An optimization method that combines user groups with the pre-deployment advertising performance.

Abstract

This invention discloses an optimization method combining user groups and pre-deployment advertising effectiveness, belonging to the field of computational advertising technology. The method includes: real-time acquisition of user behavior data streams; detection of user feature distribution drift using Wasserstein distance; triggering an adversarial generative network to dynamically segment user groups when a drift threshold is exceeded; and outputting user group labels and user-group membership matrices. It also involves collecting real-time conversion data after the execution of a multi-objective optimization strategy, evaluating the degree of user group shift and counterfactual prediction error using Mahalanobis distance, and triggering cluster model reconstruction and causal model calibration. This invention achieves continuous improvement in advertising resource utilization and conversion rate through closed-loop optimization of dynamic clustering, counterfactual gain calculation, multi-objective strategy collaboration, and feedback calibration, effectively reducing user fatigue and shortening the strategy iteration cycle to the minute level.

Description

Technical Field

[0001] This invention relates to the field of computational advertising technology, and in particular to an optimization method that combines user groups with the pre-deployment advertising effect. Background Technology

[0002] Current advertising technology has gradually shifted from extensive targeting to precision marketing based on user segmentation. Existing methods often employ static clustering combined with historical conversion rate prediction models, utilizing collaborative filtering or matrix factorization to mine user interests, and employing multi-armed slot machine algorithms for dynamic optimization. In the field of causal inference, a dual machine learning framework has been introduced into advertising effectiveness evaluation, improving the robustness of processing effect estimation by decoupling confounding variables. Simultaneously, reinforcement learning-based policy optimization systems are beginning to incorporate contextual information to achieve joint decision-making on bidding and frequency control.

[0003] However, existing technologies have significant drawbacks: static user segmentation models struggle to capture real-time feature distribution shifts, leading to the invalidation of historical segmentation rules (such as cluster center shifts caused by user interest migration); traditional counterfactual gain estimation ignores the potential interaction effects between groups and the nonlinear superposition characteristics of strategy combinations, resulting in aggregation errors in individual processing effects; multi-objective optimization strategies lack joint modeling of dynamic segmentation structures and causal gain confidence intervals, easily leading to action space dimension explosion and strategy oscillations. These problems cause advertising delivery systems to exhibit strategy lag and resource misallocation when facing highly dynamic user behavior. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides an optimization method that combines user groups with the pre-deployment advertising effect to solve the resource mismatch problem caused by strategy lag and counterfactual gain estimation bias due to the drift of dynamic user feature distribution.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides an optimization method that combines user groups with the pre-deployed advertising effect, which includes: acquiring user behavior data streams in real time, detecting user feature distribution drift through Wasserstein distance, triggering adversarial generative networks to dynamically divide user groups when the drift threshold is exceeded, and outputting user group labels and user-group membership matrices.

[0008] A dual machine learning approach is used to construct a causal model, calculate the potential gain of each user group under different advertising strategies, and generate a counterfactual gain matrix.

[0009] By inputting the user-group membership matrix and the counterfactual gain matrix into the multi-agent reinforcement learning framework, a multi-objective optimization strategy including bidding coefficients, frequency control, and creative selection is dynamically generated.

[0010] Real-time conversion data after the execution of multi-objective optimization strategies is collected, and the degree of user group shift and counterfactual prediction error are evaluated by Mahalanobis distance, triggering the reconstruction of the cluster model and the calibration of the causal model.

[0011] As a preferred embodiment of the optimization method for combining user groups with the pre-deployment advertising effect described in this invention, the specific steps for outputting user group tags and user-group membership matrices are as follows:

[0012] A sliding window captures the user behavior stream, the behavior sequence is encoded and standardized using a time-aware Transformer, and the current / baseline feature matrix is ​​output.

[0013] Calculate the entropy-regularized Wasserstein distance between the current distribution and the baseline distribution, and dynamically adjust the drift judgment threshold by combining it with the exponentially weighted moving average.

[0014] If drift is detected, the adversarial generative network is activated to generate virtual features. The generation process is constrained by the cluster center alignment loss, and the original membership matrix is ​​output.

[0015] The original membership matrix is ​​subjected to Top-3 sparsification, adjacent clusters with a center distance of less than K are merged, and the compressed data is stored in a distributed database.

[0016] As a preferred embodiment of the optimization method for combining user groups and pre-deployed advertising effects described in this invention, wherein: the drift determination threshold refers to the critical value for determining whether the current distribution has undergone a significant change;

[0017] The dynamically adjusted drift determination threshold includes,

[0018] The initial threshold is set based on the mean and standard deviation of the historical distribution distance;

[0019] The real-time drift determination threshold is calculated by combining the current distance and the threshold of the previous time step with an exponentially weighted moving average.

[0020] As a preferred embodiment of the optimization method combining user groups and pre-deployment advertising effectiveness described in this invention, the specific steps for generating the counterfactual gain matrix are as follows:

[0021] The user behavior features are weighted and concatenated with the sparse cluster matrix, and then combined with the group statistics interaction terms to generate enhanced features.

[0022] Individual treatment effects are calculated by orthogonalizing residuals, and weighted aggregation is performed with the introduction of a time decay factor to correct the potential population gain.

[0023] Construct a three-dimensional counterfactual gain tensor across policies and groups, and apply KL divergence constraints to merge policy differences among similar groups;

[0024] Significance is evaluated using BCa Bootstrap sampling, and the gain matrix after confidence interval correction is output.

[0025] As a preferred embodiment of the optimization method for combining user groups and pre-deployed advertising effectiveness described in this invention, the steps of weightedly concatenating user behavior features with a sparse dynamic clustering matrix and generating enhanced features by combining group statistical interaction terms are as follows:

[0026] Standardize user behavior characteristics and concatenate them with the group membership degree according to preset weights;

[0027] Calculate the group mean and activity decay value, generate behavioral statistics indicators, and interact with user characteristics item by item;

[0028] After merging, dimensionality reduction is performed using PCA, and Gaussian noise is added to form enhanced features.

[0029] As a preferred embodiment of the optimization method combining user groups and pre-deployment advertising effectiveness described in this invention, the dynamic generation includes a multi-objective optimization strategy encompassing bid coefficients, frequency control, and creative selection. The specific steps are as follows:

[0030] The user-group membership matrix and the counterfactual gain matrix are concatenated into a joint state vector, which is then superimposed with real-time context features.

[0031] Define the bidding coefficient as a continuous action, the frequency control as a three-level discrete action, and the creative selection as a discrete action of the Top 10 creatives;

[0032] The agent dynamically couples sub-policy weights by using LSTM, and updates the policy in parallel using the FQI algorithm.

[0033] Using the lower bound of the confidence interval of the counterfactual gain matrix as a hard constraint, a policy candidate set is generated by the Lagrange multiplier method.

[0034] The VCG mechanism is used to screen for Pareto optimal strategies.

[0035] As a preferred embodiment of the optimization method combining user groups and pre-deployment advertising effectiveness described in this invention, the specific steps for generating the strategy candidate set are as follows:

[0036] Map the lower bound of the confidence interval sampled by BCa Bootstrap to the hard boundary of the bid action;

[0037] Construct a constrained optimization model and solve it using the Lagrange multiplier method;

[0038] Generate bidding actions within the feasible region and combine them with frequency control and creative strategies via a Cartesian product.

[0039] As a preferred embodiment of the optimization method combining user groups and pre-deployed advertising effectiveness described in this invention, the following steps are taken: Real-time conversion data is collected after the execution of the multi-objective optimization strategy; the degree of user group shift and counterfactual prediction error are evaluated using Mahalanobis distance; and the cluster model reconstruction and causal model calibration are triggered.

[0040] Align real-time transformed data with the clustering matrix and calculate the Mahalanobis distance offset of user group distribution;

[0041] Calculate the global error index by comparing the actual conversion gain with the predicted value;

[0042] If the offset exceeds the offset judgment threshold and the error exceeds the tolerance range, the cluster model reconstruction and causal model calibration will be triggered immediately.

[0043] If a single indicator exceeds the limit, it will be triggered after monitoring multiple time windows continuously.

[0044] In a second aspect, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, it implements any step of the optimization method for combining user groups with the pre-deployed advertising effect as described in the first aspect of the present invention.

[0045] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the optimization method for combining user groups with the pre-deployed advertising performance as described in the first aspect of the present invention.

[0046] The beneficial effects of this invention are as follows: This invention encodes user behavior sequences using a time-decaying Transformer, dynamically detects feature distribution drift by combining entropy-regularized Wasserstein distance, and generates a sparse user-group membership matrix using an adversarial generative network to solve the lag of traditional static grouping; subsequently, it constructs enhanced features by combining group statistical interaction terms, removes confounding variable interference through dual machine learning and orthogonalized residual calculation, and generates a counterfactual gain matrix with confidence interval correction to eliminate causal inference bias; furthermore, it uses the sparse structure of the user-group membership matrix to isolate multi-agent parameter training, couples bidding, frequency control, and creative selection strategies through an attention mechanism, and selects Pareto optimal solutions by combining a VCG mechanism; based on Mahalanobis distance to quantify the correlation between user group offset and prediction error, it dynamically triggers model reconstruction and calibration, forming a "detection-decision-feedback" closed loop, realizing continuous optimization of advertising resource utilization and conversion rate, effectively reducing user fatigue and shortening the strategy iteration cycle to the minute level. Attached Figure Description

[0047] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of the optimization method in Example 1 that combines user groups with the pre-deployment advertising effect.

[0049] Figure 2 This is a flowchart of generating a policy candidate set in Example 1. Detailed Implementation

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0051] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0052] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0053] Example 1, referring to Figure 1 and Figure 2 This is the first embodiment of the present invention, which provides an optimization method that combines user groups with the pre-deployment advertising effect, including the following steps:

[0054] S1: Real-time acquisition of user behavior data stream, detection of user feature distribution drift through Wasserstein distance, triggering adversarial generative network to dynamically divide user groups when the drift threshold is exceeded, and outputting user group labels and user-group membership matrix.

[0055] Specifically, it includes the following steps:

[0056] S1.1: The user behavior flow is captured by a sliding window, the behavior sequence is encoded by a time-aware Transformer model, PCA is used to reduce the dimensionality to a fixed dimension and standardize it, and the current feature matrix and the baseline feature matrix are output.

[0057] Sliding windows are a method for dynamically capturing data streams, obtaining sequences of user behavior at fixed time intervals or with a fixed number of events. Their advantage lies in their ability to capture real-time trends in user behavior, providing a dynamic data foundation for subsequent feature extraction.

[0058] S1.2: Calculate the entropy-regularized Wasserstein distance between the current distribution and the benchmark distribution, and dynamically adjust the drift judgment threshold by combining the exponentially weighted moving average to determine whether the distribution has drifted.

[0059] The drift determination threshold is the critical value used to determine whether the current distribution has changed significantly.

[0060] Specifically, the steps for dynamically adjusting the drift detection threshold are as follows:

[0061] The current feature matrix and the baseline feature matrix are regarded as the current probability distribution and the baseline probability distribution, respectively.

[0062] It should be noted that the "current probability distribution" reflects the latest distribution of real-time user behavior characteristics and is used to capture dynamic changes in user behavior; the "baseline probability distribution" represents the stable distribution of historical user behavior characteristics and is used as a reference benchmark for comparison to detect whether the current distribution has drifted significantly.

[0063] The distance between the current probability distribution and the baseline probability distribution is calculated using the entropy-regularized Wasserstein distance formula.

[0064] In contrast, entropy-regularized Wasserstein distance can robustly detect shifts in user feature distribution, while dynamic threshold adjustment adapts to changes in data distribution through exponentially weighted moving averages, avoiding misjudgments.

[0065] Within a historical time period, the entropy-regularized Wasserstein distance calculated based on historical data is used as the historical distribution distance.

[0066] It should be noted that historical data includes user behavior data, attribute data, and contextual data within a historical time period. Specifically, this includes user behavior type, behavior time, user age, gender, region, and the time, location, and context of the behavior. Through feature extraction, data cleaning, and standardization, a feature matrix is ​​generated for calculating the entropy-regularized Wasserstein distance, thus obtaining the historical distribution distance. This step has significant application value in tasks such as user behavior analysis and distribution drift detection.

[0067] An initial threshold is set using the mean and standard deviation of historical distribution distances.

[0068] Historical data refers to user behavior data within a certain period of the past. It is used to generate historical feature matrices and calculate historical distribution distances, to initialize thresholds, and to provide a reference for dynamically adjusting thresholds.

[0069] Example explanation: Assume the historical time period is the past 30 days. Calculate an entropy-regularized Wasserstein distance each day to obtain 30 historical distribution distances D1, D2, ..., D 30 Calculate the mean μ history =1.2, standard deviation σ history =0.3. Adjustment coefficient α = 2. Then the initial threshold T0 is: T0 = 1.2 + 2 × 0.3 = 1.8.

[0070] Using the exponentially weighted moving average (EWMA), the real-time drift threshold is calculated based on the distance between the previous probability distribution and the baseline probability distribution and the drift threshold value of the previous time step.

[0071] Among them, the "drift determination threshold" refers to the statistical boundary value calculated by the dynamic algorithm, which is used to determine whether the distribution of user behavior has shifted significantly. Its technical essence is equivalent to the dynamic drift detection threshold.

[0072] S1.3: If drift is detected, activate the adversarial generative network.

[0073] Ideally, Generative Activation Adversarial Networks (GANs) can synthesize high-quality virtual features through a generator when user behavior distribution drift is detected, thus alleviating the data sparsity problem. At the same time, the discriminator outputs a user-group membership matrix to ensure the accuracy of the grouping results. By constraining the generation process with cluster center alignment loss, the generated virtual features are made consistent with the distribution of real features, further improving the robustness and generalization ability of the grouping model. This allows for dynamic adaptation to changes in user behavior and provides a more accurate user grouping basis for subsequent multi-objective optimization strategies.

[0074] S1.4: The generator synthesizes virtual features, the discriminator outputs user-group membership, and the original membership matrix is ​​output through the cluster center alignment loss constraint generation process.

[0075] Specifically, the generator receives user behavior features as input and synthesizes virtual features through a multi-layer neural network. The discriminator receives the virtual features output by the generator and real user features, distinguishes between virtual and real features through adversarial learning, and outputs a user-group membership matrix. During the generation process, cluster center alignment loss is used to calculate the distance between the cluster centers of the generated features and the real features. The generator is optimized through backpropagation to ensure that the distribution of generated features is consistent with that of real features. Finally, the original membership matrix is ​​output, which contains the membership degree of each user to all user groups, and is used for subsequent sparsification and compression processing.

[0076] S1.5: Perform Top-3 sparsification on the original membership matrix, merge adjacent clusters with a center distance less than K, compress them with floating-point precision and store them in the distributed database, output the final user-group labels and membership matrix, and store them in the distributed database.

[0077] "Top-3 sparsity" is a data sparsity technique for user-group membership matrices. Its core idea is to retain the membership degree of each user to the top three user groups, while setting the membership degree of the remaining user groups to zero or ignoring it. This significantly reduces the complexity of data storage and computation, while retaining the most important user group information.

[0078] It should be noted that "K" is a preset cluster center distance threshold used to determine whether two clusters are close enough to be merged. Specifically, when the center distance between two clusters is less than K, they are considered to belong to the same category, and thus they are merged into one cluster to reduce redundancy in user groups and improve the simplicity and practicality of the clustering results.

[0079] The specific steps to obtain the "preset cluster center distance threshold" are as follows: First, run a clustering algorithm (such as K-means or hierarchical clustering) based on historical data or sample data to calculate the distance distribution between different cluster centers; then, according to business needs and data characteristics, select a suitable distance threshold as the K value, such as the median or a specific percentile of the distance distribution; finally, evaluate the rationality of the K value through cross-validation or actual application effects, and adjust it as necessary to optimize the clustering results.

[0080] S2: Use dual machine learning to build a causal model, calculate the potential gain of each user group under different advertising strategies, and generate a counterfactual gain matrix.

[0081] Specifically, it includes the following steps:

[0082] S2.1: Read the user-group membership matrix from the distributed database, decompress it, and reconstruct it into a sparse dynamic grouping matrix.

[0083] During the decompression process, the ZFP floating-point precision compression algorithm is used to decompress the data and restore the original structure of the sparse dynamic clustering matrix, ensuring that it can be used directly in subsequent calculations.

[0084] S2.2: The user behavior features are weighted and concatenated with the sparse dynamic clustering matrix, and enhanced features are generated by combining the interaction terms of the group statistics. The propensity score model and the potential outcome prediction model are fitted respectively, and the individual treatment effect is calculated by orthogonalizing the residuals.

[0085] Specifically, based on the user's unique identifier, user behavior features (such as click-through rate, conversion rate, and activity decay value) are extracted from the user behavior data stream, and the membership degree of the user's group is obtained from the sparse dynamic clustering matrix.

[0086] It should be noted that only the membership degree of the top-3 user groups is retained in the sparse dynamic clustering matrix.

[0087] Standardize user behavior characteristics (e.g., Z-score standardization);

[0088] The standardized user behavior features are matched with the sparse dynamic clustering matrix according to the user's unique identifier to ensure a one-to-one correspondence between features and membership degrees.

[0089] The user's behavioral characteristics and the degree of membership of the user's group are concatenated according to a preset weight ratio to generate a preliminary joint feature matrix;

[0090] Specifically, user behavior characteristics (weight 0.4) and the membership degree of the user's group (weight 0.6) are concatenated column by column.

[0091] Based on the historical user behavior data of the user's group, calculate the group mean and the group activity decay value, and generate behavioral statistical indicators.

[0092] Specifically, firstly, based on the historical user behavior data of the user's group, the behavioral characteristics of each user within a historical time period are extracted, such as behavior frequency and behavior type distribution; then, the mean of all user behavioral characteristics within each group is calculated to obtain the group mean; next, the activity decay value of each user's behavior is calculated based on the time decay function, and the group activity decay value is calculated by combining the activity decay values ​​of users within the group; finally, the group mean and the group activity decay value are weighted and combined to generate behavioral statistical indicators, which are used to describe the behavioral characteristics and activity status of the group.

[0093] Example: Suppose a group contains 100 users. Calculate the average click frequency of these users over the past 30 days as the group average. At the same time, calculate the click activity decay value of each user based on the exponential decay function, and take the average as the group activity decay value. Finally, weight the two to generate a behavioral statistical indicator to describe the click behavior and activity level of the group.

[0094] By interacting user behavior characteristics with the corresponding group's behavioral statistics indicators item by item according to feature dimensions, group-enhanced interaction items are generated.

[0095] Specifically, user behavior characteristics and corresponding group behavior statistics are extracted, and each characteristic dimension is multiplied or weighted and summed to generate group-enhanced interaction items. For example, the user's click frequency is multiplied by the group's average click frequency, or the user's purchase amount is weighted and summed with the group's purchase amount attenuation value, ultimately generating enhanced interaction items that reflect the interaction relationship between user behavior and group characteristics.

[0096] Example: Suppose a user clicks 5 times, and the average click frequency of their group is 10 times. Multiply the two to generate a group-enhanced interaction term 50. If the user's purchase amount is 200 yuan, and the group's purchase amount decay value is 150 yuan, then sum the two by weight to generate a group-enhanced interaction term 350, which is used to describe the interaction relationship between user behavior and group characteristics.

[0097] The preliminary joint feature matrix and the group enhancement interaction terms are merged column by column, and the enhanced features are formed by dimensionality reduction through PCA (preserving 95% variance) and adding Gaussian noise (standard deviation 0.01).

[0098] It should also be noted that the model objective of the propensity score model is to estimate the conditional probability that a user will accept a specific advertising strategy given a particular context, as expressed below:

[0099] v(X) = P(T = t|C);

[0100] Where v(C) represents the propensity score, with a value range of [0,1], T is the intervention variable (such as whether to apply the high-price strategy), which is usually a binary variable T∈{0,1}, and X is the input feature, which includes enhanced features such as user behavior features, group membership degree, and group statistical interaction terms.

[0101] Propensity score model input layer architecture:

[0102] hierarchy Dimension illustrate User Behavior Characteristics 128 Standardized click-through rate / conversion rate, etc. Group membership degree 3 Top-3 Sparsed Membership Vector Group statistical interaction 64 Hadamard product of user characteristics and group mean Noise layer 0.01σ Gaussian noise injection

[0103] Furthermore, the model objective of the potential outcome prediction model is to predict the potential outcomes for users under different advertising strategies, expressed as follows:

[0104]

[0105] in, Let f represent the orthogonalization potential outcome under strategy a. a (X) represents the linear prediction function for policy a. For a set of strategies (such as bid coefficient tiers), This indicates that for the set of policies Any strategy a in the list;

[0106] Input layer architecture:

[0107]

[0108]

[0109] S2.3: Combine the individual treatment effect (ITE) with the weighted aggregation of the user-group membership matrix, and use the time decay factor to correct the potential group gain.

[0110] Specifically, the time decay factor is used to correct the population potential gain, and the expression is as follows:

[0111]

[0112] Among them, G k,a Let α be the original potential gain of group k under advertising strategy a, representing the average conversion rate increase of the group under advertising strategy a. k represents the group number, and α represents the timeliness adjustment coefficient, with a value range of [0, 0.5], determined by grid search (the optimal value for the customer scenario is 0.3).

[0113] Furthermore, Recency k The activity decay factor of group k is expressed as:

[0114]

[0115] Where, Δt i Indicates user u i The time interval between the most recent action (in days), N k Let i represent the number of users in group k, and i be the index number of a user in group k, with a value range of i∈{1,2,…,N}. k}, c k Let k represent the set of users in group k, that is, all users belonging to group k. Expressing user u i The behavioral freshness decay value;

[0116] Example explanation: Suppose that group k has 3 users, and the time intervals of their most recent actions are as follows: User u1: Δt1 = 1 day; User u2: Δt2 = 1 day; User u3: Δt3 = 1 day.

[0117] Then Recency k The calculation process is as follows:

[0118] Calculate the behavioral freshness decay value for each user: User u1: e -0.1×1 =e -0.1 ≈0.9048; User u2: e -0.1×3 =e -0.3 ≈0.7408; User u3: e -0.1×5 =e -0.5 ≈0.6065.

[0119] Summing and averaging:

[0120] Superiorly, compared to traditional static gain aggregation methods, this step effectively addresses the prediction bias caused by insufficient timeliness of historical data by dynamically correcting the potential gain of the group using a time decay factor. Based on an exponential decay function of the time interval of recent user behavior, the freshness of user behavior can be quantified. For example, the weight of user behavior from 3 days ago is retained higher, while the weight of behavior from 30 days ago is significantly reduced, thus effectively suppressing interference from outdated data and improving real-time performance. Simultaneously, by adjusting the decay intensity and combining it with confidence interval calibration, the gain correction effect for the long-tail user group is made more stable. In scenarios where user behavior changes rapidly, prediction accuracy is significantly improved, significantly enhancing the model's adaptability to dynamic environments.

[0121] S2.4: Construct a three-dimensional counterfactual gain tensor across policies and groups, and apply KL divergence constraints to merge policy differences among similar groups.

[0122] It should be noted that after calculating the group gain, a three-dimensional counterfactual gain tensor across policies and the group is further constructed to quantify the substitution relationship between policies and support multi-policy joint optimization.

[0123] Specifically, the expression is:

[0124]

[0125] in, In the three-dimensional counterfactual gain matrix, the group k under policy a and a ′ The gain difference between a and a ′ This represents two different strategies in an advertising strategy pair, used to compare group k under strategies a and a. ′ The gain difference is given by k, which represents the group number, and G. k,a Let G be the original gain of group k under strategy a, representing the improvement in the average conversion rate of the group. τ is a temperature parameter (τ = 0.1) used to control the discriminative power of strategy differences. k,a -Gk,a′ | represents strategy a and a ′ The absolute difference in gain is used to preserve order-of-magnitude information;

[0126] It should be noted that cross-policy joint optimization is supported by converting policy differences into probability distributions. For example, if Representing strategy a and a ′ The probability of a better result is 80%.

[0127] Furthermore, by using KL divergence constraints to merge the policy differences of similar groups, a balance between group granularity and policy stability is ensured.

[0128] Specifically, calculate k and k for the two groups. ′ The KL divergence is expressed as:

[0129] D KL (M k ||M k′ );

[0130] Among them, D KL M represents the Kullback-Leibler divergence, also known as the KL divergence. k M represents the user-group membership distribution vector of group k. k′ Representing group k ′ User-group membership distribution vector, k ′ The group number of the other group compared with group k;

[0131] If D KL If the difference is less than 0.1, the policy differences between the two groups are forcibly merged to ensure a balance between group granularity and policy stability. The formula is as follows:

[0132] Parameter values ​​are based on: α = 0.3: Customer A / B testing shows that this value increases long-tail user coverage by 52% with minimal prediction error. τ = 0.1: Experiments show that this temperature parameter effectively distinguishes differences in head strategies (Top-3 strategy discrimination improved by 40%). KL threshold 0.1: When D KL When the similarity is less than 0.1, the similarity of user behavior distribution between the two groups exceeds 95% (KS test results).

[0133] In contrast to the isolated analysis of conventional two-dimensional gain matrices, this step constructs a three-dimensional counterfactual gain tensor across strategies and groups to quantify the differences between different strategy pairs within a group. Normalization is then used to convert the absolute gain differences into a probability distribution, thus more intuitively reflecting the relative advantages between strategies. Furthermore, the introduction of KL divergence constraints to automatically merge similar groups not only reduces the number of clusters in advertising budget allocation scenarios but also maintains high strategy discriminability, significantly reducing computational complexity and improving strategy stability. This fundamentally avoids the risk of strategy oscillations caused by excessive group segmentation.

[0134] S2.5: Use BCa Bootstrap sampling to evaluate statistical significance, output the counterfactual gain matrix after confidence interval correction, and store it in a distributed database.

[0135] Among them, "BCa Bootstrap sampling" is an improved bootstrap sampling method that can more accurately assess the distribution and significance of statistics by adjusting the confidence interval through bias correction and acceleration factor.

[0136] Specifically, BCa Bootstrap sampling is performed based on the counterfactual gain matrix. Multiple sample matrices are generated through multiple resampling, and statistics (such as mean and variance) are calculated for each sample matrix. The distribution of the statistics is constructed, and the bias correction factor and acceleration factor are calculated based on the distribution. The upper and lower limits of the confidence interval are adjusted, and the counterfactual gain matrix is ​​corrected based on the adjusted confidence interval. The counterfactual gain matrix after confidence interval correction is output and stored in a distributed database for subsequent multi-objective optimization strategy generation.

[0137] S3: Input the user-group membership matrix and the counterfactual gain matrix into the multi-agent reinforcement learning framework to dynamically generate a multi-objective optimization policy that includes bidding coefficients, frequency control, and creative selection.

[0138] Specifically, it includes the following steps:

[0139] S3.1: Concatenate the user-group membership matrix and the counterfactual gain matrix into a joint state vector, and superimpose real-time context features to form a multi-dimensional state space.

[0140] Specifically, the user-group membership matrix is ​​expanded into a vector form along the user dimension and concatenated with the counterfactual gain matrix along the same user dimension to generate a joint state vector. Real-time context features (such as time, location, and scene) are extracted and encoded into numerical vectors. These real-time context features are then superimposed with the joint state vector to form a multi-dimensional state space, which is used as input for subsequent multi-agent reinforcement learning frameworks and supports the dynamic generation of multi-objective optimization strategies.

[0141] The superior approach of "forming a multi-dimensional state space" can unify and integrate user segmentation information, counterfactual gains, and real-time contextual features, providing comprehensive input information for the multi-agent reinforcement learning framework. This enhances the decision-making model's ability to perceive dynamic changes in user behavior and the contextual environment, thereby generating more accurate multi-objective optimization strategies that include bidding coefficients, frequency control, and creative selection, thus improving advertising effectiveness and user experience.

[0142] S3.2: Define the bidding coefficient as a continuous action, frequency control as a three-level discrete action, and creative selection as a Top 10 material discrete action, and dynamically couple the relationship between the three through an attention mechanism.

[0143] The “Top 10” refers to the top 10 creative materials selected from the candidate creative material library. These materials are usually ranked based on metrics such as user interests, click-through rate, and conversion rate to optimize creative selection strategies and ensure that the advertising materials are the most attractive and effective.

[0144] S3.3: Utilize joint state vectors to calculate click pass rate (CTR), user lifetime value (LTV), and fatigue penalty in real time. Use the group potential gain in the counterfactual gain matrix to calibrate the reward weights, ensuring that the reward signal is strictly aligned with the causal inference results.

[0145] S3.4: Deploy a Long Short-Term Memory Neural Network (LSTM) to coordinate agents, and coordinate the policy weights of each sub-agent according to the temporal dependency relationship of the user-group membership matrix to achieve multi-objective collaborative optimization.

[0146] Specifically, the temporal data of the user-group membership matrix is ​​input into a Long Short-Term Memory (LSTM) neural network to train the LSTM model to learn the dynamic changes of the user group and output the weight coefficients of each sub-agent (such as bidding, frequency, and creativity agents). The hidden state of the LSTM captures the temporal dependency of the user-group membership, dynamically adjusts the policy weights of the sub-agents, and generates a coordinated policy combination by combining the current state and temporal information for subsequent execution of multi-objective optimization strategies.

[0147] In a better way, coordinating the strategy weights of each sub-agent can dynamically balance the priority of multi-objective strategies (such as bidding, frequency, and creative) according to the temporal changes of the user group, avoid strategy conflicts, and improve the synergy of strategy combinations. By modeling temporal dependencies through LSTM, it ensures that the strategy weights are consistent with user behavior trends, enhances the dynamic adaptability of ad placement, and thus optimizes real-time conversion results and user experience.

[0148] S3.5: The coordinated policy weights of each sub-agent are used to guide the FQI algorithm for parallel policy updates. The sparse structure of the user-group membership matrix isolates the training process of different group parameters, resulting in the original policy action set.

[0149] It should be noted that the specific steps for parallel policy updates are as follows: Based on the coordinated policy weights of each sub-agent, the FQI algorithm is distributed to multiple computing nodes. Each node independently calculates the policy update for a specific sub-agent. The intermediate results between nodes are synchronized using a distributed computing framework, and the updated policy weights are aggregated to achieve parallel policy updates.

[0150] Example Description: Assume there are three sub-agents responsible for bidding coefficient, frequency control, and creative selection, respectively. The policy update task of each sub-agent is assigned to an independent computing node. Node 1 loads user data and model parameters related to bidding coefficient, Node 2 loads user data and model parameters related to frequency control, and Node 3 loads user data and model parameters related to creative selection. Each node independently calculates the policy update based on the FQI algorithm. Intermediate results are synchronized through a distributed computing framework (such as Spark or Ray), and the updated policy weights are aggregated to complete the parallel policy update.

[0151] S3.6: Using the lower bound of the confidence interval of the counterfactual gain matrix as a hard boundary, the feasible region of bidding actions is reconstructed by the Lagrange multiplier method to generate a set of policy candidates.

[0152] Specifically, the lower bound estimate of the confidence interval of the counterfactual gain matrix is ​​extracted using the BCa Bootstrap method and mapped to a hard constraint for the bidding action;

[0153] Specifically, BCa Bootstrap sampling is performed based on the counterfactual gain matrix to generate multiple sample matrices. The statistics (such as the mean) of each sample matrix are calculated, and the statistical distribution is constructed. The bias correction factor and acceleration factor are calculated based on the distribution, the upper and lower limits of the confidence interval are adjusted, the lower bound estimate of the confidence interval is extracted, and it is mapped as a hard constraint on the bidding action to limit the bidding range and ensure the robustness and reliability of the bidding strategy.

[0154] A better approach is to "map the bid action to a hard constraint" that can directly transform the lower bound estimate of the confidence interval of the counterfactual gain matrix into the upper or lower bound of the bid range, ensuring that the bidding strategy is executed within a robust range and avoiding excessively high or low bids that could lead to wasted advertising budget or insufficient exposure. This hard constraint provides a reliable foundation for subsequent multi-objective optimization strategies, improving the effectiveness of ad placement and budget utilization, while enhancing the interpretability and controllability of the strategy.

[0155] Define a basic revenue function based on business objectives, and construct a constrained optimization model by combining hard constraints;

[0156] Specifically, the basic revenue function is defined based on business objectives, and the steps are as follows: First, determine the key variables of the revenue function based on the business objectives (such as maximizing ad click-through rate or conversion rate), such as click count, conversion count, or ad revenue. Second, quantify the key variables into mathematical expressions, such as click count multiplied by cost-per-click, or conversion count multiplied by cost-per-conversion. Third, introduce weighting coefficients to balance the priority of different business objectives, such as a click-through rate weight of 0.6 and a conversion rate weight of 0.4. Finally, generate the basic revenue function for subsequent optimization model construction.

[0157] The steps to construct a constrained optimization model by incorporating hard constraints are as follows: Map the lower bound estimate of the confidence interval of the counterfactual gain matrix to hard constraints on bidding actions, such as upper and lower bounds for bidding; transform the hard constraints into mathematical inequalities, such as bidding range constraints:

[0158] b min ≤b≤b max ;

[0159] Where b represents the variable for the bidding action, that is, the bid value set when the advertisement is placed. min This indicates the lower limit of a bid, i.e., the minimum allowed bid value. max This indicates the upper limit of the bidding action, that is, the maximum allowed bid value.

[0160] Example: Suppose b min =0.5 yuan, b max = 5 yuan, then the value range of bid action b is [0.5, 5] yuan, ensuring that the bid is neither lower than 0.5 yuan (to avoid insufficient exposure) nor higher than 5 yuan (to avoid wasting budget).

[0161] Furthermore, the basic revenue function is integrated with hard constraints to construct a constrained optimization problem, such as maximizing the revenue function under bidding constraints. The optimization problem is solved using the Lagrange multiplier method or the penalty function method, and a bidding strategy that meets the business objectives and constraints is output for subsequent advertising execution.

[0162] The optimal solution that satisfies the constraints is obtained by solving the optimization model using the Lagrange multiplier method.

[0163] Within the feasible region defined by hard constraints, action generation (including sampling and enumeration) is performed, and combined with frequency control and creative selection strategies through a Cartesian product to generate a set of policy candidates.

[0164] Specifically, within the feasible bidding region defined by hard constraints, a set of bidding actions is generated through sampling or enumeration, for example, within the bidding range [b min ,b max100 bid values ​​are uniformly sampled within the range; the set of bid actions is combined with the frequency control strategy (such as high, medium and low levels) and the creative selection strategy (such as Top 10 creative materials) by Cartesian product to generate a strategy candidate set. For example, 100 bid values ​​× 3 levels of frequency control × 10 creative materials = 3000 strategy combinations, which are used for subsequent evaluation and selection of multi-objective optimization strategies.

[0165] Ideally, the generated strategy candidate set provides a comprehensive strategy combination foundation for the evaluation and selection of subsequent multi-objective optimization strategies, ensuring the synergy of bidding, frequency control, and creative selection strategies; through Cartesian product combination, it covers all possible strategy combinations, improving the flexibility and accuracy of strategy optimization, thereby maximizing the advertising effect within hard constraints, while meeting the business needs of frequency control and creative selection.

[0166] S3.7: Use BCa Bootstrap (bias-corrected accelerated bootstrap) to validate the policy candidate set, and use the Vickrey-Clarke-Groves (incentive-compatible auction mechanism design framework) mechanism to screen Pareto optimal policies as multi-objective optimization policies.

[0167] It should be noted that the Pareto optimal strategy execution results will be written to the user behavior log in real time via Kafka stream, triggering incremental updates to the user-group membership matrix (update period ≤ 5 minutes).

[0168] S4: Collect real-time conversion data after the execution of multi-objective optimization strategies, evaluate the degree of user group shift and counterfactual prediction error through Mahalanobis distance, and trigger the reconstruction of the cluster model and the calibration of the causal model.

[0169] Specifically, it includes the following steps:

[0170] S4.1: Capture user conversion events through the real-time log stream of the ad delivery engine, associate the sparse dynamic clustering matrix and the counterfactual gain matrix by user ID, align the time window and standardize the features, and output the aligned real-time conversion dataset.

[0171] Among them, the real-time logs of the ad delivery engine are the core component of the advertising system, responsible for making real-time decisions on ad display strategies (such as bidding, frequency control, and creative selection).

[0172] S4.2: Load the baseline feature matrix and the current feature matrix, calculate the covariance matrix and mean vector, and quantify the user group distribution offset using Mahalanobis distance.

[0173] S4.3: Input the historical Mahalanobis distance sequence into the exponentially weighted moving average model, dynamically generate the offset judgment threshold based on the historical mean and volatility, and determine whether the current user group distribution offset exceeds the stable range.

[0174] Here, the historical mean refers to the average entropy-regularized Wasserstein distance over a historical period. Volatility refers to the standard deviation of the historical distribution distance, used to quantify the dispersion of historical data.

[0175] S4.4: Extract the actual conversion gain of the user group from the real-time conversion dataset, compare it with the predicted value in the counterfactual gain matrix, calculate the group-level error and weighted aggregate it into a global prediction error index.

[0176] S4.5: If the user group distribution offset exceeds the offset judgment threshold and the global prediction error index exceeds the tolerance range (0.15), the cluster model reconstruction and causal model calibration will be triggered immediately. If a single index exceeds the limit (such as only the global prediction error exceeding 0.15), it will be triggered after monitoring multiple time windows.

[0177] It should be noted that the tolerance range (0.15) is specifically based on the global prediction error index distribution in historical data, calculating the mean and standard deviation of the error, and determining the tolerance range in conjunction with business needs; by analyzing the impact of the error on business objectives (such as ad click-through rate and conversion rate), an error tolerance threshold is set; for example, if the historical error mean is 0.10 and the standard deviation is 0.05, the mean plus one standard deviation (0.15) is selected as the tolerance range; through experiments to verify the impact of the 0.15 threshold on model performance and business indicators, ensuring that the triggering conditions for model reconstruction and calibration are reasonable and effective, 0.15 is finally determined as the tolerance range.

[0178] Among them, the group model reconstruction refers to the retraining of the adversarial generative network and the updating of the user-group membership matrix; the causal model calibration refers to the iterative correction of the propensity score and potential outcome prediction model.

[0179] It should be noted that the clustering model is the core component for dynamically dividing user groups. Based on user behavior data streams, it extracts behavior sequences through a sliding window, utilizes time-decaying Transformer encoding features, and employs a Generative Adversarial Network (GAN) to generate a user-group membership matrix. The core output of this clustering model is a sparse dynamic clustering matrix, used to describe the membership relationship between users and groups (e.g., user A has a probability of 0.8 belonging to group 1 and a probability of 0.2 belonging to group 2).

[0180] The causal model is the core component of counterfactual gain calculation. Based on user segmentation results, it constructs a propensity score model and a potential outcome prediction model through dual machine learning (DML) to estimate the causal effect of different advertising strategies on each user group (e.g., strategy A increases the conversion rate of group 1 by 15%). The core output is the counterfactual gain matrix, which is used to guide the generation of multi-objective optimization strategies.

[0181] It should also be noted that cluster model reconstruction refers to the process of retraining the generative adversarial network (GAN) to generate a new cluster matrix and overwrite the original result when a user group distribution shift is detected (Mahanobis distance exceeds the threshold). The reconstruction process directly relies on GAN generation, sparsification, and cluster merging rules to ensure that the cluster structure is dynamically updated according to the data distribution.

[0182] Causal model calibration refers to the process where, when the counterfactual prediction error exceeds the tolerance range of the global prediction error index (0.15), enhanced features (weighted concatenation and interaction terms) need to be regenerated based on the latest clustering matrix, the propensity score model (Focal Loss) is iteratively updated, and the potential outcome prediction model (SHAP constraint) is corrected. The calibration process directly calls the gain calculation flow and feature enhancement rules to ensure that causal inference is consistent with real-time data.

[0183] S4.6: Retrain the adversarial generative network based on the current feature matrix, use the cluster center alignment loss to generate a new sparse dynamic clustering matrix, and cover the original clustering results through sparsification and merging rules.

[0184] Specifically, the adversarial generative network is retrained based on the current feature matrix to generate new virtual features. The distribution consistency between virtual and real features is constrained by the cluster center alignment loss. The new feature matrix is ​​clustered using the K-means clustering algorithm to generate a new sparse dynamic clustering matrix. Clusters that are close to each other are merged according to the cluster center distance threshold to cover the original clustering results. The updated user-group membership matrix is ​​then output for subsequent optimization and execution of advertising strategies.

[0185] This embodiment also provides a computer device applicable to the optimization method combining user groups and the pre-deployment advertising effect, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the optimization method combining user groups and the pre-deployment advertising effect proposed in the above embodiment.

[0186] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.

[0187] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the optimization method proposed in the above embodiments for combining user groups with the pre-deployed advertising effect. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0188] In summary, this invention encodes user behavior sequences using a time-decaying Transformer, dynamically detects feature distribution drift using entropy-regularized Wasserstein distance, and generates a sparse user-group membership matrix using a generative adversarial network to address the lag in traditional static grouping. Subsequently, it constructs enhanced features by combining group statistics interaction terms, and removes confounding variables through dual machine learning and orthogonalized residual calculation, generating a counterfactual gain matrix with confidence interval correction to eliminate causal inference bias. Furthermore, it utilizes the sparse structure of the user-group membership matrix to isolate multi-agent parameter training, couples bidding, frequency control, and creative selection strategies through an attention mechanism, and uses a VCG mechanism to screen for Pareto optimal solutions. Based on Mahalanobis distance, it quantifies the correlation between user group offset and prediction error, dynamically triggering model reconstruction and calibration to form a "detection-decision-feedback" closed loop, achieving continuous optimization of advertising resource utilization and conversion rates, effectively reducing user fatigue and shortening the strategy iteration cycle to the minute level.

[0189] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An optimization method combining user groups and pre-deployment advertising effectiveness, characterized in that: include, Real-time acquisition of user behavior data streams; detection of user feature distribution drift using Wasserstein distance; triggering adversarial generative network to dynamically divide user groups when the drift threshold is exceeded; outputting user group labels and user-group membership matrix. A dual machine learning approach is used to construct a causal model, calculate the potential gain of each user group under different advertising strategies, and generate a counterfactual gain matrix. By inputting the user-group membership matrix and the counterfactual gain matrix into the multi-agent reinforcement learning framework, a multi-objective optimization strategy including bidding coefficients, frequency control, and creative selection is dynamically generated. Real-time conversion data after the execution of multi-objective optimization strategies is collected, and the degree of user group shift and counterfactual prediction error are evaluated by Mahalanobis distance, triggering the reconstruction of the cluster model and the calibration of the causal model.

2. The optimization method as described in claim 1, which combines user groups with the pre-deployment advertising effect, is characterized in that: The specific steps for outputting user group labels and user-group membership matrices are as follows. A sliding window captures the user behavior stream, the behavior sequence is encoded and standardized using a time-aware Transformer, and the current feature matrix and the baseline feature matrix are output. Calculate the entropy-regularized Wasserstein distance between the current distribution and the baseline distribution, and dynamically adjust the drift judgment threshold by combining it with the exponentially weighted moving average. If drift is detected, the adversarial generative network is activated to generate virtual features. The generation process is constrained by the cluster center alignment loss, and the original membership matrix is ​​output. The original membership matrix is ​​subjected to Top-3 sparsification, adjacent clusters with a center distance of less than K are merged, and the compressed data is stored in a distributed database.

3. The optimization method as described in claim 2, which combines user groups with the pre-deployment advertising effect, is characterized in that: The drift determination threshold refers to the critical value used to determine whether the current distribution has undergone a significant change. The dynamically adjusted drift determination threshold includes, The initial threshold is set based on the mean and standard deviation of the historical distribution distance; The real-time drift determination threshold is calculated by combining the current distance and the threshold of the previous time step with an exponentially weighted moving average.

4. The optimization method for combining user groups and pre-deployment advertising effectiveness as described in claim 3, characterized in that: The specific steps for generating the counterfactual gain matrix are as follows: The user behavior features are weighted and concatenated with the sparse cluster matrix, and then combined with the group statistics interaction terms to generate enhanced features. Individual treatment effects are calculated by orthogonalizing residuals, and weighted aggregation is performed with the introduction of a time decay factor to correct the potential population gain. Construct a three-dimensional counterfactual gain tensor across policies and groups, and apply KL divergence constraints to merge policy differences among similar groups; Significance is evaluated using BCa Bootstrap sampling, and the gain matrix after confidence interval correction is output.

5. The optimization method as described in claim 4, which combines user groups with the pre-deployment advertising effect, is characterized in that: The steps for generating enhanced features by weighted concatenation of user behavior features with a sparse dynamic clustering matrix and combining them with group statistical interaction terms are as follows: Standardize user behavior characteristics and concatenate them with the group membership degree according to preset weights; Calculate the group mean and activity decay value, generate behavioral statistics indicators, and interact with user characteristics item by item; After merging, dimensionality reduction is performed using PCA, and Gaussian noise is added to form enhanced features.

6. The optimization method as described in claim 5, which combines user groups with the pre-deployment advertising effect, is characterized in that: The dynamic generation includes a multi-objective optimization strategy encompassing bidding coefficients, frequency control, and creative selection. The specific steps are as follows: The user-group membership matrix and the counterfactual gain matrix are concatenated into a joint state vector, which is then superimposed with real-time context features. Define the bidding coefficient as a continuous action, the frequency control as a three-level discrete action, and the creative selection as a discrete action of the Top 10 creatives; The agent dynamically couples sub-policy weights by using LSTM, and updates the policy in parallel using the FQI algorithm. Using the lower bound of the confidence interval of the counterfactual gain matrix as a hard constraint, a policy candidate set is generated by the Lagrange multiplier method. The VCG mechanism is used to screen for Pareto optimal strategies.

7. The optimization method as described in claim 6, which combines user groups with the pre-deployment advertising effect, is characterized in that: The specific steps for generating the policy candidate set are as follows: Map the lower bound of the confidence interval sampled by BCa Bootstrap to the hard boundary of the bid action; Construct a constrained optimization model and solve it using the Lagrange multiplier method; Generate bidding actions within the feasible region and combine them with frequency control and creative strategies via a Cartesian product.

8. The optimization method as described in claim 7, which combines user groups with the pre-deployment advertising effect, is characterized in that: The real-time conversion data collected after the multi-objective optimization strategy is executed is used to assess the degree of user group shift and counterfactual prediction error through Mahalanobis distance, triggering cluster model reconstruction and causal model calibration. The specific steps are as follows. Align real-time transformed data with the clustering matrix and calculate the Mahalanobis distance offset of user group distribution; Calculate the global error index by comparing the actual conversion gain with the predicted value; If the offset exceeds the offset judgment threshold and the error exceeds the tolerance range, the cluster model reconstruction and causal model calibration will be triggered immediately. If a single indicator exceeds the limit, it will be triggered after monitoring multiple time windows continuously.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the optimization method according to any one of claims 1 to 8, which combines user groups with the pre-deployment advertising effect.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the optimization method according to any one of claims 1 to 8, which combines user groups with the pre-deployment advertising effect.

Images