A short play scene causality feature-based advertisement real-time bidding method and system

Abstract

The application discloses a kind of based on short play scene causal feature real-time bidding method and system of advertisement, it is related to artificial intelligence field.Constitute the basic state vector including user side, content side, bidding side dimension;Based on basic state vector, extract the causal feature of short play scene;Causal feature includes user time sequence feature when pursuing drama, content conversion feature and bidding adaptation feature;Introduce the non-zero short play potential income under short play scene to correct ATE;Fitting short play advertisement payment logarithmic normal delay distribution, calculate delay feedback correction coefficient and carry out ATE correction;Calculate double-price auction adaptation factor, establish the dynamic mapping relationship of corrected ATE value and optimal bid, calculate optimal bid.The application realizes the accurate calculation of ATE by extracting short play scene exclusive causal feature, dynamically adapts optimal bid in combination with the delay characteristics of short play payment conversion and platform auction format, realizes the dual maximization of long-term conversion income and return on investment under short play scene.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence, and more specifically, to a real-time bidding method and system for advertisements based on the causal features of short drama scenes. Background Technology

[0002] In the field of computational advertising, real-time bidding (RTB) is the mainstream delivery model. Currently, the advertising real-time bidding model based on general scenario causal inference (BITS) constructs contextual state by extracting general user, traffic, and bidding features, calculates the conditional average treatment effect (ATE) value of ad exposure, and establishes a direct mapping between ATE and the optimal bid based on the second-price (GSP) / first-price (FPA) auction theory.

[0003] However, directly applying the real-time bidding model for advertising based on causal inference in general scenarios to short drama scenarios has significant drawbacks: First, feature loss leads to biased ATE estimation. The general causal inference model only extracts GSCF (such as user basic profile, bidding eCPM, and traffic competition intensity), ignoring the causal features specific to short dramas, such as the order of watching the series, the completion rate (CSCR) of the bidding scenario, the type of series payment, and the ad unlocking node, resulting in misjudgment of traffic value and pseudo-optimization of bidding. Second, the underlying assumption is wrong. It assumes that the potential revenue of users is zero when there is no exposure, ignoring the natural payment behavior (SPY) of short drama users, which leads to errors in the calculation logic, resulting in underestimation of conversion value and loss of control over bidding costs. Summary of the Invention

[0004] In view of this, the present invention provides a real-time advertising bidding method and system based on the causal characteristics of short drama scenes, which is used to solve the problems of bid optimization distortion and uncontrolled bidding costs in existing real-time advertising bidding methods.

[0005] To achieve the above objectives, the following solution is proposed: A real-time bidding method for advertisements based on the causal features of short drama scenes includes: Receive advertising bidding requests from short drama platforms and construct a basic state vector that includes user-side, content-side, and bidding-side dimensions; Based on the basic state vector, causal features of short drama scenes are extracted and encoded into high-dimensional short drama causal features through Transformer; the high-dimensional short drama causal features include user watching time sequence features, content conversion features and bidding adaptation features; The conditional average treatment effect value calculated based on the causal characteristics of high-dimensional short dramas is corrected by introducing non-zero potential revenue of short dramas in the short drama scenario to obtain the initial ATE value; Fit the log-normal delay distribution of short drama advertising payment, calculate the delay feedback correction coefficient according to the preset payment observation window period, correct the initial ATE value, and obtain the target ATE value; Based on the auction format identifier and traffic competition intensity of the short drama platform, the dual-price auction adaptation factor is calculated, and the optimal bid is calculated based on the dynamic mapping relationship between the target ATE value and the optimal bid.

[0006] Preferably, it further includes: After the bidding ends, the bidding samples and delayed conversion data are collected, and the model parameters are periodically iterated and updated.

[0007] Preferably, the process of correcting the conditional average treatment effect value calculated based on the causal characteristics of high-dimensional short dramas by introducing non-zero potential benefits of short dramas in short drama scenarios includes: Quantifying the potential revenue of non-zero short dramas based on the causal characteristics of high-dimensional short dramas: ; in, Causal characteristics of high-dimensional short dramas The probability of users naturally paying. The value per episode of the series The revenue sharing ratio for advertisers' paid conversions of short dramas; By combining short dramas with a hybrid monetization model, ad click revenue and paid conversion revenue sharing are integrated into the calculation of overall conversion revenue, resulting in the initial ATE value after scene correction for short dramas: ; in, Causal characteristics of high-dimensional short dramas The overall conversion benefits, For ad click revenue, Revenue sharing based on paid conversions driven by ad exposure.

[0008] Preferably, the log-normal delay distribution of short drama advertising payment is fitted, and the delay feedback correction coefficient is calculated based on the preset payment observation window period to correct the initial ATE value, including: Based on the delay distribution, the probability of paid conversion at the current moment is calculated to obtain the delay feedback correction coefficient: ; in, For delayed feedback correction coefficient, T represents the set paid observation window period. Causal characteristics of high-dimensional short dramas The probability density function of the delayed distribution; The initial ATE value is corrected based on the delay feedback correction coefficient to obtain the target ATE value: ; in, This is the initial ATE value.

[0009] Preferably, based on the auction format identifier and traffic competition intensity of the short drama platform, a dual-price auction adaptation factor is calculated, including: When the auction format is a second-price auction, the double-price auction fit factor is: ; in, The intensity coefficient of traffic competition. ; When the auction format is a first-price auction, the fit factor for a two-price auction is: ; Among them, based on the cost characteristics of the first-price auction, the base coefficient is reduced to 0.8.

[0010] Preferably, the optimal bid is calculated based on the dynamic mapping relationship between the target ATE value and the optimal bid, including: When the auction format is a second-price auction, the optimal bid is calculated based on the target ATE value after delay correction: ; in, This is the optimal bid in a second-price auction format. For the target ATE value after delay correction, when If the traffic is deemed to have no conversion value, the bid is set to 0, and the bidding is abandoned. When the auction format is a first-price auction, the optimal bid is calculated based on the target ATE value after delay correction and the platform's competitive bid distribution characteristics: ; ; in, This is the best bid in a first-price auction format; Causal characteristics of high-dimensional short dramas The random variable of competitor bids, b The specific value to be used when bidding against competitors; Causal characteristics of high-dimensional short dramas The probability density function of competitors' bids was obtained by fitting historical bidding data of the same category and time period on the short drama platform.

[0011] Preferably, the basic state vector is obtained by concatenating user-side basic features, content-side basic features, and bidding-side basic features; Among them, the basic user features include: one-hot encoded vectors of user age, gender, short drama viewing history and / or paid behavior tags; The basic features of the content side include: feature vectors of series category, number of episodes, payment model and / or ad unlocking nodes; The basic characteristics of the bidding side include: the average bid price of the short drama platform at the current moment, the intensity of traffic competition, and / or the numerical characteristics of the advertiser's budget consumption rate.

[0012] Preferably, the user's viewing time sequence characteristics include: the user's current viewing progress, the historical episode completion rate, the click conversion rate of unlocking episodes through ads, the probability of paying naturally, and / or the willingness to purchase ad-free coupons; Content conversion characteristics include: completion rate in bidding scenarios, paid episode milestones, ad unlock duration, and / or the value of a single episode. Bidding adaptation features include: the auction format identifier of the current short drama platform, the peak traffic coefficient, and / or the average bid and conversion ratio of similar dramas.

[0013] A real-time advertising bidding system based on the causal features of short drama scenes includes: The request acquisition unit receives advertising bidding requests from the short drama platform and constructs a basic state vector containing dimensions of the user side, content side, and bidding side. The causal feature extraction unit extracts causal features of short drama scenes based on the basic state vector and encodes them into high-dimensional short drama causal features through Transformer; the high-dimensional short drama causal features include user watching time sequence features, content conversion features and bidding adaptation features; The revenue correction unit introduces the non-zero potential revenue of short dramas in the short drama scenario to correct the conditional average treatment effect value calculated based on the causal characteristics of high-dimensional short dramas, and obtains the initial ATE value. The delay correction unit fits the log-normal delay distribution of short drama advertising payment, calculates the delay feedback correction coefficient according to the preset payment observation window, corrects the initial ATE value, and obtains the target ATE value. The bid calculation unit calculates the dual-price auction adaptation factor based on the auction format identifier and traffic competition intensity of the short drama platform, and calculates the optimal bid based on the dynamic mapping relationship between the target ATE value and the optimal bid.

[0014] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects: (1) The present invention is a real-time bidding method for advertising based on the causal features of short drama scenarios. It extracts the causal feature exclusive feature (SSCF) system of short drama scenarios and integrates the core causal variables unique to short dramas, such as the time sequence of watching the series, the completion rate of the bidding scenario, the advertising unlocking node, and the hybrid monetization mode, into the ATE calculation from the three dimensions of users, content, and bidding. At the same time, it models the non-zero potential revenue (SPY) in the short drama scenario and corrects the calculation logic of ATE from the bottom layer, so that the ATE value is highly consistent with the actual conversion value of short drama advertising. It completely solves the core problems of misjudging traffic value and pseudo-optimization of bidding in the short drama scenario of the general model.

[0015] (2) This invention addresses two major industry characteristics of short drama advertising payment delay feedback and platform auction format dynamic switching. It designs a delay feedback causal correction algorithm and a dual-price auction adaptation factor (BAAF) mechanism respectively: The former calculates the correction coefficient by fitting the delay distribution of short drama payment and performs secondary correction on the ATE value, eliminating the value underestimation problem caused by delayed conversion and making the ATE estimate more in line with the conversion law of short drama; The latter dynamically adjusts the mapping relationship between ATE and the optimal bid based on the auction format identifier and traffic competition intensity of the short drama platform, realizing seamless adaptation of the two auction formats GSP / FPA, which not only avoids the cost out-of-control due to excessive bids in the FPA mode, but also solves the problem of missing high-quality traffic due to low bids in the GSP mode.

[0016] (3) This invention constructs a closed-loop mechanism for the entire chain of “feature extraction - ATE calculation - delay correction - auction adaptation - bid execution - sample feedback - model iteration”, which can collect all samples such as traffic characteristics, bidding results, and user conversion (including delayed conversion) of short drama bidding in real time and store them in the replay memory block for online iteration of model parameters. This design allows the model to actively capture the dynamic changes in the short drama industry and achieve self-optimization and dynamic adaptation of the model without manual intervention. It solves the problem of “one-time training and long-term failure” of general models in short drama scenarios and ensures the long-term stability and adaptability of advertising bidding strategies. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0018] Figure 1 A flowchart of a real-time advertising bidding method based on causal features of short drama scenes is provided for an embodiment of the present invention; Figure 2 This is an architecture diagram of a real-time advertising bidding system based on the causal characteristics of short drama scenes, provided for an embodiment of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Based on a distributed stream computing architecture and a causal inference learning cluster, this invention constructs a causal perception bidding system specifically for short drama platforms, taking into account the hybrid monetization model of "advertising + payment", the time sequence characteristics of watching dramas, and the characteristics of payment delay.

[0021] First, to standardize the expression of subsequent causal calculations and bidding decisions, the core variables involved in the embodiments of this invention are uniformly defined as follows: (1) Action variables : Represents the actual bid amount output by the system at time t for the current bidding request, and is a general action variable for bidding decisions; the optimal bid for the second-price auction is derived in subsequent steps. First-price auction: best bid These are the optimal values ​​of the action variable that maximize profits in the corresponding auction scenario.

[0022] (2) Instantaneous gain : Indicates that the system uses a bid After winning traffic, the instant total revenue generated by a single ad impression includes two parts: ad click revenue and user payment conversion revenue sharing. This variable corresponds to the comprehensive conversion revenue in subsequent steps and is the core revenue variable for calculating the Conditional Average Treatment Effect (ATE).

[0023] (3) Commercial conversion threshold of short dramas This indicates the advertiser's preset minimum acceptable conversion value threshold for a single traffic stream, used as a safety net for determining the value of the traffic. When the estimated conversion value is lower than this threshold, the system determines that the current traffic does not meet the business requirements, the bid is set to 0, and the bidding is abandoned.

[0024] Next, combined Figure 1 This invention introduces a real-time advertising bidding method based on causal features of short drama scenes, as provided in an embodiment of the present invention. Figure 1 As shown, the method includes: Step S1: Receive advertising bidding requests from the short drama platform and construct a basic state vector that includes user-side, content-side, and bidding-side dimensions.

[0025] Specifically, it receives bidding requests from the short drama platform's Ad Exchange system in real time. Based on the core characteristics of short dramas, such as "unlocking episodes by watching ads, paying per episode, and delayed payment after ad exposure", it defines the bidding interaction within a continuous time period as a discrete state.

[0026] Basic state vector It is composed of multimodal features from short drama scenes, providing a foundation for subsequent causal feature extraction in short drama scenes. Here, K is the total feature dimension of the basic state vector. The basic state vector in this embodiment mainly includes three core dimensions: (1) User-side basic features: unique hot encoded vectors of user age, gender, short drama viewing history and / or paid behavior tags.

[0027] (2) Basic features of content: feature vectors of series category, number of episodes, payment mode (unlock by watching ads / direct payment) and / or ad unlock node (AUN).

[0028] (3) Basic characteristics of the bidding side include: the average bid price (eCPM) of the short drama platform at the current moment, the intensity of traffic competition and / or the advertiser's budget consumption rate. The numerical characteristics, among which, for t This indicates the amount of advertising budget that the advertiser has consumed within the current campaign period. This represents the advertiser's total advertising budget for the current campaign period.

[0029] Step S2: Based on the basic state vector, extract the causal features of the short drama scene and encode them into high-dimensional short drama causal features using Transformer.

[0030] Specifically, based on the extracted base state vector, causal features (SSCF) specific to short drama scenarios are extracted from three dimensions: user side, content side, and bidding side. All causal features are encoded into high-dimensional dense vectors using a Transformer to eliminate feature sparsity, resulting in high-dimensional short drama causal features. Among these, the high-dimensional short drama causal features include user viewing time sequence features, content conversion features, and bidding adaptation features.

[0031] (1) User-side SSCF: The core is the user's time sequence characteristics of watching dramas, including the user's current drama watching progress (e.g., episode 5 / 20 episodes in total), historical drama completion rate, click conversion rate of dramas unlocked by ads, natural payment probability (fitted value based on user's historical behavior) and / or willingness to purchase ad-free coupons.

[0032] (2) Content-side SSCF: The core is content conversion characteristics, including the completion rate of the bidding scenario (CSCR, i.e. the actual completion rate of the series under the bidding traffic), the paid nodes of the series (e.g., episode 8 is the paid starting episode), the ad unlocking duration (e.g., 15-second ad unlocks 1 episode) and / or the value of a single episode of the series (payment amount per episode / ad revenue sharing amount).

[0033] (3) Bidding-side SSCF: The core is the bidding adaptation feature, including the auction format identifier (GSP / FPA) of the current short drama platform, traffic peak coefficient (e.g., 20-22 at night is the peak, coefficient = 1.5) and / or the average bid and conversion ratio of the same type of drama.

[0034] By concatenating the three types of causal features, the final high-dimensional causal features of the short drama are obtained. ,in,M Causal feature dimension. Causal features of high-dimensional short dramas. As the core context variable for subsequent ATE calculations.

[0035] This invention constructs a short drama scene causal feature system (SSCF), extracting short drama-specific causal features from three dimensions: user side, content side, and bidding side. These features are used as the core context variables for ATE calculation, replacing the general scene causal feature (GSCF), thus achieving accurate alignment between ATE value and the actual conversion value of the short drama.

[0036] Step S3: Introduce the non-zero potential revenue of short dramas in the short drama scenario to correct the conditional average treatment effect value calculated based on the causal features of high-dimensional short dramas.

[0037] Specifically, using high-dimensional short drama causal feature vectors With this as the core, we break the assumption of "Y(0)=0" in the general ATE calculation model, model the non-zero potential revenue (SPY) in the short drama scenario, correct the ATE calculation logic from the bottom up, and achieve precise alignment between the ATE value and the actual conversion value of the short drama.

[0038] Define the basic ATE formula for short drama scenarios:

[0039] in, To understand the causal characteristics of high-dimensional short dramas Below is the overall conversion revenue of users after ad exposure (including ad click revenue and revenue sharing from paid conversions driven by ad exposure). To understand the causal characteristics of high-dimensional short dramas Below, the user's natural paid revenue when the advertisement is not displayed (i.e., non-zero short drama potential revenue SPY, including revenue sharing from users directly purchasing dramas and ad-free coupons). .

[0040] Based on the "natural probability of payment" and "value per episode" in the causal characteristics of high-dimensional short dramas, and combined with the revenue sharing ratio of short drama platforms, the potential revenue of non-zero short dramas can be quantified: ; in, Causal characteristics of high-dimensional short dramas The probability of users naturally paying. The value per episode of the series The revenue sharing ratio for advertisers' paid conversions of short dramas (as agreed upon by the platform).

[0041] Combining the "advertising + paid" hybrid monetization model of short dramas, revenue from ad clicks and paid conversions is integrated into the overall conversion revenue. The calculation yields the initial ATE value after scene correction for the short drama. : ; in, Causal characteristics of high-dimensional short dramas The overall conversion benefits, For ad click revenue, Revenue sharing based on paid conversions driven by ad exposure.

[0042] Step S4: Fit the log-normal delay distribution of short drama ad payment, calculate the delay feedback correction coefficient according to the preset payment observation window period, and correct the initial ATE value.

[0043] Specifically, targeting the core characteristic of "post-exposure delayed payment" for short-drama advertisements, this step fits the delayed distribution of short-drama payment and calculates the Delay Feedback Correction Coefficient (DFCC). A second correction is performed to eliminate the deviation in ATE value caused by delayed feedback and obtain the target ATE value.

[0044] Based on historical conversion data from short drama platforms, a log-normal delayed distribution of short drama ad payment is fitted. ,in, τ This is a delayed timeframe. This distribution aligns with the industry characteristic of short dramas where "most users complete payment within 1-3 hours of ad exposure."

[0045] Based on the delay distribution, the probability of paid conversion at the current moment is calculated to obtain the delay feedback correction coefficient: ; in, For delayed feedback correction coefficient, T represents the set paid observation window period (e.g., 24 hours). Causal characteristics of high-dimensional short dramas The probability density function of the delayed distribution; The initial ATE value is corrected based on the delay feedback correction coefficient to obtain the target ATE value: ; in, This is the initial ATE value.

[0046] Step S5: Calculate the dual-price auction adaptation factor based on the auction format identifier and traffic competition intensity of the short drama platform, and calculate the optimal bid based on the dynamic mapping relationship between the target ATE value and the optimal bid.

[0047] Specifically, considering the auction format switching patterns of short drama platforms (mini-program short dramas mainly use GSP, while short video platform short dramas switch to FPA during peak traffic periods), a dual-price auction adaptation factor (BAAF) is defined, and a short drama scenario-adjusted system is established. The dynamic mapping relationship with the optimal bid enables cost control and profit maximization under different auction formats.

[0048] The dual-price auction adaptation factor is jointly determined by the "auction format identifier" and "traffic competition intensity" in the bidding side SSCF, and is dynamically updated.

[0049] When the auction format is a second-price auction (GSP), the fit factor for the two-price auction is: ; in, The intensity coefficient of traffic competition. The more intense the competition for traffic, The larger.

[0050] When the auction format is first-price auction (FPA), the fit factor for a two-price auction is: ; Among them, based on the cost characteristics of the first-price auction, the base coefficient is reduced to 0.8.

[0051] When the auction format is a second-price auction, the optimal bid is calculated based on the corrected initial ATE value: ; in, This is the optimal bid in a second-price auction format. For the target ATE value after delay correction, when If the traffic is deemed to have no conversion value, the bid is set to 0, and the bidding is abandoned. When the auction format is a first-price auction, the optimal bid is calculated based on the corrected initial ATE value and the platform's competitive bidding distribution characteristics: ; in, This is the best bid in a first-price auction format; Based on the characteristics of competitive bidding distribution Calculated high-dimensional short drama causal features The average bid of competitors. Causal characteristics of high-dimensional short dramas Random variables representing competitors' bids; average competitor bids The calculation process is as follows:

[0052] in, To obtain the expected bid of competitors under the current feature context, we integrate the probability density function of competitors' bids over the entire effective bid range. Causal characteristics of high-dimensional short dramas The probability density function of competitors' bids was obtained by fitting historical bidding data of the same category and time period on the short drama platform. b The specific value to be bid for the competitor.

[0053] The system participates in the bidding based on the calculated optimal bid. If the traffic is won, the ad is exposed, and user behavior data (such as completion, clicks, and payments) is collected in real time after the ad exposure. If the traffic is not won, data such as competitors' bids and traffic characteristics are collected.

[0054] This invention achieves accurate calculation of the Conditional Average Treatment Effect (ATE) by extracting causal features specific to short drama scenarios. It combines the delayed characteristics of paid conversion in short dramas with the platform's auction format to dynamically adapt the optimal bid, breaking the adaptability defects of general causal inference bidding models in short drama scenarios. This enables advertisers to maximize both long-term conversion revenue and return on investment (ROI) in short drama scenarios.

[0055] This invention, for the first time, integrates a hybrid monetization model of "advertising + paid subscriptions" for short dramas into its bidding model design. By quantitatively calculating the natural user payment revenue before ad exposure, it achieves a comprehensive evaluation of the conversion value of short drama ads. This modeling method can accurately identify high-potential, high-quality user groups and adjust bidding strategies accordingly. Without increasing ad bidding costs, it fully explores the ad revenue sharing under the hybrid monetization model for short dramas, making bidding decisions more aligned with the actual business logic of short drama platforms.

[0056] Furthermore, to ensure that the model can always adapt to changes in traffic and conversion characteristics of short drama platforms, this embodiment of the invention can also achieve online iteration of model parameters through Bayesian updates, as follows: After the bidding ends, the bidding samples and delayed conversion data are collected, and the model parameters are periodically iterated and updated.

[0057] Specifically, after each bidding round (e.g., 100 times), the actual conversion data for delayed payments is collected, and the delay distribution is updated using the MCMC (Markov Chain Monte Carlo) method. With DFCC coefficient This enables online iterative optimization of the ATE calculation model.

[0058] Bidding requests, SSCF features, bid amounts, winning bids, and user conversion data (including delayed payment data) are integrated into short drama bidding samples and stored in a large-capacity replay memory block. Based on the samples in the replay memory block, the SSCF feature weights, ATE calculation model, and latency distribution are periodically adjusted. Dual-price auction adaptation factor Iterative updates are performed to ensure that the model is always aligned with the actual conversion characteristics of short drama scenarios.

[0059] This invention addresses the industry characteristic of "post-exposure delayed payment" in short-drama advertising. It fits the log-normal delayed distribution of short-drama payment and proposes a Delay Feedback Correction Coefficient (DFCC) to correct the delayed causal relationship of the ATE value, thus solving the problem of underestimating conversion value caused by delayed feedback. Simultaneously, considering the auction format switching patterns of short-drama platforms, it defines a dual-price auction adaptation factor (BAAF) and establishes a dynamic mapping relationship between the ATE value and the GSP / FPA auction formats, achieving bid adaptation and cost control under different auction formats. Finally, it constructs a closed-loop system encompassing "feature extraction - ATE calculation - delay correction - auction adaptation - bid execution - sample iteration," ensuring the model's long-term adaptability and maximized revenue in short-drama scenarios.

[0060] This invention adopts an "offline training + online inference" algorithm architecture, which places the core complex calculations such as causal feature encoding, ATE model training, and latency distribution fitting in the offline / near-line GPU distributed training process. In the online inference stage, only lightweight feature extraction, ATE value calculation and optimal bid mapping need to be completed. The overall inference latency is strictly controlled at the millisecond level, which is far lower than the bidding timeout threshold of short drama platforms. It is perfectly compatible with the high-concurrency and low-latency bidding infrastructure of short drama platforms such as mini programs and short videos, and has direct industrialization feasibility.

[0061] The following describes the real-time advertising bidding system based on the causal features of short drama scenes provided by the embodiments of the present invention. The real-time advertising bidding system based on the causal features of short drama scenes described below can be referred to in correspondence with the real-time advertising bidding method based on the causal features of short drama scenes described above.

[0062] First, combine Figure 2 This paper introduces a real-time bidding system for advertisements based on the causal characteristics of short drama scenes, such as... Figure 2 As shown, the real-time bidding system for advertisements based on the causal characteristics of short drama scenes may include: Request Acquisition Unit 1 receives advertising bidding requests from the short drama platform and constructs a basic state vector containing dimensions of the user side, content side, and bidding side; Causal feature extraction unit 2 extracts causal features of short drama scenes based on the basic state vector and encodes them into high-dimensional short drama causal features through Transformer; the high-dimensional short drama causal features include user watching time sequence features, content conversion features and bidding adaptation features; Income correction unit 3, non-zero potential income from short dramas in the short drama scenario is introduced to correct the conditional average treatment effect value calculated based on the causal characteristics of high-dimensional short dramas, so as to obtain the initial ATE value. Delay correction unit 4 fits the log-normal delay distribution of short drama advertising payment, calculates the delay feedback correction coefficient according to the preset payment observation window period, corrects the initial ATE value, and obtains the target ATE value; Bidding calculation unit 5 calculates the dual-price auction adaptation factor based on the auction format identifier and traffic competition intensity of the short drama platform, and calculates the optimal bid based on the dynamic mapping relationship between the target ATE value and the optimal bid.

[0063] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0064] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0065] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A real-time bidding method for advertisements based on the causal characteristics of short drama scenes, characterized in that, include: Receive advertising bidding requests from short drama platforms and construct a basic state vector that includes user-side, content-side, and bidding-side dimensions; Based on the basic state vector, causal features of short drama scenes are extracted and encoded into high-dimensional short drama causal features through Transformer; the high-dimensional short drama causal features include user watching time sequence features, content conversion features and bidding adaptation features; The conditional average treatment effect value calculated based on the causal characteristics of high-dimensional short dramas is corrected by introducing non-zero potential revenue of short dramas in the short drama scenario to obtain the initial ATE value; Fit the log-normal delay distribution of short drama advertising payment, calculate the delay feedback correction coefficient according to the preset payment observation window period, correct the initial ATE value, and obtain the target ATE value; Based on the auction format identifier and traffic competition intensity of the short drama platform, the dual-price auction adaptation factor is calculated, and the optimal bid is calculated based on the dynamic mapping relationship between the target ATE value and the optimal bid.

2. The real-time advertising bidding method based on the causal features of short drama scenes according to claim 1, characterized in that, Also includes: After the bidding ends, the bidding samples and delayed conversion data are collected, and the model parameters are periodically iterated and updated.

3. The real-time advertising bidding method based on the causal features of short drama scenes according to claim 1, characterized in that, The process of correcting the conditional average treatment effect value calculated based on the causal characteristics of high-dimensional short dramas by introducing non-zero potential benefits of short dramas in short drama scenarios includes: Quantifying the potential revenue of non-zero short dramas based on the causal characteristics of high-dimensional short dramas: ； in, Causal characteristics of high-dimensional short dramas The probability of users naturally paying. The value per episode of the series The revenue sharing ratio for advertisers' paid conversions of short dramas; By combining short dramas with a hybrid monetization model, ad click revenue and paid conversion revenue sharing are integrated into the calculation of overall conversion revenue, resulting in the initial ATE value after scene correction for short dramas: ； in, Causal characteristics of high-dimensional short dramas The overall conversion benefits, For ad click revenue, Revenue sharing based on paid conversions driven by ad exposure.

4. The real-time advertising bidding method based on the causal features of short drama scenes according to claim 1, characterized in that, The log-normal delay distribution of short drama ad payment is fitted, and the delay feedback correction coefficient is calculated based on the preset payment observation window to correct the initial ATE value, including: Based on the delay distribution, the probability of paid conversion at the current moment is calculated to obtain the delay feedback correction coefficient: ； in, For delayed feedback correction coefficient, T represents the set paid observation window period. Causal characteristics of high-dimensional short dramas The probability density function of the delayed distribution; The initial ATE value is corrected based on the delay feedback correction coefficient to obtain the target ATE value: ； in, This is the initial ATE value.

5. The real-time advertising bidding method based on the causal features of short drama scenes according to claim 1, characterized in that, Based on the auction format identifiers and traffic competition intensity of short drama platforms, a dual-price auction suitability factor is calculated, including: When the auction format is a second-price auction, the double-price auction fit factor is: ； in, The intensity coefficient of traffic competition. ; When the auction format is a first-price auction, the fit factor for a two-price auction is: ； Among them, based on the cost characteristics of the first-price auction, the base coefficient is reduced to 0.

8.

6. The real-time advertising bidding method based on the causal features of short drama scenes according to claim 5, characterized in that, Based on the dynamic mapping relationship between the target ATE value and the optimal bid, the optimal bid is calculated, including: When the auction format is a second-price auction, the optimal bid is calculated based on the target ATE value after delay correction: ； in, This is the optimal bid in a second-price auction format. For the target ATE value after delay correction, when If the traffic is deemed to have no conversion value, the bid is set to 0, and the bidding is abandoned. When the auction format is a first-price auction, the optimal bid is calculated based on the target ATE value after delay correction and the platform's competitive bid distribution characteristics: ； ； in, This is the best bid in a first-price auction format; Causal characteristics of high-dimensional short dramas The random variable of competitor bids, b The specific value to be used when bidding against competitors; Causal characteristics of high-dimensional short dramas The probability density function of competitors' bids was obtained by fitting historical bidding data of the same category and time period on the short drama platform.

7. The real-time advertising bidding method based on the causal features of short drama scenes according to claim 1, characterized in that, The basic state vector is obtained by concatenating user-side basic features, content-side basic features, and bidding-side basic features; Among them, the basic user features include: one-hot encoded vectors of user age, gender, short drama viewing history and / or paid behavior tags; The basic features of the content side include: feature vectors of series category, number of episodes, payment model and / or ad unlocking nodes; The basic characteristics of the bidding side include: the average bid price of the short drama platform at the current moment, the intensity of traffic competition, and / or the numerical characteristics of the advertiser's budget consumption rate.

8. The real-time advertising bidding method based on the causal features of short drama scenes according to any one of claims 1-7, characterized in that, User's TV series watching time sequence characteristics include: the user's current TV series watching progress, the historical TV series completion rate, the click conversion rate of TV series unlocked by ads, the probability of natural payment and / or the willingness to purchase ad-free coupons; Content conversion characteristics include: completion rate in bidding scenarios, paid episode milestones, ad unlock duration, and / or the value of a single episode. Bidding adaptation features include: the auction format identifier of the current short drama platform, the peak traffic coefficient, and / or the average bid and conversion ratio of similar dramas.

9. A real-time bidding system for advertising based on the causal characteristics of short drama scenes, characterized in that, include: The request acquisition unit receives advertising bidding requests from the short drama platform and constructs a basic state vector containing dimensions of the user side, content side, and bidding side. The causal feature extraction unit extracts causal features of short drama scenes based on the basic state vector and encodes them into high-dimensional short drama causal features through Transformer; the high-dimensional short drama causal features include user watching time sequence features, content conversion features and bidding adaptation features; The revenue correction unit introduces the non-zero potential revenue of short dramas in the short drama scenario to correct the conditional average treatment effect value calculated based on the causal characteristics of high-dimensional short dramas, and obtains the initial ATE value. The delay correction unit fits the log-normal delay distribution of short drama advertising payment, calculates the delay feedback correction coefficient according to the preset payment observation window, corrects the initial ATE value, and obtains the target ATE value. The bid calculation unit calculates the dual-price auction adaptation factor based on the auction format identifier and traffic competition intensity of the short drama platform, and calculates the optimal bid based on the dynamic mapping relationship between the target ATE value and the optimal bid.

Images