory forecasting results to search for bidding strategies that can optimize key performance indicators (KPIs) for ad placement. In one embodiment, the bidding strategy includes two parts: (1) a bidding function, which determines the auction value based on the output of the conversion rate model and the KPI model, and is used to calculate the value-volume ratio score for each auction; and (2) a scoring threshold (i.e., the value-volume ratio threshold), used to determine whether to participate in the bidding for that auction. In some embodiments, the strategy search must meet hard constraints (e.g., the strategy's budget usage must be accurate, and it must meet the constraints of user / viewer ad exposure frequency).

[0074] In one embodiment, the strategy search comprises two processes (or can be implemented using a single process): the outer process searches for possible bidding functions, and the inner process uses inventory forecasts to evaluate the performance of each bidding function through combination optimization. The outer process tracks and records the bidding function that achieves the best forecast performance under the constraints.

[0075] Strategy Search - Outer Process

[0076] The number of auction entries in inventory forecasting data typically reaches tens of millions. Even after quantifying the bid value, the possible combinations of bid values ​​constitute a vast search space. Therefore, to make the search of the bid value space feasible, the outer process performs the search within a family of parameterized functions F. This family of functions maps the outputs of the auction / bid request's corresponding transaction rate model and KPI model (such as k, λ, pconv in some embodiments) to bid values. Examples of this family of functions F or its functional forms are as follows: Value = F(k, λ, p) conv | theta (θ)), or Value delivered = Surplus maximization of value delivered + F (k,λ,pconv|θ) The outer process searches for possible values ​​of parameter θ to find the bidding strategy that achieves the best results while meeting the constraints of the advertising campaign, after evaluation by the inner process.

[0077] In some embodiments, black-box optimization methods (such as genetic algorithms or Bayesian optimization) can be used to search for values ​​for parameter θ. For each θ value, the inner process evaluates the performance of the corresponding bidding strategy and feeds the result back to the outer process. The outer process then determines the next θ value to be tested based on this result. Finally, the θ value that achieves the best predictive performance (such as the lowest expected last-touch attribution action cost) and satisfies the constraints of the advertising campaign (such as matching the predicted cost distribution of the winning auction with the budget of the campaign strategy) is selected as the final bidding strategy.

[0078] In one embodiment, the function family F can take the form of a multilayer perceptron that incorporates Fourier features, where the parameter θ represents the weights of the network; in another embodiment, the function family can take the form of a multivariate polynomial, where the parameter θ represents the coefficients of the polynomial.

[0079] In some embodiments, the functions in the function family can round or modify the calculated output value to a value in the discrete value set, in which case the parameter θ may contain the discrete value set.

[0080] If a strategy has multiple key performance indicators (KPIs) for optimization, the parameter θ can include weight parameters in the inner process used to measure the relative importance of each indicator, thereby evaluating the effect of different weight allocations.

[0081] The function family F searched by the outer process is preferably determined based on empirical results. That is, it combines the inventory forecasting method, KPI model, and transaction rate model used to select a function family / function that can consistently find the optimal bidding strategy. However, the number of free parameters θ of the selected function family F should not be too large, otherwise it will lead to excessive computation in the outer process and lose its practicality.

[0082] Strategy Search - Inner Process

[0083] The inner process evaluates the predictive performance of the bidding strategy to be searched in the outer process, determined by parameter θ. As mentioned earlier, θ defines the bid value (bid function) of the auction and can optionally define the relative weights of each key performance indicator. Inventory forecasting provides the basis for this evaluation. Inventory forecasting transforms the real-time bidding problem into a combinatorial optimization problem: given a set of auctions (inventory forecasting results), which subset (combination) of auctions to participate in bidding will achieve the optimal effect. Specifically, this combinatorial optimization problem is highly related to the classic knapsack problem, with the unique characteristic that only auctions where the bid is won (not the ones where it is not won) are included in the "knapsack" cost.

[0084] In one embodiment, this particular knapsack problem is solved by an improved greedy approximation heuristic. One advantage of this heuristic is that, once applied to inventory forecasting data, it can be directly adapted to low-latency, near real-time decision-making environments.

[0085] When applying the greedy approximation heuristic algorithm, three values ​​need to be defined for each auction: the "quantity" when included in the "knapsack", the "value" when included in the "knapsack", and the total capacity of the "knapsack".

[0086] Based on the three values ​​mentioned above, this heuristic algorithm sorts auctions in descending order of their value-to-quantity ratio, prioritizing the auctions with the highest ratios for inclusion in the "knapsack" to maximize the total value within the "knapsack." It should be noted that the selection of the auction's "quantity," "value," and the total capacity of the "knapsack" can be adjusted according to the business objectives and constraints of a specific campaign strategy; this is a typical characteristic of combinatorial optimization problems. Overall, the auction's "value" is defined as the maximization objective of the campaign strategy / advertising activity, while the total capacity of the "knapsack" and the auction's "quantity" are defined as constraints.

[0087] For example, for an advertising campaign that aims to minimize the cost of last-touch attribution actions and requires a specific budget, the auction value can be set as the "quantity" of the auction, the conversion probability pconv can be set as the "value" of the auction, and the total dollar budget of the advertising campaign (which can be scaled according to the ratio of the inventory forecast period to the advertising campaign period and the downsampling rate of the inventory forecast) can be set as the total capacity of the "knapsack".

[0088] In another scenario, if the advertising campaign limits the number of impressions won by the bidder, the "quantity" of the auction is always 1. In this scenario, the campaign strategy may need to balance two key performance indicators: minimizing both conversion cost and impression cost. In this case, the "value" of the auction can be set as a weighted scaling combination of the conversion probability pconv and the bid value, where the parameter θ defines the weight and scaling factor. The outer process optimizes this weight to achieve the ideal trade-off between the indicators.

[0089] After defining the auction's value, quantity, and total knapsack capacity based on the business problem, an improved greedy approximation heuristic algorithm can be used to select the optimal auction combination to maximize value. The following steps are performed for each auction in inventory forecasting: First, calculate the value using the bidding function based on the model output (e.g., k, λ, pconv); second, calculate the probability of winning the bid corresponding to that value using the transaction rate model output; next, calculate the auction's value, quantity, and value-quantity ratio; finally, sort the auctions in inventory forecasting in descending order of value-quantity ratio.

[0090] Traditional greedy heuristic algorithms work by sequentially adding bids to a knapsack based on their value-to-quantity ratio from highest to lowest until the knapsack is full, thus selecting the optimal combination of bids. However, due to the randomness of auction results, this algorithm needs improvement, requiring the introduction of a simulation step.

[0091] The improved algorithm proceeds as follows: Auctions are sorted from highest to lowest value-to-quantity ratio. A random number generator is used to simulate the winning / unwinning results of each auction using biased random trials, based on the winning probability of each auction. During the simulation, cumulative indicators are added to each row of the inventory forecast table, including: cumulative number of winning auctions, cumulative cost of winning auctions (i.e., cumulative value of winning auctions), cumulative quantity of winning auctions, cumulative value of winning auctions, and optional other cumulative indicators related to the output of each KPI model (these indicators can also be obtained through random simulation).

[0092] The simulation process described above is repeated using the Monte Carlo method, and the cumulative metrics for each simulation are recorded. The number of simulations can be determined based on computational resources and / or time constraints; 500 simulations is a reasonable reference number. It should be noted that if the delivery strategy includes frequency control constraints, these constraints can be introduced during the simulation process to reject winning bids that violate these constraints.

[0093] After simulation, a value-to-volume ratio threshold is determined using cumulative metrics. This threshold is used to define which auctions participants bid on and which they abstain from bidding, thus completing the portfolio optimization process. Specifically, bidding is only initiated for auctions with a value-to-volume ratio higher than this threshold. The selection of this threshold should ensure that the cumulative metrics obtained from the simulation meet the business constraints of the advertising campaign as much as possible.

[0094] For example, if business constraints require the budget to be exhausted with a 95% confidence level, the threshold selection criterion is: in 95% of the simulations, the cumulative winning auction volume (e.g., cost) corresponding to the last auction with a value-to-volume ratio higher than this threshold is not less than the total backpack capacity (e.g., budget). It should be noted that if the bidding function setting results in the constraint being unsatisfactory even with bids from all participating auctions, this information will be fed back to the bidding process, and the outer process will determine that the θ value is unavailable.

[0095] After determining (or selecting through other means) the value-to-quantity ratio threshold, the predicted performance corresponding to the threshold and parameter θ is fed back to the outer process, while the inner process simultaneously feeds back the cumulative index of all simulations under the threshold.

[0096] The outer process determines which strategy performs best from a business perspective based on the distribution of these metrics. For example, if the conversion probability pconv is set as the auction value, then the cumulative winning auction value metric represents the expected number of last-touch attribution actions in each simulation. The outer process combines the cost of each simulation to calculate the median cost of the last-touch attribution action in the simulation, thereby selecting the θ value of the corresponding optimal bidding strategy.

[0097] If the strategy search is implemented using a single process, in one embodiment, the method of this disclosure may operate as follows: Construct a model with inputs k, λ, pconv (or one or more other KPI metrics) and output as the bid value (an output of 0 indicates abandoning the bid). Train the model using gradient descent, with the optimization objective being the sum of several terms: the first term incentivizes the model to maximize the key performance indicators; for the example model above, this term is the product of the success rate and the conversion probability pconv corresponding to the model's output bid value; the other terms are regularization terms used to ensure that the bidding strategy satisfies all hard constraints.

[0098] For example, to ensure that the budget is neither overspent nor surplus, regularization terms representing budget consumption can be added—such as outputting the product of the winning bid rate and the output value, with each regularization term multiplied by a weighting coefficient.

[0099] To obtain the final strategy, the model is trained using gradient descent. If the constraints are not met, the weight of the regularization term corresponding to the constraint is increased and the model is retrained. This process of adjusting the weight of the regularization term is repeated until the model converges to a bidding strategy that satisfies the constraints of the advertising campaign.

[0100] Real-time deployment of strategies

[0101] Once the bidding strategy is determined, it can be deployed to the real-time bidding stage of the online auction. In programmatic advertising scenarios, when the demand-side platform receives an auction request, it can use the information in the bidding request (such as URL, user information, IP address, etc.) to query the features required for model inference from the real-time feature library.

[0102] Examples of bid request information include: user ID 103019390129923i, IP address 10.0.0.127, URL https: / / www.businessinsider.com / ai-is-set-to-dominate-cannes-lions-ad-festival-2024-6, device brand Apple, and ad creative size 300×250. Other numerical features can also be extracted from the URL and user ID.

[0103] Based on the extracted features, the reasoning process of the transaction rate model and KPI model is executed to obtain the parameter set representing the auction (such as k, λ, pconv; it should be noted that the KPI indicators (such as pconv) corresponding to different strategies or objectives may be different); the value is calculated using the formula F (k,λ,pconv|θ), and the value-volume ratio is calculated (as in this example embodiment, based on the knapsack heuristic algorithm, value-volume ratio = pconv / value); if the ratio is higher than the value-volume ratio threshold determined by the strategy search process, the bid is made based on the calculated value; otherwise, the bid is abandoned.

[0104] In some embodiments, the selected bidding strategy is not deployed as a system that directly participates in auction bidding, but rather as a system function that pushes bidding suggestions (which auctions to participate in and what the bid value is) to relevant parties. This function can be implemented by interfacing with supply-side platforms, demand-side platforms, or other advertising technology partners of advertisers.

[0105] For example, some supply-side platforms allow partners to create private marketplaces (PMPs), which tag subsets of auctions with deal IDs, allowing advertisers to participate in bidding on their chosen demand-side platforms. In this scenario, the bidding strategy is deployed as follows: based on the bid value determined by the strategy, auctions recommended for bidding are tagged with different deal IDs and included in the corresponding private marketplace.

[0106] After a bidding strategy is deployed, its performance needs to be monitored and adjusted to ensure that it achieves the expected results. By combining the bidding strategy and inventory forecast results, the expected number of qualified bid requests, bids, and successful bids at each time point within the strategy period can be obtained, as well as the expected value of the winning bids (such as the conversion probability pconv or the distribution of other KPI indicators). These expected values ​​can be used to monitor whether the actual performance of the strategy meets expectations.

[0107] If the actual number of bids and / or the number of winning bids continues to be slightly lower than expected, the value-volume ratio threshold for bids can be appropriately lowered (and vice versa); if this adjustment cannot solve the problem, and there is a significant deviation between the actual number of bids / winning bids and the expected value of the deployment strategy, then the underlying model and inventory forecast results need to be updated (such as retraining one or more models), and the bidding strategy needs to be redesigned.

[0108] Solution Validation

[0109] Dataset

[0110] To verify the effectiveness of the bidding strategy optimization method proposed in this disclosure, and to demonstrate its superior performance in optimizing one or more key performance indicators (KPIs) compared to traditional bidding adjustment methods, the performance of different bidding strategies can be simulated based on auction data samples, assuming the accuracy of the transaction rate model and KPI model. The inventors evaluated the performance of three different bidding strategies on two independent datasets using this verification process: Earnings-Optimized Bidding Adjustment Strategy This strategy, representative of traditional bid adjustment methods, prioritizes bidding on auctions with the highest advertiser value (i.e., auctions with the highest conversion probability pconv or the highest relevant KPI), and only bids on a necessary amount of inventory to exhaust the budget. This traditional strategy does not consider a conversion rate model when selecting inventory for bidding, but it adjusts the bid value using a surplus optimization method based on the conversion rate model before submitting the bid. In this example, the pre-adjustment bid value of the inventory is the product of the conversion probability pconv and the campaign's attribution cost of action (CPA).

[0111] Combined with a profit-optimized bidding adjustment strategy

[0112] This strategy moves the transaction rate model forward into the decision-making process, and the bid value is still determined by the surplus optimization method (rather than the strategy search method proposed in this disclosure). However, by combining the transaction rate model and inventory forecast results, a greedy knapsack heuristic algorithm is used to screen the inventory that can participate in the bidding.

[0113] This strategy employs the combinatorial optimization method proposed in this disclosure, but does not introduce a strategy search step. Therefore, it is a variant of the scheme disclosed in this disclosure. It outperforms the traditional approach, but does not integrate all the features and techniques disclosed in this disclosure.

[0114] Strategy Search Based on Combinatorial Optimization

[0115] This strategy, the core solution disclosed in this paper, combines the transaction rate model and inventory forecast results to search for the auction subset and bid value that maximizes the target key performance indicators.

[0116] This embodiment also integrates inventory forecasting functionality for achieving combined optimization, as well as a strategy search process.

[0117] Dataset 1

[0118] This dataset represents inventory forecasting results for a specific campaign strategy with a 27-day campaign period and a budget of $35,000. The core performance indicator (KPI) is the cost of last-touch attribution action (CPA). Both the conversion rate model and the KPI model were trained and calibrated based on approximately three months of data. The inventory forecast table contains 4,281,770 rows of auction data, representing 0.15% of the weekly ad inventory. Considering the cycle ratio and sampling rate, the budget corresponding to this forecast is $13.61 ($35,000 × 7 days / 27 days × 0.15%).

[0119] Dataset 2

[0120] This dataset represents inventory forecasting results for another campaign strategy with a 30-day campaign, a budget of $44,000, and a core performance indicator (KPI) of attributable action cost per last reach (CPA). Both the conversion rate model and the KPI model were trained and calibrated based on approximately two months of data. The inventory forecast table contains 7,190,005 rows of auction data, representing 0.15% of the week's ad inventory. Considering the cycle proportion and sampling rate, the budget corresponding to this forecast is $15.40.

[0121] The three bidding strategies described above are applied to two datasets. Based on the selected bid value and the output results of the conversion rate model and KPI model, the expected exposure and conversion rate are determined.

[0122] Evaluation / validation results (as shown in Table 1) demonstrate that the strategy search strategy based on combinatorial optimization is optimal in optimizing key performance indicators. As shown in Table 1, this strategy (i.e., the proposed approach) outperforms the other two strategies on both datasets, achieving the lowest attribution action cost: $8.781 for dataset 1 and $15.895 for dataset 2. This stable performance across datasets further validates the effectiveness of the proposed strategy search strategy based on combinatorial optimization in reducing attribution action costs compared to traditional bid adjustment methods (and variants that do not incorporate strategy search).

[0123] Table 1. Comparison of Attribution Cost of Action (CPA) for different strategies in Dataset 1 and Dataset 2

[0124] Figure 1 is a flowchart illustrating, for example, the process, operation, method, or functional steps for implementing the optimal bidding strategy in a scenario of repeatedly participating in a primary price auction. This process / operation may include one or more of the following steps disclosed and / or described in this disclosure: Based on historical auction data, a success rate model is constructed or invoked. In one embodiment, the model represents the auction success probability as a function of the bid value, specifically using a Weibull distribution characterized by the shape parameter k and the scale parameter λ (as shown in step 102). A KPI model is built or invoked based on historical auction data. In one embodiment, the model characterizes the impact of the winning auction on specific KPI indicators (such as conversion probability pconv or other indicators, as shown in step 104). Generate inventory forecast results, that is, predict the joint histogram of the output results of the transaction rate model and the KPI model (i.e., predict the joint histogram of k, λ, and pconv, as shown in step 106). Using inventory forecast results, a strategy search is performed to determine a bidding strategy that optimizes a specific KPI indicator. The bidding strategy includes: (1) a bidding function, which determines the auction value based on the output of the transaction rate model and the KPI model, and is used to calculate the value-volume ratio of the auction; (2) a value-volume ratio threshold, which is used to determine whether to participate in the bidding (as shown in step 108). In some embodiments, the strategy search includes an outer process and an inner process: the outer process searches for possible bidding functions, and the inner process evaluates the performance of each bidding function using inventory forecast results. Monitor model performance. If the solution obtained from the policy search is unacceptable, retrain one or more models. If the solution is acceptable, deploy the model (as shown in step 110). If the model performance is acceptable, the bidding strategy is deployed; if the model performance is unacceptable, control is transferred to the model retraining (update) control module and / or the inventory forecast update generation control module (as shown in steps 120 and 122). In one embodiment, model retraining can introduce a feedback mechanism, using underperforming cases as training data and inputting them into the machine learning algorithm; Deploy the determined bidding strategy, apply the bidding function described above, and determine whether to participate in the bidding based on the value-volume ratio threshold (as shown in step 112). In one embodiment, bidding behavior and bidding results can be monitored and compared with expected values ​​based on inventory forecasts (as shown in step 114). In this embodiment, a logical judgment is used to determine whether the deviation between the actual value and the expected value is "slight" (corresponding to the "yes" branch in step 116); if the deviation is slight, the bidding strategy parameter adjustment process is executed (as shown in step 118), and the bidding strategy is redeployed after adjustment (i.e., control is returned to step 112). If the deviation between the actual value and the expected value is not “slight” (corresponding to the “No” branch of step 116), then control is transferred to the model retraining and / or inventory forecast update generation control module (component, assembly or process 120 and 122). Monitoring functions / processes are typically executed at fixed time intervals, so the significance of deviations is rechecked after a period of time.

[0125] Figure 2 is a schematic diagram illustrating, exemplarily showing the components of a computing device, server, or system 200 configured to implement the methods, processes, functions, or operations of embodiments of the present invention. As previously described, in some embodiments, the systems and methods of this disclosure may be implemented by means comprising processing elements and a set of computer-executable instructions that are part of a software application and organized according to a specific software architecture.

[0126] In general, embodiments of the present invention can be implemented through a set of software instructions designed to execute on appropriately programmed processing elements (such as graphics processors, central processing units, microprocessors, processors, controllers, state machines, or computing devices). In complex applications or systems, these instructions are typically organized into "modules," each module (or submodule) typically performing a specific task, process, function, or operation, with the operation of all modules controlled and coordinated by an operating system or other form of management platform.

[0127] Each application module or submodule corresponds to a specific function, method, process, or operation it implements. Such function / method / process / operation may include steps for implementing one or more aspects of the system and method disclosed herein.

[0128] Application modules and / or submodules 202 may contain suitable computer-executable code or a set of instructions (such as code that can be executed on a properly programmed processor, microprocessor, or central processing unit), for example, computer-executable code corresponding to a programming language. For example, the source code of the programming language may be compiled into computer-executable code; an interpreted programming language, such as a scripting language, may also be used.

[0129] Module 202 may contain one or more sets of instructions for performing the functions and operations disclosed in the accompanying drawings and this specification. These modules may include the modules shown in the drawings, or may include more or fewer modules.

[0130] As mentioned above, each module may contain a set of computer-executable instructions that can be executed by a programmed processor in a server, client device, network element, system, platform or other component. The computer-executable instructions in the same module or different modules may be executed by the same processor or different processors. In addition, the computer-executable instructions in a single module may be executed, in whole or in part, by one or more processors.

[0131] Instructions in a module or submodule can be executed by processors distributed across multiple servers, client devices, network elements, systems, platforms, or other components. In some embodiments, multiple electronic processors (each belonging to an independent device, server, or system) can collaboratively execute all or part of the software instructions contained in a module shown in the figures.

[0132] Therefore, although Figure 2 shows a set of modules that work together to perform multiple functions or operations, these functions or operations can be performed by different devices or system elements, with specific modules (or instructions in the modules) associated with these devices or system elements.

[0133] As shown in Figure 2, system 200 may represent a server or other form of computing / data processing system, platform, or device. Each module 202 contains a set of computer-executable instructions that, when executed by a suitable electronic processor (“physical processor 230” in the figure), system (or server, platform, device) 200 performs specific processes, operations, functions, or methods.

[0134] Module 202 is stored in (non-volatile) memory 220, which typically contains operating system module 204. Instructions in this module are used (including other functions) to access and control the execution of instructions in other modules. Module 202, stored in memory 220, achieves data transfer and instruction execution via a "bus" or communication line 216. This bus / communication line also enables processor 230 to communicate with each module to access and execute instructions.

[0135] The bus or communication line 216 also enables the processor 230 to interact with other components of the system 200, such as input / output devices 222, communication elements 224 for exchanging data and information with external devices, and additional storage devices 226.

[0136] In some embodiments, module 202 may include computer-executable software instructions that, when executed by one or more electronic processors, cause the processor or a system / device containing the processor to perform one or more of the following steps or stages: Based on historical auction data, a success rate model is constructed or invoked. In one embodiment, the model represents the auction success probability as a function of the bid value, specifically using a Weibull distribution characterized by the shape parameter k and the scale parameter λ (as shown in module 206). A KPI model is built or invoked based on historical auction data. In one embodiment, the model characterizes the impact of the winning auction on specific KPI indicators (such as the conversion probability pconv, as shown in module 208). Generate inventory forecast results, that is, predict the joint histogram of the output results of the transaction rate model and the KPI model (i.e., predict the joint histogram of k, λ, and pconv, as shown in module 210). Using inventory forecast results, a strategy search is performed to determine a bidding strategy that optimizes a specific KPI indicator. The bidding strategy includes: (1) a bidding function, which determines the auction value based on the output of the transaction rate model and the KPI model, and is used to calculate the value-volume ratio of the auction; (2) a value-volume ratio threshold, which is used to determine whether to participate in the bidding (as shown in module 212). In some embodiments, the strategy search includes an outer process and an inner process: the outer process searches for possible bidding functions, and the inner process evaluates the performance of each bidding function using inventory forecast results. Monitor model performance. If the solution obtained from strategy search is unacceptable, retrain one or more models. If the solution is acceptable, deploy the determined bidding strategy, apply the above bidding function, and determine whether to participate in the bidding based on the value-volume ratio threshold (as shown in module 214). If the model performance is acceptable, deploy the bidding strategy; if the model performance is unacceptable, transfer control to the model retraining (updating) control module and / or the inventory forecast update generation control module. In one embodiment, model retraining can introduce a feedback mechanism, using underperforming cases as training data and inputting them into the machine learning algorithm; In some embodiments, policy deployment may include one or more of the following steps, phases, operations, functions, or processes (based on the KPI indicators and heuristic algorithms assumed above): Deployment on the demand-side platform: In this scenario, when the demand-side platform receives a bid request, the following steps can be performed: 1) Map bid requests to a set of features using an online feature library; 2) Execute the reasoning process of the conversion rate model and KPI model; 3) Calculate the value using the function F and parameter θ selected in the strategy search process; 4) Calculate the value-to-quantity ratio; 5) If the calculated ratio exceeds the threshold, then participate in the bidding; otherwise, abandon the bidding.

[0137] The above steps typically need to be completed within milliseconds and must be executed on all bid requests that meet the delivery strategy criteria (in some cases, millions of times per second). Alternatively, the same (or equivalent) logic can be used to mark bid requests instead of directly participating in the auction. For example, based on the bid value and value-volume ratio obtained from the steps disclosed herein, bid requests can be marked as specific transaction identifiers and included in the corresponding private trading market.

[0138] In some embodiments, the functions and services provided by the systems, apparatuses, and methods of this disclosure can be made available to multiple users through accounts maintained by an access server or service platform. Such servers or service platforms may be referred to as Software as a Service (SaaS) platforms. Figures 3, 4, and 5 are architectural schematic diagrams, exemplarily illustrating a multi-tenant platform or SaaS platform architecture that can be used to implement embodiments of the systems, apparatuses, and methods of this disclosure. Figure 3 is a schematic diagram of a SaaS system that can implement embodiments of the present invention; Figure 4 is a schematic diagram of the components of an exemplary operating environment that can implement embodiments of the present invention; Figure 5 is a detailed schematic diagram of the components of the multi-tenant distributed computing service platform in Figure 4.

[0139] In some embodiments, the system or service used in this disclosure to determine and execute the optimal bidding strategy can be implemented as a microservice, process, workflow, or function that is executed after receiving a set of input data. Such microservices, processes, workflows, or functions can be executed by a server, data processing element, platform, or system.

[0140] In some embodiments, data processing and other services may be provided by a service platform deployed in the "cloud," accessible via application programming interfaces (APIs) and software development kits (SDKs). The functions, processes, and capabilities disclosed and / or described in this disclosure (in some cases, in conjunction with the accompanying drawings) may be provided as microservices within this platform, with interfaces defined via Representational State Transmission (REST) ​​and GraphQL endpoints. A management console allows users or administrators secure access to the underlying request and response data, management accounts and access permissions, and in some cases, modification of processing workflows or configurations.

[0141] It should be noted that although Figures 3, 4 and 5 illustrate a multi-tenant or SaaS architecture that provides business-related or other applications and services to multiple accounts / users, this architecture can also be used to provide other types of data processing services and application access, for example, to implement one or more processes, functions, features or operations disclosed and / or described in this disclosure.

[0142] In some embodiments, the platform or system shown in the accompanying drawings and described in this disclosure may be operated by an entity that provides a particular set of services or applications; in another embodiment, the platform may be operated by a first entity, while another entity provides applications or services to users through the platform.

[0143] Figure 3 is a schematic diagram of a system 300 that can implement embodiments of the present invention or access embodiments of the services disclosed herein. Leveraging the advantages of commercial service systems hosted by application service providers (ASPs), such as multi-tenant data processing platforms, users of this service may include individuals, enterprises, or organizations.

[0144] Users can access this service through applicable clients. Typically, client devices with internet access can provide the platform with data to be processed and evaluated, and can also obtain output information or execute relevant policies based on the data processing results. Users interact with the service platform via the internet 308 or other applicable communication networks or network combinations. Applicable client devices include desktop computers 303, smartphones 304, tablets 305, or laptops 306, etc.

[0145] System 310 may include a set of data analysis and other related services 312 to assist in determining and / or executing the optimal bidding strategy for online repeated first-level price auctions, and also includes a web interface server 314. The connection relationship of the components is shown in Figure 3. It should be noted that although the data analysis and other services 312 and the web interface server 314 are presented as a single unit in Figure 3, they can be deployed individually or simultaneously on one or more different hardware systems and components. Service 312 may include one or more data processing and model building related functions or operations to generate and / or execute the optimal bidding strategy.

[0146] For example, in some embodiments, the functions, operations, or services provided by platform or system 310 may include: Account management service 316, including but not limited to: Provide users with an authentication process or service; A process or service that generates containers or instances of data analysis and network assessment services for users.

[0147] Data processing services 318 include, but are not limited to: A success rate model is retrieved or constructed based on historical auction data. In one embodiment, the model uses a Weibull distribution characterized by shape parameter k and scale parameter λ to represent the auction success probability as a function of the bid value. Key Performance Indicator (KPI) models are retrieved or constructed based on historical auction data. In one embodiment, this model is used to characterize the impact of winning auctions on specific KPIs (such as conversion probability pconv or other indicators). Generate inventory forecast results, which involves forecasting the joint histogram of the outputs of the turnover rate model and the key performance indicator model (i.e., forecasting the joint histogram of parameters k, λ, and pconv). The strategy search is performed using inventory forecast results to determine the bidding strategy that optimizes specific key performance indicators. The bidding strategy includes: (1) a bidding function, which determines the auction value based on the output of the transaction rate model and the key performance indicator model, and is used to calculate the value-volume ratio of the auction; (2) a value-volume ratio threshold, which is used to determine whether to participate in the auction.

[0148] In one embodiment, the strategy search described above includes an outer process and an inner process: the outer process searches for possible bidding functions, and the inner process evaluates the performance of each bidding function using inventory forecast results. Monitor model performance. If the solution obtained from the strategy search is unacceptable, retrain one or more models; if the solution is acceptable, deploy the determined bidding strategy, apply the above bidding function, and determine whether to participate in the bidding based on the value-volume ratio threshold.

[0149] Management service 320, including but not limited to: Processes or services that provide management support for platforms and services, such as enabling service and / or platform providers to manage and configure processes and services offered to users.

[0150] The platform or system shown in Figure 3 can be deployed on a distributed computing system, which contains at least one server, and in practice, usually multiple servers. A server is a dedicated physical computer that provides data storage and a runtime environment for one or more software applications or services to meet the needs of other computer users communicating with that server. Communication can be achieved through public networks such as the Internet. The server providing the service is called the "host," and the remote computer receiving the service and the software application running on it are called the "client." Depending on the type of computing service provided, servers can be classified as database servers, data storage servers, file servers, mail servers, print servers, or web servers, etc.

[0151] Figure 4 is a schematic diagram of the components of an exemplary operating environment 400, in which embodiments of the present invention can be implemented. As shown, various clients 402 integrated into or combined with various computing devices can communicate with a multi-tenant service platform 408 through one or more networks 414. For example, the clients can be integrated into or combined with client applications (such as software), which are implemented by at least one or more computing devices.

[0152] Applicable computing devices include personal computers, server computers 404, desktop computers 406, laptop computers 407, handheld computers, tablet computers or personal digital assistants 410, smartphones 412, ordinary mobile phones, and consumer electronic devices that integrate one or more computing device components, such as electronic processors, microprocessors, central processing units, or controllers. Applicable networks 414 include networks employing wired and / or wireless communication technologies, as well as networks operating in accordance with any applicable network and / or communication protocols (such as the Internet).

[0153] The distributed computing service / platform (also known as a multi-tenant data processing platform) 408 may contain multiple processing layers, including a user interface layer 416, an application server layer 420, and a data storage layer 424. The user interface layer 416 may maintain multiple user interfaces 417, including graphical user interfaces and / or web-based interfaces. These user interfaces include the platform's default user interface, used by platform users or "tenants" to access applications and data (labeled "Platform User Interface" in the diagram), and one or more dedicated user interfaces customized according to specific user needs (labeled "Tenant A User Interface"..."Tenant Z User Interface" in the diagram), which can be accessed through one or more application programming interfaces (APIs).

[0154] The platform's default user interface includes various interface components. Tenants or platform administrators can use these components to manage tenants' access to and usage rights of the functions provided by the service platform, including accessing tenant data, launching instances of specific applications, or performing specific data processing operations.

[0155] Each application server or processing layer 422 shown in the diagram can be implemented by a set of computers and / or components including computer servers and processors, which can perform various functions, methods, processes, or operations by executing software applications or instruction sets. The data storage layer 424 may include one or more data stores, including platform data store 425 and one or more tenant data stores 426. The data stores can be implemented using any applicable data storage technology, including relational database management systems based on Structured Query Language.

[0156] Service Platform 408 is a multi-tenant architecture, allowing the operating entity to provide a set of business-related or other types of data processing applications, data storage services, and related functions to multiple tenants. For example, these applications and functions may include providing web-based access to enterprises, enabling them to provide services to end users, allowing users with a browser and internet or intranet access to view, enter, process, or modify specific types of information.

[0157] The aforementioned functions or applications are typically implemented by one or more software code / instruction modules, which are stored in and run by one or more servers 422 of the platform application server layer 420. As described above in Figure 3, the platform system shown in Figure 4 can be deployed on a distributed computing system, which contains at least one server, and in practice, usually multiple servers.

[0158] As mentioned above, businesses do not need to build and maintain such platforms or systems themselves; they can directly use systems provided by other entities (such as third parties). Operating entities can build commercial systems / platforms within a multi-tenant platform architecture, providing multiple users with independent instances of one or more data processing workflows (such as the data analysis, optimal bidding strategy generation, and / or execution services disclosed and / or described in this disclosure). Each business (such as an advertiser or demand-side platform) is a tenant of this platform.

[0159] One advantage of a multi-tenant platform is that each tenant can customize data processing workflow instances based on their specific business needs or operational methods. Each tenant can be an enterprise or organization that provides business services and related functions to multiple users through the multi-tenant platform.

[0160] Figure 5 is a detailed schematic diagram of the components of the multi-tenant distributed computing service platform shown in Figure 4, in which embodiments of the present invention can be implemented. Typically, embodiments of the present invention can be implemented through a set of software instructions designed to run on appropriately programmed processing elements (such as central processing units, microprocessors, processors, controllers, state machines, or computing devices, etc., which are not intended to limit the scope of this example). In complex systems, these instructions are typically organized into "modules," each module performing a specific task, process, function, or operation, with the operation of all modules controlled and coordinated by an operating system or other form of management platform.

[0161] As previously described, Figure 5 is a detailed schematic diagram of the component 500 of a multi-tenant distributed computing service platform, in which embodiments of the present invention can be implemented. This exemplary architecture includes a user interface layer 502, within which one or more user interfaces 503 are provided, including a graphical user interface and an application programming interface. Each user interface may include one or more interface elements 504, through which users can interact to access the functions and / or data provided by the application layer and / or data storage layer in this exemplary architecture.

[0162] Graphical user interface (GUI) components include buttons, menus, checkboxes, drop-down lists, scroll bars, sliders, adjusters, text boxes, icons, labels, progress bars, status bars, toolbars, windows, hyperlinks, and dialog boxes. Application programming interfaces (APIs) can be local or remote, and their components include parameterized procedure calls, programmatic objects, and message passing protocols.

[0163] Application layer 510 may include one or more application modules 511, each application module having one or more sub-modules 512. Each application module 511 or sub-module 512 corresponds to a function, method, process, or operation it implements (e.g., providing data processing and service-related functions or processes for platform users). Such functions, methods, processes, or operations may include steps for implementing aspects of the systems and methods disclosed and / or described in this disclosure, such as one or more processes or functions described above in conjunction with the accompanying drawings: A success rate model is retrieved or constructed based on historical auction data. In one embodiment, the model uses a Weibull distribution characterized by shape parameter k and scale parameter λ to represent the auction success probability as a function of the bid value. Key performance indicator (KPI) models are retrieved or constructed based on historical auction data. In one embodiment, the model is used to characterize the impact of winning bids on specific KPIs (such as conversion probability pconv or other indicators). Generate inventory forecast results, which involves forecasting the joint histogram of the outputs of the turnover rate model and the key performance indicator model (i.e., forecasting the joint histogram of parameters k, λ, and pconv). The strategy search is performed using inventory forecast results to determine the bidding strategy that optimizes specific key performance indicators. The bidding strategy includes: (1) a bidding function, which determines the auction value based on the output of the transaction rate model and the key performance indicator model, and is used to calculate the value-volume ratio of the auction; (2) a value-volume ratio threshold, which is used to determine whether to participate in the auction.

[0164] In one embodiment, the strategy search described above includes an outer process and an inner process: the outer process searches for possible bidding functions, and the inner process evaluates the performance of each bidding function. Monitor model performance. If the solution obtained from the strategy search is unacceptable, retrain one or more models; if the solution is acceptable, deploy the determined bidding strategy, apply the above bidding function, and determine whether to participate in the bidding based on the value-volume ratio threshold.

[0165] Application modules and / or submodules may contain applicable computer-executable code or instruction sets (e.g., code that can run on a properly programmed processor, microprocessor, or central processing unit), such as computer-executable code corresponding to a programming language. For example, the source code of a programming language may be compiled into computer-executable code; an interpreted programming language, such as a scripting language, may also be used directly.

[0166] Each application server (as shown in element 422 in Figure 4) can contain all application modules, or different application servers can each contain different sets of application modules. These sets can be non-overlapping or overlapping.

[0167] The data storage layer 520 may contain one or more data objects 522, each data object having one or more data object components 521, such as attributes and / or behaviors. For example, a data object may correspond to a data table in a relational database, and data object components may correspond to columns or fields in the data table; it may also correspond to a data record containing fields and related services; or it may correspond to a persistent instance of a programmatic data object such as a structure or class. Each data storage device in the data storage layer may contain all data objects, or different data storage devices may each contain different sets of data objects, and these sets may be non-overlapping or overlapping.

[0168] It should be noted that the exemplary computing environments shown in Figures 3, 4, and 5 are not limiting examples of the present invention. Embodiments of the present invention can also be implemented, in whole or in part, in other environments, including various devices (including mobile devices), software applications, systems, apparatuses, networks, software-as-a-service platforms, infrastructure-as-a-service platforms, or other configurable components that allow multiple users to perform operations such as data entry, data processing, application execution, or data auditing.

[0169] This disclosure includes the following terms and embodiments: 1. A method for generating a repeating first-level price auction bidding strategy, comprising: Retrieve the success rate model constructed based on historical auction data, which expresses the probability of winning an auction as a function of the bid value; Retrieve the key performance indicator (KPI) model constructed based on historical auction data. The KPI model is used to characterize the impact of winning auctions on specific key performance indicators. Generate inventory forecast results, which involves forecasting the combined histogram of the outputs from the turnover rate model and the key performance indicator model. The strategy search process is executed to determine the bidding strategy used to optimize the key performance indicators, and to identify the auction sessions to participate in and the corresponding bid value. A determined bidding strategy is deployed, which includes a bidding function, an auction value-volume ratio, and a value-volume ratio threshold. The bidding function determines the auction value based on the output of the transaction rate model and the key performance indicator model. The value-volume ratio threshold is used to determine whether to participate in the auction.

[0170] 2. The method described in Clause 1, wherein the value-to-volume ratio of the auction is the ratio of the key performance indicator value to the value of the auctioned item.

[0171] 3. The method according to Clause 1, wherein the strategy search comprises an outer process and an inner process, the outer process searching for possible bidding functions and the inner process evaluating the performance of each bidding function.

[0172] 4. The method according to Clause 1, wherein the success rate model is constructed by applying a machine learning algorithm to at least a portion of historical auction data.

[0173] 5. The method according to Clause 1, wherein the key performance indicator model is constructed by applying a machine learning algorithm to at least a portion of historical auction data.

[0174] 6. The method according to Clause 5, wherein the key performance indicator model is constructed using conversion probability as the key performance indicator.

[0175] 7. The method according to Clause 1, wherein the inventory forecast result is a prediction of the distribution of the output results of the clearance rate model and the key performance indicator model in future auctions based on historical auction data.

[0176] 8. The method according to Clause 1, wherein the performance of a given bidding strategy is simulated using inventory forecast results, and a set of parameters is selected to optimize the simulation performance.

[0177] 9. The method according to Clause 8, wherein a bidding strategy is deployed using selected parameters to determine the auction value and whether to participate in the auction in an online and near real-time manner.

[0178] 10. The method according to Clause 1, wherein the success rate model constructed based on historical auction data expresses the auction winning probability as a function of the bid value using a Weibull distribution characterized by shape parameter k and scale parameter λ, where k is the shape parameter of the distribution and λ is the scale parameter of the distribution.

[0179] 11. The method according to Clause 1, wherein the specific key performance indicator is a combination of multiple indicators, and the key performance indicator model is used to generate the impact value of the winning bid auction on each combined indicator.

[0180] 12. The method described in Clause 1 further includes monitoring the performance of the closing rate model and / or key performance indicator model: if the model performance meets the requirements, deploying a determined bidding strategy; if the model performance does not meet the requirements, transferring control to a pre-defined process or component to control the retraining of the closing rate model and / or key performance indicator model, or to control the generation of updated inventory forecast results.

[0181] 13. A system comprising: One or more electronic processors configured to execute a set of computer-executable instructions; One or more non-volatile electronic data storage media storing the computer-executable instructions, which, when executed, cause one or more electronic processors to perform the following operations: Retrieve the success rate model constructed based on historical auction data, which expresses the probability of winning an auction as a function of the bid value; Retrieve the key performance indicator (KPI) model constructed based on historical auction data. The KPI model is used to characterize the impact of winning auctions on specific key performance indicators. Generate inventory forecast results, which involves forecasting the combined histogram of the outputs from the turnover rate model and the key performance indicator model. The strategy search process is executed to determine the bidding strategy used to optimize the key performance indicators, and to identify the auction sessions to participate in and the corresponding bid value. A determined bidding strategy is deployed, which includes a bidding function, an auction value-volume ratio, and a value-volume ratio threshold. The bidding function determines the auction value based on the output of the transaction rate model and the key performance indicator model. The value-volume ratio threshold is used to determine whether to participate in the auction.

[0182] 14. One or more non-volatile computer-readable media storing a set of computer-executable instructions, which, when executed by one or more programmed electronic processors, cause the processors to perform the following operations: Retrieve the success rate model constructed based on historical auction data, which expresses the probability of winning an auction as a function of the bid value; Retrieve the key performance indicator (KPI) model constructed based on historical auction data. The KPI model is used to characterize the impact of winning auctions on specific key performance indicators. Generate inventory forecast results, which involves forecasting the combined histogram of the outputs from the turnover rate model and the key performance indicator model. The strategy search process is executed to determine the bidding strategy used to optimize the key performance indicators, and to identify the auction sessions to participate in and the corresponding bid value. A determined bidding strategy is deployed, which includes a bidding function, an auction value-volume ratio, and a value-volume ratio threshold. The bidding function determines the auction value based on the output of the transaction rate model and the key performance indicator model. The value-volume ratio threshold is used to determine whether to participate in the auction.

[0183] The embodiments disclosed and / or described herein can be implemented as control logic in a modular or integrated form using computer software. Based on the disclosure and technical teachings provided herein, those skilled in the art can recognize and employ other methods to implement one or more embodiments of the present invention in hardware or a combination of hardware and software.

[0184] In some embodiments, some of the methods, models, or functions disclosed and / or described in this disclosure may be implemented by a trained neural network or machine learning model that operates by executing a set of computer-executable instructions or otherwise. The instruction set may be stored in a non-volatile computer-readable medium and executed by a programmed processor or processing element.

[0185] The instruction set can be provided to users via instruction transmission or by an application that executes the instruction set (e.g., through a network such as the Internet). End users can use the instruction set or application by accessing a Software as a Service (SaaS) platform or the services provided by that platform. The specific implementation of the method, model, or function can be used to define one or more operations, functions, processes, or methods in the development, training, and operation of neural networks, the application of machine learning techniques, and the development and implementation of corresponding decision-making processes.

[0186] It should be noted that a neural network or deep learning model can be represented as a data structure that stores data representing a multi-layered network containing nodes. Connections are established between nodes in different layers, and decision results or values ​​are output by performing operations on the input data.

[0187] In general, a neural network can be viewed as a system composed of interconnected artificial "neurons" that can transmit information between each other. The connections between neurons have numerical weights, which are "tuned" during training so that the trained network can make an accurate response when it receives an image or pattern to be recognized.

[0188] From this perspective, a neural network consists of multiple layers of feature detection "neurons," each responding to different combinations of inputs from the layer above. The network training process requires a labeled input dataset containing representative input patterns of various types, each associated with a predicted output response. Training iteratively determines the weights of the feature neurons in intermediate and final layers using a general method. From a computational model perspective, each neuron calculates the dot product of the input and weights, adds a bias term, and then obtains the output through a non-linear triggering function or activation function (such as a sigmoid response function).

[0189] Machine learning models are composed of multiple interconnected neurons that infer or generate decision results (such as classification results) from input data samples. The typical training method involves inputting multiple sets of input data samples, along with the correct "response" or decision result for each set. That is, each input data sample is associated with a label or other identifier to indicate the correct response the trained model should output. After training the model with these samples and labels, when the connection weights between neurons converge and stabilize, or fluctuate within an acceptable range, the model can output the "correct" label or classification result based on the input data samples.

[0190] Based on this, as a non-limiting example, the function F in this disclosure can be implemented by a neural network, the turnover rate model and the key performance indicator model can also be implemented by a neural network, and the advanced time series forecasting method based on inventory forecasting can also be implemented by one or more neural networks.

[0191] In some examples, neural networks can be implemented based on a variety of different topologies and / or architectures, including deep neural networks with fully connected layers (such as dense layers), long short-term memory layers, convolutional layers, temporal convolutional layers, and other applicable deep neural network topologies and / or architectures, or combinations of various architectures.

[0192] The output layer of a neural network can take many forms, including but not limited to: output layers with logistic sigmoid activation functions, output layers with hyperbolic tangent activation functions, linear unit output layers, rectified linear unit output layers, and other applicable nonlinear unit output layers, as well as combinations of various types. In some examples, the neural network can be configured to represent the probability distribution of a single or multiple models.

[0193] One or more software components, processes, or functions disclosed and / or described in this disclosure may be written as software code in any applicable computer language (such as Python, Java, JavaScript, C++, or Perl) using conventional programming techniques or object-oriented programming techniques, and executed by a processor. This software code may be stored as a series of instructions or commands on a non-volatile computer-readable medium such as random access memory, read-only memory, a hard disk, or an optical disc read-only memory.

[0194] In this document, non-volatile computer-readable media refers to almost any medium suitable for storing data or instruction sets, excluding transient waveforms. Such computer-readable media can be deployed on a single computing device or distributed across multiple different computing devices within a system or network. Furthermore, the instruction set can be provided to users (e.g., via instruction transmission or an application that executes the instruction set) through an instruction set, and end users can use the instruction set or application by accessing a Software as a Service (SaaS) platform or services provided by that platform.

[0195] In one exemplary implementation, the processing element or processor described in this document may be a central processing unit (CPU) or a device that can be considered a CPU (such as a virtual machine). In this implementation, the CPU or device integrating a CPU may be connected, coupled, and / or communicate with one or more peripheral devices such as a display. In another exemplary implementation, the processing element or processor may be integrated into a mobile computing device such as a smartphone or tablet.

[0196] The non-volatile computer-readable storage media described in this document may include a variety of physical storage devices, including independent redundant disk arrays, flash memory, universal serial bus flash drives, external hard drives, thumb flash drives, pen flash drives, key fob flash drives, high-density digital video optical disc drives, internal hard drives, Blu-ray disc drives, holographic digital data storage optical disc drives, synchronous dynamic random access memory, and other similar devices or storage forms based on similar technologies.

[0197] Such computer-readable storage media enable processing elements or processors to access computer-executable process steps and applications stored in removable and non-removable storage media, enabling data export or import of the device. As described above in the embodiments of this disclosure, non-volatile computer-readable media refers to almost all storage media employing any structure, technology, or method, except for transient waveforms or similar media.

[0198] This document describes some implementations of the technology disclosed herein, in conjunction with system block diagrams and / or flowcharts of functions, operations, processes, and methods. It should be understood that one or more modules in the block diagram, one or more steps in the flowchart, and any combination of block diagram modules and flowchart steps can be implemented using computer-executable program instructions. It should be noted that in some embodiments, some modules in the block diagram or some steps in the flowchart do not necessarily need to be executed in the order described herein, and may even be omitted.

[0199] The computer-executable program instructions can be loaded into a general-purpose computer, a special-purpose computer, a processor, or other programmable data processing device to create a special-purpose device, such that the instructions, which execute on the computer, processor, or other programmable data processing device, can implement one or more of the functions, operations, processes, or methods described in this document. Alternatively, the computer program instructions can be stored in a computer-readable storage medium, which can cause the computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium can produce an article of manufacture comprising instruction means for implementing one or more of the functions, operations, processes, or methods disclosed and / or described in this disclosure.

[0200] Although this disclosure has described in conjunction with what are currently considered to be the most practical and diverse implementations, it should be understood that this disclosure is not limited to the disclosed embodiments; rather, it is intended to cover various modifications and equivalent configurations falling within the scope of the appended claims. While specific terminology is used in this document, such terminology is for general and descriptive purposes only and does not constitute limitation.

[0201] This document describes some implementations of this disclosure through examples, enabling those skilled in the art to implement the relevant solutions of this disclosure, including manufacturing and using the relevant equipment or system, and performing the included methods. The patent protection scope of some implementations of this disclosure is defined by the claims, and may also include other examples that may occur to those skilled in the art. If the structure and / or functional elements of these examples are not substantially different from the wording of the claims, or if their structure and / or functional elements are not substantially different from the wording of the claims, then these examples shall all fall within the protection scope of the claims.

[0202] All references cited in this document, including publications, patent applications and patents, are incorporated into this document in their entirety by reference to the same extent that each reference is individually and explicitly stated to be incorporated by reference or fully elaborated in this document.

[0203] Unless otherwise stated herein, or in obvious contradiction in the context, terms such as “a,” “an,” “the,” and “the” used in this specification and subsequent claims shall be understood to cover both singular and plural forms. Unless otherwise stated herein, terms such as “having,” “comprising,” and “including” used in this specification and claims shall be understood as open-ended terms, meaning “including but not limited to.”

[0204] Unless otherwise stated herein, the numerical ranges described herein are for convenience only, and each individual value is considered to have been separately described herein. Unless otherwise stated herein, or in obvious contradiction in the context, all methods described herein may be performed in any applicable order.

[0205] All examples or exemplary statements provided in this document (such as "for example") are merely for the purpose of more clearly illustrating embodiments of this disclosure and, unless otherwise claimed, do not constitute a limitation on the scope of protection of this invention. Nothing in this specification should be construed as treating elements not recited in the claims as essential elements of the embodiments of this disclosure.

[0206] The term "or" as used in this specification, drawings, and claims is an inclusive term that refers to both alternative relationships and parallel relationships.

[0207] The components shown in the accompanying drawings and those described above may be arranged in different ways, and additional components and steps not shown or described may be added. Similarly, some features and sub-combinations of this disclosure have independent practical value and can be used independently of other features and sub-combinations. The embodiments of this disclosure are merely illustrative and do not constitute a limitation; readers in the art can conceive of other alternative embodiments. Therefore, this disclosure is not limited to the embodiments described in the drawings, and various implementations and modifications can be made to this invention without departing from the scope of the following claims.

Claims

1. A method for generating a repeating first-level price auction bidding strategy, comprising the following steps: Retrieve the success rate model constructed based on historical auction data, which expresses the probability of winning an auction as a function of the bid value; Retrieve a key performance indicator (KPI) model constructed based on historical auction data, wherein the KPI model characterizes the impact of winning auctions on specific key performance indicators; Generate inventory forecast results, that is, predict the joint histogram of the output results of the transaction rate model and the key performance indicator model; The strategy search process is executed to determine a bidding strategy that specifies the auction sessions to be bid on and the corresponding bid value, in order to optimize the key performance indicators. The bidding strategy is determined by the deployment, wherein the bidding strategy includes a bidding function, an auction value-volume ratio, and a value-volume ratio threshold. The bidding function determines the auction value based on the output of the transaction rate model and the key performance indicator model. The value-volume ratio threshold is used to determine whether to participate in the auction.

2. The method according to claim 1, characterized in that, The value-to-quantity ratio of the auction is the ratio of the value of the key performance indicator to the value of the auctioned item.

3. The method according to claim 1, characterized in that, The strategy search process includes an outer process and an inner process. The outer process searches for various possible bidding functions, while the inner process evaluates the performance of each bidding function.

4. The method according to claim 1, characterized in that, The transaction rate model is constructed by applying machine learning algorithms to at least a portion of historical auction data.

5. The method according to claim 1, characterized in that, The key performance indicator model was constructed by applying machine learning algorithms to at least a portion of historical auction data.

6. The method according to claim 5, characterized in that, The key performance indicator model is constructed using conversion probability as the key performance indicator.

7. The method according to claim 1, characterized in that, The inventory forecast results are predictions based on historical auction data, forecasting the distribution of outputs from the success rate model and key performance indicator model in future auctions.

8. The method according to claim 1, characterized in that, The performance of various parameterized bidding strategies is simulated using the inventory forecast results, and the set of parameters that optimizes the simulation performance is selected.

9. The method according to claim 8, characterized in that, The selected parameters are used to deploy a bidding strategy to determine the auction value and whether to participate in the auction in an online and near real-time manner.

10. The method according to claim 1, characterized in that, The transaction rate model constructed based on historical auction data uses a Weibull distribution characterized by shape parameter k and scale parameter λ to express the auction winning probability as a function of the bid value, where k is the shape parameter of the distribution and λ is the scale parameter of the distribution.

11. The method according to claim 1, characterized in that, The specific key performance indicator is a combination of multiple key performance indicators, and the key performance indicator model is used to generate the impact value of the winning bid auction on each key performance indicator in the combination.

12. The method according to claim 1, characterized in that, It also includes monitoring the performance of the transaction rate model and / or key performance indicator model; if the model performance meets the requirements, the determined bidding strategy is deployed; if the model performance does not meet the requirements, control is transferred to a preset process or component, which controls the retraining of the transaction rate model and / or key performance indicator model, or controls the generation of updated inventory forecast results.

13. A system comprising: One or more electronic processors are configured to execute a set of computer-executable instructions; One or more non-volatile electronic data storage media storing the computer-executable instructions, which, when executed, cause the one or more electronic processors to perform the following operations: Retrieve the success rate model constructed based on historical auction data, which expresses the probability of winning an auction as a function of the bid value; Retrieve a key performance indicator (KPI) model constructed based on historical auction data, wherein the KPI model characterizes the impact of winning auctions on specific key performance indicators; Generate inventory forecast results, that is, predict the joint histogram of the output results of the transaction rate model and the key performance indicator model; The strategy search process is executed to determine a bidding strategy that specifies the auction sessions to be bid on and the corresponding bid value, in order to optimize the key performance indicators. The bidding strategy is determined by the deployment, wherein the bidding strategy includes a bidding function, an auction value-volume ratio, and a value-volume ratio threshold. The bidding function determines the auction value based on the output of the transaction rate model and the key performance indicator model. The value-volume ratio threshold is used to determine whether to participate in the auction.

14. The system according to claim 13, characterized in that, The value-to-quantity ratio of the auction is the ratio of the value of the key performance indicator to the value of the auctioned item.

15. The system according to claim 13, characterized in that, The strategy search process includes an outer process and an inner process. The outer process searches for various possible bidding functions, and the inner process evaluates the performance of each bidding function. The conversion rate model is constructed by applying machine learning algorithms to at least a portion of historical auction data. The key performance indicator model is constructed by applying machine learning algorithms to at least a portion of historical auction data, and the key performance indicator model is constructed using conversion probability as the key performance indicator.

16. The system according to claim 13, characterized in that, The transaction rate model constructed based on historical auction data uses a Weibull distribution characterized by shape parameter k and scale parameter λ to express the auction winning probability as a function of the bid value, where k is the shape parameter of the distribution and λ is the scale parameter of the distribution.

17. One or more non-volatile computer-readable media storing a set of computer-executable instructions, which, when executed by one or more programmed electronic processors, cause the processors to perform the following operations: Retrieve the success rate model constructed based on historical auction data, which expresses the probability of winning an auction as a function of the bid value; Retrieve a key performance indicator (KPI) model constructed based on historical auction data, wherein the KPI model characterizes the impact of winning auctions on specific key performance indicators; Generate inventory forecast results, that is, predict the joint histogram of the output results of the transaction rate model and the key performance indicator model; The strategy search process is executed to determine a bidding strategy that specifies the auction sessions to be bid on and the corresponding bid value, in order to optimize the key performance indicators. The bidding strategy is determined by the deployment, wherein the bidding strategy includes a bidding function, an auction value-volume ratio, and a value-volume ratio threshold. The bidding function determines the auction value based on the output of the transaction rate model and the key performance indicator model. The value-volume ratio threshold is used to determine whether to participate in the auction.

18. One or more non-volatile computer-readable media according to claim 17, characterized in that, The value-to-quantity ratio of the auction is the ratio of the value of the key performance indicator to the value of the auctioned item.

19. One or more non-volatile computer-readable media according to claim 17, characterized in that, The strategy search process includes an outer process and an inner process. The outer process searches for various possible bidding functions, and the inner process evaluates the performance of each bidding function. The conversion rate model is constructed by applying machine learning algorithms to at least a portion of historical auction data. The key performance indicator model is constructed by applying machine learning algorithms to at least a portion of historical auction data, and the key performance indicator model is constructed using conversion probability as the key performance indicator.

20. One or more non-volatile computer-readable media according to claim 17, characterized in that, The transaction rate model constructed based on historical auction data uses a Weibull distribution characterized by shape parameter k and scale parameter λ to express the auction winning probability as a function of the bid value, where k is the shape parameter of the distribution and λ is the scale parameter of the distribution.

Images