Virtual resource intelligent scheduling method based on big data user portrait and ai prediction model

Abstract

The application discloses a virtual resource intelligent scheduling method based on big data user portrait and an AI prediction model, which comprises the following steps: performing unified cost mapping on each resource type, unifying resource strength parameters of different dimensions into costs measured by money, and constructing a historical incentive event sequence; training a hierarchical binary classification user usage probability prediction model, calculating the incentive response elasticity coefficient of each user under each business line and resource type; constructing a resource cancellation cost prediction model, outputting the point prediction value and confidence interval of the cancellation cost in each future time window, generating a guaranteed quota and an elastic quota parameter; dividing the available quota of the day into a guaranteed quota layer and a global elastic quota layer, performing target function value operation on the candidate resource combination, and generating a resource allocation instruction; performing inner loop consumption rate correction and outer loop model calibration, updating the adjustment parameter, and realizing probability calibration or model retraining. The application realizes precise personalized allocation and automatic iterative optimization of virtual resources.

Description

Technical Field

[0001] This invention belongs to the field of big data processing and artificial intelligence technology, specifically relating to a virtual resource intelligent scheduling method based on big data user profiling and AI prediction models. Background Technology

[0002] With the rapid development of the mobile internet and digital content industry, various digital content distribution and service platforms commonly use virtual resources as a core operational tool for user incentives and retention. The effectiveness of virtual resource deployment directly impacts the platform's user activity, paid conversion rate, and overall revenue performance. In these platforms, virtual resources typically include various types such as electronic vouchers, discounts, digital gift packs, activity points, and limited-time benefits; their allocation and distribution are crucial aspects of platform operation.

[0003] In existing virtual resource scheduling technologies, resource allocation strategies are typically quite broad, lacking personalized and fine-grained matching. Common practices include issuing vouchers of the same value to all users, or simply stratifying users based on simple rules such as registration duration or membership level and then offering fixed discounts. This allocation method fails to consider individual differences among users in terms of spending power, activity level, and incentive sensitivity, easily leading to a large allocation of resources to low-responsive users, resulting in resource waste. Simultaneously, the incentives for high-value users may be insufficient, making it difficult to effectively drive their paid conversions.

[0004] Furthermore, existing resource scheduling systems typically employ a "responsive" design, meaning resource allocation logic is only executed when a user triggers a specific action (such as logging in or entering an activity page). Such systems generally cannot predict future user activity levels, peak willingness to pay, or resource consumption trends across different business lines. Consequently, platforms may face temporary resource budget shortages or hastily implemented allocation plans during peak operating periods (such as holidays or large-scale events), while during off-peak periods, resources may be idle and budgets may be underutilized.

[0005] In large-scale digital content platforms, multiple business units (such as content channels of different categories, special activity zones, etc.) are typically operated simultaneously, and the virtual resource pools of each business unit are generally managed independently. Existing technologies usually lack a global resource coordination and optimization mechanism across business lines, making it difficult to achieve optimal resource allocation among business lines under the constraint of the total budget. This can easily lead to a structural imbalance of "insufficient resources in popular business lines and redundant resources in less popular business lines."

[0006] Regarding resource deployment effectiveness evaluation and strategy iteration, existing systems typically rely solely on simple summary statistics (such as claim rate and usage rate) as the basis for effectiveness evaluation after resource deployment, lacking real-time tracking and attribution analysis of individual user response behavior. Iterative updates to deployment strategies generally depend on the experience and judgment of operations personnel and periodic (usually weekly or monthly) manual reviews, making it difficult to achieve automated strategy optimization based on real-time feedback data. Consequently, the update frequency and accuracy of scheduling strategies cannot meet the needs of refined operations.

[0007] In summary, existing virtual resource scheduling technologies are insufficient in terms of personalized matching, forward-looking prediction, cross-business line collaboration, and automated iteration, making it difficult to meet the requirements of digital content distribution platforms for refined, automated, and real-time scheduling in scenarios with high concurrency, multiple business lines, and multiple resource types. Summary of the Invention

[0008] This invention aims to address the shortcomings of existing virtual resource scheduling technologies in terms of personalized matching, forward-looking prediction, cross-business line collaboration, and automated iteration. It provides a virtual resource intelligent scheduling method based on big data user profiles and AI prediction models to meet the requirements of high concurrency, multiple business lines, and multiple resource types in digital content distribution platforms for refined, automated, and real-time scheduling.

[0009] To address the aforementioned technical problems, this invention provides a method for intelligent virtual resource scheduling based on big data user profiling and AI prediction models, the method comprising: S1: Collect user historical resource deployment records, user behavior logs, user context feature data, resource type configuration data, and historical resource write-off cost amount time series; perform a unified cost mapping operation on the resource intensity parameters of each resource type, and uniformly map resource intensity parameters of different dimensions to a unified cost amount measured in monetary terms; construct a historical incentive event sequence based on the unified cost amount; S2: Construct and train a user usage probability prediction model, which is a hierarchical binary classification model; based on the trained user usage probability prediction model, calculate the incentive response elasticity coefficient of each user under each business line and each resource type through finite difference numerical differentiation operation. S3: Based on the historical resource write-off cost amount time series of each business line, construct a resource write-off cost prediction model; use the resource write-off cost prediction model to output the point prediction value and confidence interval of the resource write-off cost amount of each business line in each future time window; generate the minimum quota parameter and flexible quota parameter of each business line based on the point prediction value and the confidence interval. S4: Based on the guaranteed quota parameter and the flexible quota parameter, the total available quota for the day is divided into a guaranteed quota layer and a global flexible quota layer; when a real-time user request is received, based on the usage probability value output by the user usage probability prediction model, the marginal gain coefficient after write-off, the global resource adjustment parameter, and the business line resource adjustment parameter, the objective function value calculation is performed on the candidate resource combination, the resource combination that maximizes the objective function value is selected to generate a resource allocation instruction, and quota deduction and status update are performed; S5: Perform inner-loop consumption rate correction operation, and adaptively update the global resource adjustment parameters and the business line resource adjustment parameters based on the deviation between the actual net consumption rate and the predicted consumption rate; perform outer-loop model calibration operation, and perform probabilistic calibration mapping or model retraining based on the deviation between the actual utilization rate and the predicted utilization rate of the calibration unit.

[0010] The unified cost mapping operation performs differentiated processing based on resource type: Electronic vouchers are directly used as the uniform cost amount based on their face value; discount amounts are obtained by multiplying the actual redeemed order amount or the average historical order amount by the discount percentage; activity points are obtained by dividing the number of points by the preset points redemption value coefficient; digital gift packs or limited-time benefits are obtained by querying the fixed cost configuration table based on the resource type identifier and resource strength parameters.

[0011] For unused discount quota records, the average historical order amount of the user within the preset time window of the business line is used first for calculation. If there is no historical order record, the average reimbursement order amount within the preset time window of the business line is used instead, thereby avoiding the use of future information and ensuring consistency between online inference and training.

[0012] The user usage probability prediction model adopts a hierarchical binary classification model that includes a global intercept term, a user random effect term, a business line random effect term, a resource type random effect term, a global resource intensity sensitivity coefficient, a personalized intensity sensitivity adjustment term, and a context feature weight vector. The user random effects term, business line random effects term, and personalized intensity sensitive adjustment term are parameterized by the sum of the linear projection of the observable features and the Gaussian prior residuals with a mean of zero.

[0013] Before training, a weight construction operation is performed on each sample. The sample weight is the product of the time decay weight and the causal debiasing weight. The time decay weight is estimated in layers according to resource type and user activity level. The causal debiasing weight is based on the propensity score and applies a stabilization factor and truncation processing. After the weights are constructed, the effective sample size is checked. If the effective sample size is lower than the preset threshold, the weights are rescaled or the truncation threshold is adjusted.

[0014] For user-business-resource type combinations with sparse samples, a rollback process is performed, setting the personalized intensity sensitivity adjustment term of the combination to zero, and retaining only the global resource intensity sensitivity coefficient and the prior mean of the group level as the basis for prediction.

[0015] After calculating the excitation response elasticity coefficient, users can be grouped and labeled according to the comparison results between the elasticity coefficient and the preset high and medium elasticity thresholds. The elasticity coefficient, group labels and associated model parameters are written into key-value storage to form a user response elasticity parameter dataset.

[0016] The resource write-off cost prediction model performs a logarithmic transformation on the time series of historical resource write-off cost amounts, and then uses STL decomposition based on local weighted regression to obtain the trend term, seasonal term, and residual term. The trend term is fitted using a Holt linear trend model or locally weighted linear regression, the seasonal term is parametrically represented using Fourier series expansion, and the residual term is modeled using an autoregressive integral moving average model. The model can also incorporate external event regression terms to describe the impact of single-day or multi-day activities on write-off costs using an event impact function.

[0017] When generating minimum quota parameters and flexible quota parameters based on point prediction values ​​and confidence intervals, the lower limit of the confidence interval prediction is used as the minimum quota parameter, and the difference between the upper and lower limits of the prediction is used as the flexible quota parameter. Time-sharing window splitting and consistency verification are performed to ensure that time-sharing predictions are aligned with daily predictions.

[0018] In S4, when dividing the total available quota for the day into a minimum quota layer and a global flexible quota layer, the basic quota for the day is first calculated based on the estimated amount of the predicted write-off cost for the entire platform and the total budget for the activity period. The net surplus from the previous day is then added to obtain the total available quota for the day. The minimum quota parameters for each business line are truncated and scaled proportionally under the constraint of the upper limit of the business line resource quota to obtain the total amount of the minimum quota layer. The difference between the total available quota for the day and the total amount of the minimum quota layer is truncated by combining the flexible quota parameters of each business line and the upper limit of the business line resource quota. The minimum value is taken to obtain the total amount of the global flexible quota layer. The remaining part is used as a reserve quota balance and frozen.

[0019] When performing objective function value calculation on candidate resource combinations, the objective function value is the product of the usage probability value and the marginal gain coefficient after write-off, minus the expected consumption amount weighted by the global resource adjustment parameter and the business line resource adjustment parameter; the model input cost and security pre-occupancy amount of the discount quota type are calculated separately and used for objective function value calculation and quota verification, respectively.

[0020] When quota verification fails, the combination with the largest objective function value is reselected from the affordable candidate set that meets the requirements of the security pre-allocation not exceeding the remaining amount of the global elastic quota layer and not exceeding the available elastic quota of the business line. When the affordable candidate set is empty or the objective function value of all combinations is not greater than zero, the decision result is set to an empty combination and transferred to the minimum quota layer for processing or the issuance instruction is rejected. When there are multiple candidate combinations with equal and positive objective function values, the one with the smaller security pre-allocation is selected first. If the security pre-allocation is the same, the one with the higher efficiency ratio is selected.

[0021] In S5, the inner-loop consumption rate correction calculates the relative deviation between the actual net consumption rate and the predicted consumption rate at both the business line level and the global level. When the relative deviation of the business line exceeds the inner-loop trigger threshold, the correction term of the business line resource adjustment parameter is updated and subjected to exponential smoothing and range truncation. The business line resource adjustment parameter adopts a structure of superimposed benchmark value and correction term. The global resource adjustment parameter is obtained by multiplicatively updating the global relative deviation after exponential moving average smoothing, and range truncation protection is performed on the multiplicative factor and the update result.

[0022] The outer loop model calibration defines calibration units based on a three-dimensional combination of user groups, business lines, and resource types. It calculates the deviation between the actual and predicted usage rates of each calibration unit. When the absolute value of the deviation exceeds the outer loop trigger threshold, it fits an ordinal regression or Pratt scaling mapping to the calibration unit to obtain the probability update mapping function parameters. During online inference, it queries the probability update mapping function parameters in the following priority order: user group × business line × resource type, business line × resource type, resource type, user group, and global default. If a high priority does not exist, it falls back to a low priority.

[0023] When the number of calibration units that trigger probability calibration or the average deviation exceeds a preset threshold, the user is triggered to retrain the probability prediction model; when the resource write-off cost prediction model shows continuous deviations across multiple business lines, the cost prediction model is triggered to retrain.

[0024] Compared with the prior art, the present invention has the following advantages: (1) By unifying the resource intensity parameters of different dimensions into a cost measured in monetary terms through unified cost mapping, and by using a hierarchical binary classification user usage probability prediction model combined with causal biased sample weights to perform personalized incentive response estimation and elasticity coefficient calculation for user-business line-resource type combination, the precise matching of resource allocation and user incremental response capability is achieved, reducing resource waste for low-response users and improving incentive efficiency for high-elasticity users, thereby solving the problem of extensive resource allocation and lack of personalized fine matching in the existing technology.

[0025] (2) By performing logarithmic transformation, STL decomposition and itemized modeling on the historical write-off cost time series through the resource write-off cost prediction model, the point prediction value and confidence interval of each future time window are output, and the minimum quota and flexible quota parameters are generated accordingly. This realizes the forward prediction of resource consumption and dynamic planning of budget, avoiding insufficient resource budget during peak periods or idle resources during off-peak periods, thus solving the problem that the existing responsive scheduling system lacks the ability to predict in advance.

[0026] (3) By dividing the total available quota for the day into a minimum quota layer and a global elastic quota layer, and by scoring and verifying the candidate resource combination in real time based on the objective function including the probability of use, the marginal gain coefficient after cancellation, the global and business line resource adjustment parameters and the expected consumption, request-level resource scheduling is carried out under the global budget constraint. This realizes the unified configuration and budget consistency control of resources across multiple business lines, reduces the risk of structural imbalance of resources between business lines, and solves the problem of lack of global overall planning and optimization across business lines in the existing technology.

[0027] (4) By adaptively updating the global and business line resource adjustment parameters through the inner loop consumption rate correction operation, and performing probability calibration mapping or model retraining based on the deviation between the actual utilization rate and the predicted utilization rate of the calibration unit, an automated closed-loop iterative mechanism of "deployment-evaluation-calibration-re-deployment" is formed, which improves the update frequency and accuracy of the scheduling strategy and reduces the reliance on manual review. This solves the problems of strategy iteration relying on manual experience, long update cycle and inability to adapt to real-time feedback optimization in the existing technology.

[0028] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0029] Figure 1 This is a flowchart illustrating an intelligent virtual resource scheduling method based on big data user profiling and AI prediction models, provided in an embodiment of the present invention.

[0030] Figure 2 This is a schematic diagram of a two-tier quota allocation process provided in an embodiment of the present invention.

[0031] Figure 3 This is a schematic diagram of a two-tier quota collaborative allocation process provided in an embodiment of the present invention.

[0032] Figure 4 This is a schematic diagram of a dual-loop adaptive parameter iterative closed loop provided in an embodiment of the present invention. Detailed Implementation

[0033] This detailed description of the embodiments will be provided in conjunction with the accompanying drawings. It should be understood that the embodiments described below are only a part of the embodiments of the present invention, not all of them, and are intended to explain the technical principles of the present invention, and do not constitute any limitation on the scope of protection of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0034] The following are explanations and definitions of the terms used in the embodiments of this invention: Virtual resources: Non-physical resources used to incentivize users in internet service platforms, including but not limited to electronic vouchers, discount quotas, digital gift packs, activity points, and limited-time benefits.

[0035] Resource strength: The specific quantitative tiers used when distributing virtual resources, such as the face value of vouchers, the discount percentage, the number of points, the gift package tier number, and the duration of benefits.

[0036] Standardized cost amount: This refers to the cost amount obtained by converting different resource types and resource intensity according to a standardized cost function, and the unit is denoted as yuan. Standardized cost amount is used for model training, online decision-making, and quota control to make different resource types comparable.

[0037] Shared feature model: A unified prediction model trained on the entire platform or a large group of users. It reduces the estimation instability caused by the sparsity of single-user samples by sharing user features, business line features and resource features.

[0038] Stratified random effects: In statistical models, random parameters are set for grouping variables such as users, business lines, and resource types, and a contraction balance is achieved between the group prior and the individual posterior through a prior distribution. When the sample size of a certain group is insufficient, the posterior estimate of the stratified random effects automatically contracts towards the group mean.

[0039] Incentive response elasticity coefficient: refers to the elasticity of user usage probability with respect to changes in uniform cost amount, representing the relative change in usage probability caused by changes in resource intensity. In this application, it is calculated using finite difference numerical differentiation.

[0040] User Usage Probability Prediction Model: A hierarchical binary classification model used to predict the probability of a user using a resource given a user, business line, resource type, and uniform cost amount. The model includes a global intercept term, a user / business line / resource type random effects term, global and personalized sensitivity coefficients, and contextual feature weights.

[0041] Resource write-off cost prediction model: This refers to a prediction model built based on the time series of historical resource write-off cost amounts and external event information on a logarithmic scale, through time series decomposition and component modeling. It is used to output the point prediction value and confidence interval of the resource write-off cost amount of each business line within a future time window.

[0042] Calibration Unit: The smallest statistical unit defined by a three-dimensional combination of user group, service line, and resource type, used for calculating the deviation between actual and predicted utilization rates and for probabilistic calibration in outer loop calibration.

[0043] Inner-loop consumption rate correction: refers to the calculation process of adaptively updating global resource adjustment parameters and business line resource adjustment parameters based on the relative deviation between the actual net consumption rate and the predicted consumption rate, used to control the quota consumption rhythm.

[0044] Outer loop model calibration: refers to the process of performing probabilistic calibration mapping and / or model retraining operations at preset intervals based on the deviation between the actual and predicted usage rates of calibration units, in order to correct systematic biases in the model.

[0045] Delayed feedback correction: refers to a method that corrects the observed write-off rate by estimating the conditional probability of completing write-off within the remaining validity period for samples that have not yet expired, in the face of a time delay in resource write-off.

[0046] The implementation of this invention will be described in detail below with reference to the aforementioned terminology.

[0047] This invention proposes a virtual resource intelligent scheduling method based on big data user profiling and AI prediction models. The method is executed on a computer device, which includes at least one processor, a memory, and a network communication interface. The memory stores a computer program and structured data such as a user behavior log table, a user historical resource allocation record table, a quota status data table, a user response elasticity parameter dataset, and an algorithm parameter configuration table. When the processor executes the computer program, it implements the following steps, and data transfer between steps is completed through a shared data table or message queue in the memory.

[0048] refer to Figure 1 The method includes the following steps: S1: Historical Data Collection and Feature Construction: Collect user historical resource delivery records, user behavior logs, user context feature data, and resource type configuration data; perform a unified cost mapping operation on the resource intensity parameters of each resource type in the user's historical resource delivery records to convert resource intensity parameters of different dimensions into a unified cost amount with a unified dimension; construct a historical incentive event sequence based on the unified cost amount, usage flag, delivery timestamp, and user context feature vector; S2: User Response Elasticity Parameter Modeling: Construct a user usage probability prediction model, which is a hierarchical binary classification model including a global intercept term, a user random effect term, a business line random effect term, a resource type random effect term, a global resource intensity sensitivity coefficient, a personalized intensity sensitivity adjustment term, and a context feature weight term; train the user usage probability prediction model; based on the trained user usage probability prediction model, calculate the incentive response elasticity coefficient of each user under each business line and each resource type through finite difference numerical differentiation operation; S3: Forward-looking resource demand forecasting and pre-scheduling parameter generation: Based on the historical resource write-off cost amount time series of each business line, construct a resource write-off cost forecasting model; use the resource write-off cost forecasting model to output the point forecast value and confidence interval of the resource write-off cost amount of each business line in each future time window; generate the minimum quota parameter and flexible quota parameter of each business line based on the point forecast value and the confidence interval. S4: Online Request Processing and Two-Tier Resource Quota Coordination: Based on the guaranteed quota parameter and the elastic quota parameter, the total available quota for the day is divided into a guaranteed quota layer and a global elastic quota layer; when a real-time user request is received, based on the usage probability value output by the user usage probability prediction model, the marginal gain coefficient after write-off, the global resource adjustment parameter, and the business line resource adjustment parameter, the objective function value calculation is performed on the candidate resource combination, the resource combination that maximizes the objective function value is selected to generate a resource allocation instruction, and quota deduction and status update are performed; S5: Resource allocation feedback and dual-loop adaptive parameter iteration: Perform inner loop consumption rate correction operation, and adaptively update the global resource adjustment parameters and the business line resource adjustment parameters based on the deviation between the actual net consumption rate and the predicted consumption rate; perform outer loop model calibration operation, and perform probabilistic calibration mapping and / or model retraining based on the deviation between the actual utilization rate and the predicted utilization rate of the calibration unit.

[0049] Among them, S1 to S3 are executed in batches by the processor in the offline stage, S4 is executed in real time in response to user requests in the online stage, and the inner loop consumption rate correction operation in S5 is triggered and executed on a timed basis according to a preset time window, while the outer loop model calibration operation is triggered and executed on a daily or weekly basis.

[0050] S1: Historical data collection and feature construction.

[0051] S1 provides standardized training data for subsequent user response resilience parameter modeling and forward-looking resource demand forecasting. Its role is to unify multi-source data into a structured feature representation and eliminate dimensional differences between different resource types, enabling subsequent model training and online inference to be performed on a unified numerical scale. S1 specifically includes the following sub-steps.

[0052] S11: Collect raw data.

[0053] Through a pre-defined data acquisition interface, user behavior log data is obtained from a distributed log storage system, historical resource deployment records are read from a transaction database, user contextual feature data is read from a user profile database, resource type dictionaries, basic business line information, resource fixed cost configuration tables, and activity budget configuration data are read from a platform configuration database, external event calendar data is obtained from an external data interface, and historical resource write-off cost amount time series are read from a time series database. All these raw data types are then aligned and format-validated before being stored in corresponding data tables in the form of structured records.

[0054] The raw data includes: User historical resource deployment records include the following fields: user ID, business line ID, resource type ID, resource intensity parameter, deployment timestamp, claim flag, usage flag, usage timestamp, expiration timestamp, order amount, actual discount amount, payment amount, and actual redemption consumption.

[0055] User behavior logs are event sequences sorted by timestamps. Each event record includes an event type identifier, user identifier, timestamp, associated business line identifier, and associated amount. The event types include login, browsing, clicking, searching, adding to favorites, adding to cart, placing an order, and payment.

[0056] The resource type dictionary includes fields such as: resource type identifier, resource type name, resource intensity parameter, validity period, cost conversion factor, quota deduction method, and quota consumption indicator. The quota deduction method includes two values: "deduct immediately upon issuance" and "deduct after verification." "Deduct immediately upon issuance" means that the actual consumption is deducted from the corresponding quota level when the resource is issued; "deduct after verification" means that only the quota is pre-allocated when the resource is issued, and the actual deduction is performed after the verification event is confirmed. The quota consumption indicator is either "yes" or "no." Resource types with a value of "no" (such as pure information display rights, cost-free points reminders, etc.) do not undergo quota deduction or pre-allocation calculations when issued.

[0057] User context features include fields such as: user device type, operating system, city, number of active days in the past 30 days, amount paid in the past 90 days, average order amount in the past 90 days, membership level, historical resource utilization rate, and current business line identifier.

[0058] The historical resource write-off cost amount time series is stored at the hourly and daily levels, and auxiliary indicators such as issuance and receipt are stored synchronously.

[0059] The external event calendar includes fields such as: event type identifier, event start date, event end date, and event name; the event type identifier includes "major promotion", "holiday", and "new product launch".

[0060] Basic information for the business line includes fields such as: business line identifier, product category, historical average daily write-off cost level, business line resource quota limit, average write-off order amount in the past 90 days, and 90th percentile of write-off order amount in the past 90 days.

[0061] The resource fixed cost configuration table includes fields such as: resource type identifier, resource strength parameter, and fixed cost amount; the fixed cost amount is the preset unit cost value corresponding to fixed cost resources such as digital gift packs and limited-time benefits.

[0062] The activity budget configuration data includes fields such as: activity identifier, total budget for the activity period, budget effective date, budget expiration date, amount already disbursed, and remaining available budget; the total budget for the activity period is the upper limit constraint on the total cost of this resource deployment activity.

[0063] The randomized controlled trial grouping table includes the following fields: experiment identifier, user identifier, experiment group identifier (treatment group / control group), experiment start time, experiment end time, and attribution window duration.

[0064] The algorithm parameter configuration table includes fields such as: parameter name, parameter value, applicable business line identifier, applicable resource type identifier, effective timestamp, update timestamp, and parameter description.

[0065] S12: Perform a unified cost mapping operation on the resource strength parameters.

[0066] Because the original intensity parameters for different resource types have different dimensions and data types (vouchers are face value, discounts are percentages, activity points are integer counts, and digital gift packs are fixed costs), directly inputting them into the subsequent model will lead to inconsistent feature scales, affecting model parameter convergence. To solve this technical problem, this step modifies the resource intensity parameters for each historical campaign record. Perform uniform cost mapping operation Convert it into a uniform cost amount with a unified dimension. The transformation result is then written back to the cost field of the corresponding record. The unified cost mapping operation is performed as follows: (1) When the resource type is identified as an electronic voucher, the voucher value is directly assigned to the unified cost amount field.

[0067] (2) When the resource type is identified as a discount amount, the resource strength parameter The discount percentage ranges from [value missing]. First, read the "use flag" of the record: if the use flag is "used", then read the actual reimbursement order amount in the record and execute. The calculation is performed, and the result is written to the unified cost amount field; if the flag is "not used", the average historical order amount of the user in this business line over the past 90 days is read. ,implement The calculation results are written to the unified cost amount field. This is the arithmetic average of the payment amounts for all completed orders (status: 'Transaction Successful' or 'Completed') within the specified time window, and must include at least one valid order record; if the user has no historical order records, then the average amount of orders cleared over the past 90 days for that business line is used. Alternative Perform the above calculations. The calculation method is the arithmetic mean of the payment amounts of all completed orders by users within the same time window for that business line. This approach ensures that unredeemed samples in the discount amount type do not introduce the actual order amount after redemption, eliminates time-series label leakage in the training data, and prevents positive bias in model parameter estimation.

[0068] (3) When the resource type is identified as activity points, the resource intensity parameter For the integral quantity value, according to Calculate the uniform cost amount and write it into the corresponding field, where, The preset points redemption value coefficient is read from the resource fixed cost configuration table in the platform configuration database according to the resource type identifier, representing the cash cost value corresponding to a unit of points.

[0069] (4) When the resource type is identified as a digital gift pack or a limited-time benefit, the fixed cost configuration table read from the platform configuration database by S11 is queried according to the resource type identifier and resource strength parameter, and the value obtained from the table is written into the unified cost amount field.

[0070] S13: Construct a sequence of historical stimulus events.

[0071] For each user Business lines and resource type Using "user identifier + business line identifier + resource type identifier" as the index key, retrieve all deployment records of this user for type k resources under this business line from the user's historical resource deployment record table, sort them in ascending order by deployment timestamp, and form a historical incentive event sequence:

[0072] in, For the first The uniform cost amount for each deployment (obtained from S12); The label is a binary label, taken from the "use flag" of the record. Indicates that it has been used. (Indicates not used); For the delivery timestamp; This is the user context feature vector for this campaign.

[0073] The historical event sequence is stored in structured data form, providing sample inputs in the user-business-resource type dimension for model training in S2.

[0074] S14: Sample sparsity detection and backtracking labeling.

[0075] For each user-business line-resource type combination, perform the following sparse detection operation: The number of records for this combination in the historical stimulus event sequence is counted, and the difference between the maximum and minimum uniform cost amounts within the combination is calculated. If the number of records is less than a preset threshold, or the difference is lower than a preset change range threshold, the sample label field of this combination is set to "rollback" status, and the personalized random effects item corresponding to this combination is skipped in the subsequent model training in S2. The independent estimate sets its personalized strength sensitivity adjustment term to zero, retaining only the global strength sensitivity coefficient. The prior mean of the group hierarchy is used as the basis for predicting this combination.

[0076] The sample label field also records the fallback level, which includes: falling back to a level that does not contain the random effect of the business line, or falling back to a level that only contains the random effect of the resource type.

[0077] By employing the aforementioned sparse detection and backtracking labeling, the problem of excessively large variance in parameter estimation caused by insufficient sample size for a single user is suppressed, enabling the model to output population-level predicted values ​​under sparse data conditions.

[0078] S2: Modeling of user response elasticity parameters.

[0079] S2 is used to construct a personalized response quantification model for users' virtual resource incentives. Its function is to stably estimate the usage probability and incentive response elasticity coefficient of each user under different business lines and resource types under highly sparse data conditions through hierarchical parameter sharing and prior contraction operations. The estimation results are output in the form of a structured parameter dataset for use in online inference of S4 and model calibration of S5. S2 specifically includes the following sub-steps.

[0080] S21: Build and train a user usage probability prediction model.

[0081] To address the technical problem of unreliability in user-specific independent modeling due to sample sparsity, instead of directly fitting independent logistic regression models for each user, business line, and resource type, a globally shared hierarchical binary classification model (i.e., a user usage probability prediction model) is constructed. Through parameter sharing and hierarchical prior structures, the variance of parameter estimation is controlled within an acceptable range.

[0082] For users Business lines Resource types and uniform cost amount User usage probability satisfy:

[0083] in, This is the global intercept term; For users The user random effects term represents the basic usage tendency of users; For business lines The random effects term of the business line; For resource type The random effects term characterizes the differences in usage tendencies across resource type dimensions; This is the sensitivity coefficient for global resource intensity. For users In business lines For resource types Personalized intensity-sensitive adjustment options; For context feature weight vectors; This is the user context feature vector collected by S11; This is the uniform cost amount calculated by S12.

[0084] By sharing global parameters and a hierarchical random effects structure, the user-based probabilistic prediction model automatically shrinks to the population prior when individual samples are insufficient, thereby suppressing the problem of excessive variance in parameter estimation caused by sample sparsity and improving the model's generalization performance on unseen combinations.

[0085] To provide stable estimation under sparse sample conditions , and By employing a feature prior parameterization approach, the random effects at each level are expressed as the sum of linear projections of observable features and residuals:

[0086]

[0087]

[0088] in, For users The feature vector is obtained by aggregated statistical calculation of user context features and user behavior logs stored in S11, including the quantile of paid amount in the past 90 days, the quantile of active days in the past 30 days, historical resource utilization rate, membership level, device price range, and city level. For business lines The feature vector is obtained by aggregation and statistical calculation of the basic information of business lines and the time series of historical resource write-off cost amounts stored in S11, including business line type, average write-off cost over the past 90 days, user overlap, and historical resource write-off rate; the user overlap is the business line... b The ratio of the intersection of the user set and the user set of the entire platform is obtained based on the user identifier statistics in the user's historical resource delivery record table; the historical resource redemption rate is calculated based on the proportion of the number of records with the usage flag 'used' in the user's historical resource delivery record table to the total number of records; For users In business lines and resource type The interaction feature vector is obtained by aggregate statistical operations on the historical stimulus event sequence constructed in S13, including the user's historical interactions with type Resource utilization rate, number of days since last use, and median historical uniform cost of coupons collected; These are the global learning parameters to be learned; For the random effects residuals, it is assumed that they follow a Gaussian prior distribution with a mean of 0.

[0089] The above parameterization method associates random effects with observable features through linear projection. When individual samples are insufficient, the model parameters automatically shrink to the prior mean determined by the population features, thereby suppressing overfitting and taking into account the sample efficiency of parameter estimation.

[0090] Resource type random effects item Employ a feature prior parameterization method similar to that used for user random effects: ,in For resource type k The feature vector includes the category code of the resource type name, the average value of the denomination, and the historical global redemption rate; These are the global learning parameters to be learned; Follows a mean of 0 and a variance of The Gaussian prior distribution. When the historical samples of a certain resource type are insufficient to reach a preset threshold, the corresponding... Set to zero. rollback to .

[0091] If a business line has no historical data or the number of records is less than a preset threshold, the random effect term of that business line is set to zero, and the model is reverted to a shared model that does not include the random effect of that business line. The revert level is recorded in the parameter label.

[0092] To address the technical issues of historical sample distribution drift over time and selection bias introduced by historical delivery strategies, a weight construction operation is performed on each training sample in the historical stimulus event sequence constructed in S13 during the training of the user usage probability prediction model, resulting in sample weights. It is defined as the time decay weight. Causal bias weighting The product of:

[0093] Through this weighted operation, the contributions of recent samples and debiased samples in the loss function are amplified, while the contributions of distant samples and samples affected by policy bias are attenuated, thereby enabling the trained model to have higher fitting accuracy to the current data distribution.

[0094] The time decay weight is calculated using the following formula:

[0095] in, For the first The difference between the delivery timestamp of the record and the current timestamp is converted into the number of days. For the first The decay coefficient for each tiered group (divided by resource type and user activity level).

[0096] Estimation is performed uniformly by tier group rather than per user. Specifically, the estimation method is as follows: extract usage event records of the same resource type within the same tier group from the historical stimulus event sequence, calculate the timestamp difference sequence between two adjacent usage events, and take the median of the sequence. ,but:

[0097] In one embodiment, when the sample size of the stratified group is insufficient, take... This allows the time decay effect to reduce the weight of old samples to half within about two weeks, while avoiding excessive decay that would result in an insufficient number of effective training samples.

[0098] The causal debiasing weights are calculated using a stable inverse probability weighting operation, with a maximum weight truncation. The calculation method is as follows:

[0099] in, For the sample i The tendency score value assigned to the current resource intensity parameter level in the historical delivery records is used to fit a multi-classification model (based on the historical resource delivery records of users) The corresponding resource intensity is used as the target variable, with a covariate vector. Calculated using (as input features); For the sample i The covariate vector includes: users u Posting timestamp Previous contextual features (user contextual features from S11), business lines b Type encoding (from S11 business line basic information), resource type identifier k The type code (from the resource type dictionary of S11), the cumulative number of times the user has received resources under this service line (up to...). (Previous), the user's historical average resource allocation in this business (up to) forward); The result variable is not included. and any related Subsequent time-related future information is included to prevent tag leakage; As a stabilization factor, the marginal probability of this resource intensity parameter level is taken. Or a unified constant of 1; To prevent the denominator from being too small, a cutoff value of 0.05 is used in one embodiment; The maximum weight truncation threshold is set to 20 in one embodiment to limit the maximum contribution of a single sample to the gradient of the loss function and prevent gradient explosion during training.

[0100] After the weights are constructed, an effective sample size detection operation is performed. The effective sample size is calculated as follows: .

[0101] If the effective sample size is lower than a preset threshold, then a rescaling or scaling operation is performed on all weights. The cutoff range is set to reduce weight variance and ensure the stability of gradient updates during training. For Abnormal samples that are greater than the mean of all weights plus three times the standard deviation are removed from the training set and written to the abnormal log.

[0102] Through the above weight construction and anomaly detection operations, the excessive perturbation of model parameters by a few extreme weight samples is suppressed, while the training loss function is made to give higher gradient contributions to recent samples and debiased samples.

[0103] During model training and solving, the weighted maximum a posteriori log-likelihood is used as the training objective.

[0104] in, For the user, use a probability prediction model to predict the first... i Forward inference output value of each sample; The regularization coefficient is 0.01, 0.01, 0.01, and 0.1 respectively in one embodiment; Context feature weight vector The regularization coefficient, Resource type characteristic parameters The regularization coefficients are set to 0.01 and 0.01 respectively in one embodiment. These regularization coefficients are determined through cross-validation to achieve a balance between model fitting ability and parameter complexity, thereby suppressing overfitting.

[0105] The training objective is solved through iterative optimization: in each iteration, forward inference is performed on the current batch of samples to obtain predicted probability values. Based on the predicted probability values ​​and the true labels, a weighted cross-entropy loss and regularization term are calculated. The gradient vectors of each parameter are calculated using the backpropagation algorithm, and then the optimizer updates the parameter values ​​based on these gradient vectors. The optimizer employs either the Adaptive Moments Estimator (Adam) or the Limited-Memory BFGS Quasi-Newton Optimization Algorithm (L-BFGS). The former is suitable for large-scale sparse feature scenarios, while the latter converges faster on small to medium-scale data; the choice depends on the size of the training data. Simultaneously, variational approximation is performed to obtain the posterior distribution of each random effect parameter.

[0106] The iteration continues until the loss value converges or the preset maximum number of iterations is reached. After training, the converged model parameters are serialized and stored. During the online inference phase, the model parameters are loaded, and forward inference is performed on the input user-business line-resource type combination and uniform cost amount, outputting the corresponding usage probability value. .

[0107] S22: Calculate the elastic coefficient of the excitation response.

[0108] To address the technical issue that probability values ​​cannot directly reflect a user's sensitivity to changes in resource intensity, this step, after the user usage probability prediction model has been trained, uses finite difference numerical differentiation to batch calculate the incentive response elasticity coefficient for each user under various business lines and resource types. The calculation is performed in a vectorized manner, with a single traversal covering all user-business line-resource type combinations.

[0109] For users Business lines Resource types and recommendation strength Using the user's most recent user context feature vector c as a fixed input, the user's probability prediction model is invoked to process the input. Perform forward inference to obtain the probability value of use. Then, the elastic coefficients of the excitation response are approximated using finite difference:

[0110] in, In practice, the perturbation step size is taken. Perform the following finite difference operation:

[0111] The calculated excitation response elastic coefficient is truncated to a reasonable range; in one embodiment, it is truncated to... This is to eliminate abnormal extreme values ​​caused by insufficient samples or model extrapolation.

[0112] If a user-business-resource type combination results in an elasticity coefficient calculation exceeding the truncation range due to insufficient samples (the combination is marked as "rollback" in the sparsity detection in S14, or the effective sample size is less than the preset effective sample threshold when the elasticity coefficient is calculated in batches in S22), [5,5], then the elasticity coefficient of this combination is replaced with the median of the elasticity coefficient of the same business line or the same hierarchical group, and the fallback level is recorded in the parameter label. If the reference strength is... If the value is unknown, the unified cost amount is obtained by mapping the resource intensity parameter level with the highest historical reception frequency for that user through S12; if it is still missing, the median of the unified cost amount for that resource type across the entire platform is obtained.

[0113] S23: Perform cluster labeling operations and generate a dataset of user response elasticity parameters.

[0114] Based on the elasticity coefficient dataset output by S22, the elasticity coefficient value of each record is read one by one and compared with the preset high elasticity threshold. And the elastic threshold Perform a comparison operation and write the corresponding code into the cluster label field of the record based on the comparison result: Highly elastic users ( ), medium elastic users ( Low elasticity users Negative elasticity users ( ).

[0115] and The initial values ​​are taken from the 70th and 30th percentiles of the historical elasticity coefficient distribution, respectively, and obtained by sorting and quantile lookup operations on the elasticity coefficient array. The threshold is recalculated at a preset period (monthly in one embodiment) to adapt to the periodic changes in data distribution; during the online control experiment, the threshold can be iteratively updated based on experimental feedback.

[0116] After cluster labeling, the elasticity coefficients, cluster labels, and associated model parameters are assembled into structured records. Using a concatenated string of "user identifier + business line identifier + resource type identifier" as the index key, these records are written to a key-value store, forming the user response elasticity parameter dataset. Each record includes the following fields: stimulus response elasticity coefficient value, model posterior parameters (…). , , (posterior mean), using probability calibration parameters, cluster label encoding, backtracking level marking, and updating timestamps.

[0117] Simultaneously, a sorting index for the incentive response elasticity coefficients is generated, arranged in descending order of elasticity coefficient values, and stored. It should be noted that this sorting index is not the primary input for the global elastic quota request-level scoring calculation in S4, but is used in the following scenarios: candidate recall and sorting during minimum quota allocation; and triggering candidate request processing sequentially from high to low index priority when the global elastic quota balance is insufficient. The incentive response elasticity coefficients themselves do not directly participate in the objective function value calculation of S4; the objective function value calculation relies on the calibrated usage probability and resource adjustment parameters. This design avoids the estimation error of the elasticity coefficients being amplified step-by-step in online decision-making, improving the numerical stability of online decision-making.

[0118] S3: Forward-looking resource demand forecasting and pre-scheduling parameter generation.

[0119] S3 is used to proactively predict the resource needs of each business line within multiple future time windows. Its role is to provide the estimated demand and uncertainty range for each business line within each future time window for the two-tier quota allocation calculation in S4, enabling quota allocation to be performed based on data-driven prediction results, rather than relying on fixed proportions or manual settings. This step constructs and trains a resource write-off cost prediction model. This model, based on the time series of historical resource write-off cost amounts on a logarithmic scale and external event information, outputs point prediction values ​​and confidence intervals for the future resource write-off cost amounts of each business line through time series decomposition and component modeling calculations. S3 specifically includes the following sub-steps.

[0120] S31: Construct a resource write-off cost prediction model.

[0121] The system reads the historical resource write-off cost amount time series and external event calendar data of each business line stored in S11 from the time series database. Logarithmic transformation, STL decomposition, component modeling, and external event regression operations are then performed sequentially to construct a resource write-off cost prediction model. This model outputs the point prediction values ​​and distribution information of future resource write-off cost amounts for each business line, which are then used in the quota allocation calculation in S4.

[0122] Since the original resource write-off cost amount time series is usually right-skewed and contains extreme values ​​with large magnitude ranges, directly modeling at the original scale will cause the loss function to be dominated by a few extreme samples, affecting parameter convergence. To solve this technical problem, a logarithmic transformation operation is first performed on the historical resource write-off cost amount time series to obtain the logarithmic write-off cost value. :

[0123] in, For business lines At the point of time The resource write-off cost amount is read from the time-series database by business line identifier and timestamp index.

[0124] Then to The sequence is decomposed into a trend term sequence by performing an STL decomposition operation based on locally weighted regression. Seasonal item sequence and residual sequence Three components:

[0125] Seasonal cycle Based on business characteristics, the following settings are used: 24 for hourly data and 7 for daily data. The decomposition operation transforms the original non-stationary sequence into three stationary or quasi-stationary subsequences that can be modeled separately, reducing the complexity of a single model needing to simultaneously fit multi-scale patterns.

[0126] Then, prediction sub-models are established for each of the three components.

[0127] Trend Term Prediction Sub-model: The trend term sequence obtained from the decomposition Holt's linear trend model or locally weighted linear regression is used to extract slowly changing long-term growth or decay trend components. The fitting operation is performed by recursive state updates or sliding window weighted least squares solutions on the trend sequence.

[0128] Seasonal term prediction sub-model: the seasonal term sequence obtained from the decomposition Fourier series expansion is used for parameterization to extract periodic fluctuation components. Taking daily data with a 7-day period as an example:

[0129] in, The harmonic order, in one embodiment, is taken as 2 to 3, determined by performing a Fast Fourier Transform (FFT) on the historical seasonal term sequence and detecting the dominant frequency component, thus covering the main intra-weekly fluctuation patterns with a lower number of parameters; coefficients , The solution is obtained by performing a least-squares fitting operation on the seasonal term sequence.

[0130] Residual term prediction sub-model: the sequence of residual terms obtained from decomposition The modeling operation was performed using an autoregressive integral moving average (ARIMA) model to extract the short-term autocorrelation structure after removing trend and seasonal components.

[0131] in, The white noise error term; the order of the ARIMA model. By performing fitting operations on each candidate order combination and calculating the Akaike Information Criterion (AIC) value or the Bayesian Information Criterion (BIC) value, the order combination that minimizes the information criterion is selected, without the need to manually specify the order parameters.

[0132] To incorporate the impact of external events on resource write-off costs into the model, fields such as event type, start and end dates are read from the external event calendar data stored in S11 to construct an event impact function. And it is added to the model as a regression variable to obtain the complete model expression:

[0133] Among them, the event impact function The definition is as follows: If the event If it is a single-day event, then when For the date of the event Otherwise, it is 0; If the event For multi-day events, a trapezoidal function is used: the value linearly increases from 0 to 1 for the first 3 days of the event, remains at 1 during the event, and linearly decreases to 0 after the event ends.

[0134] Regression coefficient This is obtained by performing a least-squares regression operation on the residual data within the historical event window. This regression term enables the model to explicitly encode structured information from an external event calendar, improving the accuracy of predictions within the event window.

[0135] S32: Perform forward prediction and confidence interval estimation.

[0136] Using the resource write-off cost prediction model constructed in S31, a forward prediction operation is performed on the resource write-off cost amount at future points in time, and the prediction confidence interval is estimated, providing an uncertainty measurement input for the quota allocation operation in S4.

[0137] Each component at time point The predicted values ​​are summed to obtain the predicted logarithmic write-off cost value:

[0138] Perform an inverse exponential transform on the logarithmic scale prediction value to obtain the original scale prediction value:

[0139] Confidence intervals are constructed on a uniform logarithmic scale. This is based on historical prediction error sequences. Perform sorting operations on the error sequence and extract the 5th percentile. and the 95th percentile Then the lower and upper bounds for prediction at the original scale are respectively:

[0140]

[0141] The interval provides the prediction range at a 90% confidence level, providing a quantitative basis for the allocation calculation of the minimum and flexible quotas in S4. When more accurate probability prediction is required, the logarithmic scaling model can be replaced by a probability prediction model that directly outputs quantiles (such as the DeepAR model or the Prophet model). After the replacement, the construction method of the confidence interval is adjusted accordingly to the quantile values ​​output by the model.

[0142] S33: Generate the pre-scheduled parameter configuration table and the time-sharing window prediction data table.

[0143] For each business line b and future time window h Based on the lower and upper bounds of the prediction output by S32, the following parameter generation calculation is performed: Assign the predicted lower limit value as the minimum quota parameter: ; Calculate the difference between the upper and lower forecast limits to obtain the elastic quota parameters: ; Record prediction point estimation And generate timestamps.

[0144] The above parameters are written into the pre-scheduling parameter configuration table in the form of structured records, so that S4 can directly read and use them for quota allocation calculation.

[0145] Since the online decision-making in S4 and the inner-loop parameter update in S5 require hourly-level forecasts, and daily forecasts are insufficient to support fine-grained quota control, this step further performs a time-sharing window splitting operation. The splitting operation is obtained by multiplying the daily forecast point estimate, lower limit, and upper limit using historical time-sharing proportion coefficients. Set up business lines b The estimated daily forecast point is The historical time-sharing window ratio is (Obtained through statistical calculations from historical data of the same period), then: The estimated time-sharing window point is: ; The lower bound of time-sharing window prediction is ; The upper limit of the time-sharing window prediction is .

[0146] After the data split, a consistency check is performed: the lower limit of each time-sharing window is summed, and the summation result is compared with the daily lower limit; the same summation and difference calculation is performed on the upper limit of each time-sharing window. If the absolute value of the difference between the summation result of any time-sharing interval and the corresponding daily value exceeds 1% of the daily value, a proportional correction calculation is performed on each time-sharing value to align the time-sharing interval with the daily interval. This check ensures the numerical consistency between the time-sharing forecast data and the daily forecast data, avoiding accumulated errors when S4 reads the time-sharing data to perform quota allocation.

[0147] S3 ultimately outputs two structured data tables: The pre-scheduling parameter configuration table includes fields such as: business line identifier, time window identifier, minimum quota parameter value, elastic quota parameter value, prediction point estimate, confidence interval lower limit, confidence interval upper limit, and generation timestamp. The time-sharing window forecast data table includes fields such as: business line identifier, window start timestamp, window end timestamp, forecast point estimate, forecast lower limit, and forecast upper limit.

[0148] The two data tables are stored in a structured record format for use in the quota allocation calculation of S4 and the consumption rate correction calculation of S5.

[0149] S4: Online request processing and two-tiered resource quota collaborative allocation.

[0150] S4 is used to perform low-latency, personalized resource allocation decision calculations under the constraints of total resource quota and multi-service line quota when a user's real-time request arrives. Its function is to transform the usage probability parameters output by S2, the pre-scheduling parameter configuration table output by S3, and the real-time request context into specific resource allocation instructions. Simultaneously, it ensures the numerical consistency of quota data at all levels through quota deduction and status update operations. This step maintains a quota status data table in memory and provides decision result records and quota change records for the feedback loop of S5. S4 specifically includes the following sub-steps.

[0151] S41: Calculate the marginal gain coefficient after write-off.

[0152] Using the user historical resource delivery record table and user behavior log data stored in S1, causal increment estimation calculations are performed on the randomized controlled experiment data to obtain the post-recovery marginal gain coefficient required for online decision-making. This coefficient represents the business line b Next resource type k The average incremental revenue estimate corresponding to each write-off event is used as the revenue weight parameter in the S44 objective function calculation.

[0153] Specifically, from the randomized controlled trial grouping table stored in S11, using the experiment identifier and user identifier as index keys, users in the treatment group and control group are retrieved; then, the user's historical resource allocation record table stored in S11 is linked to obtain their resource allocation and cancellation event records within the attribution window. The treatment group refers to the group assigned the resource type in the randomized controlled trial. k The user set; the control group consists of users who have not been allocated resources. k Furthermore, the set of users whose user context feature vector distance is within a preset similarity threshold is identified. Transaction amount data for these users before and after the attribution window is retrieved from the user behavior log table stored in S11.

[0154] For each user The incremental revenue value is obtained by performing a difference calculation based on the transaction amount data before and after the attribution window:

[0155] in, .

[0156] Based on the data from the randomized controlled trial, the incremental benefit values ​​of the treatment group and the control group were respectively calculated by mean operation, and then the difference between the two group means was calculated to obtain the average treatment effect estimate:

[0157] in, To handle the number of users in the group, This represents the number of users in the control group. When the sample size of a randomized controlled experiment is insufficient, the average treatment effect estimate can be calculated based on the observed data using the stabilization inverse probability weighting operation described in S21, which will not be elaborated here.

[0158] Based on the user historical resource delivery record table stored in S11, a counting operation is performed on the cancellation event records of users in the attribution window of the processing group, and then divided by the number of users in the processing group to obtain the average number of cancellation events per user:

[0159] in, For users i The number of times this resource type was actually written off due to this resource allocation, as shown in the attribution window. If so, then assign it the value 0.01.

[0160] The marginal gain coefficient after write-off is calculated based on the average treatment effect value and the average number of write-off events per person. The calculation method is as follows: .

[0161] when At that time, The upper limit is truncated to 3 times the global average for this resource type or 3 times the average across all platforms (whichever is smaller), and written to the exception flag field. If If the value is negative or exceeds the preset reasonable range, it will be truncated to 0 or the historical average. The initial value is set in the range of 2.0 to 10.0.

[0162] S42: Calculate the available quota for the day and perform two-level quota splitting and status initialization.

[0163] This step is re-executed when the daily scheduled task is triggered or at a preset cycle. (Refer to...) Figure 2 Its function is to read the pre-scheduling parameter configuration table and the quota status data of the previous day output by S3, calculate the total available quota for the day, divide it into a minimum quota layer and a global elastic quota layer, allocate elastic quotas to each business line, initialize the quota status data table for the day, and provide quota constraint boundaries for subsequent request-level decision operations.

[0164] Specifically, the first parameter is read from the pre-scheduling parameter configuration table output by S3. t Daily platform forecasts estimated write-off cost amount. Combined with the total budget for the activity period read from the activity budget configuration data stored in S11 Perform the following operations to obtain the first... t Daily basic quota:

[0165] Read the confirmed consumption amount from the previous day from the quota status data table. The total available quota for the previous day With confirmed write-off consumption The difference is used to calculate the net balance of the previous day:

[0166] If it is the first day of the activity or the quota status data table does not contain records from the previous day, then The value is set to 0. The value is also set to 0; Then, the total available quota for the day is calculated from the basic quota and the previous day's net surplus:

[0167] like If the daily quota availability flag is set to "paused," including the minimum quota layer and the global elastic quota layer, an alarm record will be written to the exception log. Quota allocation calculations are only allowed to continue within a preset percentage (5% in one embodiment) of the remaining total quota for the activity period if there is an explicitly configured over-quota authorization. Before performing the quota splitting calculation, the quota allocation is first... Perform a conditional check; if the condition is not met, skip the splitting operation.

[0168] Subsequently, the minimum allowance parameter values ​​for each business line within the current rolling 24-hour window are read from the pre-scheduling parameter configuration table output by S3. And the elastic quota parameter value, as well as the business line resource quota limit in the business line basic information obtained from S11. A two-stage quota projection calculation is performed to divide the total available quota for the day into a minimum quota layer and a global flexible quota layer.

[0169] The two-stage quota projection calculation includes: (1) For each business line Set the initial guaranteed minimum quota value to .like Then it is truncated to ,otherwise .

[0170] (2) Perform a summation operation on the minimum guaranteed quota values ​​after truncation for each business line to obtain the total temporary minimum guaranteed quota. .like Then, a proportional scaling operation is performed on the minimum allowance value for each business line: If compression still exists In the case of scaling, iterative scaling operations are performed until all business lines meet the requirements. .

[0171] Sum the scaled minimum allowance values ​​for each business line to obtain the total minimum allowance tier:

[0172] Perform a difference calculation between the total available quota for the day and the total guaranteed quota to obtain the original allocable amount for the global elastic quota layer:

[0173] The elastic quota parameter values ​​for each business line are truncated and summed under the constraint of the difference between the upper limit and the minimum guarantee quota value of the business line resource quota, to obtain the theoretical total elastic quota:

[0174] The minimum value is calculated by subtracting the original allocatable amount from the theoretical total amount of the global elastic quota layer to obtain the total amount of the global elastic quota layer.

[0175] like Then, perform a difference calculation on the two to obtain the reserve quota balance. Write the balance to the standby quota field and set its status flag to "frozen".

[0176] The conditions for unfreezing and activating the reserve quota are: the global actual net consumption rate is lower than the rate threshold corresponding to the current time-sharing window prediction lower limit; or the S5 inner loop parameter update calculation detects a positive global quota balance for three consecutive time windows. When activated, the difference between the elastic quota parameter value and the actual consumption of each business line is normalized, and the reserve quota balance is allocated to the global elastic quota layer according to the normalized ratio, and the elastic quota field of each business line is updated synchronously.

[0177] Based on the elastic quota parameter values, resource quota upper limits, and scaled minimum quota values ​​for each business line, a truncation calculation is performed to obtain the elastic quota for each business line:

[0178] After completing the above segmentation, write the following fields into the daily quota status data table to complete the initialization: Total remaining quota. Remaining quotas for each business line Remaining amount of the minimum quota layer Global elastic quota layer remaining amount Reserve quota balance Minimum quota balance for each business line Flexible quotas for each business line Each business line's flexible quota has been pre-allocated. (Initially 0), confirmed write-off consumption of elastic quotas for each business line. (Initial value: 0), the minimum quota tier for each business line has been pre-allocated. (Initial value: 0) Allocated but unused minimum allowance tiers for each business line (Initially 0), Total confirmed write-off consumption across all business lines' full quota tiers (Initially 0).

[0179] The quota status data table retains the following aggregated fields according to the time window dimension: the amount of write-offs and deductions confirmed by each business line in the current window. The current window releases pre-allocated quotas for each business line. The current window shows all confirmed write-off and deduction amounts. Release the pre-occupied space globally in the current window The aforementioned window aggregation field is accumulated and written by the quota deduction and status update operation of S46 at the end of each time window, and is then read by S522 and S523.

[0180] S43: Read the real-time request context and decision input data.

[0181] When the server receives a real-time user request (including login requests, page access requests, transaction requests, etc.) through the network communication interface, it reads the following decision input data from the corresponding data storage: Real-time user request context, including user ID, current behavior type, current business line ID, and request time; User response elastic profile, including usage probability values ​​output by S2 ; The current quota balance and consumed amount at each level in the quota status data table; The set of candidate resource combinations, i.e., the types of resources currently available for allocation and their resource strength parameters. Combination list; Global resource adjustment parameters And business line resource adjustment parameters .

[0182] The candidate resource combination set is generated as follows: All resource types and their denominations associated with the current business line are read from the resource type dictionary stored in S11. Types with "Whether to consume quota identifier" set to "No" are filtered out (this type of resource does not participate in quota allocation decisions). Resource types that are currently suspended from allocation are then filtered out. The remaining resource types are then enumerated and combined with their corresponding denominations to form the set. The set; this set is preloaded in memory as configuration data and is refreshed synchronously according to the resource type dictionary update cycle.

[0183] S44: Perform global elastic quota layer request-level decision calculation.

[0184] When a user request arrives and the global elastic quota layer is "available", the objective function value is calculated for each candidate resource combination based on the decision input data read by S43, and the resource combination that maximizes the objective function value is selected.

[0185] refer to Figure 3 Specifically, based on the user identifier, business line identifier, and request timestamp in the real-time user request context, as well as the user's historical allocation records in the quota status data table, the following user-level frequency constraint and mutual exclusion constraint detection operations are performed sequentially: (1) Query the most recent release timestamp of the resource type for the user under the business line. If the difference between the timestamp and the current request timestamp is less than the preset cooling-off period threshold (7 days in one embodiment), then mark it as a frequency constraint trigger. (2) If the current request is associated with a resource provisioning record, then mark it as a single request constraint trigger; (3) Query the records of the same type of resources that have not expired and held by the user. If there are valid records, mark them as triggering the duplicate issuance constraint. (4) Count the number of times the user has issued resources across resource types in the past 7 days in this business line. If the number exceeds the preset cumulative frequency threshold of this business line, it is marked as a cumulative frequency constraint trigger. (5) Count the number of times the user has been issued in all business lines in the past 7 days. If the number exceeds the preset global frequency threshold, it is marked as a global frequency constraint trigger. (6) Query the user’s unexpired and unused resource records in this business line. If there are valid records and the current candidate resource type is not in the preset superimposed license list, then mark it as a mutual exclusion constraint trigger. (7) Check if the user ID exists in the preset filter list, or if the current business line ID is not in the permission list. If so, mark it as a list constraint trigger.

[0186] If any of the above constraint detection operations returns a trigger flag, then the decision result is set to an empty combination. Alternatively, it may be transferred to the minimum quota layer for processing according to the preset routing configuration; when transferring to the minimum quota layer, the above constraint detection calculation needs to be re-executed. Empty combination This indicates that no resource allocation instructions will be generated.

[0187] For requests that do not trigger constraints, iterate through the candidate combination set. For each candidate combination, the model input cost calculation, expected consumption calculation, and safety pre-occupancy calculation are performed sequentially.

[0188] Model input cost The calculation method is as follows: When the resource type is electronic voucher, activity points, digital gift pack, or limited-time benefit: , This represents the uniform cost amount for the resource mapped from S12; When the resource type is a discount amount: ,in For users In business lines Average order amount over the past 90 days (read from user context feature data; if missing, it will be the average amount of orders cleared by the business line over the past 90 days). ), d The percentage of discount offered.

[0189] Expected consumption The calculation method is as follows:

[0190] in, For users who call S21 to use the probabilistic prediction model, The input is the usage probability value obtained by performing forward inference. If S532 has subsequently generated probability update mapping function parameters for the corresponding user segment × business line × resource type or higher priority, then the original predicted probability is first converted into a calibrated probability according to the mapping rules of S532. Subsequent calculations of expected consumption, safety reserve, and objective function value all use [the following method / method]. .

[0191] Safety Reserved Quantity The calculation method is as follows:

[0192] in, This represents the average order amount for the user over the past 90 days. This represents the 90th percentile of the order amount reimbursed over the past 90 days for the business line.

[0193] The discount quota type calculates the model input cost and the safety pre-allocation separately. The former is used for objective function value calculation to ensure prediction accuracy, while the latter is used for quota verification calculation to control the risk of quota over-issuance.

[0194] Subsequently, based on the usage probability value, the marginal gain coefficient after write-off, the global resource adjustment parameter, the business line resource adjustment parameter, and the expected consumption, the objective function value is calculated for each candidate combination:

[0195] For candidate combination set Compare the objective function values ​​of all combinations and select the combination that maximizes the objective function value. ,Right now .

[0196] If all objective function values ​​in the candidate combination set are not greater than 0, then the decision result is set to... .

[0197] When multiple candidate combinations have equal and positive objective function values, the selection operation is performed according to the following priority: first, the safe pre-occupancy is selected. Smaller combinations; if the safety reserve is the same, select the combination with the higher efficiency ratio. The efficiency ratio is calculated as follows:

[0198] This selection operation further optimizes resource quota utilization efficiency and reduces quota occupation risk under the condition that the objective function value is the same.

[0199] when At this time, a quota verification calculation is performed to prevent quota over-issuance. The remaining amount of the global elastic quota layer, the elastic quota of each business line, the pre-allocated amount of the elastic quota layer of each business line, the confirmed write-off consumption, the allocated but unused amount of the minimum quota layer, and the pre-allocated amount of each layer are read from the quota status data table. The following calculations are then performed to obtain the available elastic quota for each business line:

[0200] in, For business lines b Flexible quota availability Flexible quotas for business lines. For business lines b The flexible quota tier has already been pre-allocated. For business lines b The flexible quota tier has confirmed the write-off consumption. This represents the upper limit of resource allocation for each business line. For business lines b The guaranteed quota tier has been allocated but not yet used. For business lines b The minimum quota tier has already been reserved. For business lines b Total confirmed write-off consumption across all quota tiers (net value after deducting refund reversals).

[0201] In the above formula, the first term represents the remaining amount within the business line's elastic quota layer; the second term represents the remaining amount within the business line's total quota limit. The smaller of the two terms is taken to ensure that resource allocation instructions neither exceed the elastic quota layer constraints nor the business line's total quota limit constraints.

[0202] Safety pre-occupancy Remaining amount of global elastic quota layer and the available amount of flexible quotas for business lines Perform comparison operations separately: If both conditions are met and Then generate combinations The corresponding resource allocation instructions.

[0203] If the above conditions are not met, a simple descending traversal is not performed on the safe pre-occupancy dimension. Instead, a set of remaining affordable candidates is constructed and the objective function value maximization selection operation is re-executed.

[0204] exist Select the combination that maximizes the objective function value and is greater than 0. If If the objective function is empty or all combinations thereof have objective function values ​​not greater than 0, then the decision result is set to... The application will be transferred to the minimum quota layer for processing or the issuance instruction will be refused.

[0205] This reselection operation automatically searches for feasible suboptimal solutions when quotas are insufficient, avoiding inefficient allocations that may result from traversing in descending order along a single dimension.

[0206] S45: When the decision result of the global elastic quota layer is not to issue quotas or the global elastic quota layer is unavailable, perform the minimum quota layer allocation operation.

[0207] Specifically, the guaranteed quota layer applies to user requests that reach the business line but for which no issuance instruction has been generated at the global elastic quota layer. The allocated resource strength parameter is the lowest resource strength parameter in the current business line's candidate resource combination set. .

[0208] The same request is first processed by the global elastic quota layer; if the decision result is... And the remaining amount of the guaranteed quota layer is greater than The corresponding safety reserve is generated based on the minimum guarantee quota layer. The corresponding resource allocation instructions.

[0209] If the remaining quota at the minimum guarantee level is insufficient, the issuance instruction will be rejected. Alternatively, if the "Quota Consumption Flag" for the resource in the resource type dictionary is "No", the resource type will be downgraded to a no-quota-consumption type for issuance (this issuance will not result in quota deduction or pre-allocation). The minimum guarantee quota level and the global elastic quota level will not generate issuance instructions repeatedly for the same request. For the same user and the same business line, a single request will generate an issuance instruction for at most one resource type and one resource strength parameter. Before generating an issuance instruction, the same user-level frequency constraint and mutual exclusion constraint detection calculations as S44 must be re-executed. If any constraint is triggered, the issuance instruction will be rejected.

[0210] At the end of each time window, the remaining amount of the guaranteed quota layer, the confirmed consumption amount, and the effective pre-occupancy amount are read from the quota status data table. The difference is calculated to obtain the unconsumed guaranteed quota amount, which is then added to the remaining amount field of the guaranteed quota layer for the next time window.

[0211] When the cumulative consumption (including pre-allocated amount) of a certain business line reaches When this happens, the available status flag of all quota layers in the business line is set to "paused". The pre-allocated quota is released or cancelled after the quota deduction and status update operation described in S46.

[0212] S46: Perform budget deductions and status updates.

[0213] Combination of generating and issuing instructions Safety pre-occupancy calculated according to S44 In the quota status data table, increase the pre-allocated amount field of the corresponding quota layer by this value and decrease the available amount field by this value. If the resource type is a resource that is issued and deducted immediately, then the pre-allocated amount is the actual consumption amount.

[0214] Each time a quota deduction or release operation is executed, the change is simultaneously added to the corresponding field of the current time window in the quota status data table.

[0215] When a subsequent feedback event arrives, read the fields from the feedback event, such as the claim flag, usage flag, usage timestamp, actual discount amount, and transaction amount, and update the quota status data table according to the following quota deduction and status update process: When the usage flag of the feedback event is "used", a confirmation deduction operation is performed based on the actual consumption amount in the feedback event: if the actual consumption amount is greater than the pre-allocated amount, a supplementary deduction difference operation is performed on the quota remaining amount field; if the actual consumption amount is less than the pre-allocated amount, a release difference operation is performed on the pre-allocated amount field, and the released amount is added back to the quota remaining amount field.

[0216] When resources expire, the claim flag is "not claimed", or the use flag is "not used", a release operation is performed on the pre-occupied quantity field, and the released quantity is added back to the remaining quota field.

[0217] If a refund or transaction cancellation occurs after the discount quota resource is redeemed, the refund amount and discount percentage in the feedback event are multiplied to obtain the reversal amount. The reversal amount is then added back to the remaining amount field of the quota layer (global elastic quota layer or guaranteed quota layer) that was actually deducted when the resource was redeemed. At the same time, the status flag of the corresponding resource record is set to "invalidated" or "reclaimed", and the corresponding business line quota is released.

[0218] When performing the supplementary deduction difference calculation, if the remaining quota of the business line or the remaining total quota is insufficient to cover the supplementary deduction amount, a security control operation is performed: the supplementary deduction amount is compared with a preset excess threshold, where the excess threshold is min(2% of the total quota and 5% of the upper limit of the business line resource quota). If the supplementary deduction amount does not exceed the excess threshold, temporary excess deduction is allowed, and the status field of the record is marked as "temporary excess" in the quota status data table, to be reversed during the end-of-day batch verification operation. If the supplementary deduction amount exceeds the excess threshold, or if at least one "temporary excess" record already exists on the same day and supplementary deduction is triggered again, a security circuit breaker operation is performed: the available status flag of the global elastic quota layer of the business line is set to "paused," the quota of the business line is frozen, only the issuance permission of the lowest tier of the guaranteed quota layer is retained, the background difference correction task is started, and an alarm record is written to the exception log.

[0219] The end-of-day batch verification task recalculates the quota changes for all transactions offline, performing correction operations on the differences arising from "temporary over-quota" records and the circuit breaker operation. The reversal operation adjusts the quota ledger records for the actual write-off or refund date, not the next day's ledger records. The circuit breaker is lifted when the end-of-day batch verification operation confirms that the quota data values ​​are consistent and there are no incomplete reversal records. If the background difference correction task fails, it automatically retryes and writes an alarm record to the exception log. The result of the end-of-day batch verification operation will be used as input for calculating the previous day's net balance in S42 on the next day.

[0220] S47: Output decision results and quota status snapshot.

[0221] Combining resources The corresponding resource allocation instruction is written to the allocation queue, and the downstream service executes the operation of allocating resources to the user account.

[0222] At the same time, the following data should be written to the decision result log and quota status data table: Real-time resource allocation decision records include fields such as: user identifier, business line identifier, resource type identifier, resource intensity parameter, timestamp, security pre-allocation quantity, validity period expiration timestamp, quota deduction method, and quota tier identifier (minimum / flexible).

[0223] Quota status snapshot record, fields include: total remaining quota, remaining quota for each business line, remaining quota for the minimum guarantee layer, remaining quota for the global elastic quota layer, reserve quota balance, elastic quota and remaining quota for each business line, and pre-allocated quota.

[0224] S5: Resource allocation feedback and dual-loop adaptive parameter iteration.

[0225] S5 is used to establish a closed-loop feedback mechanism for real-time effects after resource allocation. (Refer to...) Figure 4Its function is to automatically update resource adjustment parameters and model parameters through inner-loop consumption rate correction calculation and outer-loop model calibration calculation, so that the constraint parameters in online decision-making calculation continuously track changes in actual data distribution and suppress the accumulation of prediction bias caused by data drift. S5 specifically includes the following steps.

[0226] S51: Read feedback event streams and decision-related data.

[0227] The resource allocation records output by S47 are read from the decision result log and loaded into memory in the form of a feedback event stream. Each record in the feedback event stream includes the following fields: user identifier, business line identifier, resource type identifier, resource intensity parameter, allocation timestamp, validity period expiration timestamp, claim flag, usage flag, usage timestamp, transaction amount, actual discount amount, actual redemption consumption, quota deduction status flag, quota deduction method, and quota layer identifier. Read the number of confirmed write-off deductions, released pre-occupied quotas, and quota status change records for each business line within the current time window from the quota status data table. Read the estimated prediction point value of the current time window for each business line from the time-sharing window prediction data table output by S3; Read the predicted usage probability value of the corresponding user from the user response elasticity parameter dataset output by S2.

[0228] S52: Inner loop consumption rate correction calculation. This includes the following sub-steps: S521: Calculate the observed write-off rate and generate a write-off rate deviation reference signal.

[0229] Because resource write-offs involve a time delay (an uncertain interval exists between allocation and actual use), directly calculating usage flags at the end of the observation window would systematically underestimate the true write-off rate due to some unexpired samples not yet being written off. To address this technical issue, this step involves each business line... Within the observation window, samples are stratified according to their delivery timestamps, and a weighted observation cancellation rate is calculated using a delay feedback correction operation.

[0230] Specifically, from the feedback event stream loaded by S51, the service lines within the current observation window are filtered using the service line identifier as the index key. b All issuance records are divided into three subsets based on the usage flag and the expiration time stamp: the used sample set. (Use the "used" flag) Expired but unused sample set (A set of samples whose expiration timestamp is earlier than the current time and whose usage flag is "unused") and samples that have not yet expired. (The expiry date timestamp is later than the current time).

[0231] For each unexpired sample Based on the projected final write-off rate Estimate its remaining validity period The conditional probability of completing the write-off within the specified time.

[0232] The projected final write-off rate The calculation is performed as follows: Based on the revocation timestamp sequence of historical samples from the same business line and resource type, a Kaplan-Meier survival function is constructed, and the remaining validity period is queried. The conditional write-off probability value; or calling a pre-trained quantile regression model to determine the remaining validity period. Resource type identifier and business line identifier are used as input features to perform forward inference operations and output the estimated write-off rate value. When the number of new resource types, new business lines, or historical sample records is less than the preset threshold, the following priority is applied to perform a fallback: historical average of all resource types in the same business line → historical average of the same resource type across the entire platform → default value 0.3. The estimated final write-off rate parameter is updated daily via offline batch tasks, and the update results are written to the algorithm parameter configuration table as input data for this step.

[0233] Based on the partitioning results of the above three subsets and the expected final write-off rate parameters, the following weighted summation and normalization operations are performed to obtain the weighted observed write-off rate:

[0234] in, For business lines within the window b The total number of issuance records. To suppress interference from single-window random fluctuations, an exponential moving average smoothing calculation is performed on the weighted observation write-off rate:

[0235] Based on the smoothed weighted observation write-off rate The average predicted usage probability within the window of this service line output by S2. Calculate the reference signal for the write-off rate deviation:

[0236] Average predicted usage probability within this business line window Based on the user usage probability prediction model trained by S21, the current window's business lines... b The results are obtained by performing forward inference operations on each of the distributed samples and then taking the arithmetic mean of the predicted probability values.

[0237] It should be noted that the write-off rate deviation reference signal is only used as a reference input for the S53 outer loop model calibration calculation and does not participate in the update calculation of resource adjustment parameters in S522 and S523. If the proportion of expired samples in the current window is less than 30% or the absolute number of records is less than 50, the write-off rate deviation reference signal field of this window will be set to null. If the number of records is insufficient for three consecutive windows, the probability reference update calculation based on the write-off rate deviation will be suspended, but the consumption rate correction calculation will not be affected.

[0238] S522: Calculate the net consumption rate deviation of the business line and update the business line resource adjustment parameters.

[0239] Based on the quota status data table and the predicted data output by S3, the net consumption rate deviation of the business line is calculated, and the resource adjustment parameters of the business line are updated accordingly. .

[0240] Specifically, from the quota status data table maintained by S46, the business lines within the observation window Δ are read using the business line identifier and time window identifier as index keys. b Confirmed write-off deductions and release pre-occupied amount The actual net consumption rate is obtained by performing the following calculations:

[0241] in, For business lines b The actual net consumption rate, Δ is the observation window duration.

[0242] Read the business lines from the time-sharing window forecast data table output by S3. b Current time-sharing window prediction point estimate The predicted consumption rate is obtained by performing the following calculations:

[0243] in, To predict the consumption rate.

[0244] The following difference calculation is performed between the actual net consumption rate and the predicted consumption rate, and then divided by the predicted consumption rate to obtain the relative deviation of the business line's net consumption rate:

[0245] If the consumption rate is predicted If it is less than a preset minimum threshold (0.01 in one embodiment), then... If the value is set to 0, the correction operation will be skipped in this window. Simultaneously, the business line resource adjustment parameters... The update trigger condition is decoupled from the write-off rate confidence judgment, requiring only that the number of quota status change records in the current window is not less than the minimum record count threshold (50 records in one embodiment); if the number of valid records is less than 50, the previous window's record is maintained. The value remains unchanged.

[0246] The resource adjustment parameters for business lines adopt a superposition structure of "baseline value + correction term":

[0247] in, The baseline value for adjusting the resource parameters of the business line is set to 0.1 in one embodiment; This is the correction term, and its initial value is set to 0.

[0248] right With inner loop trigger threshold (In one embodiment, the value is 0.2) Perform a comparison operation: when If the number of valid records meets the minimum threshold, then a correction term update operation is performed based on the relative deviation of the net consumption rate:

[0249]

[0250] in, This is the inner loop correction gain coefficient, which is set to 0.3 in one embodiment; The exponential smoothing coefficient is set to 0.7 in one embodiment to suppress the interference of single-window random fluctuations on the correction term.

[0251] The new resource adjustment parameter value is obtained by summing the baseline value and the updated correction term: And perform a truncation operation on the result, restricting it to... Within the range. If truncated... Exceeding If the range is specified, it is written back as the truncation boundary value. After the update... Write the corresponding field to the quota status data table.

[0252] This correction operation enables the resource adjustment parameters of the business line to be adaptively adjusted according to the actual consumption deviation: when the actual consumption rate is higher than the predicted value, the parameter is increased. To tighten quota allocation for this business line, reducing it when the actual consumption rate is lower than the forecast. This will allow for a more relaxed allocation of quotas, thereby balancing the rate at which quotas are consumed across different business lines.

[0253] S523: Calculate the global net consumption rate deviation and update the global resource adjustment parameters.

[0254] Based on global quota status data and predicted data output by S3, the global net consumption rate deviation is calculated, and the global resource adjustment parameters are updated accordingly. .

[0255] At the start of the observation window, the initial assignment of global resource adjustment parameters is performed first. The actual net consumption rate and predicted consumption rate for the most recent time window are read from the quota status data table, and a division operation is performed to obtain the rate ratio. ,according to The comparison result with the preset threshold sets the initial global resource adjustment parameter value: like ,but ; like ,but ; like ,but .

[0256] When a new activity begins, there are no identical historical windows, or the historical rate is 0, The value is assigned to 0.5; if a preset initial value exists in the algorithm parameter configuration table, then that value is read from the algorithm parameter configuration table as the initial value. . The minimum number of records required for calculation is the quota records from at least the three most recent windows with the same record. If the number of records is insufficient, the default value of 0.5 will be used.

[0257] From the quota status data table maintained by S46, read the globally confirmed write-off deduction amount within the observation window Δ. and global release pre-allocation The following calculation yields the global actual net consumption rate:

[0258] in, This represents the global actual net consumption rate. This is the duration of the observation window.

[0259] Read the estimated value of the current time-sharing window prediction point for each business line from the time-sharing window prediction data table output by S3, perform a summation operation, and divide by the window duration to obtain the global prediction consumption rate:

[0260] in, This represents the global predicted consumption rate value.

[0261] The difference between the actual global net consumption rate and the predicted global consumption rate is calculated and divided by the predicted value to obtain the relative deviation of the global net consumption rate:

[0262] Perform an exponential moving average smoothing operation on the relative deviation of the global net consumption rate:

[0263] in, This represents the relative deviation of the global net consumption rate after smoothing in the previous window. This is the relative deviation of the global net consumption rate calculated for the current window. This is the relative deviation of the global net consumption rate after smoothing the current window.

[0264] Based on the smoothed relative deviation of the global net consumption rate, the following multiplicative update operation is performed to obtain the new global resource adjustment parameter values:

[0265] in, This represents the inner loop correction gain coefficient. This is the global resource adjustment parameter value from the previous window. Adjusts global resource parameters for the current window after the update.

[0266] After the update operation is completed, perform the following protective check operation: Multiplicative factors Perform a range check and truncate it to the range [0.2, 2.5]. right Perform a truncation operation, limiting it to the range [0.05, 2.0]. like If the value exceeds a reasonable range (e.g., less than -1.5 or greater than 2.0) or the actual net consumption rate is negative and there is no corresponding release pre-reservation record, then skip this operation. Update, keep The error remains unchanged, and an alarm record is written to the exception log.

[0267] Updated and The data is written to the global parameter area of ​​the quota status data table as online decision input for subsequent S4 requests. Simultaneously, the write-off rate deviation reference signal is included. As input data for the calibration calculation of the S53 outer loop model.

[0268] S53: Outer loop model calibration calculation. This includes the following sub-steps: The outer loop model calibration operation is performed offline in batches at a preset period (daily or weekly in one embodiment). Its purpose is to correct the systematic biases of the probability prediction model and the resource write-off cost prediction model, so that the model output continuously tracks the actual data distribution.

[0269] S531: Calculate the deviation between the actual and predicted utilization rates of the calibration unit.

[0270] Calibration units are grouped by user Business lines With resource type The three-dimensional combination definition, namely For each calibration unit, the actual utilization rate is calculated from the feedback event stream. The mean of the model's predicted utilization rate is calculated from the predicted probability data output by S2. Perform the following deviation calculation:

[0271] Among them, user groups It can be determined by the cluster label code output by S23; The calculation method is based on the user usage probability prediction model trained by S21. Forward inference is performed on all samples in the current calibration unit one by one, and the arithmetic mean of the obtained predicted probability values ​​is calculated. To ensure statistical reliability, the number of records in each calibration unit is not less than 100; if it is less than 100, adjacent cluster labels are merged upwards to the available level that meets the minimum number of records.

[0272] actual utilization rate of the outer ring The calculation uses the same delayed feedback correction operation as S521 (i.e., for each unexpired sample, use...). Calculate the expected write-off probability, and obtain the result after sample normalization. Alternatively, calculations can be performed using only expired samples. The default method is delayed feedback correction.

[0273] Simultaneously, a minimum overdue sample ratio check is performed: if the proportion of unexpired samples in the current calibration unit exceeds 60% or the number of expired sample records is less than 100, the calibration operation for this calibration unit in this cycle is skipped, and it is marked as "delayed to the next cycle". This minimum sample requirement ensures the statistical reliability of the deviation calculation.

[0274] S532: Trigger probability calibration operation to update probability update mapping function parameters.

[0275] Deviation value for each calibration unit With outer ring trigger threshold (In one embodiment, Perform comparison operations: like If this occurs, a probabilistic calibration operation for that calibration unit will be triggered.

[0276] The probability calibration operation is performed on the calibration dataset corresponding to the calibration unit: the model output probability is used as the basis for the calibration operation. As input, fit an isotonic regression or Platt scaling mapping function:

[0277] in, and The parameters of the probability-updating mapping function are obtained by performing maximum likelihood estimation on the calibration dataset. The calibrated probability values ​​are then calculated. This updated usage probability is used by S4 during online inference. During online request processing, S4 calculates... Afterwards, before performing the objective function value calculation, the corresponding probability update mapping function is called first. Perform the mapping transformation. Simultaneously, re-perform the elasticity coefficient calculation in S22 based on the calibrated probability values, and update the cluster labels in S23. This calibration operation eliminates systematic biases in the model predictions through nonparametric (ordination-preserving regression) or parametric (Platform scaling) mapping.

[0278] When processing online requests, the probability update mapping function parameter records are queried in the following order of priority: user segmentation. ×Business Line ×Resource Type Business lines ×Resource Type Resource type User segmentation The global default probability update mapping function parameters are used. If a high-priority probability update mapping function parameter record does not exist, the process falls back to a low-priority one; if the global default probability update mapping function parameter also does not exist, the calibration operation is skipped, and the original model probability values ​​are used directly. This priority fallback mechanism ensures that valid calibration results are output under different data coverage conditions.

[0279] S533: Triggers model retraining and updates parameter dataset and prediction model.

[0280] When the number of calibration units triggering the calibration operation exceeds a preset threshold, or the global mean deviation exceeds a preset threshold, a retraining operation of the user's probabilistic prediction model is triggered. This includes: appending the latest feedback event stream data to the training dataset; re-executing the hierarchical binary classification model training operation and causal bias removal weight construction operation in S21; writing the updated model parameters and probability update mapping function parameters into the user response elasticity parameter dataset; and re-executing the batch calculation operation of elasticity coefficients in S22. This retraining operation continuously improves the model's generalization performance on unseen combinations by incorporating the latest data.

[0281] If the resource write-off cost prediction model exhibits continuous deviations across multiple business lines (e.g., prediction errors exceeding a preset threshold within three consecutive calibration periods), it triggers a retraining operation or hyperparameter adjustment operation in S3, and regenerates the pre-scheduled parameter configuration table and time-sharing window prediction data table for S3. This operation ensures the prediction model's adaptability to changes in data distribution.

[0282] S5 ultimately outputs the following data: Updated probability update mapping function parameter table; Updated user response resilience parameter dataset; Updated user usage probability prediction model weight coefficients and causal bias removal weight calculation parameters; Updated parameters for the resource write-off cost prediction model; The calibration log includes fields such as: trigger calibration unit identifier, deviation value, parameter change before and after calibration, probability update mapping function parameter index key, and rollback level marker.

[0283] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for intelligent scheduling of virtual resources based on big data user profiling and AI prediction models, characterized in that, The method includes the following steps: S1: Collect user historical resource deployment records, user behavior logs, user context feature data, resource type configuration data, and historical resource write-off cost amount time series; perform a unified cost mapping operation on the resource intensity parameters of each resource type, and uniformly map resource intensity parameters of different dimensions to a unified cost amount measured in monetary terms; construct a historical incentive event sequence based on the unified cost amount; S2: Construct and train a user usage probability prediction model, which is a hierarchical binary classification model; based on the trained user usage probability prediction model, calculate the incentive response elasticity coefficient of each user under each business line and each resource type through finite difference numerical differentiation operation. S3: Based on the historical resource write-off cost amount time series of each business line, construct a resource write-off cost prediction model; use the resource write-off cost prediction model to output the point prediction value and confidence interval of the resource write-off cost amount of each business line in each future time window; generate the minimum quota parameter and flexible quota parameter of each business line based on the point prediction value and the confidence interval. S4: Based on the guaranteed quota parameter and the flexible quota parameter, the total available quota for the day is divided into a guaranteed quota layer and a global flexible quota layer; when a real-time user request is received, based on the usage probability value output by the user usage probability prediction model, the marginal gain coefficient after write-off, the global resource adjustment parameter, and the business line resource adjustment parameter, the objective function value calculation is performed on the candidate resource combination, the resource combination that maximizes the objective function value is selected to generate a resource allocation instruction, and quota deduction and status update are performed; S5: Perform inner-loop consumption rate correction operation, and adaptively update the global resource adjustment parameters and the business line resource adjustment parameters based on the deviation between the actual net consumption rate and the predicted consumption rate; perform outer-loop model calibration operation, and perform probabilistic calibration mapping or model retraining based on the deviation between the actual utilization rate and the predicted utilization rate of the calibration unit.

2. The method according to claim 1, characterized in that, The unified cost mapping operation is performed in the following manner: When the resource type is identified as an electronic voucher, the voucher value will be directly assigned to the unified cost amount field. When the resource type is identified as a discount amount, the usage flag in the user's historical resource delivery record is read. If the usage flag indicates that the resource has been used, the actual redeemed order amount in the record is read, and the actual redeemed order amount is multiplied by the discount percentage. The result is then written to the unified cost amount field. If the usage flag indicates that the resource has not been used, the average historical order amount of the user within the recent preset time window of this business line is read, and the average historical order amount is multiplied by the discount percentage. The result is then written to the unified cost amount field. If the user has no historical order records, the average redeemed order amount within the recent preset time window of this business line is used instead of the average historical order amount for the above calculation. When the resource type is identified as activity points, the point quantity value is divided by the preset point redemption value coefficient to obtain a uniform cost amount. When the resource type is identified as a digital gift pack or a limited-time benefit, the fixed cost configuration table is queried based on the resource type identifier and resource strength parameters, and the value obtained from the table is used as the unified cost amount.

3. The method according to claim 1, characterized in that, The user usage probability prediction model satisfies the following relationship: in, For users u In business lines b Resource types k Standardize cost amount x And the probability of use under the user context feature vector c; This is the global intercept term; For users The user random effects term represents the basic usage tendency of users; For business lines The random effects term of the business line; For resource type The random effects term characterizes the differences in usage tendencies across resource type dimensions; This is the sensitivity coefficient for global resource intensity. For users In business lines For resource types Personalized intensity-sensitive adjustment options; For context feature weight vectors; The user random effect term, the business line random effect term, and the personalized intensity sensitive adjustment term are respectively expressed as the sum of the linear projection and residual of the observable features through feature prior parameterization. Among them, the observable characteristics of the user random effects term include the quantile of the amount paid in the near preset time window, the quantile of the number of active days in the near preset time window, the historical resource utilization rate, and the membership level; The observable characteristics of the random effects term for business lines include business line type, mean write-off cost within the near-preset time window, user overlap, and historical resource write-off rate; The observable characteristics of personalized intensity-sensitive adjustment items include the user's historical usage rate of this resource type, the number of days since the last use, and the median historical uniform cost amount of coupons collected; The residuals follow a Gaussian prior distribution with a mean of zero.

4. The method according to claim 1, characterized in that, Before the user trains the probability prediction model, a weight construction operation is performed on each training sample to obtain the sample weight, which is the product of the time decay weight and the causal debiasing weight. The time decay weight is calculated according to the following formula: in, For the first The difference between the delivery timestamp of the record and the current timestamp is converted into the number of days. Let be the attenuation coefficient for the g-th hierarchical group, which is divided according to resource type and user activity level, and the attenuation coefficient is estimated for each hierarchical group respectively; The causal bias removal weight is calculated using the following formula: in, As a stabilization factor, take the marginal probability or unified constant 1 of the corresponding resource strength parameter; For the first i The tendency score of each sample is assigned to the current resource intensity parameter, which is calculated by fitting a multi-classification model to the user's historical resource allocation records; Let i be the resource strength of the i-th sample; User context feature vector; This is a covariate vector, which does not include outcome variables or time-related information after the delivery timestamp; To prevent the denominator from being too small, a cutoff value is used; The maximum weight truncation threshold; After the weight construction operation is completed, the effective sample size detection operation is performed, and the effective sample size is calculated according to the following formula: When the effective sample size is lower than a preset threshold, a rescaling operation is performed on all weights or the maximum weight truncation threshold is increased to reduce the weight variance.

5. The method according to claim 1, characterized in that, In S2, a sample sparsity detection operation is performed on each user-business-resource type combination before training: Count the number of records for this combination in the historical stimulus event sequence, and calculate the difference between the maximum and minimum values ​​of the uniform cost amount within this combination; If the number of records is less than a preset threshold, or the difference is lower than a preset change range threshold, the sample label field of the combination is set to a fallback state. In model training, the independent estimation of the personalized intensity sensitivity adjustment term corresponding to the combination is skipped, and the personalized intensity sensitivity adjustment term is set to zero. Only the global resource intensity sensitivity coefficient and the group level prior mean are retained as the prediction basis for the combination. The sample label field also records the fallback level. In S2, after calculating the elastic coefficients of the excitation response, the cluster labeling operation is performed: The incentive response elasticity coefficient value of each user-business line-resource type combination is compared with the preset high elasticity threshold and medium elasticity threshold. Based on the comparison result, the user is labeled as a high elasticity user, medium elasticity user, low elasticity user, or negative elasticity user. The preset high elasticity threshold and medium elasticity threshold are taken as preset quantiles of the historical excitation response elasticity coefficient distribution; After the cluster labeling is completed, the incentive response elasticity coefficient, cluster label and associated model parameters are assembled into structured records, and written into key-value storage using the concatenated string of user identifier, business line identifier and resource type identifier as the index key to form the user response elasticity parameter dataset.

6. The method according to claim 1, characterized in that, The construction of the resource write-off cost prediction model includes: Perform a logarithmic transformation operation on the time series of historical resource write-off cost amounts to obtain a logarithmic write-off cost value series. Perform STL decomposition operation based on local weighted regression on the logarithmically transformed sequence to decompose it into a trend term sequence, a seasonal term sequence, and a residual term sequence; The trend term sequence is fitted using a Holt linear trend model or locally weighted linear regression. The seasonal term sequence is parameterized using Fourier series expansion, and the harmonic order of the Fourier series expansion is determined by performing a fast Fourier transform on the historical seasonal term sequence and detecting the dominant frequency component. The residual term sequence is modeled using an autoregressive integral moving average model. The order of the autoregressive integral moving average model is determined by performing fitting operations on each candidate order combination and selecting the order combination that minimizes the information content criterion. The resource write-off cost prediction model also includes an external event regression term: The event types and start and end dates are read from external event calendar data, an event impact function is constructed, and it is added to the model as a regression variable; The event impact function is defined as follows: for a single-day event, the value is 1 on the event day and 0 for the rest; for a multi-day event, a trapezoidal function is used, which linearly increases from 0 to 1 within a preset number of days before the event, remains at 1 during the event, and linearly decays to 0 after the event ends. The regression coefficients of the event impact function are obtained by performing least squares regression on the residual data within the historical event window.

7. The method according to claim 1, characterized in that, In S3, the minimum quota parameters and flexible quota parameters for each business line are generated based on the predicted point value and the confidence interval, including: The predicted lower limit of the confidence interval is assigned as the minimum quota parameter, and the difference between the predicted upper limit and the predicted lower limit is used as the elastic quota parameter and written into the pre-scheduling parameter configuration table. Further perform time-sharing window splitting operation, multiply the estimated value of the daily forecast point, the lower limit of the forecast, and the upper limit of the forecast by the historical time-sharing proportion coefficient respectively, to obtain the time-sharing window forecast data table; A consistency check operation is performed on the predicted values ​​of each time window. When the absolute value of the difference between the summation result of any time interval and the corresponding daily value exceeds the preset ratio of the corresponding daily value, a proportional correction operation is performed on each time value to align the time interval with the daily interval.

8. The method according to claim 1, characterized in that, The total available quota for the day is divided into a minimum quota tier and a global flexible quota tier, including: The basic quota for the day is calculated based on the estimated cost amount of the full-platform predicted write-off and the total budget for the activity period. The total available quota for the day is obtained by summing the basic quota for the day with the net balance of the previous day. The minimum quota parameter value of each business line is truncated under the constraint of the upper limit of the business line resource quota. When the total minimum quota after truncation exceeds the total available quota for the day, the minimum quota value of each business line is scaled proportionally until all business lines meet the constraint of not exceeding the upper limit of the business line resource quota, and the total minimum quota layer is obtained. The original allocable amount of the global elastic quota layer is obtained by performing a difference calculation between the total available quota for the day and the total guaranteed quota layer. The theoretical total amount of elastic quota is obtained by combining the elastic quota parameter values ​​of each business line and the upper limit constraint of the resource quota of the business line. The minimum value of the original allocable amount of the global elastic quota layer and the theoretical total amount of elastic quota is taken to obtain the total amount of the global elastic quota layer. The difference between the original allocatable amount of the global elastic quota layer and the total amount of the global elastic quota layer is used as the reserve quota balance, and the status flag of the reserve quota balance is set to the frozen state.

9. The method according to claim 1, characterized in that, Performing objective function value calculations on candidate resource combinations includes: For each candidate resource combination, the model input cost is calculated based on the resource type. Electronic vouchers, activity points, digital gift packs, and limited-time benefits are used as model input costs with a uniform cost amount. The discount amount is calculated by multiplying the average order amount of users in the near preset time window by the discount percentage as the model input cost. The user usage probability prediction model is invoked, and forward inference is performed with the input cost of the model as input to obtain the usage probability value. Calculate the following objective function values ​​for each candidate combination: in, For users u In business lines b Resource types k Resource strength parameters d The objective function value is below. To use probability values, This is the marginal gain coefficient after write-off. λ These are global resource adjustment parameters. Adjust parameters for business line resources. This represents the expected consumption amount. Perform comparison calculations to verify the quotas by comparing the pre-allocated security quota with the remaining quota at the global elastic quota layer and the available elastic quota for each business line. The model input cost and security pre-allocation amount of the discount quota type are calculated separately. The model input cost is used for objective function value calculation, and the security pre-allocation amount is used for quota verification calculation.

10. The method according to claim 9, characterized in that, When the quota verification fails, a set of remaining affordable candidates is constructed that satisfies that the security pre-allocation amount does not exceed the remaining amount of the global elastic quota layer and does not exceed the available amount of the elastic quota of the business line. The objective function value maximization selection operation is re-executed in the set of remaining affordable candidates. When the remaining affordable candidate set is empty or the objective function value of all combinations therein is not greater than zero, the decision result is set to an empty combination and transferred to the minimum quota layer for processing or the issuance instruction is rejected. When multiple candidate combinations have the same and positive objective function value, the combination with the smaller safety pre-occupancy is selected first. If the safety pre-occupancy is the same, the combination with the higher efficiency ratio is selected. The efficiency ratio is the product of the usage probability value and the post-write marginal gain coefficient divided by the expected consumption.

11. The method according to claim 1, characterized in that, The update of business line resource adjustment parameters in the inner loop consumption rate correction calculation includes: Read the confirmed write-off deduction and release pre-occupancy amount of this business line from the quota status data table in the observation window, and calculate the actual net consumption rate. Read the estimated value of the current time-sharing window forecast point for this business line from the time-sharing window forecast data table, and calculate the predicted consumption rate; The difference between the actual net consumption rate and the predicted consumption rate is calculated and divided by the predicted consumption rate to obtain the relative deviation of the net consumption rate of the business line. When the absolute value of the relative deviation of the net consumption rate of the business line exceeds the inner loop trigger threshold, the correction term of the business line resource adjustment parameter is updated based on the relative deviation of the net consumption rate of the business line. The business line resource adjustment parameter adopts a structure of superimposed benchmark value and correction term. Exponential smoothing operation is performed on the updated correction term, and range truncation operation is performed on the business line resource adjustment parameter. The update of global resource adjustment parameters in the inner loop consumption rate correction operation includes: Read the globally confirmed write-off deduction and the globally released pre-occupied amount from the quota status data table within the observation window, and calculate the global actual net consumption rate. Read the estimated values ​​of the current time-sharing window prediction points for each business line from the time-sharing window prediction data table and sum them up to calculate the global prediction consumption rate. The difference between the actual global net consumption rate and the global predicted consumption rate is calculated and divided by the global predicted consumption rate to obtain the relative deviation of the global net consumption rate. Perform an exponential moving average smoothing operation on the relative deviation of the global net consumption rate; Based on the smoothed relative deviation of the global net consumption rate, a new global resource adjustment parameter value is obtained through a multiplicative update operation; The range of the multiplicative factors used in the multiplicative update operation is truncated, and the range of the updated global resource adjustment parameters is protected by a truncation operation.

12. The method according to claim 1, characterized in that, The probability calibration mapping in the outer loop model calibration operation includes: The calibration unit is defined by a three-dimensional combination of user segmentation, business line, and resource type; For each calibration unit, the actual usage rate is statistically analyzed from the feedback event stream, the mean of the model-predicted usage rate is calculated from the predicted probability data, and the deviation between the actual usage rate and the mean of the predicted usage rate is calculated. When the absolute value of the deviation exceeds the outer loop trigger threshold, a probabilistic calibration operation is performed on the calibration unit to fit an ordinal regression or Pratt scaling mapping function to obtain the probabilistic update mapping function parameters. During online inference, the probability update mapping function parameters are queried sequentially according to the priority order of user segmentation and business line and resource type combination, business line and resource type combination, resource type, user segmentation, and global default. If a high-priority record does not exist, it will fall back to a low-priority record. If the global default probability update mapping function parameter does not exist, the calibration operation will be skipped and the original model probability value will be used directly. The outer ring model calibration calculation also includes: When the number of calibration units that trigger the probability calibration operation exceeds a preset threshold, or the average value of the deviation values ​​of all calibration units exceeds a preset threshold, the retraining operation of the user's probability prediction model is triggered, including appending the latest feedback event stream data to the training dataset, re-executing the training operation and causal bias removal weight construction operation of the hierarchical binary classification model, writing the updated model parameters and probability update mapping function parameters into the user response elasticity parameter dataset, and re-executing the batch calculation operation of the stimulus response elasticity coefficient. When the resource write-off cost prediction model shows continuous deviations across multiple business lines, the retraining operation of the resource write-off cost prediction model is triggered, and the pre-scheduling parameter configuration table and time-sharing window prediction data table are regenerated.

Images