RTB optimization method, system and equipment based on pyramid data model

Abstract

The invention provides an RTB optimization method, system and device based on a pyramid data model, and the method comprises the steps: collecting the historical bidding data of online advertisements, carrying out the data preprocessing, generating the original data of each advertisement in each time period, defining a time scale sequence, carrying out the data aggregation processing of the original data, and carrying out the data aggregation processing of the original data; the method comprises the steps of forming pyramid type multi-scale structure data from thin to thick and from bottom to top, constructing a pyramid data model, dynamically selecting a reference time scale, extracting short-term data and long-term data from the pyramid data model, fusing the short-term data and the long-term data with time decay weight, and inputting a bidding optimization model under budget constraint. And generating a real-time bidding strategy of the online advertisement, and carrying out retraining and fine tuning on the bidding optimization model to form closed-loop optimization. According to the invention, by constructing a closed-loop system of data acquisition, multi-scale processing, strategy optimization and effect feedback, important support is provided for improving the advertising effect of advertisers.

Description

Technical Field

[0001] The present invention belongs to the technical field of real-time bidding strategy optimization, and in particular relates to an RTB optimization method, system and device based on a pyramid data model. Background Art

[0002] Real-time bidding (RTB) is a key mechanism in display advertising, allowing advertisers to evaluate and set bids for each ad impression. The core component of RTB is the demand-side platform (DSP), which represents the interests of advertisers and aims to achieve automated optimization of display ads to maximize the effectiveness of ad delivery. In this process, the primary problem faced by DSP is how to design an efficient and real-time bidding strategy to maximize advertising revenue, but the current online advertising bidding scheme still has obvious shortcomings in the following aspects:

[0003] (1) Over-reliance on historical data of a single time scale, failure to reasonably utilize data features of multiple time scales, and lack of understanding and utilization of the long-term laws of advertising placement have resulted in the model’s inability to adapt to market changes and poor reference value;

[0004] (2) Real-time advertising bidding strategies are relatively simple, usually relying on static rules or overly basic algorithms, lacking the ability to flexibly respond to market changes and accurately adjust, and have poor adaptability in a dynamic market environment. They are unable to quickly optimize based on real-time data and advertising effects, which in turn affects the effectiveness of advertising and maximizing revenue.

[0005] (3) There is a general lack of refined management and optimization of budget constraints, making it difficult to effectively balance the relationship between budget allocation and advertising effectiveness during the advertising delivery process and to maximize click-through effects within a limited budget.

[0006] Therefore, we need to develop an RTB optimization method, system and equipment based on the pyramid data model, which can consider multi-scale data and optimize the real-time advertising bidding strategy to realize the intelligence and precision of online advertising delivery, providing important support for improving the advertising effect of advertisers. Summary of the invention

[0007] The purpose of the present invention is to provide an RTB optimization method, system and device based on a pyramid data model to solve the problems mentioned in the above background technology of the existing online advertising bidding scheme, such as a single reference data time scale, a simple bidding strategy, and inaccurate budget control.

[0008] To achieve the above object, on the one hand, the present invention provides an RTB optimization method based on a pyramid data model, and the method is specifically as follows:

[0009] Collect historical bidding data of online advertisements, including advertisement ID, bidding time, bidding amount, advertisement consumption, and click data. After data preprocessing, perform preliminary aggregation according to key dimensions to generate statistical indicators for each advertisement in each time period, which are recorded as raw data.

[0010] Defining a time scale sequence, performing data aggregation processing on the original data according to the time dimension to form pyramid-type multi-scale structure data from fine to coarse and from bottom to top, and constructing a pyramid data model based on the pyramid-type multi-scale structure data;

[0011] Dynamically select a reference time scale based on the characteristics of the current time period, extract short-term data and long-term data from the pyramid data model, fuse them with time-decayed weights, input them into a bidding optimization model under budget constraints, and generate a real-time bidding strategy for online advertising;

[0012] The final effect data of the advertisement delivery is recorded, and the final effect data is regularly used for retraining and fine-tuning the bidding optimization model, updating the model parameters, and forming a closed-loop optimization.

[0013] Based on the above scheme, the step of defining a time scale sequence includes:

[0014] Each layer of the pyramid data model corresponds to a specific time scale, represents the original data under the specific time scale, defines a time scale sequence, defines an aggregation function for each element in the time scale sequence, and aggregates the original data according to the time dimension of each element in the time scale sequence.

[0015] Based on the above scheme, the data aggregation processing includes: based on the time scale sequence and the aggregation function, aggregating the original data layer by layer from bottom to top to form the pyramid multi-scale structure data, and completing the construction of the pyramid data model, wherein the first layer: contains the original data and has the finest granularity.

[0016] Based on the above scheme, the selection of the benchmark time scale is specifically as follows: for each real-time bidding request, n benchmark time scales are determined as key reference objects for subsequent feature extraction and real-time bidding strategy generation. The benchmark time scale is selected based on the first balance principle to ensure the timeliness and stability of the data.

[0017] Based on the above scheme, the extraction of short-term data and long-term data is specifically as follows: after determining the benchmark time scale, the corresponding short-term data and the long-term data are selected from the pyramid data model as inputs of the bidding optimization model, and the specific selection follows the second balance principle, which includes: short-term data selection principle, long-term data selection principle, and timeliness and data volume balance principle.

[0018] Based on the above scheme, the fusion with time decay weight is specifically: based on the time decay weight, the weight of each layer in the short-term data and the long-term data is calculated by a decreasing function, and an exponential decay function is used for calculation. The formula is as follows: ,in For the current moment, is the timestamp of the data, is the decay rate parameter; for each layer The data is given weight ,and > > ... > , is the weight given to the i-th layer of data, is the weight given to the nth layer of data; let The data vector is , the data of each layer The data vector Multiply by the corresponding weight Get the weighted data vector of each layer , and merge the weighted data vectors of all layers as the final feature input of the bidding optimization model: .

[0019] Based on the above solution, the construction of the bidding optimization model includes:

[0020] A model of budget constraint and click-through rate is established as the bidding optimization model. The goal of bidding optimization is to maximize the number of clicks on the advertisement under the budget constraint, which can be described as the following bidding optimization problem: ,in is the bid price of the jth auction, is the estimated click rate of the j-th bidding, For the total budget.

[0021] Based on the above solution, the retraining and fine-tuning of the bidding optimization model specifically includes:

[0022] Cleaning, merging, and organizing the final effect data to generate a structured effect data set;

[0023] Set retraining trigger conditions, including daily or weekly regular triggers, and triggers when data increments reach a certain scale;

[0024] Using a stratified sampling method, the latest effect data set is divided into a training set and a validation set, and the bidding optimization model is retrained to obtain updated model parameters;

[0025] Evaluating the prediction effect of the bidding optimization model after updating the model parameters through the validation set;

[0026] Perform version management and update of the bidding optimization model parameters.

[0027] On the one hand, the present invention provides an RTB optimization system based on a pyramid data model, comprising:

[0028] The data collection and preprocessing module is used to collect historical bidding data of online advertisements, including advertisement ID, bidding time, bidding amount, advertisement consumption, and click data. After data preprocessing, the historical bidding data is initially aggregated according to key dimensions to generate statistical indicators for each advertisement in each time period, which are recorded as raw data.

[0029] A pyramid data model construction module is used to define a time scale sequence, perform data aggregation processing on the original data according to the time dimension, form pyramid-type multi-scale structure data from fine to coarse and from bottom to top, and construct a pyramid data model based on the pyramid-type multi-scale structure data;

[0030] A real-time bidding strategy generation module is used to dynamically select a reference time scale based on the characteristics of the current time period, extract short-term data and long-term data from the pyramid data model, fuse them with time-decayed weights, input them into a bidding optimization model under budget constraints, and generate a real-time bidding strategy for online advertising;

[0031] The bidding optimization closed-loop module is used to record the final effect data of advertising delivery, and regularly use the final effect data to retrain and fine-tune the bidding optimization model, update model parameters, and form a closed-loop optimization.

[0032] On the other hand, the present invention also provides an RTB optimization device based on a pyramid data model, comprising:

[0033] Memory, used to store programs;

[0034] The processor is used to load a program to execute the RTB optimization method based on the pyramid data model described in the above solution.

[0035] Compared with the prior art, the present invention has the following advantages and effects:

[0036] (1) It adopts a pyramid-like multi-scale data processing method, which can effectively integrate and analyze data from different time scales, capture the long-term laws and immediate changes of advertising in different time periods, and make advertising strategies more dynamically adjustable;

[0037] (2) A bidding optimization method based on real-time data analysis was proposed. The click-through rate prediction model was used to solve the optimal click-through rate prediction value, and the click-through rate of advertisements was predicted in real time. This provided a basis for dynamically adjusting the bidding amount and improved the accuracy and effectiveness of advertising delivery.

[0038] (3) A bidding optimization model based on budget constraints is proposed. The time-decayed weighted fusion method and the Lagrange multiplier method are introduced to manage the budget in a refined manner, ensuring that the total cost of advertising does not exceed the set budget and maximizing the advertising effect within the budget constraint. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for the specific embodiments or the description of the prior art. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn according to the actual scale.

[0040] Figure 1 It is a flow chart of an RTB optimization method based on a pyramid data model provided by an embodiment of the present invention;

[0041] Figure 2 It is a structural diagram of an RTB optimization system based on a pyramid data model provided by an embodiment of the present invention;

[0042] Figure 3 It is a structural diagram of an RTB optimization device based on a pyramid data model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0043] In order to more clearly explain the purpose, technical solutions and advantages of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all of the embodiments. The example implementation methods can be implemented in various forms and should not be understood as being limited to the examples described herein. On the contrary, these implementation methods are provided to make the present invention more comprehensive and complete, and to fully convey the concepts of the example implementation methods to those skilled in the art.

[0044] In addition, the described features, structures or characteristics may be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present invention. However, those skilled in the art will appreciate that the technical solution of the present invention can be practiced without one or more of the specific details, or other methods, components, devices, steps, etc. may be adopted. In other cases, known methods, devices, implementations or operations are not shown or described in detail to avoid blurring various aspects of the present invention.

[0045] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0046] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to actual conditions.

[0047] The present invention is described in detail below in conjunction with specific embodiments:

[0048] Example 1

[0049] As attached Figure 1 As shown, Embodiment 1 of the present invention provides an RTB optimization method based on a pyramid data model, and the specific steps of the method are as follows:

[0050] It should be noted that real-time bidding for online advertising generally includes: advertisers register an account on the advertising platform and set up a delivery plan and budget, and provide advertising materials; when a user visits the media's website or application, the media sends a display request to the advertising platform; after receiving the request, the advertising platform triggers real-time bidding and sends a bid request to advertisers that meet the delivery conditions; advertisers return bids and materials based on real-time algorithm strategies; the advertising platform compares the bids of all parties and pushes the winner's advertising content to the media in real time for display; users' clicks and conversion behaviors are recorded for advertising settlement and effect evaluation.

[0051] Step S1, collection and preprocessing of historical bidding data: collecting historical bidding data of online advertisements, including but not limited to advertisement ID, bidding time, bidding amount, advertisement consumption, and click data, and performing data preprocessing on the historical bidding data;

[0052] Step S2, constructing a pyramid data model: defining a time scale sequence, performing data aggregation processing on the original data according to the time dimension, forming pyramid-type multi-scale structure data from fine to coarse and from bottom to top, and constructing a pyramid data model based on the pyramid-type multi-scale structure data;

[0053] Step S3, generating a real-time bidding strategy based on the bidding optimization model: dynamically selecting a reference time scale based on the current time period characteristics, extracting short-term data and long-term data from the pyramid data model, fusing them with time decay weights, and inputting them into the bidding optimization model under budget constraints to generate a real-time bidding strategy for online advertising;

[0054] Step S4, updating the bidding optimization model: recording the final effect data of the advertisement delivery, and regularly using the final effect data for retraining and fine-tuning the bidding optimization model, updating the model parameters, and forming a closed-loop optimization.

[0055] Specifically, the advertising platform will return the result data of each advertising auction, namely the bidding log, to the advertiser. The main fields of the bidding log include but are not limited to: auction_id: the unique ID of the bidding request, which can be associated with the media log; timestamp: the timestamp of the bidding, namely the bidding time, accurate to milliseconds; : The ID of the advertisement participating in the bidding, i.e., the advertisement ID; bid: the bid of the advertiser, i.e., the bidding amount, in cents; cost: the advertising fee that the advertiser needs to pay if the bidding is successful, i.e., the advertising consumption; click: whether the advertisement is clicked, i.e., the click data. When the data value is 0, it means no click, and when it is 1, it means click; conversion: indicates whether a conversion behavior occurs. When the data value is 0, it means no, and when it is 1, it means yes;

[0056] Preferably, in step S1, the historical bidding data is preprocessed, including but not limited to: removing abnormal or incomplete records, including data with negative bids and zero display numbers; encrypting or masking sensitive information, including user IDs, to protect privacy; and preliminarily aggregating the historical bidding data after data preprocessing according to key dimensions to generate statistical indicators for each advertisement in each time period, recorded as raw data, wherein the key dimensions include but are not limited to the advertisement ID, and the statistical indicators include but are not limited to the bidding amount, advertisement consumption, display data, and click data. It should be noted that this step is the basis for building a pyramid data model.

[0057] Specifically, the specific data structure in the original data includes but is not limited to: : The ID of the advertisement participating in the bidding, which is also the unique identifier of the advertisement, is a string type; time: the time dimension of aggregation, which can be aggregated by hour, day or smaller time dimensions. The time format includes but is not limited to yyyymmddhh (year-month-day-time), such as 2024010108 represents the 8th hour of January 1, 2024; impression: display data, the number of times the advertisement is displayed to the user in a certain time period, which is an integer; click: the number of clicks on the advertisement in a certain time period, which refers to the number of times the user clicks on the advertisement, which is an integer; cost: advertising cost, which is a floating point type; ctr: click-through rate, that is, the number of times the advertisement is clicked / the number of impressions, which is a floating point type; cpc: the average click price, cost / number of clicks, which is a floating point type; the raw data can be stored in a relational database, Hive or other structured storage engines to facilitate subsequent efficient retrieval and calculation of the data.

[0058] Preferably, the time scale sequence defined in step S2 is specifically:

[0059] Step S201: Each layer of the pyramid data model corresponds to a specific time scale and represents the original data under the specific time scale. Exemplarily, in this embodiment, a time scale sequence is defined: scale = [1h ， 1d ， 7d ， 30d] , where "1h" represents 1 hour, "1d" represents 1 day, "7d" represents 7 days, i.e., a week, and "30d" represents 30 days, i.e., a month. The specific time scales in the sequence are arranged from small to large, reflecting a pyramid structure from the bottom to the top, and the pyramid data model supports efficient query and analysis under different time dimensions;

[0060] Specifically, the selection of the specific time scale can be adjusted according to the actual business scenario and data volume, including but not limited to: adding a "15min" scale to capture traffic fluctuations in a shorter period, or adding a "1q" scale, i.e., a quarterly scale, to depict trends over a longer span;

[0061] Step S202: for each element in the time scale sequence scale scale[i] , define the aggregate function agg(data ， scale[i]) , the original data ,according to scale[i] Aggregation is performed along the time dimension of scale[0]="1h" For example:

[0062]

[0063]

[0064] It should be noted that is the ad ID; the aggregate time dimension yyyymmddhh is formatted as hourly level (for example, 2024070312 means 12:00 on July 3, 2024), recorded as 𝑡; the total number of ad impressions within the hour is recorded as ; The total number of clicks on the ad during that hour, recorded as ; Advertising consumption during that hour, recorded as ; Click-through rate (CTR), defined as the ratio of clicks to impressions, denoted by ; Cost per click (CPC), defined as the ratio of total consumption to total number of clicks, denoted as ; "|" means "for", that is, for each and t for grouping and aggregation calculation; specifically, the aggregation logic is expressed by a formula statement as follows: the original data data is grouped according to the advertising ID and hour dimension, the total of the display, click, and consumption of each group is calculated, and the aggregated click rate and click price are calculated accordingly; in more implementation cases, the high-level aggregation function can be implemented by adjusting the time dimension (such as day, week, month) based on the results of the lower layer;

[0065] Preferably, the data aggregation processing in step S2 includes:

[0066] Step S203: Based on the time scale sequence scale and the aggregation function, the original data is aggregated layer by layer from bottom to top to form the pyramid multi-scale structure data, and the construction of the pyramid data model is completed. Exemplarily: the first layer ( ): contains the original data, with the finest granularity; the second layer ( ): result data aggregated by hour, with the time dimension of "yyyymmddhh"; the third layer ( ): result data aggregated by day, with the time dimension "yyyymmdd00"; the fourth layer ( ): Result data aggregated by week, with the time dimension "yyyyww0000"; the fifth layer ( ): result data aggregated by month, with the time dimension being "yyyymm0000";

[0067] Specifically, the result data of each layer in the pyramid-type multi-scale structure data are stored in an independent table, including but not limited to "ad_hourly_stats" and "ad_daily_stats", for the convenience of subsequent query and use. The structure of the table is consistent with the Schema of the pyramid-type multi-scale structure data, but the time dimension is different;

[0068] It should be noted that, since data aggregation processing involves data traversal and calculation, when constructing the pyramid data model, the distributed computing capabilities of the big data platform, such as MapReduce, Spark, etc., are fully utilized to improve the aggregation efficiency; at the same time, considering that the data in the pyramid data model will be frequently queried, it is necessary to establish indexes for key fields, such as ( , time) joint index, etc.

[0069] Preferably, step S3 specifically includes:

[0070] Step S301: for each real-time bidding request, n reference time scales scale_ref are determined as key reference objects for subsequent feature extraction and real-time bidding strategy generation. The scale_ref is selected based on the first balance principle to ensure the timeliness and stability of the data. The first balance principle includes but is not limited to:

[0071] (1) Selecting a time scale based on the speed of traffic change: When the traffic or bidding intensity changes in the recent period exceeds the first change threshold, for example, the change rate in the time window exceeds 10%, a smaller time scale is preferred to provide a sensitive response to market fluctuations and facilitate rapid adjustment of bidding strategies. For example, data corresponding to the time scale of "1 hour" or "1 day" is selected to cope with the frequently changing bidding environment;

[0072] When the change in traffic or bidding intensity is lower than the second change threshold, such as a change rate lower than 5%, a larger time scale is selected to smooth the data and enhance the stability of the strategy. For example, data of "7 days" or "30 days" is selected to use stable statistical features in a long period to guide bidding decisions;

[0073] (2) Based on the amount of historical data of the advertisement: For advertisements with more historical data, a higher level time scale is selected to improve the robustness of the strategy with the help of long-term accumulated data; for new advertisements or advertisements with less data, a lower level time scale is selected to refer more to recent performance;

[0074] (3) According to the advertising budget: advertisements with larger budgets can choose higher-level time scales, emphasizing long-term and stable delivery rhythm; advertisements with smaller budgets can choose higher and lower time scales, flexibly seizing short-term opportunities;

[0075] Step S302: After determining the reference time scale scale_ref, select the corresponding short-term data and the long-term data from the pyramid data model as inputs of the bidding optimization model. The specific selection follows the second balance principle, which includes but is not limited to:

[0076] (1) Principle of short-term data selection: Select recent data from the corresponding layer of the benchmark time scale and the layers below it to capture the recent performance of advertisements in each segmented time period. The recent time period is preferably 1 to 7 days in the past. The recent performance includes but is not limited to click-through rate and consumption per hour.

[0077] (2) Principle of long-term data selection: select data above the benchmark time scale to reveal the regular characteristics of advertising over a long period of time, including but not limited to the click distribution of each day of the week and the conversion trend of each hour of the day;

[0078] (3) Principle of balancing timeliness and data volume: The selected time window should be as close to the current time point as possible to reflect the latest market status; at the same time, each advertisement should have sufficient data within the selected window to support statistical analysis.

[0079] It should be noted that the short-term data and the long-term data selected from different layers of the pyramid data model represent historical bidding data of different timeliness and granularity, and need to be smoothly fused. Step S303 will describe the fusion of data through time decay weights.

[0080] Step S303: Based on the time decay weight, the weight of each layer in the short-term data and the long-term data is calculated by a decreasing function. Exemplarily, in this embodiment, an exponential decay function is used for calculation, and the formula is as follows: ,in For the current moment, is the timestamp of the data, is the decay rate parameter; for each layer The data is given weight ,and > > ... > , is the weight given to the i-th layer of data, is the weight given to the data in the nth layer (i.e. the last layer); let The data vector is , the data of each layer The data vector Multiply by the corresponding weight Get the weighted data vector of each layer , and merge the weighted data vectors of all layers as the final feature input of the bidding optimization model: .

[0081] It should be noted that after weighted merging, short-term data has a large weight, which can keenly grasp the current trend; long-term data has a small weight, but can provide a stable background reference; and , the bidding optimization model can flexibly control the attention paid to short-term and long-term signals;

[0082] Step S304: Establish a model of budget constraint and click-through rate as the bidding optimization model. The goal of bidding optimization is to maximize the number of clicks on the advertisement under the budget constraint. For example, it can be described as the following bidding optimization problem: ,in is the bid price of the jth auction, is the estimated click rate of the j-th bidding, is the total budget; the key to solving the bidding optimization problem is the click-through rate estimate In this embodiment, a logistic regression (LR) model is established as a click rate prediction model to estimate the click rate. To make an estimate, the formula is as follows: ,in is the final feature input, is the weight parameter to be learned; the logistic regression (LR) model is trained using maximum likelihood estimation: , which maximizes the click-through rate estimate The accuracy of the bidding optimization problem is solved by the Lagrange multiplier method to obtain the optimal bid value. , specifically: Introduce the Lagrange multiplier , we get the unconstrained optimization problem: , ask L about The partial derivative of and set it equal to zero, that is = 0, we get: ;Will Substitute the Constraints, find , and then get the best bid for the ad ;

[0083] It should be noted that in an actual system, the above optimization process needs to be performed in real time each time a bidding request arrives, which places high demands on engineering implementation. In this embodiment, a method of pre-calculating piecewise functions is adopted to discretize the continuous bidding space and select from a limited number of options, thereby improving computing efficiency while satisfying the effect.

[0084] Step S305: Combine the historical bidding data, bidding optimization problem, budget constraint, and click-through rate estimation model to form an end-to-end real-time bidding strategy generation strategy. Exemplarily, the real-time bidding strategy generation process includes the following:

[0085] The system receives a bidding request and parses the request fields, including but not limited to user ID, ad space, and device information; based on the parsed field values, the system searches for the corresponding historical bidding data in the pyramid data model, extracts and generates the final feature input x; based on the click-through rate prediction model, the system uses the input x to predict the click-through rate ctr of the bidding request; based on the click-through rate ctr and the remaining budget, the system calculates the optimal bid value bid through the bidding optimization model; the calculated optimal bid is returned to the advertising platform as a bidding result, and the bidding feedback from the advertising platform is waited for; after receiving the bidding result, the system updates the remaining budget of the advertisement, and records the display data, click data, and consumption log.

[0086] Preferably, the final effect data in step S4, including but not limited to exposure data, click data, and conversion data, is an important feedback signal of the optimization model. The final effect data is usually transmitted back to the advertiser by the advertising platform in real time; a complete bidding log and accumulation mechanism is established to facilitate the subsequent use of the final effect data.

[0087] Specifically, a reasonable log format is designed, and key fields including but not limited to bid request, bid response, exposure, click, and consumption are set; the final effect data is cleaned, merged, and sorted to generate a structured effect data set, which corresponds to the bidding log but pays more attention to effect indicators; the effect data set is imported into a data warehouse and merged with the original historical bidding data to form a continuous and complete time series of log data; the log data is statistically analyzed regularly to compare the effect differences of various types of advertisements, and provide directional guidance for the optimization of the real-time bidding strategy;

[0088] It should be noted that the scale of the effect data set may be much larger than the historical bidding data, and it is necessary to reasonably design the tiered storage and lifecycle management strategies to ensure the real-time availability of hot data, while also taking into account the persistent storage of cold data, and at the same time pay special attention to the privacy and security of the data.

[0089] Preferably, the retraining and fine-tuning of the bidding optimization model in step S4 is specifically as follows:

[0090] It should be noted that as new final effect data is continuously generated, the original bidding optimization model parameters may gradually lose their timeliness. In order to maintain the effectiveness of the model, a periodic model retraining strategy should be formulated;

[0091] Step S401, set retraining trigger conditions, including but not limited to daily or weekly regular triggering, and triggering when the data increment reaches a specific scale; use a stratified sampling method to divide the latest effect data set into a training set and a validation set to ensure that the distribution of samples in each advertisement and each time period is balanced, avoid training data skew, and use them for retraining and effect evaluation of the bidding optimization model. The division ratio is adjusted according to actual conditions; retrain the bidding optimization model to obtain updated model parameters; evaluate the prediction effect of the bidding optimization model after updating the model parameters through the validation set. If the effect is better than the existing model, start the model replacement process and replace the original bidding optimization model with the bidding optimization model after updating the model parameters; otherwise, retain the original bidding optimization model, adjust the training parameters or increase the data volume and retrain again;

[0092] Exemplarily, in addition to the retraining, incremental learning, online learning and other methods are used to achieve real-time fine-tuning of the bidding optimization model parameters in view of the streaming characteristics of the effect data set. Automatic machine learning (AutoML) technology is introduced to allow the algorithm to automatically search for the optimal model hyperparameter combination, reducing the complexity of manual parameter adjustment.

[0093] Step S402: perform version management and update of the bidding optimization model parameters. In the scenario where the bidding optimization model is frequently iterated, a strict version management mechanism is established to support rapid update and rollback backup of the bidding optimization model.

[0094] Exemplarily, each version of the bidding optimization model is assigned a unique version number, and is managed in a unified manner with the corresponding code, data, configuration, etc. to ensure the reproducibility of the training environment; a grayscale release strategy is adopted to conduct A / B testing on the new model of the bidding optimization model in a low-traffic scenario, and the full model is released only after confirming that the effect is better than the bidding optimization model; an early warning monitoring mechanism is established to track the key indicators of the new model in real time, including but not limited to prediction accuracy, response delay, etc., and an alarm is triggered immediately once an abnormality is found; an emergency rollback plan is set up to ensure that when problems arise with the new model, it can be restored to the previous stable version in time to avoid affecting the advertising effect; historical versions of the bidding optimization model are regularly cleaned and optimized to reduce storage and computing overhead; a model knowledge base is established to facilitate rapid reuse during subsequent optimization.

[0095] It should be noted that the RTB optimization method based on the pyramid data model proposed in this embodiment can be applied to real-time bidding scenarios of various online advertising platforms, which include but are not limited to three roles: advertisers, advertising platforms, and media parties.

[0096] In this embodiment, historical bidding data of online advertisements are collected and preprocessed, raw data of each advertisement in each time period is generated, a time scale sequence is defined, data aggregation processing is performed on the raw data, pyramid-type multi-scale structure data from fine to coarse and from bottom to top is formed, and a pyramid data model is constructed, a reference time scale is dynamically selected, short-term data and long-term data are extracted from the pyramid data model, and after fusion with time decay weights, the data are input into a bidding optimization model under budget constraints, a real-time bidding strategy for online advertisements is generated, and the bidding optimization model is retrained and fine-tuned to form a closed-loop optimization. This embodiment constructs a closed-loop system of "data collection-multi-scale processing-strategy optimization-effect feedback", introduces a pyramid-type multi-scale data processing model, optimizes the advertising strategy of advertisers, improves advertising delivery efficiency and conversion effect, structure and fusion mechanism, enhances the dynamic adaptability and stability of the strategy, and provides important support for improving the advertising effect of advertisers.

[0097] Example 2

[0098] As attached Figure 2 As shown, Embodiment 2 of the present invention provides an RTB optimization system S200 based on a pyramid data model, which includes but is not limited to:

[0099] Data collection and preprocessing module 201: used to collect historical bidding data of online advertisements, including but not limited to advertisement ID, bidding time, bidding amount, advertisement consumption, and click data, and perform data preprocessing on the historical bidding data;

[0100] Specifically, the historical bidding data is pre-processed by cleaning, filtering, aggregation, etc. to prepare for the subsequent construction of the pyramid data model.

[0101] The pyramid data model construction module 202 is used to define a time scale sequence, perform data aggregation processing on the original data according to the time dimension, form pyramid-type multi-scale structure data from fine to coarse and from bottom to top, and construct a pyramid data model based on the pyramid-type multi-scale structure data;

[0102] Exemplarily, according to the defined time scale sequence, the preprocessed historical bidding data is aggregated according to different time dimensions to form pyramid-structured multi-scale data. The granularity of the pyramid decreases from fine to coarse from bottom to top, and data of different aggregation degrees are stored to facilitate subsequent selection of data from multiple time scales.

[0103] The real-time bidding strategy generation module 203 is used to dynamically select a reference time scale based on the characteristics of the current time period, extract short-term data and long-term data from the pyramid data model, fuse them with time-decayed weights, input them into the bidding optimization model under budget constraints, and generate a real-time bidding strategy for online advertising;

[0104] Exemplarily, each time a bidding request arrives, one or more benchmark time scales are dynamically selected according to the current situation, short-term data and long-term data are selected from the corresponding levels of the pyramid, and fused into model input features through time decay weighted fusion. Combined with budget constraints, the data is input into the click-through rate prediction model, and a real-time bidding strategy is generated through the bidding optimization model.

[0105] Bidding optimization closed-loop module 204: used to record the final effect data of advertisement delivery, and regularly use the final effect data for retraining and fine-tuning the bidding optimization model, update model parameters, and form a closed-loop optimization;

[0106] For example, the performance data of each link of advertising delivery, such as exposure, clicks, conversions, etc., is recorded in real time to analyze the performance differences of different advertisements and provide feedback for strategy optimization. It is also necessary to design a reasonable data storage and management mechanism, use the newly generated performance data, and regularly retrain and fine-tune the bidding optimization model to maintain the timeliness of the model; formulate model iteration and update strategies, and establish a version management mechanism; improve optimization efficiency through incremental learning, online learning, AutoML and other technical means.

[0107] Example 3

[0108] As attached Figure 3 As shown, Embodiment 3 of the present invention provides an RTB optimization device S300 based on a pyramid data model, which is specifically an electronic device, including but not limited to: one or more processors; a storage device for storing programs; when the program is executed by the processor, the processor implements an RTB optimization method based on a pyramid data model described in Embodiment 1; the specific components of the electronic device are as follows:

[0109] A central processing unit (CPU) 301 can execute programs stored in a read-only memory (ROM) 302 or loaded from a storage unit 308 to a random access memory (RAM) 303;

[0110] Bus 304, used to connect CPU 301, ROM 302, RAM 303 and input / output (I / O) interface 305;

[0111] Components connected to the I / O interface 305: input part 306 (such as keyboard, mouse), output part 307 (such as display, speaker), storage part 308 (such as hard disk), communication part 309 (such as network card), drive 310 and removable media 311 (such as disk, CD-ROM).

[0112] Specifically, the main steps of the method described in Example 1 can be implemented by a computer program, which can be stored on a computer-readable medium; computer-readable media include but are not limited to computer-readable storage media, computer-readable signal media; computer-readable storage media include but are not limited to electrical, magnetic, optical, electromagnetic, infrared or semiconductor systems, devices or equipment, such as electrical connections, disks, RAM, ROM, flash memory, optical disks, etc.; computer-readable signal media include but are not limited to data signals propagated in baseband or carrier waves, carrying program codes that can be used by the instruction execution system, and can be transmitted in electromagnetic, optical, etc. forms.

[0113] In further implementations, a computer-readable storage medium is provided, which may be installed in the electronic device or may exist independently, and carries one or more programs. When the programs are executed by the electronic device, the electronic device is supported to execute the method steps described in Embodiment 1, including but not limited to:

[0114] The advertising platform collects historical advertising bidding data from various channels, mainly including information on multiple dimensions such as advertising ID, bidding time, amount, consumption, and clicks;

[0115] Perform pre-processing operations such as cleaning, deduplication, and aggregation on the raw data to remove abnormal data, and make preliminary summaries by ad ID and time dimension to prepare for subsequent analysis;

[0116] Define a time scale sequence, aggregate the pre-processed bidding data according to different time dimensions (such as hours, days, weeks, and months), and form a pyramid data model, including pyramid-style multi-time scale data: from the bottom to the top, the data granularity becomes coarser layer by layer, and is stored in the corresponding data table respectively; based on the distributed computing capabilities of the big data platform, improve the efficiency of data aggregation and retrieval;

[0117] For each real-time bidding request, dynamically select the appropriate benchmark time scale based on current traffic trends, historical data volume, delivery budget and other factors, select the benchmark time scale and related hierarchical data in the pyramid data model, take into account both short-term and long-term data, and set the time decay weight;

[0118] The weighted fusion features are input into the click-through rate prediction model. The budget constraint factor is also taken into consideration, and a real-time bidding strategy is generated through the bidding optimization model.

[0119] Real-time recording of advertising performance data of each link, including key indicators such as exposure, clicks, and conversions; regular use of newly accumulated performance data to retrain and optimize existing bidding strategies, and continuously improve the effectiveness and timeliness of the model; use incremental learning, online learning, AutoML and other technologies in the optimization process to improve optimization efficiency; establish a strict model version management mechanism to support grayscale release, effect monitoring, emergency rollback and other risk control measures.

[0120] Those skilled in the art will readily appreciate other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. This application is intended to cover any modification, use or adaptation of the present invention, which follows the general principles of the present invention and includes common knowledge or customary techniques in the art that are not disclosed by the present invention. The specification and examples are intended to be exemplary only, and the true scope and spirit of the present invention are indicated by the claims. It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from the scope thereof. The scope of the present invention is limited only by the appended claims.

Claims

1. An RTB optimization method based on a pyramid data model, characterized in that: include: Collect historical bidding data of online advertisements, including advertisement ID, bidding time, bidding amount, advertisement consumption, and click data. After data preprocessing, perform preliminary aggregation according to key dimensions to generate statistical indicators for each advertisement in each time period, which are recorded as raw data. Defining a time scale sequence, performing data aggregation processing on the original data according to the time dimension to form pyramid-type multi-scale structure data from fine to coarse and from bottom to top, and constructing a pyramid data model based on the pyramid-type multi-scale structure data; Dynamically select a reference time scale based on the characteristics of the current time period, extract short-term data and long-term data from the pyramid data model, fuse them with time-decayed weights, input them into a bidding optimization model under budget constraints, and generate a real-time bidding strategy for online advertising; The final effect data of the advertisement delivery is recorded, and the final effect data is regularly used for retraining and fine-tuning the bidding optimization model, updating the model parameters, and forming a closed-loop optimization.

2. A RTB optimization method based on a pyramid data model according to claim 1, characterized in that: The time scale sequence is defined, including: Each layer of the pyramid data model corresponds to a specific time scale, represents the original data under the specific time scale, defines a time scale sequence, defines an aggregation function for each element in the time scale sequence, and aggregates the original data according to the time dimension of each element in the time scale sequence.

3. A RTB optimization method based on a pyramid data model according to claim 2, characterized in that: The data aggregation processing includes: based on the time scale sequence and the aggregation function, aggregating the original data layer by layer from bottom to top to form the pyramid-like multi-scale structure data, and completing the construction of the pyramid data model, wherein the first layer: contains the original data and has the finest granularity.

4. A RTB optimization method based on a pyramid data model according to claim 1, characterized in that: The selected reference time scale is specifically: For each real-time bidding request, n benchmark time scales are determined as key reference objects for subsequent feature extraction and real-time bidding strategy generation. The benchmark time scales are selected based on the first balance principle to ensure the timeliness and stability of the data.

5. A RTB optimization method based on a pyramid data model according to claim 1, characterized in that: The extraction of short-term data and long-term data is specifically as follows: After determining the benchmark time scale, the corresponding short-term data and the long-term data are selected from the pyramid data model as inputs of the bidding optimization model. The specific selection follows the second balance principle, which includes: short-term data selection principle, long-term data selection principle, and timeliness and data volume balance principle.

6. A RTB optimization method based on a pyramid data model according to claim 1, characterized in that: The time-attenuated weight fusion is specifically: Based on the time decay weight, the weight of each layer in the short-term data and the long-term data is calculated by a decreasing function, and an exponential decay function is used for calculation, and the formula is as follows: ,in For the current moment, is the timestamp of the data, is the decay rate parameter; for each layer data, assign weights ,and > > ... > , is the weight given to the i-th layer of data, is the weight given to the nth layer of data; let The data vector is , the data of each layer The data vector Multiply by the corresponding weight Get the weighted data vector of each layer , and merge the weighted data vectors of all layers as the final feature input of the bidding optimization model: .

7. A RTB optimization method based on a pyramid data model according to claim 1, characterized in that: The construction of the bidding optimization model includes: A model of budget constraint and click-through rate is established as the bidding optimization model. The goal of bidding optimization is to maximize the number of clicks on the advertisement under the budget constraint, which can be described as the following bidding optimization problem: ,in is the bid price of the jth auction, is the estimated click rate of the j-th auction, For the total budget.

8. A RTB optimization method based on a pyramid data model according to claim 6, characterized in that: The retraining and fine-tuning of the bidding optimization model specifically includes: Cleaning, merging, and organizing the final effect data to generate a structured effect data set; Set retraining trigger conditions, including daily or weekly regular triggers, and triggers when data increments reach a certain scale; Using a stratified sampling method, the latest effect data set is divided into a training set and a validation set, and the bidding optimization model is retrained to obtain updated model parameters; Evaluating the prediction effect of the bidding optimization model after updating the model parameters through the validation set; Perform version management and update of the bidding optimization model parameters.

9. An RTB optimization system based on a pyramid data model, characterized in that: include: The data collection and preprocessing module is used to collect historical bidding data of online advertisements, including advertisement ID, bidding time, bidding amount, advertisement consumption, and click data. After data preprocessing, the historical bidding data is initially aggregated according to key dimensions to generate statistical indicators for each advertisement in each time period, which are recorded as raw data. A pyramid data model construction module is used to define a time scale sequence, perform data aggregation processing on the original data according to the time dimension, form pyramid-type multi-scale structure data from fine to coarse and from bottom to top, and construct a pyramid data model based on the pyramid-type multi-scale structure data; A real-time bidding strategy generation module is used to dynamically select a reference time scale based on the characteristics of the current time period, extract short-term data and long-term data from the pyramid data model, fuse them with time-decayed weights, input them into a bidding optimization model under budget constraints, and generate a real-time bidding strategy for online advertising; The bidding optimization closed-loop module is used to record the final effect data of advertising delivery, and regularly use the final effect data to retrain and fine-tune the bidding optimization model, update model parameters, and form a closed-loop optimization.

10. An RTB optimization device based on a pyramid data model, characterized in that: include: Memory, used to store programs; A processor, used for loading a program to execute the RTB optimization method based on a pyramid data model as described in any one of claims 1-8.

Images