Explanatable real-time bidding reference price prediction method based on causal inference

Abstract

The invention discloses a causal inference-based interpretable real-time bidding reference price prediction method. The method comprises the steps of receiving a bidding request of an advertisement transaction platform, constructing a structured causal model, inputting a historical bidding feature matrix into the structured causal model, and obtaining a causal graph among variables influencing a transaction price through learning by the structured causal model; performing reference price prediction on the bidding request by using a dual-stage dual machine learning method; based on a prediction result of the dual machine learning model and uncertainty quantization, generating a price prediction interval taking an estimated value as a center; generating three-level interpretable information including a node level, a path level and a global level; and returning a bid response containing the reference price prediction and the prediction interval and / or interpretable information thereof to the advertisement transaction platform. According to the invention, high-precision causal inference, interpretable analysis and low-delay online inference of price prediction in real-time bidding are realized.

Description

Technical Field

[0001] This application relates to the field of advertising technology, and in particular to an interpretable real-time bidding benchmark price prediction method based on causal inference. Background Technology

[0002] Real-Time Bidding (RTB) has become the core model of programmatic advertising trading on the internet. In this model, an ad impression is auctioned off through an Ad Exchange (ADX) at an incredibly fast pace, within 120 milliseconds, to hundreds of Demand-Side Platforms (DSPs) for bidding. The core competitiveness of a DSP lies in its bidding algorithm, and one of the key inputs to this algorithm is the accurate estimation of the benchmark price (or reserve price) for the current ad impression. If the benchmark price is estimated too high, high-value ad impressions will fail to sell, harming the media's revenue. If the estimate is too low, even if the DSP wins the bid, its profit margin will be squeezed due to the excessively high bid.

[0003] Currently, mainstream benchmark price prediction methods in the industry mainly rely on machine learning models that target click-through rate (CTR), conversion rate (CVR), or expected value (eCPM). These models are typically based on techniques such as gradient boosting decision trees (GBDT) or deep neural networks (DNN) to learn the statistical correlation between features (such as user profiles, contextual information, and ad placement attributes) and transaction prices from massive historical bidding logs, and then perform point prediction.

[0004] However, these correlation-based prediction methods have the following inherent drawbacks, which limit the further optimization of DSP strategies and the compliance of the system:

[0005] 1. Confounding Bias and Spurious Correlation Issues: The RTB ecosystem contains complex variable relationships and numerous confounding factors. For example, "user activity" can both increase "advertiser bids" and directly boost "final transaction prices." Traditional correlation models cannot effectively distinguish between genuine causal effects and spurious correlations caused by confounding factors. The patterns they learn may be biased, leading to decreased prediction accuracy in out-of-distribution data or dynamically changing market environments. Current technologies lack modeling and analysis of the causal structure behind bidding data.

[0006] 2. Black-box decision-making and lack of interpretability: While models such as GBDT and DNN have strong predictive capabilities, their decision-making process is typically a "black box." When the system makes a price prediction, DSP operators, advertisers, and even regulatory agencies cannot know "why this price was chosen." This lack of interpretability makes it difficult to optimize internal DSP strategies based on a solid foundation, and hinders the identification and debugging of problems.

[0007] 3. Rigid Single-Point Prediction Strategy: Existing methods generally output a single benchmark price point, which lacks flexibility and cannot quantify the uncertainty of the prediction itself. DSPs struggle to make flexible and robust bid adjustments based on their real-time strategies (such as pursuing exposure or profit margins) and market competition, thus missing potential revenue or incurring unnecessary expenditure.

[0008] 4. The Challenge of Balancing Performance and Complexity: Although existing research has attempted to introduce complex causal models or interpretable artificial intelligence (XAI) techniques, these methods often incur enormous computational costs, making it difficult to meet the extreme real-time requirements of RTB bidding at the millisecond level (P99 latency <100ms). Combining high-precision causal inference, interpretable analysis, and low-latency online inference remains a technical challenge that has not yet been effectively solved.

[0009] Therefore, there is an urgent need in this field for a novel technical solution that can simultaneously address the four challenges of accuracy (eliminating bias through causal inference), transparency (providing human-readable explanations), flexibility (providing room for strategy adjustment), and real-time performance (meeting RTB latency constraints), thereby providing core impetus for the next generation of intelligent, reliable, and efficient RTB bidding systems. Summary of the Invention

[0010] The purpose of this application is to provide an interpretable real-time auction benchmark price prediction method based on causal inference, which enables high-precision causal inference, interpretable analysis, and low-latency online inference for price prediction in real-time auctions.

[0011] In a first aspect of this application, an interpretable real-time bidding benchmark price prediction method based on causal inference is provided, comprising:

[0012] Step (a): Receive bidding requests from the advertising exchange platform, construct a structured causal model, input the historical bidding feature matrix X into the structured causal model, and the structured causal model learns to obtain a causal graph between variables that affect the transaction price. The discovery process of the causal graph combines business prior knowledge for constraints and corrections.

[0013] Step (b) involves using a two-stage, dual-machine learning approach to predict the benchmark price for the bidding request. The first-stage machine learning model estimates the propensity score and the benchmark result. The second-stage machine learning model estimates the personalized causal effect based on R-learner residual regression and combines it with the winning probability to obtain the benchmark price p by maximizing the expected profit. * ;

[0014] Step (c): Based on the prediction results and uncertainty quantification of the dual machine learning model, generate an estimated value p.* The price forecast range centered on;

[0015] Step (d) generates three levels of interpretable information, including node-level, path-level, and global-level, for the benchmark price prediction or prediction range;

[0016] Step (e) involves returning a bid response containing the benchmark price forecast and its forecast range and / or interpretable information to the advertising exchange platform.

[0017] In a preferred embodiment, the step of the structured causal model obtaining a causal graph between variables affecting transaction prices through learning further includes:

[0018] The NOTEARS algorithm is used to initially estimate the adjacency matrix W of the directed acyclic graph, and constraints based on business priors are imposed. These business priors include prohibiting reverse edges from results to features, enforcing temporal causality, directional constraints from budget to bid to exposure, and predefined whitelist or blacklist edge sets.

[0019] The V-structure rule of the PC algorithm and the Meek rule are combined to orient the edges.

[0020] In a preferred embodiment, the update of the causal graph is further triggered by at least one of the following conditions: periodic batch update, stability index PSI greater than 0.2, KL divergence greater than 0.1, A / B test key index deviation exceeding 2σ, or ADWIN algorithm detecting data distribution drift.

[0021] In a preferred embodiment, the predictive performance of the machine learning model in the first stage is dynamically monitored using the area under the curve (AUC) of the propensity score m(x), the expected calibration error (ECE), and the effective sample size (ESS). When the ECE is greater than 3% or the AUC is less than 0.65, the model is deemed to be underperforming and retraining or model rollback is triggered.

[0022] In a preferred embodiment, the step of generating the price prediction range further includes: constructing an estimated value p based on the winning probability function Pwin(p|x). * The price prediction interval centered at [p*(1-β), p*(1+β)] is determined adaptively by the bootstrap confidence interval or the risk tolerance parameter α.

[0023] In a preferred embodiment, the step of generating three levels of interpretable information, including node-level, path-level, and global-level information, further includes:

[0024] Node-level explanation: The contribution of each input feature to the prediction is calculated and displayed using the SHAP method;

[0025] Path-level explanation: Combining the aforementioned cause-effect graph, analyze and demonstrate the impact of key features on the final price through different causal paths;

[0026] Global Interpretation: Provides the overall causal graph of the model and the global importance ranking of features.

[0027] In a preferred embodiment, the following steps are also included: online learning and optimization:

[0028] The discrete price segment actions are fine-tuned using a deep Q-network and trained using an experience replay buffer with a capacity of no more than 100k records.

[0029] The inference model is accelerated by using INT8 quantization technology, achieving quantization inference with a P99 latency of less than 10 milliseconds in an RTX A800 GPU and TensorRT environment.

[0030] In a preferred embodiment, privacy protection and compliance processing steps are also included:

[0031] It supports updating model parameters locally on each data source through a federated learning framework, aggregating model updates using only the FedAvg algorithm instead of the original data;

[0032] Injecting Laplacian or Gaussian noise into the output of model training or inference results can achieve differential privacy protection.

[0033] In a second aspect of this application, a computer program product is provided, the computer program product comprising computer-executable instructions that, when executed by a processor, implement the steps of the aforementioned method.

[0034] In a third aspect of this application, a computer-readable storage medium is provided, wherein computer-executable instructions are stored therein, which, when executed by a processor, implement the steps of the aforementioned method.

[0035] It should be understood that, within the scope of this invention, the above-described technical features of this invention and the technical features specifically described below (such as in the embodiments) can be combined with each other to form new or preferred technical solutions. Due to space limitations, they will not be described in detail here. Attached Figure Description

[0036] Figure 1 This is a flowchart illustrating an interpretable real-time bidding benchmark price prediction method based on causal inference according to the first embodiment of this application. Detailed Implementation

[0037] In the following description, many technical details are presented to help the reader better understand this application. However, those skilled in the art will understand that the technical solutions claimed in this application can be implemented even without these technical details and various variations and modifications based on the following embodiments.

[0038] Through extensive and in-depth research, the inventors have proposed an interpretable real-time bidding benchmark price prediction method based on causal inference. This method introduces an integrated framework of structured causal modeling (SCM), dual machine learning (DML), and interpretable output fusion into the advertising real-time bidding scenario. Compared to traditional solutions based on statistical regression or black-box deep learning models, it aims to achieve significant improvements in accuracy, transparency, compliance, and performance efficiency (low latency). First, it uses a structured causal model to separate the true causal path and eliminate spurious correlations. Second, it generates three levels of human-readable explanations—nodes, paths, and a global definition—to achieve real-time interpretability of price predictions. Third, it achieves millisecond-level inference with an end-to-end P99 latency of ≤35ms, as confirmed by internal proof-of-concept (PoC) testing on a single RTX A800 card. This invention achieves significant improvements in accuracy, transparency, compliance, and performance efficiency.

[0039] This application has at least the following beneficial effects and advantages:

[0040] 1. The Structured Causal Model (SCM) of this invention constructs a directed acyclic causal graph from the features of real-time bidding advertising requests. Combining business prior constraints and causal discovery algorithms, it outputs a causal parent set of bids and outcome variables, used to eliminate spurious correlations and improve the interpretability and stability of the prediction model. This model achieves a computable expression of structured causal relationships in pre-advertising benchmark price prediction scenarios and is a fundamental component of an integrated framework for dual machine learning and interpretable output.

[0041] 2. The dual machine learning module of this invention, based on the causal parent set output by the structured causal model, eliminates confounding factors through two-stage residualization and cross-fitting, resulting in a robust causal estimate of the benchmark price. This module combines statistical unbiasedness with industrial-grade real-time performance, and achieves controllable price risk and interpretable strategy through adaptive interval output, significantly improving the pricing accuracy and transparency of the real-time bidding system.

[0042] 3. The interpretable output layer of this invention combines the SHAP algorithm with a structured causal model. Through a three-level interpretation mechanism involving nodes, paths, and the global layer, it achieves transparency and auditability in the price prediction decision-making process. This mechanism ensures that the model not only possesses numerical accuracy but also complies with compliance and regulatory requirements. While maintaining millisecond-level inference performance, it achieves causal interpretable output for complex advertising pricing behavior, significantly improving the system's credibility and compliance.

[0043] The following table provides an explanation of some of the concepts:

[0044]

[0045] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings.

[0046] The first embodiment of this application relates to an interpretable real-time bidding benchmark price prediction method based on causal inference, the process of which is as follows: Figure 1 As shown, the method includes the following steps:

[0047] Step (a) involves receiving bidding requests from the advertising exchange platform, constructing a structured causal model, and inputting the historical bidding feature matrix X into the structured causal model. The structured causal model learns to obtain a causal graph between variables affecting the transaction price. The discovery process of the causal graph incorporates business prior knowledge for constraint and correction. The acyclicity of the causal graph is guaranteed by h(W) = tr(exp(W°W)) - d = 0; v-structure and Meek rules are used for direction determination and edge resolution; and white / black lists and temporal causality (t→t+1) priors are supported.

[0048] Specifically, firstly, the NOTEARS algorithm is used to initially estimate the adjacency matrix W of the directed acyclic graph (DAG), and constraints based on business priors are imposed. These business priors include prohibiting reverse edges from results to features, enforcing temporal causality, directional constraints from budget to bid to exposure, and predefined whitelist or blacklist edge sets. Secondly, the V-structure rule and Meek rule of the PC algorithm are combined to orient the edges. The NOTEARS algorithm is a causal relationship inference method based on linear algebra and optimization theory. It estimates the causal relationships between variables by minimizing the loss function and constraining the weight matrix to satisfy the DAG condition.

[0049] The causal graph update is triggered by at least one of the following conditions: periodic batch update, a stability index (PSI) greater than 0.2, a KL divergence greater than 0.1, a deviation of more than 2σ in the A / B test key indicators, or the ADWIN (Adaptive Sliding Wind) algorithm detecting data distribution drift. A minimum revaluation "cool-down window" is also set to avoid oscillations.

[0050] Step (b) involves using the causal graph as a guide to predict the benchmark price of bidding requests using a two-stage dual machine learning method. The first-stage machine learning model estimates the bid variable based on the causal graph, while the second-stage model estimates the outcome variable. Unbiased causal effect estimates are obtained through residual regression, and the causal estimate p of the benchmark price is then derived. * The first stage of training is used to estimate the propensity score m(x) = Pr(a = 1|x) and the baseline result g(x) = E[r|x]; the second stage estimates the personalized causal effect based on R-learner residual regression. And by combining the winning probability Pwin(p|x), the expected profit is maximized. The benchmark price p* is calculated. The predictive performance of the machine learning model in the first stage is dynamically monitored using the area under the curve (AUC) of the propensity score m(x), the expected calibration error (ECE), and the effective sample size (ESS). When ECE is greater than 3% or AUC is less than 0.65, the performance is deemed substandard, triggering retraining or model rollback. In addition to AUC and ECE, Brierscore ≤ 0.18 and ESS / N ≥ 0.6 are added as thresholds for entering the second stage; failure to meet these thresholds triggers model rollback / retraining.

[0051] Step (c) generates an estimated value p based on the prediction results and uncertainty quantification of the dual machine learning model. * The price prediction range is centered on the target price. Specifically, a prediction range is constructed based on the winning probability function Pwin(p|x) ​​with the estimated value p. * The price prediction interval centered at [p*(1-β), p*(1+β)] is obtained by fitting the historical bivalent distribution and performing iso-monotropic regression or Platt calibration; β is adaptively determined by approximating β≈z_{1-α / 2}·σ_p / p* through the bootstrap confidence interval or the risk tolerance parameter α.

[0052] Step (d) generates three levels of interpretable information—node-level, path-level, and global-level—for the benchmark price prediction or prediction interval. Node-level interpretation calculates and displays the contribution of each input feature to the prediction using the SHAP (SHapley Additive exPlanations) method. Path-level interpretation, combined with the causal graph, analyzes and displays the impact of key features on the final price through different causal paths. Global interpretation provides the overall causal graph of the model and a global ranking of feature importance.

[0053] Step (e) returns the bid response, which includes the benchmark price forecast and its forecast range and / or interpretable information, to the advertising exchange platform.

[0054] In one embodiment, the method further includes an online learning step, specifically:

[0055] The discrete price segment actions are fine-tuned using a Deep Q-Network (DQN) (K=16), and trained using an experience replay buffer with a capacity of no more than 100k records.

[0056] In one embodiment, the method further includes an optimization step, specifically:

[0057] The inference model is accelerated by using INT8 quantization technology, achieving quantization inference with a P99 latency of less than 10 milliseconds in an RTX A800 GPU and TensorRT environment.

[0058] In one embodiment, the method further includes privacy protection and compliance processing steps, specifically:

[0059] It supports updating model parameters locally on each data source through a federated learning framework, aggregating model updates using only the FedAvg algorithm instead of the original data;

[0060] Laplacian or Gaussian noise is injected into the output of model training or inference results to achieve differential privacy protection. The differential privacy budget ranges ε∈[0.5,3], δ≤1e-6; Secure Aggregation is used to aggregate model updates from all parties.

[0061] The core of this invention lies in constructing an integrated framework encompassing causal discovery, causal estimation, interval output, and interpretability. By introducing a structured causal model, confounding factors are fundamentally identified and controlled, elevating pricing from correlation prediction to causal inference. Through dual machine learning and price interval generation, an unbiased, robust, and flexible pricing strategy is achieved. A three-level interpretability mechanism meets the needs of compliance auditing and strategy debugging. And through a series of engineering optimizations, end-to-end inference of the entire complex system is ensured within millisecond-level latency.

[0062] To better understand the technical solution of this application, a specific example is provided below. The details listed in this example are mainly for ease of understanding and are not intended to limit the scope of protection of this application.

[0063] Real-time bidding (RTB) advertising transactions involve bidding decisions for hundreds of DSPs within a total bidding window of ≤120ms. Common relevance prediction methods targeting CTR (click-through rate) or CVR (conversion rate) struggle to eliminate confounders and lack interpretability, hindering DSP strategy optimization and compliance audits. Examples of confounders include: User-side: region, time period, device type, operating system (OS), network conditions, historical frequency, frequency control, implicit interests; Media-side: viewability, page type, first screen / fold-down, loading latency; Auction-side: DSP budget margin, bidding pace, frequency control strategy, competitor bidding distribution, PMP / Deal-ID restrictions; System-side: timeouts, packet loss, strategy version, cold start, etc. To eliminate spurious relevance, the solution uses only causal parent sets Pa(a) and Pa(r) in the SCM and performs residualization during the DML phase to isolate confounders. In a causal parent set, the set of confounding variables that serve as common causes for both the treatment variable (bid variable) and the outcome variable (transaction price) in a causal graph is called the causal parent set.

[0064] In this application, unless otherwise stated, the following terms shall have the meanings indicated:

[0065] Processing variable: refers to the core variable intended to quantify its causal effect. In the real-time bidding scenario of this invention, it usually refers to the advertiser's bid or the system's suggested bid.

[0066] Outcome variable: refers to the final result of interest, which in this invention is the final transaction price.

[0067] Causal parent set: refers to the set of confounding factor variables that serve as common causes of treatment and outcome variables in a directed acyclic graph discovered through structured causal models.

[0068] Structured causal graph discovery (Causal Discovery)

[0069] Construct a structured causal model and use NOTEARS to estimate the directed acyclic graph W;

[0070] ① The model input is: ad exposure logs

[0071] Where: x i For contextual features such as user, page position, time slot, and device, a i For bidding actions (Binary or pricebucket), r i This represents the transformation value or binary transformation label. Where i is the sample index and N is the total number of samples.

[0072] NOTEARS optimization:

[0073] Ensure that the adjacency matrix W corresponds to an acyclic causal graph, and h(W) is a smooth acyclic constraint. Where X∈ The characteristic matrix; λ represents the adjacency weight; L(X; W) represents the fitting loss (negative log-likelihood of linear least squares or generalized linear model); λ represents L1 sparsity regularization. For Hadamard product; There are no cyclic constraints; d is the number of variables.

[0074] ②Impose business priors (prohibit reverse edges from result to feature; enforce temporal causality; directional constraints of budget → bid → exposure; white / black list edge set);

[0075] ③ Combining the v-structure of the PC (Peter–Clark) algorithm with Meek rule orientation. The PC algorithm uses conditional independence tests in orientation. Gaussian variables are handled using Fisher-Z, Z = √(n-|S|-3)*0.5*ln((1+ρ_{XY·S}) / (1-ρ_{XY·S})); |Z|>Φ^{-1}(1-α / 2) indicates correlation; discrete variables are handled using G... 2 / χ 2 Verification. Subsequently, direction inference is performed using v-structure and Meek rules to reduce false positives and negatives.

[0076] ④ Only retain the causal parent set with bid a and result r for estimation to weaken spurious correlations. Update strategy: Weekly batch revaluation; if PSI>0.2, KL divergence>0.1, A / B key indicator deviation>2σ, or ADWIN detection shows drift, then revaluation is performed immediately.

[0077] Two-stage Double Machine Learning (DML)

[0078] The goal of DML is to continuously estimate personalized causal effects.

[0079] Phase 1: Propensity Score and Baseline Conversion Rate

[0080] Propensity Score Both multilayer perceptron (MLP) and gradient boosting decision tree (GBDT) estimation methods are feasible.

[0081] Benchmark conversion rate MLP estimation is used.

[0082] Composite loss function

[0083] Phase 2: Regression of the R-learner effect

[0084] Residual:

[0085] Target:

[0086] Cross-fitting improves unbiasedness: theoretical error order O(N) -1 / 2 ).

[0087] Benchmark price range generation (Floor-Band)

[0088] Let the execution cost be c (base price, matching fee, cloud resource amortization), and the win function Pwin(p|x) ​​be estimated by the historical second-price distribution;

[0089] Expected profit:

[0090] Numerical solution: p*=argmax_pπ(p|x), Newton-Raphson iteration, convergence in 2-3 steps;

[0091] Output range: [p*(1-β), p*(1+β)], where β is usually set to 3-5% for DSP adjustment.

[0092] p is the benchmark price (CPM); x is the context feature vector of a single display; Pwin(p|x) ​​is the probability of winning at price p.

[0093] β determination: The benchmark price p* is recalculated 200 times using bootstrap, and the relative half-width β is taken as β = (p*97.5 - p*2.5) / (2·p*median); or based on the risk tolerance α, β ≈ z_{1-α / 2}·σ_p / p*. The default β is 3–5%, which can be adaptively adjusted according to volatility and DSP risk preference.

[0094] Interpretable layer

[0095] 1. Graph Attention (GAT) weights, α ij =softmax(a(Wh) i Wh j )).

[0096] 2. SHAP-Causal algorithm: Combines Shapley value with causal sampling to output node / path contribution.

[0097] 3. Natural language template: "Because (feature) = value, the price increases by 0.37 CNY (contributing 28%)", which is easy for the DSP side to parse.

[0098] Online incremental learning and inference paths

[0099] • Data flow: BidRequest → Feature normalization → Redis cache SCM → GPU inference (DML, Floor-Band, SHAP) → BidResponse;

[0100] • Delay budget (average (ms)): Feature preparation 6 → Inference 12 → Serialization 3 → Margin 6 (network);

[0101] • DQN fine-tuning: Discretize price segments ({p_k}) as actions; experience replay buffer ≤ 100k records, INT8 quantization inference (ms);

[0102] INT8 quantization inference latency: P50≈6ms, P99≈9ms; Environment: RTX A800 (80GB) + TensorRT 10.x, batch = 1, concurrency = 1, model is a two-layer MLP (≈3.2M parameters) + lightweight SHAP sampling (32 paths). Latency will increase when concurrency > 8 or higher.

[0103] Feasibility assessment: The latency data above comes from a 5M QPS stress test environment, achieved under RTX A800 + 4 CPU vCore conditions; the hardware and throughput assumptions are consistent with current mainstream x86 / ARM server configurations.

[0104] Model Collaboration and Training Process

[0105]

[0106] This invention combines SCM and DML for the first time in pre-investment pricing of ADX, which is different from existing patents on post-investment incremental or black-box ROI pricing. It applies SCM and DML simultaneously in the pre-investment benchmark price prediction task of ADX and outputs interpretable results.

[0107] Specifically: it outputs a price corridor with bilateral valuation rather than a single-point floor, allowing the DSP to optimize its own strategy within bandwidth; it provides three levels of interpretability—node-path-global—to meet the interpretable automated decision-making requirements of IAB Transparency Center and GDPR; it adopts a Kafka-Redis-GPU intra-city link with a latency of P9932ms, achieving millisecond-level self-learning closed loop; and it supports compliant expansion with FL (FedAvg) + DP-Noise plug-and-play functionality, shielding raw logs and reducing cross-border risks.

[0108] The following table compares the indicators of the real-time bidding benchmark price prediction method in this application with those of traditional statistical models:

[0109] index Baseline (statistical model) This invention (SCM-DML) relative improvement Price absolute error (CNY) 0.083 0.072 ↓13％ DSP bid hit rate 78.4％ 90.2％ ↑15％ Traffic waste rate 11.2％ 7.6％ ↓32％ End-to-end delay (P99, ms) 48 32 ↓33％ GPU usage (inference, GB) 1.6 0.93 ↓42％ Data breach risk (expert assessment) middle Low —

[0110] Where, absolute price error = mean(|p_pred-p_ref|), and p_ref is the offline replay or post-hoc optimal binary replay price. Hit rate = number of winning bids / number of potentially winning bids. Traffic waste rate = proportion of impressions reached but filtered by DSP or invalid. PoC data comes from a single environment; production deployment requires the registration experiment report. PoC was performed on an EU server, with offline replay + 60k QPS grayscale testing of 2.3B Impression; the results showed no significant negative pullback in the 95% confidence interval.

[0111] In this application, based on INT8 TensorRT and Embedding share, inference GPU usage is reduced by approximately 42%; weekly batch structure discovery and micro-batch distillation reduce training costs by approximately 29%; and interpreting compressed transmission (protobuf≈152B) reduces bandwidth overhead by approximately 18%.

[0112] Accordingly, embodiments of this application also provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the various method embodiments of this application. Computer-readable storage media include permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device. As defined herein, computer-readable storage media do not include transient computer-readable media, such as modulated data signals and carrier waves.

[0113] Furthermore, embodiments of this application also provide an interpretable real-time bidding benchmark price prediction system based on causal inference, including a memory for storing computer-executable instructions and a processor; the processor is used to implement the steps in the above-described method embodiments when executing the computer-executable instructions in the memory. The processor may be a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Digital Signal Processor (DSP), Microcontroller Unit (MCU), Neural Processing Unit (NPU), Application Specific Integrated Circuit (ASIC), Field Programmable Gate Array (FPGA), or other programmable logic devices. The aforementioned memory may be read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or solid-state drive, etc. The steps of the methods disclosed in the embodiments of this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules in the processor.

[0114] Furthermore, embodiments of this application also provide a computer program product, including computer-executable instructions that, when executed by a processor, implement the steps in the above-described method embodiments.

[0115] All references to this specification are considered to be incorporated integrally into the disclosure of this application so that they can serve as the basis for modifications if necessary. Furthermore, it should be understood that the above descriptions are merely preferred embodiments of this specification and are not intended to limit the scope of protection of this specification. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of one or more embodiments of this specification should be included within the scope of protection of one or more embodiments of this specification.

[0116] In some cases, the actions or steps described in the claims can be performed in a different order than that shown in the embodiments and still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily require a specific or sequential order to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

Claims

1. A method for predicting interpretable real-time bidding benchmark prices based on causal inference, characterized in that, include: Step (a): Receive bidding requests from the advertising exchange platform, construct a structured causal model, input the historical bidding feature matrix X into the structured causal model, and the structured causal model learns to obtain a causal graph between variables that affect the transaction price. The discovery process of the causal graph combines business prior knowledge for constraints and corrections. Step (b) involves using a two-stage, dual-machine learning approach to predict the benchmark price for the bidding request. The first-stage machine learning model estimates the propensity score and the benchmark result. The second-stage machine learning model estimates the personalized causal effect based on R-learner residual regression and combines it with the winning probability to obtain the benchmark price p by maximizing the expected profit. * ; Step (c): Based on the prediction results and uncertainty quantification of the dual machine learning model, generate an estimated value p. * The price forecast range centered on; Step (d) generates three levels of interpretable information, including node-level, path-level, and global-level, for the benchmark price prediction or prediction interval; Step (e) involves returning a bid response containing the benchmark price forecast and its forecast range and / or interpretable information to the advertising exchange platform.

2. The method as described in claim 1, characterized in that, The structured causal model further includes the following steps in learning a causal graph between variables affecting transaction prices: The NOTEARS algorithm is used to initially estimate the adjacency matrix W of the directed acyclic graph, and constraints based on business priors are imposed. These business priors include prohibiting reverse edges from results to features, enforcing temporal causality, directional constraints from budget to bid to exposure, and predefined whitelist or blacklist edge sets. The V-structure rule of the PC algorithm and the Meek rule are combined to orient the edges.

3. The method as described in claim 2, characterized in that, Also includes: The update of the causal graph is triggered by at least one of the following conditions: periodic batch update, stability index PSI greater than 0.2, KL divergence greater than 0.1, A / B test key index deviation exceeding 2σ, or ADWIN algorithm detecting data distribution drift.

4. The method as described in claim 1, characterized in that, Also includes: The predictive performance of the machine learning model in the first stage is dynamically monitored by the area under the curve (AUC) of the propensity score m(x), the expected calibration error (ECE), and the effective sample size (ESS). When the ECE is greater than 3% or the AUC is less than 0.65, the performance is deemed unsatisfactory and retraining or model rollback is triggered.

5. The method as described in claim 1, characterized in that, The step of generating the price prediction range further includes: constructing an estimated value p based on the win probability function Pwin(p|x). * The price prediction interval centered at [p*(1-β), p*(1+β)] is determined adaptively by the bootstrap confidence interval or the risk tolerance parameter α.

6. The method as described in claim 1, characterized in that, The steps for generating three levels of interpretable information, including node-level, path-level, and global-level information, further include: Node-level explanation: The contribution of each input feature to the prediction is calculated and displayed using the SHAP method; Path-level explanation: Combining the aforementioned cause-effect graph, analyze and demonstrate the impact of key features on the final price through different causal paths; Global Interpretation: Provides the overall causal graph of the model and the global importance ranking of features.

7. The method as described in claim 1, characterized in that, Also includes: Online learning and optimization steps: The discrete price segment actions are fine-tuned using a deep Q-network and trained using an experience replay buffer with a capacity of no more than 100k records. The inference model is accelerated using INT8 quantization technology, achieving quantization inference with a P99 latency of less than 10 milliseconds in an RTX A800 GPU and Tensor RT environment.

8. The method as described in claim 1, characterized in that, It also includes privacy protection and compliance processing steps: It supports updating model parameters locally on each data source through a federated learning framework, aggregating model updates using only the FedAvg algorithm instead of the original data; Injecting Laplacian or Gaussian noise into the output of model training or inference results can achieve differential privacy protection.

9. A non-transitory computer-readable storage medium, characterized in that, The non-transitory computer-readable storage medium stores computer-executable instructions that, when executed by a processor, implement the steps of the method as described in any one of claims 1 to 8.

10. A computer program product comprising computer-executable instructions, characterized in that, When executed by a processor, the computer-executable instructions implement the steps of the method according to any one of claims 1 to 8.

Images